html_url
stringlengths 51
51
| title
stringlengths 6
280
| comments
stringlengths 67
24.7k
| body
stringlengths 51
36.2k
⌀ | comment_length
int64 16
1.45k
| text
stringlengths 159
38.3k
|
|---|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/3553 | set_format("np") no longer works for Image data | This error also propagates to jax and is even trickier to fix, since `.with_format(type='jax')` will use numpy conversion internally (and fail). For a three line failure:
```python
dataset = datasets.load_dataset("mnist")
dataset.set_format("jax")
X_train = dataset["train"]["image"]
``` | ## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
| 35 | set_format("np") no longer works for Image data
## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
This error also propagates to jax and is even trickier to fix, since `.with_format(type='jax')` will use numpy conversion internally (and fail). For a three line failure:
```python
dataset = datasets.load_dataset("mnist")
dataset.set_format("jax")
X_train = dataset["train"]["image"]
``` |
https://github.com/huggingface/datasets/issues/3553 | set_format("np") no longer works for Image data | Hi! We've recently introduced a new Image feature that yields PIL Images (and caches transforms on them) instead of arrays.
However, this feature requires a custom transform to yield np arrays directly:
```python
ddict = datasets.load_dataset("mnist")
def pil_image_to_array(batch):
return {"image": [np.array(img) for img in batch["image"]]} # or jnp.array(img) for Jax
ddict.set_transform(pil_image_to_array, columns="image", output_all_columns=True)
```
[Docs](https://huggingface.co/docs/datasets/master/process.html#format-transform) on `set_transform`.
Also, the approach proposed by @cgarciae is not the best because it loads the entire column in memory.
@albertvillanova @lhoestq WDYT? The Audio and the Image feature currently don't support the TF/Jax/PT Formatters, but for the Numpy Formatter maybe it makes more sense to return np arrays (and not a dict in the case of the Audio feature or a PIL Image object in the case of the Image feature). | ## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
| 127 | set_format("np") no longer works for Image data
## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
Hi! We've recently introduced a new Image feature that yields PIL Images (and caches transforms on them) instead of arrays.
However, this feature requires a custom transform to yield np arrays directly:
```python
ddict = datasets.load_dataset("mnist")
def pil_image_to_array(batch):
return {"image": [np.array(img) for img in batch["image"]]} # or jnp.array(img) for Jax
ddict.set_transform(pil_image_to_array, columns="image", output_all_columns=True)
```
[Docs](https://huggingface.co/docs/datasets/master/process.html#format-transform) on `set_transform`.
Also, the approach proposed by @cgarciae is not the best because it loads the entire column in memory.
@albertvillanova @lhoestq WDYT? The Audio and the Image feature currently don't support the TF/Jax/PT Formatters, but for the Numpy Formatter maybe it makes more sense to return np arrays (and not a dict in the case of the Audio feature or a PIL Image object in the case of the Image feature). |
https://github.com/huggingface/datasets/issues/3553 | set_format("np") no longer works for Image data | Yes I agree it should return arrays and not a PIL image (and possible an array instead of a dict for audio data).
I'm currently finishing some code refactoring of the image and audio and opening a PR today. Maybe we can look into that after the refactoring | ## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
| 48 | set_format("np") no longer works for Image data
## Describe the bug
`dataset.set_format("np")` no longer works for image data, previously you could load the MNIST like this:
```python
dataset = load_dataset("mnist")
dataset.set_format("np")
X_train = dataset["train"]["image"][..., None] # <== No longer a numpy array
```
but now it doesn't work, `set_format("np")` seems to have no effect and the dataset just returns a list/array of PIL images instead of numpy arrays as requested.
Yes I agree it should return arrays and not a PIL image (and possible an array instead of a dict for audio data).
I'm currently finishing some code refactoring of the image and audio and opening a PR today. Maybe we can look into that after the refactoring |
https://github.com/huggingface/datasets/issues/3548 | Specify the feature types of a dataset on the Hub without needing a dataset script | After looking into this, discovered that this is already supported if the `dataset_infos.json` file is configured correctly! Here is a working example: https://huggingface.co/datasets/abidlabs/test-audio-13
This should be probably be documented, though. | **Is your feature request related to a problem? Please describe.**
Currently if I upload a CSV with paths to audio files, the column type is string instead of Audio.
**Describe the solution you'd like**
I'd like to be able to specify the types of the column, so that when loading the dataset I directly get the features types I want.
The feature types could read from the `dataset_infos.json` for example.
**Describe alternatives you've considered**
Create a dataset script to specify the features, but that seems complicated for a simple thing.
cc @abidlabs | 30 | Specify the feature types of a dataset on the Hub without needing a dataset script
**Is your feature request related to a problem? Please describe.**
Currently if I upload a CSV with paths to audio files, the column type is string instead of Audio.
**Describe the solution you'd like**
I'd like to be able to specify the types of the column, so that when loading the dataset I directly get the features types I want.
The feature types could read from the `dataset_infos.json` for example.
**Describe alternatives you've considered**
Create a dataset script to specify the features, but that seems complicated for a simple thing.
cc @abidlabs
After looking into this, discovered that this is already supported if the `dataset_infos.json` file is configured correctly! Here is a working example: https://huggingface.co/datasets/abidlabs/test-audio-13
This should be probably be documented, though. |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | Thanks for reporting. I think this can be fixed by improving the `CachedDatasetModuleFactory` and making it look into the `parquet` cache directory (datasets from push_to_hub are loaded with the parquet dataset builder). I'll look into it | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 36 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
Thanks for reporting. I think this can be fixed by improving the `CachedDatasetModuleFactory` and making it look into the `parquet` cache directory (datasets from push_to_hub are loaded with the parquet dataset builder). I'll look into it |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | We haven't had a chance to fix this yet. If someone would like to give it a try I'd be happy to give some guidance | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 25 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
We haven't had a chance to fix this yet. If someone would like to give it a try I'd be happy to give some guidance |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | @lhoestq Do you have an idea of what changes need to be made to `CachedDatasetModuleFactory`? I would be willing to take a crack at it. Currently unable to train with datasets I have `push_to_hub` on a cluster whose compute nodes are not connected to the internet.
It looks like it might be this line:
https://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L994
Which wouldn't pick up the stuff saved under `"datasets/allenai___parquet/*"`. Additionally, the datasets saved under `"datasets/allenai___parquet/*"` appear to have hashes in their name, e.g. `"datasets/allenai___parquet/my_dataset-def9ee5552a1043e"`. This would not be detected by `CachedDatasetModuleFactory`, which currently looks for subdirectories
https://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L995-L999 | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 92 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
@lhoestq Do you have an idea of what changes need to be made to `CachedDatasetModuleFactory`? I would be willing to take a crack at it. Currently unable to train with datasets I have `push_to_hub` on a cluster whose compute nodes are not connected to the internet.
It looks like it might be this line:
https://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L994
Which wouldn't pick up the stuff saved under `"datasets/allenai___parquet/*"`. Additionally, the datasets saved under `"datasets/allenai___parquet/*"` appear to have hashes in their name, e.g. `"datasets/allenai___parquet/my_dataset-def9ee5552a1043e"`. This would not be detected by `CachedDatasetModuleFactory`, which currently looks for subdirectories
https://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L995-L999 |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | `importable_directory_path` is used to find a **dataset script** that was previously downloaded and cached from the Hub
However in your case there's no dataset script on the Hub, only parquet files. So the logic must be extended for this case.
In particular I think you can add a new logic in the case where `hashes is None` (i.e. if there's no dataset script associated to the dataset in the cache).
In this case you can check directly in the in the datasets cache for a directory named `<namespace>__parquet` and a subdirectory named `<config_id>`. The config_id must match `{self.name.replace("/", "--")}-*`.
In your case those two directories correspond to `allenai___parquet` and then `allenai--my_dataset-def9ee5552a1043e`
Then you can find the most recent version of the dataset in subdirectories (e.g. sorting using the last modified time of the `dataset_info.json` file).
Finally, we will need return the module that is used to load the dataset from the cache. It is the same module than the one that would have been normally used if you had an internet connection.
At that point you can ping me, because we will need to pass all this:
- `module_path = _PACKAGED_DATASETS_MODULES["parquet"][0]`
- `hash` it corresponds the name of the directory that contains the .arrow file, inside `<namespace>__parquet/<config_id>`
- ` builder_kwargs = {"hash": hash, "repo_id": self.name, "config_id": config_id}`
and currently `config_id` is not a valid argument for a `DatasetBuilder`
I think in the future we want to change this caching logic completely, since I don't find it super easy to play with. | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 251 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
`importable_directory_path` is used to find a **dataset script** that was previously downloaded and cached from the Hub
However in your case there's no dataset script on the Hub, only parquet files. So the logic must be extended for this case.
In particular I think you can add a new logic in the case where `hashes is None` (i.e. if there's no dataset script associated to the dataset in the cache).
In this case you can check directly in the in the datasets cache for a directory named `<namespace>__parquet` and a subdirectory named `<config_id>`. The config_id must match `{self.name.replace("/", "--")}-*`.
In your case those two directories correspond to `allenai___parquet` and then `allenai--my_dataset-def9ee5552a1043e`
Then you can find the most recent version of the dataset in subdirectories (e.g. sorting using the last modified time of the `dataset_info.json` file).
Finally, we will need return the module that is used to load the dataset from the cache. It is the same module than the one that would have been normally used if you had an internet connection.
At that point you can ping me, because we will need to pass all this:
- `module_path = _PACKAGED_DATASETS_MODULES["parquet"][0]`
- `hash` it corresponds the name of the directory that contains the .arrow file, inside `<namespace>__parquet/<config_id>`
- ` builder_kwargs = {"hash": hash, "repo_id": self.name, "config_id": config_id}`
and currently `config_id` is not a valid argument for a `DatasetBuilder`
I think in the future we want to change this caching logic completely, since I don't find it super easy to play with. |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | Hi! Is there a workaround for the time being?
Like passing `data_dir` or something like that?
I would like to use [this diffuser example](https://github.com/huggingface/diffusers/tree/main/examples/unconditional_image_generation) on my cluster whose nodes are not connected to the internet. I have downloaded the dataset online form the login node. | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 45 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
Hi! Is there a workaround for the time being?
Like passing `data_dir` or something like that?
I would like to use [this diffuser example](https://github.com/huggingface/diffusers/tree/main/examples/unconditional_image_generation) on my cluster whose nodes are not connected to the internet. I have downloaded the dataset online form the login node. |
https://github.com/huggingface/datasets/issues/3547 | Datasets created with `push_to_hub` can't be accessed in offline mode | Hi ! Yes you can save your dataset locally with `my_dataset.save_to_disk("path/to/local")` and reload it later with `load_from_disk("path/to/local")`
(removing myself from assignees since I'm currently not working on this right now) | ## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
| 30 | Datasets created with `push_to_hub` can't be accessed in offline mode
## Describe the bug
In offline mode, one can still access previously-cached datasets. This fails with datasets created with `push_to_hub`.
## Steps to reproduce the bug
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
in bash:
```
export HF_DATASETS_OFFLINE=1
```
in Python:
```
import datasets
mpwiki = datasets.load_dataset("teven/matched_passages_wikidata")
```
## Expected results
`datasets` should find the previously-cached dataset.
## Actual results
ConnectionError: Couln't reach the Hugging Face Hub for dataset 'teven/matched_passages_wikidata': Offline mode is enabled
## Environment info
- `datasets` version: 1.16.2.dev0
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.10
- PyArrow version: 3.0.0
Hi ! Yes you can save your dataset locally with `my_dataset.save_to_disk("path/to/local")` and reload it later with `load_from_disk("path/to/local")`
(removing myself from assignees since I'm currently not working on this right now) |
https://github.com/huggingface/datasets/issues/3543 | Allow loading community metrics from the hub, just like datasets | Hi ! Thanks for your message :) This is a great idea indeed. We haven't started working on this yet though. For now I guess you can host your metric on the Hub (either with your model or your dataset) and use `hf_hub_download` to download it (docs [here](https://github.com/huggingface/huggingface_hub/blob/main/docs/hub/how-to-downstream.md#cached_download)) | **Is your feature request related to a problem? Please describe.**
Currently, I can load a metric implemented by me by providing the local path to the file in `load_metric`.
However, there is no option to do it with the metric uploaded to the hub.
This means that if I want to allow other users to use it, they must download it first which makes the usage less smooth.
**Describe the solution you'd like**
Load metrics from the hub just like datasets are loaded.
In order to not break stuff, the convention can be to put the metric file in a "metrics" folder in the hub.
| 48 | Allow loading community metrics from the hub, just like datasets
**Is your feature request related to a problem? Please describe.**
Currently, I can load a metric implemented by me by providing the local path to the file in `load_metric`.
However, there is no option to do it with the metric uploaded to the hub.
This means that if I want to allow other users to use it, they must download it first which makes the usage less smooth.
**Describe the solution you'd like**
Load metrics from the hub just like datasets are loaded.
In order to not break stuff, the convention can be to put the metric file in a "metrics" folder in the hub.
Hi ! Thanks for your message :) This is a great idea indeed. We haven't started working on this yet though. For now I guess you can host your metric on the Hub (either with your model or your dataset) and use `hf_hub_download` to download it (docs [here](https://github.com/huggingface/huggingface_hub/blob/main/docs/hub/how-to-downstream.md#cached_download)) |
https://github.com/huggingface/datasets/issues/3543 | Allow loading community metrics from the hub, just like datasets | Here's the code I used, in case it can be of help to someone else:
```python
import os, shutil
from huggingface_hub import hf_hub_download
def download_metric(repo_id, file_path):
# repo_id: for models "username/model_name", for datasets "datasets/username/model_name"
local_metric_path = hf_hub_download(repo_id=repo_id, filename=file_path)
updated_local_metric_path = (os.path.dirname(local_metric_path) + os.path.basename(local_metric_path).replace(".", "_") + ".py")
shutil.copy(local_metric_path, updated_local_metric_path)
return updated_local_metric_path
metric = load_metric(download_metric(REPO_ID, FILE_PATH))
``` | **Is your feature request related to a problem? Please describe.**
Currently, I can load a metric implemented by me by providing the local path to the file in `load_metric`.
However, there is no option to do it with the metric uploaded to the hub.
This means that if I want to allow other users to use it, they must download it first which makes the usage less smooth.
**Describe the solution you'd like**
Load metrics from the hub just like datasets are loaded.
In order to not break stuff, the convention can be to put the metric file in a "metrics" folder in the hub.
| 55 | Allow loading community metrics from the hub, just like datasets
**Is your feature request related to a problem? Please describe.**
Currently, I can load a metric implemented by me by providing the local path to the file in `load_metric`.
However, there is no option to do it with the metric uploaded to the hub.
This means that if I want to allow other users to use it, they must download it first which makes the usage less smooth.
**Describe the solution you'd like**
Load metrics from the hub just like datasets are loaded.
In order to not break stuff, the convention can be to put the metric file in a "metrics" folder in the hub.
Here's the code I used, in case it can be of help to someone else:
```python
import os, shutil
from huggingface_hub import hf_hub_download
def download_metric(repo_id, file_path):
# repo_id: for models "username/model_name", for datasets "datasets/username/model_name"
local_metric_path = hf_hub_download(repo_id=repo_id, filename=file_path)
updated_local_metric_path = (os.path.dirname(local_metric_path) + os.path.basename(local_metric_path).replace(".", "_") + ".py")
shutil.copy(local_metric_path, updated_local_metric_path)
return updated_local_metric_path
metric = load_metric(download_metric(REPO_ID, FILE_PATH))
``` |
https://github.com/huggingface/datasets/issues/3533 | Task search function on hub not working correctly | Because it has dots in some YAML keys, it can't be parsed and indexed by the back-end | When I want to look at all datasets of the category: `speech-processing` *i.e.* https://huggingface.co/datasets?task_categories=task_categories:speech-processing&sort=downloads , then the following dataset doesn't show up for some reason:
- https://huggingface.co/datasets/speech_commands
even thought it's task tags seem correct:
https://raw.githubusercontent.com/huggingface/datasets/master/datasets/speech_commands/README.md | 17 | Task search function on hub not working correctly
When I want to look at all datasets of the category: `speech-processing` *i.e.* https://huggingface.co/datasets?task_categories=task_categories:speech-processing&sort=downloads , then the following dataset doesn't show up for some reason:
- https://huggingface.co/datasets/speech_commands
even thought it's task tags seem correct:
https://raw.githubusercontent.com/huggingface/datasets/master/datasets/speech_commands/README.md
Because it has dots in some YAML keys, it can't be parsed and indexed by the back-end |
https://github.com/huggingface/datasets/issues/3518 | Add PubMed Central Open Access dataset | In the framework of BigScience:
- bigscience-workshop/data_tooling#121
I have created this dataset as a community dataset: https://huggingface.co/datasets/albertvillanova/pmc_open_access
However, I was wondering that it may be more appropriate to move it under an org namespace: `pubmed_central` or `pmc`
This way, we could add other datasets I'm also working on: Author Manuscript Dataset, Historical OCR Dataset, LitArch Open Access Subset.
What do you think? @lhoestq @mariosasko | ## Adding a Dataset
- **Name:** PubMed Central Open Access
- **Description:** The PMC Open Access Subset includes more than 3.4 million journal articles and preprints that are made available under license terms that allow reuse.
- **Paper:** *link to the dataset paper if available*
- **Data:** https://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/
- **Motivation:** *what are some good reasons to have this dataset*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 64 | Add PubMed Central Open Access dataset
## Adding a Dataset
- **Name:** PubMed Central Open Access
- **Description:** The PMC Open Access Subset includes more than 3.4 million journal articles and preprints that are made available under license terms that allow reuse.
- **Paper:** *link to the dataset paper if available*
- **Data:** https://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/
- **Motivation:** *what are some good reasons to have this dataset*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
In the framework of BigScience:
- bigscience-workshop/data_tooling#121
I have created this dataset as a community dataset: https://huggingface.co/datasets/albertvillanova/pmc_open_access
However, I was wondering that it may be more appropriate to move it under an org namespace: `pubmed_central` or `pmc`
This way, we could add other datasets I'm also working on: Author Manuscript Dataset, Historical OCR Dataset, LitArch Open Access Subset.
What do you think? @lhoestq @mariosasko |
https://github.com/huggingface/datasets/issues/3518 | Add PubMed Central Open Access dataset | Why not ! Having them under such namespaces would also help people searching for this kind of datasets.
We can also invite people from pubmed at one point | ## Adding a Dataset
- **Name:** PubMed Central Open Access
- **Description:** The PMC Open Access Subset includes more than 3.4 million journal articles and preprints that are made available under license terms that allow reuse.
- **Paper:** *link to the dataset paper if available*
- **Data:** https://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/
- **Motivation:** *what are some good reasons to have this dataset*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 28 | Add PubMed Central Open Access dataset
## Adding a Dataset
- **Name:** PubMed Central Open Access
- **Description:** The PMC Open Access Subset includes more than 3.4 million journal articles and preprints that are made available under license terms that allow reuse.
- **Paper:** *link to the dataset paper if available*
- **Data:** https://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/
- **Motivation:** *what are some good reasons to have this dataset*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Why not ! Having them under such namespaces would also help people searching for this kind of datasets.
We can also invite people from pubmed at one point |
https://github.com/huggingface/datasets/issues/3510 | `wiki_dpr` details for Open Domain Question Answering tasks | Hi ! According to the DPR paper, the wikipedia dump is the one from Dec. 20, 2018.
Each instance contains a paragraph of at most 100 word, as well as the title of the wikipedia page it comes from and the DPR embedding (a 768-d vector). | Hey guys!
Thanks for creating the `wiki_dpr` dataset!
I am currently trying to use the dataset for context retrieval using DPR on NQ questions and need details about what each of the files and data instances mean, which version of the Wikipedia dump it uses, etc. Please respond at your earliest convenience regarding the same! Thanks a ton!
P.S.: (If one of @thomwolf @lewtun @lhoestq could respond, that would be even better since they have the first-hand details of the dataset. If anyone else has those, please reach out! Thanks!) | 46 | `wiki_dpr` details for Open Domain Question Answering tasks
Hey guys!
Thanks for creating the `wiki_dpr` dataset!
I am currently trying to use the dataset for context retrieval using DPR on NQ questions and need details about what each of the files and data instances mean, which version of the Wikipedia dump it uses, etc. Please respond at your earliest convenience regarding the same! Thanks a ton!
P.S.: (If one of @thomwolf @lewtun @lhoestq could respond, that would be even better since they have the first-hand details of the dataset. If anyone else has those, please reach out! Thanks!)
Hi ! According to the DPR paper, the wikipedia dump is the one from Dec. 20, 2018.
Each instance contains a paragraph of at most 100 word, as well as the title of the wikipedia page it comes from and the DPR embedding (a 768-d vector). |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work), so that for datasets that have a working dataset preview, we can remove the dummy data IMO. On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.
I don't really have an opinion regarding the JSON metadata as I don't know enough about it.
| I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 105 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work), so that for datasets that have a working dataset preview, we can remove the dummy data IMO. On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.
I don't really have an opinion regarding the JSON metadata as I don't know enough about it.
|
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | I don't know all the details, but generally I'd be in favor of unifying the metadata formats into YAML inside .md (and so deprecating the dataset_infos.json)
(Ultimately the CI can run on "HuggingFace Actions" instead of on GitHub) | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 38 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
I don't know all the details, but generally I'd be in favor of unifying the metadata formats into YAML inside .md (and so deprecating the dataset_infos.json)
(Ultimately the CI can run on "HuggingFace Actions" instead of on GitHub) |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | The dataset_infos.json file currently has these useful infos for each dataset configuration, that I think can be moved to the dataset tags:
- Size of the dataset in MB: download size, arrow file size, and total size (sum of download + arrow)
- Size of each split in MB and number of examples. Again this can be moved to the dataset tags
- Feature type of each column
- supported task templates (it defines what columns correspond to the features and labels for example)
But it also has this, which I'm not sure if it should be in the tags or not:
- Checksums of the downloaded files for integrity verifications
So ultimately this file could probably be deprecated in favor of having the infos in the tags.
> Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
To get the exact number of examples and size in MB of the dataset, one needs to download and generate it completely. IMO these infos are very important when someone considers using a dataset. Though using streaming we could do some extrapolation to have approximate values instead.
For the integrity verifications we also need the number of examples and the checksums of the downloaded files, so it requires the dataset to be fully downloaded once. This can be optional though.
> IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work)
I agree with this. Usually if a dataset works in streaming mode, then it works in non-streaming mode (the other way around is not true though).
> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.
Yes indeed, or at least make sure that it was tested on the true data. | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 355 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
The dataset_infos.json file currently has these useful infos for each dataset configuration, that I think can be moved to the dataset tags:
- Size of the dataset in MB: download size, arrow file size, and total size (sum of download + arrow)
- Size of each split in MB and number of examples. Again this can be moved to the dataset tags
- Feature type of each column
- supported task templates (it defines what columns correspond to the features and labels for example)
But it also has this, which I'm not sure if it should be in the tags or not:
- Checksums of the downloaded files for integrity verifications
So ultimately this file could probably be deprecated in favor of having the infos in the tags.
> Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
To get the exact number of examples and size in MB of the dataset, one needs to download and generate it completely. IMO these infos are very important when someone considers using a dataset. Though using streaming we could do some extrapolation to have approximate values instead.
For the integrity verifications we also need the number of examples and the checksums of the downloaded files, so it requires the dataset to be fully downloaded once. This can be optional though.
> IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work)
I agree with this. Usually if a dataset works in streaming mode, then it works in non-streaming mode (the other way around is not true though).
> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.
Yes indeed, or at least make sure that it was tested on the true data. |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | (note that if we wanted to display sizes, etc we could also pretty easily parse the `dataset_infos.json` on the hub side) | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 21 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
(note that if we wanted to display sizes, etc we could also pretty easily parse the `dataset_infos.json` on the hub side) |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | I agree that we can move the relevant parts of `dataset_infos.json` to the YAML tags.
> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data. <
> > Yes indeed, or at least make sure that it was tested on the true data.
I like the idea of testing streaming and falling back to the dummy data test if streaming does not work. Generating dummy data can be very tedious, so this would be a nice incentive for the contributors to make their datasets streamable. | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 112 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
I agree that we can move the relevant parts of `dataset_infos.json` to the YAML tags.
> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data. <
> > Yes indeed, or at least make sure that it was tested on the true data.
I like the idea of testing streaming and falling back to the dummy data test if streaming does not work. Generating dummy data can be very tedious, so this would be a nice incentive for the contributors to make their datasets streamable. |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | About dummy data, please see e.g. this PR: https://github.com/huggingface/datasets/pull/3692/commits/62368daac0672041524a471386d5e78005cf357a
- I updated the previous dummy data: I just had to rename the file and its directory
- the dummy data zip contains only a single file: `pubmed22n0001.xml.gz`
Then I discover it fails: https://app.circleci.com/pipelines/github/huggingface/datasets/9800/workflows/173a4433-8feb-4fc6-ab9e-59762084e3e1/jobs/60437
```
No such file or directory: '.../dummy_data/pubmed22n0002.xml.gz'
```
- it needs dummy data for all the 1114 files:
`_URLs = [f"ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed22n{i:04d}.xml.gz" for i in range(1, 1115)]`
- this confirms me that it never passed the test: these dummy data files were not present before my PR
- therefore, is it really useful the data test if we just ignore it when it does not pass?
In relation with JSON metadata, I'm generating the file for `pubmed` (see above) in a GCP instance: it's running for more than 3 hours and only 9 million examples generated so far (before my PR, it had 32 million, now it has more). | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 151 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
About dummy data, please see e.g. this PR: https://github.com/huggingface/datasets/pull/3692/commits/62368daac0672041524a471386d5e78005cf357a
- I updated the previous dummy data: I just had to rename the file and its directory
- the dummy data zip contains only a single file: `pubmed22n0001.xml.gz`
Then I discover it fails: https://app.circleci.com/pipelines/github/huggingface/datasets/9800/workflows/173a4433-8feb-4fc6-ab9e-59762084e3e1/jobs/60437
```
No such file or directory: '.../dummy_data/pubmed22n0002.xml.gz'
```
- it needs dummy data for all the 1114 files:
`_URLs = [f"ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed22n{i:04d}.xml.gz" for i in range(1, 1115)]`
- this confirms me that it never passed the test: these dummy data files were not present before my PR
- therefore, is it really useful the data test if we just ignore it when it does not pass?
In relation with JSON metadata, I'm generating the file for `pubmed` (see above) in a GCP instance: it's running for more than 3 hours and only 9 million examples generated so far (before my PR, it had 32 million, now it has more). |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | I mention in https://github.com/huggingface/datasets-server/wiki/Preliminary-design that the future "datasets server" could be in charge of generating both the dummy data and the dataset-info.json file if required (or their equivalent). | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 28 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
I mention in https://github.com/huggingface/datasets-server/wiki/Preliminary-design that the future "datasets server" could be in charge of generating both the dummy data and the dataset-info.json file if required (or their equivalent). |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | Hi ! I think dummy data generation is out of scope for the datasets server, since it's about generating the original data files.
That would be amazing to have it generate the dataset_infos.json though ! | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 35 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
Hi ! I think dummy data generation is out of scope for the datasets server, since it's about generating the original data files.
That would be amazing to have it generate the dataset_infos.json though ! |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | From some offline discussion with @mariosasko and especially for vision datasets, we'll probably not require dummy data anymore and use streaming instead :) This will make adding a new dataset much easier.
This should also make sure that streaming works as expected directly in the CI, without having to check the dataset viewer once the PR is merged | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 58 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
From some offline discussion with @mariosasko and especially for vision datasets, we'll probably not require dummy data anymore and use streaming instead :) This will make adding a new dataset much easier.
This should also make sure that streaming works as expected directly in the CI, without having to check the dataset viewer once the PR is merged |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | It seems that migration from dataset-info.json to dataset card YAML has been acted.
Probably it's a good idea, but I didn't find the pros and cons of this decision, so I put some I could think of:
pros:
- only one file to parse, share, sync
- it gives a hint to the users that if you write your dataset card, you should also specify the metadata
cons:
- the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
- YAML vs JSON: not sure which one is easier for users to fill and maintain
- two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.
- [low priority] besides the JSON file, we might want to support yaml or toml file if the user prefers (as [prettier](https://prettier.io/docs/en/configuration.html) and others do for their config files, for example). Inside the md, I understand that only YAML is allowed | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 210 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
It seems that migration from dataset-info.json to dataset card YAML has been acted.
Probably it's a good idea, but I didn't find the pros and cons of this decision, so I put some I could think of:
pros:
- only one file to parse, share, sync
- it gives a hint to the users that if you write your dataset card, you should also specify the metadata
cons:
- the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
- YAML vs JSON: not sure which one is easier for users to fill and maintain
- two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.
- [low priority] besides the JSON file, we might want to support yaml or toml file if the user prefers (as [prettier](https://prettier.io/docs/en/configuration.html) and others do for their config files, for example). Inside the md, I understand that only YAML is allowed |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | > the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
Note that we could simply not have the checksums in the YAML metadata at all, or maybe at one point have a pointer to another file instead.
We can also choose to hide (collapse) certain sections in the YAML by default when we open the dataset card editor.
> two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.
I think it's fine for now. Later if we really end up with too many YAML sections we can see if we need to tweak the API endpoints or the `datasets`/`huggingface_hub` tools
> YAML vs JSON: not sure which one is easier for users to fill and maintain
Regarding YAML vs JSON: I think YAML is easier to write by hand, and I also think that it's better for consistency - i.e. we're using more and more YAML to configure models/datasets/spaces | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 214 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
> the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
Note that we could simply not have the checksums in the YAML metadata at all, or maybe at one point have a pointer to another file instead.
We can also choose to hide (collapse) certain sections in the YAML by default when we open the dataset card editor.
> two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.
I think it's fine for now. Later if we really end up with too many YAML sections we can see if we need to tweak the API endpoints or the `datasets`/`huggingface_hub` tools
> YAML vs JSON: not sure which one is easier for users to fill and maintain
Regarding YAML vs JSON: I think YAML is easier to write by hand, and I also think that it's better for consistency - i.e. we're using more and more YAML to configure models/datasets/spaces |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | > the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
We can definitely work on this on the hub side to make the UX better | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 66 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
> the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums
We can definitely work on this on the hub side to make the UX better |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | Tensorflow Datasets catalog includes a community catalog where you can find and use HF datasets (see [here](https://www.tensorflow.org/datasets/community_catalog/huggingface)).
FYI I noticed today that they are using the exported dataset_infos.json files from github to get the metadata (see their code [here](https://github.com/tensorflow/datasets/blob/a482f01c036a10496f5e22e69a2ef81b707cc418/tensorflow_datasets/scripts/documentation/build_community_catalog.py#L261)) | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 39 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
Tensorflow Datasets catalog includes a community catalog where you can find and use HF datasets (see [here](https://www.tensorflow.org/datasets/community_catalog/huggingface)).
FYI I noticed today that they are using the exported dataset_infos.json files from github to get the metadata (see their code [here](https://github.com/tensorflow/datasets/blob/a482f01c036a10496f5e22e69a2ef81b707cc418/tensorflow_datasets/scripts/documentation/build_community_catalog.py#L261)) |
https://github.com/huggingface/datasets/issues/3507 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data | Metadata is now stored as YAML, and dummy data is deprecated, so I think we can close this issue. | I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw | 19 | Discuss whether support canonical datasets w/o dataset_infos.json and/or dummy data
I open this PR to have a public discussion about this topic and make a decision.
As previously discussed, once we have the metadata in the dataset card (README file, containing both Markdown info and YAML tags), what is the point of having also the JSON metadata (dataset_infos.json file)?
On the other hand, the dummy data is necessary for testing (in our CI suite) that the canonical dataset loads correctly. However:
- the dataset preview feature is already an indirect test that the dataset loads correctly (it also tests it is streamable though)
- we are migrating canonical datasets to the Hub
Do we really need to continue testing them in out CI?
Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).
Feel free to ping other people for the discussion.
CC: @lhoestq @mariosasko @thomwolf @julien-c @patrickvonplaten @anton-l @LysandreJik @yjernite @nateraw
Metadata is now stored as YAML, and dummy data is deprecated, so I think we can close this issue. |
https://github.com/huggingface/datasets/issues/3505 | cast_column function not working with map function in streaming mode for Audio features | Hi! This is probably due to the fact that `IterableDataset.map` sets `features` to `None` before mapping examples. We can fix the issue by passing the old `features` dict to the map generator and performing encoding/decoding there (before calling the map transform function). | ## Describe the bug
I am trying to use Audio class for loading audio features using custom dataset. I am able to cast 'audio' feature into 'Audio' format with cast_column function. On using map function, I am not getting 'Audio' casted feature but getting path of audio file only.
I am getting features of 'audio' of string type with load_dataset call. After using cast_column 'audio' feature is converted into 'Audio' type. But in map function I am not able to get Audio type for audio feature & getting string type data containing path of file only. So I am not able to use processor in encode function.
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset, Audio
from transformers import Wav2Vec2Processor
def encode(batch, processor):
print("Audio: ",batch['audio'])
batch["input_values"] = processor(batch["audio"]['array'], sampling_rate=16000).input_values
return batch
def print_ds(ds):
iterator = iter(ds)
for d in iterator:
print("Data: ",d)
break
processor = Wav2Vec2Processor.from_pretrained(pretrained_model_path)
dataset = load_dataset("custom_dataset.py","train",data_files={'train':'train_path.txt'},
data_dir="data", streaming=True, split="train")
print("Features: ",dataset.features)
print_ds(dataset)
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
print("Features: ",dataset.features)
print_ds(dataset)
dataset = dataset.map(lambda x: encode(x,processor))
print("Features: ",dataset.features)
print_ds(dataset)
```
## Expected results
map function not printing Audio type features be used with processor function and getting error in processor call due to this.
## Actual results
# after load_dataset call
Features: {'sentence': Value(dtype='string', id=None), 'audio': Value(dtype='string', id=None)}
Data: {'sentence': 'और अपने पेट को माँ की स्वादिष्ट गरमगरम जलेबियाँ हड़पते\n', 'audio': 'data/0116_003.wav'}
# after cast_column call
Features: {'sentence': Value(dtype='string', id=None), 'audio': Audio(sampling_rate=16000, mono=True, _storage_dtype='string', id=None)}
Data: {'sentence': 'और अपने पेट को माँ की स्वादिष्ट गरमगरम जलेबियाँ हड़पते\n', 'audio': {'path': 'data/0116_003.wav', 'array': array([ 1.2662281e-06, 1.0264218e-06, -1.3615092e-06, ...,
1.3017889e-02, 1.0085563e-02, 4.8155054e-03], dtype=float32), 'sampling_rate': 16000}}
# after map call
Features: None
Audio: data/0116_003.wav
Traceback (most recent call last):
File "demo2.py", line 36, in <module>
print_ds(dataset)
File "demo2.py", line 11, in print_ds
for d in iterator:
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 341, in __iter__
for key, example in self._iter():
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 338, in _iter
yield from ex_iterable
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 192, in __iter__
yield key, self.function(example)
File "demo2.py", line 32, in <lambda>
dataset = dataset.map(lambda x: batch_encode(x,processor))
File "demo2.py", line 6, in batch_encode
batch["input_values"] = processor(batch["audio"]['array'], sampling_rate=16000).input_values
TypeError: string indices must be integers
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-4.14.243 with-debian-bullseye-sid
- Python version: 3.7.9
- PyArrow version: 6.0.1
| 42 | cast_column function not working with map function in streaming mode for Audio features
## Describe the bug
I am trying to use Audio class for loading audio features using custom dataset. I am able to cast 'audio' feature into 'Audio' format with cast_column function. On using map function, I am not getting 'Audio' casted feature but getting path of audio file only.
I am getting features of 'audio' of string type with load_dataset call. After using cast_column 'audio' feature is converted into 'Audio' type. But in map function I am not able to get Audio type for audio feature & getting string type data containing path of file only. So I am not able to use processor in encode function.
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset, Audio
from transformers import Wav2Vec2Processor
def encode(batch, processor):
print("Audio: ",batch['audio'])
batch["input_values"] = processor(batch["audio"]['array'], sampling_rate=16000).input_values
return batch
def print_ds(ds):
iterator = iter(ds)
for d in iterator:
print("Data: ",d)
break
processor = Wav2Vec2Processor.from_pretrained(pretrained_model_path)
dataset = load_dataset("custom_dataset.py","train",data_files={'train':'train_path.txt'},
data_dir="data", streaming=True, split="train")
print("Features: ",dataset.features)
print_ds(dataset)
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
print("Features: ",dataset.features)
print_ds(dataset)
dataset = dataset.map(lambda x: encode(x,processor))
print("Features: ",dataset.features)
print_ds(dataset)
```
## Expected results
map function not printing Audio type features be used with processor function and getting error in processor call due to this.
## Actual results
# after load_dataset call
Features: {'sentence': Value(dtype='string', id=None), 'audio': Value(dtype='string', id=None)}
Data: {'sentence': 'और अपने पेट को माँ की स्वादिष्ट गरमगरम जलेबियाँ हड़पते\n', 'audio': 'data/0116_003.wav'}
# after cast_column call
Features: {'sentence': Value(dtype='string', id=None), 'audio': Audio(sampling_rate=16000, mono=True, _storage_dtype='string', id=None)}
Data: {'sentence': 'और अपने पेट को माँ की स्वादिष्ट गरमगरम जलेबियाँ हड़पते\n', 'audio': {'path': 'data/0116_003.wav', 'array': array([ 1.2662281e-06, 1.0264218e-06, -1.3615092e-06, ...,
1.3017889e-02, 1.0085563e-02, 4.8155054e-03], dtype=float32), 'sampling_rate': 16000}}
# after map call
Features: None
Audio: data/0116_003.wav
Traceback (most recent call last):
File "demo2.py", line 36, in <module>
print_ds(dataset)
File "demo2.py", line 11, in print_ds
for d in iterator:
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 341, in __iter__
for key, example in self._iter():
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 338, in _iter
yield from ex_iterable
File "/opt/conda/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 192, in __iter__
yield key, self.function(example)
File "demo2.py", line 32, in <lambda>
dataset = dataset.map(lambda x: batch_encode(x,processor))
File "demo2.py", line 6, in batch_encode
batch["input_values"] = processor(batch["audio"]['array'], sampling_rate=16000).input_values
TypeError: string indices must be integers
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-4.14.243 with-debian-bullseye-sid
- Python version: 3.7.9
- PyArrow version: 6.0.1
Hi! This is probably due to the fact that `IterableDataset.map` sets `features` to `None` before mapping examples. We can fix the issue by passing the old `features` dict to the map generator and performing encoding/decoding there (before calling the map transform function). |
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | Hi @ToddMorrill, thanks for reporting.
Three weeks ago I contacted the team who created the Pile dataset to report this issue with their data host server: https://the-eye.eu
They told me that unfortunately, the-eye was heavily affected by the recent tornado catastrophe in the US. They hope to have their data back online asap. | ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 53 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
Hi @ToddMorrill, thanks for reporting.
Three weeks ago I contacted the team who created the Pile dataset to report this issue with their data host server: https://the-eye.eu
They told me that unfortunately, the-eye was heavily affected by the recent tornado catastrophe in the US. They hope to have their data back online asap. |
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | Hi @ToddMorrill, people from the Pile team have mirrored their data in a new host server: https://mystic.the-eye.eu
See:
- #3627
It should work if you update your URL.
We should also update the URL in our course material. | ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 38 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
Hi @ToddMorrill, people from the Pile team have mirrored their data in a new host server: https://mystic.the-eye.eu
See:
- #3627
It should work if you update your URL.
We should also update the URL in our course material. |
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | The old URL is still present in the HuggingFace course here:
https://huggingface.co/course/chapter5/4?fw=pt
I have created a PR for the Notebook here: https://github.com/huggingface/notebooks/pull/148
Not sure if the HTML is in a public repo. I wasn't able to find it. | ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 38 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
The old URL is still present in the HuggingFace course here:
https://huggingface.co/course/chapter5/4?fw=pt
I have created a PR for the Notebook here: https://github.com/huggingface/notebooks/pull/148
Not sure if the HTML is in a public repo. I wasn't able to find it. |
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | Both URLs are broken now
`HTTPError: 404 Client Error: Not Found for URL: https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst`
And
`ConnectTimeout: HTTPSConnectionPool(host='mystic.the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(, 'Connection to mystic.the-eye.eu timed out. (connect timeout=10.0)'))`
| ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 34 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
Both URLs are broken now
`HTTPError: 404 Client Error: Not Found for URL: https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst`
And
`ConnectTimeout: HTTPSConnectionPool(host='mystic.the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(, 'Connection to mystic.the-eye.eu timed out. (connect timeout=10.0)'))`
|
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | I was able to find a torrent with "The Pile" dataset here: [The Pile An 800GB Dataset of Diverse Text for Language Modeling ](https://academictorrents.com/details/0d366035664fdf51cfbe9f733953ba325776e667)
The complete dataset is huge, so I would suggest you to download only the "PUBMED_title_abstracts_2019_baseline.jsonl.zst" file, which is about 7GB. You can do this by using a torrent client of your choice (I typically utilize Transmission, which is pre-installed in Ubuntu distributions).
| ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 66 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
I was able to find a torrent with "The Pile" dataset here: [The Pile An 800GB Dataset of Diverse Text for Language Modeling ](https://academictorrents.com/details/0d366035664fdf51cfbe9f733953ba325776e667)
The complete dataset is huge, so I would suggest you to download only the "PUBMED_title_abstracts_2019_baseline.jsonl.zst" file, which is about 7GB. You can do this by using a torrent client of your choice (I typically utilize Transmission, which is pre-installed in Ubuntu distributions).
|
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | @albertvillanova another issue:
```
15 experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights()
16 File "/lfs/ampere1/0/brando9/beyond-scale-language-data-diversity/src/diversity/div_coeff.py", line 474, in experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights
17 column_names = next(iter(dataset)).keys()
18 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1353, in __iter__
19 for key, example in ex_iterable:
20 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 207, in __iter__
21 yield from self.generate_examples_fn(**self.kwargs)
22 File "/lfs/ampere1/0/brando9/.cache/huggingface/modules/datasets_modules/datasets/EleutherAI--pile/ebea56d358e91cf4d37b0fde361d563bed1472fbd8221a21b38fc8bb4ba554fb/pile.py", line 236, in _generate_examples
23 with zstd.open(open(files[subset], "rb"), "rt", encoding="utf-8") as f:
24 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/streaming.py", line 74, in wrapper
25 return function(*args, download_config=download_config, **kwargs)
26 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/download/streaming_download_manager.py", line 496, in xopen
27 file_obj = fsspec.open(file, mode=mode, *args, **kwargs).open()
28 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py", line 134, in open
29 return self.__enter__()
30 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py", line 102, in __enter__
31 f = self.fs.open(self.path, mode=mode)
32 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/spec.py", line 1241, in open
33 f = self._open(
34 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py", line 356, in _open
35 size = size or self.info(path, **kwargs)["size"]
36 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 121, in wrapper
37 return sync(self.loop, func, *args, **kwargs)
38 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 106, in sync
39 raise return_result
40 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 61, in _runner
41 result[0] = await coro
42 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py", line 430, in _info
43 raise FileNotFoundError(url) from exc
44 FileNotFoundError: https://the-eye.eu/public/AI/pile_preliminary_components/NIH_ExPORTER_awarded_grant_text.jsonl.zst
```
any suggestions? | ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 182 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
@albertvillanova another issue:
```
15 experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights()
16 File "/lfs/ampere1/0/brando9/beyond-scale-language-data-diversity/src/diversity/div_coeff.py", line 474, in experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights
17 column_names = next(iter(dataset)).keys()
18 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1353, in __iter__
19 for key, example in ex_iterable:
20 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 207, in __iter__
21 yield from self.generate_examples_fn(**self.kwargs)
22 File "/lfs/ampere1/0/brando9/.cache/huggingface/modules/datasets_modules/datasets/EleutherAI--pile/ebea56d358e91cf4d37b0fde361d563bed1472fbd8221a21b38fc8bb4ba554fb/pile.py", line 236, in _generate_examples
23 with zstd.open(open(files[subset], "rb"), "rt", encoding="utf-8") as f:
24 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/streaming.py", line 74, in wrapper
25 return function(*args, download_config=download_config, **kwargs)
26 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/download/streaming_download_manager.py", line 496, in xopen
27 file_obj = fsspec.open(file, mode=mode, *args, **kwargs).open()
28 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py", line 134, in open
29 return self.__enter__()
30 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py", line 102, in __enter__
31 f = self.fs.open(self.path, mode=mode)
32 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/spec.py", line 1241, in open
33 f = self._open(
34 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py", line 356, in _open
35 size = size or self.info(path, **kwargs)["size"]
36 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 121, in wrapper
37 return sync(self.loop, func, *args, **kwargs)
38 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 106, in sync
39 raise return_result
40 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py", line 61, in _runner
41 result[0] = await coro
42 File "/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py", line 430, in _info
43 raise FileNotFoundError(url) from exc
44 FileNotFoundError: https://the-eye.eu/public/AI/pile_preliminary_components/NIH_ExPORTER_awarded_grant_text.jsonl.zst
```
any suggestions? |
https://github.com/huggingface/datasets/issues/3504 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst | this seems to work but it's rather annoying.
Summary of how to make it work:
1. get urls to parquet files into a list
2. load list to load_dataset via `load_dataset('parquet', data_files=urls)` (note api names to hf are really confusing sometimes)
3. then it should work, print a batch of text.
presudo code
```python
urls_hacker_news = [
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00000-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00001-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00002-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00003-of-00004.parquet"
]
...
# streaming = False
from diversity.pile_subset_urls import urls_hacker_news
path, name, data_files = 'parquet', 'hacker_news', urls_hacker_news
# not changing
batch_size = 512
today = datetime.datetime.now().strftime('%Y-m%m-d%d-t%Hh_%Mm_%Ss')
run_name = f'{path} div_coeff_{num_batches=} ({today=} ({name=}) {data_mixture_name=} {probabilities=})'
print(f'{run_name=}')
# - Init wandb
debug: bool = mode == 'dryrun'
run = wandb.init(mode=mode, project="beyond-scale", name=run_name, save_code=True)
wandb.config.update({"num_batches": num_batches, "path": path, "name": name, "today": today, 'probabilities': probabilities, 'batch_size': batch_size, 'debug': debug, 'data_mixture_name': data_mixture_name, 'streaming': streaming, 'data_files': data_files})
# run.notify_on_failure() # https://community.wandb.ai/t/how-do-i-set-the-wandb-alert-programatically-for-my-current-run/4891
print(f'{debug=}')
print(f'{wandb.config=}')
# -- Get probe network
from datasets import load_dataset
import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
probe_network = GPT2LMHeadModel.from_pretrained("gpt2")
device = torch.device(f"cuda:{0}" if torch.cuda.is_available() else "cpu")
probe_network = probe_network.to(device)
# -- Get data set
def my_load_dataset(path, name):
print(f'{path=} {name=} {streaming=}')
if path == 'json' or path == 'bin' or path == 'csv':
print(f'{data_files_prefix+name=}')
return load_dataset(path, data_files=data_files_prefix+name, streaming=streaming, split="train").with_format("torch")
elif path == 'parquet':
print(f'{data_files=}')
return load_dataset(path, data_files=data_files, streaming=streaming, split="train").with_format("torch")
else:
return load_dataset(path, name, streaming=streaming, split="train").with_format("torch")
# - get data set for real now
if isinstance(path, str):
dataset = my_load_dataset(path, name)
else:
print('-- interleaving datasets')
datasets = [my_load_dataset(path, name).with_format("torch") for path, name in zip(path, name)]
[print(f'{dataset.description=}') for dataset in datasets]
dataset = interleave_datasets(datasets, probabilities)
print(f'{dataset=}')
batch = dataset.take(batch_size)
print(f'{next(iter(batch))=}')
column_names = next(iter(batch)).keys()
print(f'{column_names=}')
# - Prepare functions to tokenize batch
def preprocess(examples):
return tokenizer(examples["text"], padding="max_length", max_length=128, truncation=True, return_tensors="pt")
remove_columns = column_names # remove all keys that are not tensors to avoid bugs in collate function in task2vec's pytorch data loader
def map(batch):
return batch.map(preprocess, batched=True, remove_columns=remove_columns)
tokenized_batch = map(batch)
print(f'{next(iter(tokenized_batch))=}')
```
https://stackoverflow.com/questions/76891189/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-th/76902681#76902681
https://discuss.huggingface.co/t/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-the-files-are-not-available/50555/5?u=severo | ## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1 | 319 | Unable to download PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Describe the bug
I am unable to download the PubMed dataset from the link provided in the [Hugging Face Course (Chapter 5 Section 4)](https://huggingface.co/course/chapter5/4?fw=pt).
https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
# This takes a few minutes to run, so go grab a tea or coffee while you wait :)
data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
pubmed_dataset
```
I also tried with `wget` as follows.
```
wget https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst
```
## Expected results
I expect to be able to download this file.
## Actual results
Traceback
```
---------------------------------------------------------------------------
timeout Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
158 try:
--> 159 conn = connection.create_connection(
160 (self._dns_host, self.port), self.timeout, **extra_kw
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
83 if err is not None:
---> 84 raise err
85
/usr/lib/python3/dist-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
73 sock.bind(source_address)
---> 74 sock.connect(sa)
75 return sock
timeout: timed out
During handling of the above exception, another exception occurred:
ConnectTimeoutError Traceback (most recent call last)
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
664 # Make the request on the httplib connection object.
--> 665 httplib_response = self._make_request(
666 conn,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
375 try:
--> 376 self._validate_conn(conn)
377 except (SocketTimeout, BaseSSLError) as e:
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
995 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
--> 996 conn.connect()
997
/usr/lib/python3/dist-packages/urllib3/connection.py in connect(self)
313 # Add certificate verification
--> 314 conn = self._new_conn()
315 hostname = self.host
/usr/lib/python3/dist-packages/urllib3/connection.py in _new_conn(self)
163 except SocketTimeout:
--> 164 raise ConnectTimeoutError(
165 self,
ConnectTimeoutError: (<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)')
During handling of the above exception, another exception occurred:
MaxRetryError Traceback (most recent call last)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
438 if not chunked:
--> 439 resp = conn.urlopen(
440 method=request.method,
/usr/lib/python3/dist-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
718
--> 719 retries = retries.increment(
720 method, url, error=e, _pool=self, _stacktrace=sys.exc_info()[2]
/usr/lib/python3/dist-packages/urllib3/util/retry.py in increment(self, method, url, response, error, _pool, _stacktrace)
435 if new_retry.is_exhausted():
--> 436 raise MaxRetryError(_pool, url, error or ResponseError(cause))
437
MaxRetryError: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
During handling of the above exception, another exception occurred:
ConnectTimeout Traceback (most recent call last)
/tmp/ipykernel_15104/606583593.py in <module>
3 # This takes a few minutes to run, so go grab a tea or coffee while you wait :)
4 data_files = "https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst"
----> 5 pubmed_dataset = load_dataset("json", data_files=data_files, split="train")
6 pubmed_dataset
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, script_version, **config_kwargs)
1655
1656 # Create a dataset builder
-> 1657 builder_instance = load_dataset_builder(
1658 path=path,
1659 name=name,
~/.local/lib/python3.8/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, script_version, **config_kwargs)
1492 download_config = download_config.copy() if download_config else DownloadConfig()
1493 download_config.use_auth_token = use_auth_token
-> 1494 dataset_module = dataset_module_factory(
1495 path, revision=revision, download_config=download_config, download_mode=download_mode, data_files=data_files
1496 )
~/.local/lib/python3.8/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_files, **download_kwargs)
1116 # Try packaged
1117 if path in _PACKAGED_DATASETS_MODULES:
-> 1118 return PackagedDatasetModuleFactory(
1119 path, data_files=data_files, download_config=download_config, download_mode=download_mode
1120 ).get_module()
~/.local/lib/python3.8/site-packages/datasets/load.py in get_module(self)
773 else get_patterns_locally(str(Path().resolve()))
774 )
--> 775 data_files = DataFilesDict.from_local_or_remote(patterns, use_auth_token=self.downnload_config.use_auth_token)
776 module_path, hash = _PACKAGED_DATASETS_MODULES[self.name]
777 builder_kwargs = {"hash": hash, "data_files": data_files}
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
576 for key, patterns_for_key in patterns.items():
577 out[key] = (
--> 578 DataFilesList.from_local_or_remote(
579 patterns_for_key,
580 base_path=base_path,
~/.local/lib/python3.8/site-packages/datasets/data_files.py in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)
545 base_path = base_path if base_path is not None else str(Path().resolve())
546 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)
--> 547 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)
548 return cls(data_files, origin_metadata)
549
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_origin_metadata_locally_or_by_urls(data_files, max_workers, use_auth_token)
492 data_files: List[Union[Path, Url]], max_workers=64, use_auth_token: Optional[Union[bool, str]] = None
493 ) -> Tuple[str]:
--> 494 return thread_map(
495 partial(_get_single_origin_metadata_locally_or_by_urls, use_auth_token=use_auth_token),
496 data_files,
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in thread_map(fn, *iterables, **tqdm_kwargs)
92 """
93 from concurrent.futures import ThreadPoolExecutor
---> 94 return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
95
96
~/.local/lib/python3.8/site-packages/tqdm/contrib/concurrent.py in _executor_map(PoolExecutor, fn, *iterables, **tqdm_kwargs)
74 map_args.update(chunksize=chunksize)
75 with PoolExecutor(**pool_kwargs) as ex:
---> 76 return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))
77
78
~/.local/lib/python3.8/site-packages/tqdm/notebook.py in __iter__(self)
252 def __iter__(self):
253 try:
--> 254 for obj in super(tqdm_notebook, self).__iter__():
255 # return super(tqdm...) will not catch exception
256 yield obj
~/.local/lib/python3.8/site-packages/tqdm/std.py in __iter__(self)
1171 # (note: keep this check outside the loop for performance)
1172 if self.disable:
-> 1173 for obj in iterable:
1174 yield obj
1175 return
/usr/lib/python3.8/concurrent/futures/_base.py in result_iterator()
617 # Careful not to keep a reference to the popped future
618 if timeout is None:
--> 619 yield fs.pop().result()
620 else:
621 yield fs.pop().result(end_time - time.monotonic())
/usr/lib/python3.8/concurrent/futures/_base.py in result(self, timeout)
442 raise CancelledError()
443 elif self._state == FINISHED:
--> 444 return self.__get_result()
445 else:
446 raise TimeoutError()
/usr/lib/python3.8/concurrent/futures/_base.py in __get_result(self)
387 if self._exception:
388 try:
--> 389 raise self._exception
390 finally:
391 # Break a reference cycle with the exception in self._exception
/usr/lib/python3.8/concurrent/futures/thread.py in run(self)
55
56 try:
---> 57 result = self.fn(*self.args, **self.kwargs)
58 except BaseException as exc:
59 self.future.set_exception(exc)
~/.local/lib/python3.8/site-packages/datasets/data_files.py in _get_single_origin_metadata_locally_or_by_urls(data_file, use_auth_token)
483 if isinstance(data_file, Url):
484 data_file = str(data_file)
--> 485 return (request_etag(data_file, use_auth_token=use_auth_token),)
486 else:
487 data_file = str(data_file.resolve())
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in request_etag(url, use_auth_token)
489 def request_etag(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> Optional[str]:
490 headers = get_authentication_headers_for_url(url, use_auth_token=use_auth_token)
--> 491 response = http_head(url, headers=headers, max_retries=3)
492 response.raise_for_status()
493 etag = response.headers.get("ETag") if response.ok else None
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
474 headers = copy.deepcopy(headers) or {}
475 headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
--> 476 response = _request_with_retry(
477 method="HEAD",
478 url=url,
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
408 if tries > max_retries:
--> 409 raise err
410 else:
411 logger.info(f"{method} request to {url} timed out, retrying... [{tries/max_retries}]")
~/.local/lib/python3.8/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
403 tries += 1
404 try:
--> 405 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
406 success = True
407 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/usr/lib/python3/dist-packages/requests/api.py in request(method, url, **kwargs)
58 # cases, and look like a memory leak in others.
59 with sessions.Session() as session:
---> 60 return session.request(method=method, url=url, **kwargs)
61
62
/usr/lib/python3/dist-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
531 }
532 send_kwargs.update(settings)
--> 533 resp = self.send(prep, **send_kwargs)
534
535 return resp
/usr/lib/python3/dist-packages/requests/sessions.py in send(self, request, **kwargs)
644
645 # Send the request
--> 646 r = adapter.send(request, **kwargs)
647
648 # Total elapsed time of the request (approximately)
/usr/lib/python3/dist-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
502 # TODO: Remove this in 3.0.0: see #2811
503 if not isinstance(e.reason, NewConnectionError):
--> 504 raise ConnectTimeout(e, request=request)
505
506 if isinstance(e.reason, ResponseError):
ConnectTimeout: HTTPSConnectionPool(host='the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(<urllib3.connection.VerifiedHTTPSConnection object at 0x7f06dd698850>, 'Connection to the-eye.eu timed out. (connect timeout=10.0)'))
```
## Environment info
- `datasets` version: 1.17.0
- Platform: Linux-5.11.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 6.0.1
this seems to work but it's rather annoying.
Summary of how to make it work:
1. get urls to parquet files into a list
2. load list to load_dataset via `load_dataset('parquet', data_files=urls)` (note api names to hf are really confusing sometimes)
3. then it should work, print a batch of text.
presudo code
```python
urls_hacker_news = [
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00000-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00001-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00002-of-00004.parquet",
"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00003-of-00004.parquet"
]
...
# streaming = False
from diversity.pile_subset_urls import urls_hacker_news
path, name, data_files = 'parquet', 'hacker_news', urls_hacker_news
# not changing
batch_size = 512
today = datetime.datetime.now().strftime('%Y-m%m-d%d-t%Hh_%Mm_%Ss')
run_name = f'{path} div_coeff_{num_batches=} ({today=} ({name=}) {data_mixture_name=} {probabilities=})'
print(f'{run_name=}')
# - Init wandb
debug: bool = mode == 'dryrun'
run = wandb.init(mode=mode, project="beyond-scale", name=run_name, save_code=True)
wandb.config.update({"num_batches": num_batches, "path": path, "name": name, "today": today, 'probabilities': probabilities, 'batch_size': batch_size, 'debug': debug, 'data_mixture_name': data_mixture_name, 'streaming': streaming, 'data_files': data_files})
# run.notify_on_failure() # https://community.wandb.ai/t/how-do-i-set-the-wandb-alert-programatically-for-my-current-run/4891
print(f'{debug=}')
print(f'{wandb.config=}')
# -- Get probe network
from datasets import load_dataset
import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
probe_network = GPT2LMHeadModel.from_pretrained("gpt2")
device = torch.device(f"cuda:{0}" if torch.cuda.is_available() else "cpu")
probe_network = probe_network.to(device)
# -- Get data set
def my_load_dataset(path, name):
print(f'{path=} {name=} {streaming=}')
if path == 'json' or path == 'bin' or path == 'csv':
print(f'{data_files_prefix+name=}')
return load_dataset(path, data_files=data_files_prefix+name, streaming=streaming, split="train").with_format("torch")
elif path == 'parquet':
print(f'{data_files=}')
return load_dataset(path, data_files=data_files, streaming=streaming, split="train").with_format("torch")
else:
return load_dataset(path, name, streaming=streaming, split="train").with_format("torch")
# - get data set for real now
if isinstance(path, str):
dataset = my_load_dataset(path, name)
else:
print('-- interleaving datasets')
datasets = [my_load_dataset(path, name).with_format("torch") for path, name in zip(path, name)]
[print(f'{dataset.description=}') for dataset in datasets]
dataset = interleave_datasets(datasets, probabilities)
print(f'{dataset=}')
batch = dataset.take(batch_size)
print(f'{next(iter(batch))=}')
column_names = next(iter(batch)).keys()
print(f'{column_names=}')
# - Prepare functions to tokenize batch
def preprocess(examples):
return tokenizer(examples["text"], padding="max_length", max_length=128, truncation=True, return_tensors="pt")
remove_columns = column_names # remove all keys that are not tensors to avoid bugs in collate function in task2vec's pytorch data loader
def map(batch):
return batch.map(preprocess, batched=True, remove_columns=remove_columns)
tokenized_batch = map(batch)
print(f'{next(iter(tokenized_batch))=}')
```
https://stackoverflow.com/questions/76891189/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-th/76902681#76902681
https://discuss.huggingface.co/t/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-the-files-are-not-available/50555/5?u=severo |
https://github.com/huggingface/datasets/issues/3499 | Adjusting chunk size for streaming datasets | Hi ! Data streaming uses `fsspec` to read the data files progressively. IIRC the block size for buffering is 5MiB by default. So every time you finish iterating over a block, it downloads the next one. You can still try to increase the `fsspec` block size for buffering if it can help. To do so you just need to increase `fsspec.spec.AbstractBufferedFile.DEFAULT_BLOCK_SIZE `
Currently this is unfortunately done in a single thread, so it blocks the processing to download and uncompress the next block. At one point it would be nice to be able to do that in parallel ! | **Is your feature request related to a problem? Please describe.**
I want to use mc4 which I cannot save locally, so I stream it. However, I want to process the entire dataset and filter some documents from it. With the current chunk size of around 1000 documents (right?) I hit a performance bottleneck because of the frequent decompressing.
**Describe the solution you'd like**
I would appreciate a parameter in the load_dataset function, that allows me to set the chunksize myself (to a value like 100'000 in my case). Like that, I hope to improve the processing time.
| 99 | Adjusting chunk size for streaming datasets
**Is your feature request related to a problem? Please describe.**
I want to use mc4 which I cannot save locally, so I stream it. However, I want to process the entire dataset and filter some documents from it. With the current chunk size of around 1000 documents (right?) I hit a performance bottleneck because of the frequent decompressing.
**Describe the solution you'd like**
I would appreciate a parameter in the load_dataset function, that allows me to set the chunksize myself (to a value like 100'000 in my case). Like that, I hope to improve the processing time.
Hi ! Data streaming uses `fsspec` to read the data files progressively. IIRC the block size for buffering is 5MiB by default. So every time you finish iterating over a block, it downloads the next one. You can still try to increase the `fsspec` block size for buffering if it can help. To do so you just need to increase `fsspec.spec.AbstractBufferedFile.DEFAULT_BLOCK_SIZE `
Currently this is unfortunately done in a single thread, so it blocks the processing to download and uncompress the next block. At one point it would be nice to be able to do that in parallel ! |
https://github.com/huggingface/datasets/issues/3490 | Does datasets support load text from HDFS? | Hi ! `datasets` currently supports reading local files or files over HTTP. We may add support for other filesystems (cloud storages, hdfs...) at one point though :) | The raw text data is stored on HDFS due to the dataset's size is too large to store on my develop machine,
so I wander does datasets support read data from hdfs? | 27 | Does datasets support load text from HDFS?
The raw text data is stored on HDFS due to the dataset's size is too large to store on my develop machine,
so I wander does datasets support read data from hdfs?
Hi ! `datasets` currently supports reading local files or files over HTTP. We may add support for other filesystems (cloud storages, hdfs...) at one point though :) |
https://github.com/huggingface/datasets/issues/3488 | URL query parameters are set as path in the compression hop for fsspec | I think the test passes because it simply ignore what's after `gzip://`.
The returned urlpath is expected to look like `gzip://filename::url`, and the filename is currently considered to be what's after the final `/`, hence the result.
We can decide to change this and simply have `gzip://::url`, this way we don't need to guess the filename, what do you think ? | ## Describe the bug
There is an ssue with `StreamingDownloadManager._extract`.
I don't know how the test `test_streaming_gg_drive_gzipped` passes:
For
```python
TEST_GG_DRIVE_GZIPPED_URL = "https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz"
urlpath = StreamingDownloadManager().download_and_extract(TEST_GG_DRIVE_GZIPPED_URL)
```
gives `urlpath`:
```python
'gzip://uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz::https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz'
```
The gzip path makes no sense: `gzip://uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz`
## Steps to reproduce the bug
```python
from datasets.utils.streaming_download_manager import StreamingDownloadManager
dl_manager = StreamingDownloadManager()
urlpath = dl_manager.extract("https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz")
print(urlpath)
```
## Expected results
The query parameters should not be set as part of the path.
| 61 | URL query parameters are set as path in the compression hop for fsspec
## Describe the bug
There is an ssue with `StreamingDownloadManager._extract`.
I don't know how the test `test_streaming_gg_drive_gzipped` passes:
For
```python
TEST_GG_DRIVE_GZIPPED_URL = "https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz"
urlpath = StreamingDownloadManager().download_and_extract(TEST_GG_DRIVE_GZIPPED_URL)
```
gives `urlpath`:
```python
'gzip://uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz::https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz'
```
The gzip path makes no sense: `gzip://uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz`
## Steps to reproduce the bug
```python
from datasets.utils.streaming_download_manager import StreamingDownloadManager
dl_manager = StreamingDownloadManager()
urlpath = dl_manager.extract("https://drive.google.com/uc?export=download&id=1Bt4Garpf0QLiwkJhHJzXaVa0I0H5Qhwz")
print(urlpath)
```
## Expected results
The query parameters should not be set as part of the path.
I think the test passes because it simply ignore what's after `gzip://`.
The returned urlpath is expected to look like `gzip://filename::url`, and the filename is currently considered to be what's after the final `/`, hence the result.
We can decide to change this and simply have `gzip://::url`, this way we don't need to guess the filename, what do you think ? |
https://github.com/huggingface/datasets/issues/3485 | skip columns which cannot set to specific format when set_format | You can add columns that you wish to set into `torch` format using `dataset.set_format("torch", ['id', 'abc'])` so that input batch of the transform only contains those columns | **Is your feature request related to a problem? Please describe.**
When using `dataset.set_format("torch")`, I must make sure every columns in datasets can convert to `torch`, however, sometimes I want to keep some string columns.
**Describe the solution you'd like**
skip columns which cannot set to specific format when set_format instead of raise an error.
| 27 | skip columns which cannot set to specific format when set_format
**Is your feature request related to a problem? Please describe.**
When using `dataset.set_format("torch")`, I must make sure every columns in datasets can convert to `torch`, however, sometimes I want to keep some string columns.
**Describe the solution you'd like**
skip columns which cannot set to specific format when set_format instead of raise an error.
You can add columns that you wish to set into `torch` format using `dataset.set_format("torch", ['id', 'abc'])` so that input batch of the transform only contains those columns |
https://github.com/huggingface/datasets/issues/3485 | skip columns which cannot set to specific format when set_format | Sorry, I miss `output_all_columns` args and thought after `dataset.set_format("torch", columns=columns)` I can only get specific columns I assigned. | **Is your feature request related to a problem? Please describe.**
When using `dataset.set_format("torch")`, I must make sure every columns in datasets can convert to `torch`, however, sometimes I want to keep some string columns.
**Describe the solution you'd like**
skip columns which cannot set to specific format when set_format instead of raise an error.
| 18 | skip columns which cannot set to specific format when set_format
**Is your feature request related to a problem? Please describe.**
When using `dataset.set_format("torch")`, I must make sure every columns in datasets can convert to `torch`, however, sometimes I want to keep some string columns.
**Describe the solution you'd like**
skip columns which cannot set to specific format when set_format instead of raise an error.
Sorry, I miss `output_all_columns` args and thought after `dataset.set_format("torch", columns=columns)` I can only get specific columns I assigned. |
https://github.com/huggingface/datasets/issues/3484 | make shape verification to use ArrayXD instead of nested lists for map | Hi!
Yes, this makes sense for numeric values, but first I have to finish https://github.com/huggingface/datasets/pull/3336 because currently ArrayXD only allows the first dimension to be dynamic. | As describe in https://github.com/huggingface/datasets/issues/2005#issuecomment-793716753 and mentioned by @mariosasko in [image feature example](https://colab.research.google.com/drive/1mIrTnqTVkWLJWoBzT1ABSe-LFelIep1c#scrollTo=ow3XHDvf2I0B&line=1&uniqifier=1), IMO make shape verifcaiton to use ArrayXD instead of nested lists for map can help user reduce unnecessary cast. I notice datasets have done something special for `input_ids` and `attention_mask` which is also unnecessary after this feature added. | 26 | make shape verification to use ArrayXD instead of nested lists for map
As describe in https://github.com/huggingface/datasets/issues/2005#issuecomment-793716753 and mentioned by @mariosasko in [image feature example](https://colab.research.google.com/drive/1mIrTnqTVkWLJWoBzT1ABSe-LFelIep1c#scrollTo=ow3XHDvf2I0B&line=1&uniqifier=1), IMO make shape verifcaiton to use ArrayXD instead of nested lists for map can help user reduce unnecessary cast. I notice datasets have done something special for `input_ids` and `attention_mask` which is also unnecessary after this feature added.
Hi!
Yes, this makes sense for numeric values, but first I have to finish https://github.com/huggingface/datasets/pull/3336 because currently ArrayXD only allows the first dimension to be dynamic. |
https://github.com/huggingface/datasets/issues/3480 | the compression format requested when saving a dataset in json format is not respected | Thanks for reporting @SaulLu.
At first sight I think the problem is caused because `pandas` only takes into account the `compression` parameter if called with a non-null file path or buffer. And in our implementation, we call pandas `to_json` with `None` `path_or_buf`.
We should fix this:
- either handling directly the `compression` parameter ourselves
- or refactoring to pass non-null path or buffer to pandas
CC: @lhoestq | ## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
| 67 | the compression format requested when saving a dataset in json format is not respected
## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
Thanks for reporting @SaulLu.
At first sight I think the problem is caused because `pandas` only takes into account the `compression` parameter if called with a non-null file path or buffer. And in our implementation, we call pandas `to_json` with `None` `path_or_buf`.
We should fix this:
- either handling directly the `compression` parameter ourselves
- or refactoring to pass non-null path or buffer to pandas
CC: @lhoestq |
https://github.com/huggingface/datasets/issues/3480 | the compression format requested when saving a dataset in json format is not respected | I was thinking if we can handle the `compression` parameter by ourselves? Compression types will be similar to what `pandas` offer. Initially, we can try this with 2-3 compression types and see how good/bad it is? Let me know if it sounds good, I can raise a PR for this next week | ## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
| 52 | the compression format requested when saving a dataset in json format is not respected
## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
I was thinking if we can handle the `compression` parameter by ourselves? Compression types will be similar to what `pandas` offer. Initially, we can try this with 2-3 compression types and see how good/bad it is? Let me know if it sounds good, I can raise a PR for this next week |
https://github.com/huggingface/datasets/issues/3480 | the compression format requested when saving a dataset in json format is not respected | Hi ! Thanks for your help @bhavitvyamalik :)
Maybe let's start with `gzip` ? I think it's the most common use case, then if we're fine with it we can add other compression methods | ## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
| 34 | the compression format requested when saving a dataset in json format is not respected
## Describe the bug
In the documentation of the `to_json` method, it is stated in the parameters that
> **to_json_kwargs – Parameters to pass to pandas’s pandas.DataFrame.to_json.
however when we pass for example `compression="gzip"`, the saved file is not compressed.
Would you also have expected compression to be applied? :relaxed:
## Steps to reproduce the bug
```python
my_dict = {"a": [1, 2, 3], "b": [1, 2, 3]}
```
### Result with datasets
```python
from datasets import Dataset
dataset = Dataset.from_dict(my_dict)
dataset.to_json("dic_with_datasets.jsonl.gz", compression="gzip")
!cat dic_with_datasets.jsonl.gz
```
output
```
{"a":1,"b":1}
{"a":2,"b":2}
{"a":3,"b":3}
```
Note: I would expected to see binary data here
### Result with pandas
```python
import pandas as pd
df = pd.DataFrame(my_dict)
df.to_json("dic_with_pandas.jsonl.gz", lines=True, orient="records", compression="gzip")
!cat dic_with_pandas.jsonl.gz
```
output
```
4��a�dic_with_pandas.jsonl��VJT�2�QJ��\� ��g��yƵ���������)���
```
Note: It looks like binary data
## Expected results
I would have expected that the saved result with datasets would also be a binary file
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-4.18.0-193.70.1.el8_2.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.11
- PyArrow version: 5.0.0
Hi ! Thanks for your help @bhavitvyamalik :)
Maybe let's start with `gzip` ? I think it's the most common use case, then if we're fine with it we can add other compression methods |
https://github.com/huggingface/datasets/issues/3475 | The rotten_tomatoes dataset of movie reviews contains some reviews in Spanish | Hi @puzzler10, thanks for reporting.
Please note this dataset is not hosted on Hugging Face Hub. See:
https://github.com/huggingface/datasets/blob/c8f914473b041833fd47178fa4373cdcb56ac522/datasets/rotten_tomatoes/rotten_tomatoes.py#L42
If there are issues with the source data of a dataset, you should contact the data owners/creators instead. In the homepage associated with this dataset (http://www.cs.cornell.edu/people/pabo/movie-review-data/), you can find the authors of the dataset and how to contact them:
> If you have any questions or comments regarding this site, please send email to Bo Pang or Lillian Lee.
P.S.: Please also note that the example you gave of non-English review is in Portuguese (not Spanish). ;) | ## Describe the bug
See title. I don't think this is intentional and they probably should be removed. If they stay the dataset description should be at least updated to make it clear to the user.
## Steps to reproduce the bug
Go to the [dataset viewer](https://huggingface.co/datasets/viewer/?dataset=rotten_tomatoes) for the dataset, set the offset to 4160 for the train dataset, and scroll through the results. I found ones at index 4166 and 4173. There's others too (e.g. index 2888) but those two are easy to find like that.
## Expected results
English movie reviews only.
## Actual results
Example of a Spanish movie review (4173):
> "É uma pena que , mais tarde , o próprio filme abandone o tom de paródia e passe a utilizar os mesmos clichês que havia satirizado "
| 95 | The rotten_tomatoes dataset of movie reviews contains some reviews in Spanish
## Describe the bug
See title. I don't think this is intentional and they probably should be removed. If they stay the dataset description should be at least updated to make it clear to the user.
## Steps to reproduce the bug
Go to the [dataset viewer](https://huggingface.co/datasets/viewer/?dataset=rotten_tomatoes) for the dataset, set the offset to 4160 for the train dataset, and scroll through the results. I found ones at index 4166 and 4173. There's others too (e.g. index 2888) but those two are easy to find like that.
## Expected results
English movie reviews only.
## Actual results
Example of a Spanish movie review (4173):
> "É uma pena que , mais tarde , o próprio filme abandone o tom de paródia e passe a utilizar os mesmos clichês que havia satirizado "
Hi @puzzler10, thanks for reporting.
Please note this dataset is not hosted on Hugging Face Hub. See:
https://github.com/huggingface/datasets/blob/c8f914473b041833fd47178fa4373cdcb56ac522/datasets/rotten_tomatoes/rotten_tomatoes.py#L42
If there are issues with the source data of a dataset, you should contact the data owners/creators instead. In the homepage associated with this dataset (http://www.cs.cornell.edu/people/pabo/movie-review-data/), you can find the authors of the dataset and how to contact them:
> If you have any questions or comments regarding this site, please send email to Bo Pang or Lillian Lee.
P.S.: Please also note that the example you gave of non-English review is in Portuguese (not Spanish). ;) |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | As discussed, I remember I set `decoded=False` here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the "audio" field (for Audio feature) was accessed. | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 32 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
As discussed, I remember I set `decoded=False` here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the "audio" field (for Audio feature) was accessed. |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | > I set decoded=False here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the "audio" field (for Audio feature) was accessed
https://github.com/huggingface/datasets/pull/3430 will add more control to decoding, so I think it's OK to enable decoding in `__iter__` for now. After we merge the linked PR, the user can easily disable it again. | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 61 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
> I set decoded=False here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the "audio" field (for Audio feature) was accessed
https://github.com/huggingface/datasets/pull/3430 will add more control to decoding, so I think it's OK to enable decoding in `__iter__` for now. After we merge the linked PR, the user can easily disable it again. |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | @mariosasko I wonder why there is no issue in `Audio` feature with decoding disabled in `__iter__`, whereas there is in `Image` feature.
Enabling decoding in `__iter__` will make fail Audio regressions tests: https://github.com/huggingface/datasets/runs/4608657230?check_suite_focus=true
```
=========================== short test summary info ============================
FAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_not_decoded
FAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_decoded
========================= 2 failed, 15 passed in 8.37s ========================= | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 52 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
@mariosasko I wonder why there is no issue in `Audio` feature with decoding disabled in `__iter__`, whereas there is in `Image` feature.
Enabling decoding in `__iter__` will make fail Audio regressions tests: https://github.com/huggingface/datasets/runs/4608657230?check_suite_focus=true
```
=========================== short test summary info ============================
FAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_not_decoded
FAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_decoded
========================= 2 failed, 15 passed in 8.37s ========================= |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | Please also note that the regression tests were implemented in accordance with the specifications:
- when doing a `map` (wich calls `__iter__`) of a function that doesn't access the audio field, the decoding should be disabled; this is why the decoding is disabled in `__iter__` (and only enabled in `__getitem__`). | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 50 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
Please also note that the regression tests were implemented in accordance with the specifications:
- when doing a `map` (wich calls `__iter__`) of a function that doesn't access the audio field, the decoding should be disabled; this is why the decoding is disabled in `__iter__` (and only enabled in `__getitem__`). |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | > I wonder why there is no issue in Audio feature with decoding disabled in __iter__, whereas there is in Image feature.
@albertvillanova Not sure if I understand this part. Currently, both the Image and the Audio feature don't decode data in `__iter__`, so their behavior is aligned there.
| ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 49 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
> I wonder why there is no issue in Audio feature with decoding disabled in __iter__, whereas there is in Image feature.
@albertvillanova Not sure if I understand this part. Currently, both the Image and the Audio feature don't decode data in `__iter__`, so their behavior is aligned there.
|
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | Therefore, this is not an issue, neither for Audio nor Image feature.
Could you please elaborate more on the expected use case? @lhoestq @NielsRogge
The expected use cases (in accordance with the specs: see #2324):
- decoding should be enabled when accessing a specific item (`__getitem__`)
- decoding should be disabled while iterating (`__iter__`) to allow preprocessing of non-audio/image features (like label or text, for example) using `.map`
- decoding should be enabled in a `.map` only if the `.map` function accesses the audio/image feature (implemented using `LazyDict`) | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 88 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
Therefore, this is not an issue, neither for Audio nor Image feature.
Could you please elaborate more on the expected use case? @lhoestq @NielsRogge
The expected use cases (in accordance with the specs: see #2324):
- decoding should be enabled when accessing a specific item (`__getitem__`)
- decoding should be disabled while iterating (`__iter__`) to allow preprocessing of non-audio/image features (like label or text, for example) using `.map`
- decoding should be enabled in a `.map` only if the `.map` function accesses the audio/image feature (implemented using `LazyDict`) |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | For me it's not an issue, actually. I just (mistakenly) tried to iterate over a PyTorch Dataset instead of a PyTorch DataLoader,
i.e. I did this:
`batch = next(iter(train_ds)) `
whereas I actually wanted to do
`batch = next(iter(train_dataloader))`
and then it turned out that in the first case, the image was a string of bytes rather than a Pillow image, hence Quentin opened an issue. | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 66 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
For me it's not an issue, actually. I just (mistakenly) tried to iterate over a PyTorch Dataset instead of a PyTorch DataLoader,
i.e. I did this:
`batch = next(iter(train_ds)) `
whereas I actually wanted to do
`batch = next(iter(train_dataloader))`
and then it turned out that in the first case, the image was a string of bytes rather than a Pillow image, hence Quentin opened an issue. |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | Thanks @NielsRogge for the context.
So IMO everything is working as expected.
I'm closing this issue. Feel free to reopen it again if further changes of the specs should be addressed. | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 31 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
Thanks @NielsRogge for the context.
So IMO everything is working as expected.
I'm closing this issue. Feel free to reopen it again if further changes of the specs should be addressed. |
https://github.com/huggingface/datasets/issues/3473 | Iterating over a vision dataset doesn't decode the images | Thanks for the details :)
I still think that it's unexpected to get different results when doing
```python
for i in range(len(dataset)):
sample = dataset[i]
```
and
```python
for sample in dataset:
pass
```
even though I understand that if you don't need to decode the data, then decoding image or audio data when iterating is a waste of time and resources.
But in this case users can still drop the column that need decoding to get the full speed back no ? | ## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
| 83 | Iterating over a vision dataset doesn't decode the images
## Describe the bug
If I load `mnist` and I iterate over the dataset, the images are not decoded, and the dictionary with the bytes is returned.
## Steps to reproduce the bug
```python
from datasets import load_dataset
import PIL
mnist = load_dataset("mnist", split="train")
first_image = mnist[0]["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # passes
first_image = next(iter(mnist))["image"]
assert isinstance(first_image, PIL.PngImagePlugin.PngImageFile) # fails
```
## Expected results
The image should be decoded, as a PIL Image
## Actual results
We get a dictionary
```
{'bytes': b'\x89PNG\r\n\x1a\n\x00..., 'path': None}
```
## Environment info
- `datasets` version: 1.17.1.dev0
- Platform: Darwin-20.6.0-x86_64-i386-64bit
- Python version: 3.7.2
- PyArrow version: 6.0.0
The bug also exists in 1.17.0
## Investigation
I think the issue is that decoding is disabled in `__iter__`:
https://github.com/huggingface/datasets/blob/dfe5b73387c5e27de6a16b0caeb39d3b9ded66d6/src/datasets/arrow_dataset.py#L1651-L1661
Do you remember why it was disabled in the first place @albertvillanova ?
Also cc @mariosasko @NielsRogge
Thanks for the details :)
I still think that it's unexpected to get different results when doing
```python
for i in range(len(dataset)):
sample = dataset[i]
```
and
```python
for sample in dataset:
pass
```
even though I understand that if you don't need to decode the data, then decoding image or audio data when iterating is a waste of time and resources.
But in this case users can still drop the column that need decoding to get the full speed back no ? |
https://github.com/huggingface/datasets/issues/3465 | Unable to load 'cnn_dailymail' dataset | Hi @talha1503, thanks for reporting.
It seems there is an issue with one of the data files hosted at Google Drive:
```
Google Drive - Quota exceeded
Sorry, you can't view or download this file at this time.
Too many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.
```
As you probably know, Hugging Face does not host the data, and in this case the data owner decided to host their data at Google Drive, which has quota limits.
Is there anything we could do, @lhoestq @mariosasko? | ## Describe the bug
I wanted to load cnn_dailymail dataset from huggingface datasets on Google Colab, but I am getting an error while loading it.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('cnn_dailymail', '3.0.0', ignore_verifications = True)
```
## Expected results
Expecting to load 'cnn_dailymail' dataset.
## Actual results
`NotADirectoryError: [Errno 20] Not a directory: '/root/.cache/huggingface/datasets/downloads/1bc05d24fa6dda2468e83a73cf6dc207226e01e3c48a507ea716dc0421da583b/cnn/stories'`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-5.4.104+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.12
- PyArrow version: 3.0.0
| 141 | Unable to load 'cnn_dailymail' dataset
## Describe the bug
I wanted to load cnn_dailymail dataset from huggingface datasets on Google Colab, but I am getting an error while loading it.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('cnn_dailymail', '3.0.0', ignore_verifications = True)
```
## Expected results
Expecting to load 'cnn_dailymail' dataset.
## Actual results
`NotADirectoryError: [Errno 20] Not a directory: '/root/.cache/huggingface/datasets/downloads/1bc05d24fa6dda2468e83a73cf6dc207226e01e3c48a507ea716dc0421da583b/cnn/stories'`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux-5.4.104+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.12
- PyArrow version: 3.0.0
Hi @talha1503, thanks for reporting.
It seems there is an issue with one of the data files hosted at Google Drive:
```
Google Drive - Quota exceeded
Sorry, you can't view or download this file at this time.
Too many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.
```
As you probably know, Hugging Face does not host the data, and in this case the data owner decided to host their data at Google Drive, which has quota limits.
Is there anything we could do, @lhoestq @mariosasko? |
https://github.com/huggingface/datasets/issues/3464 | struct.error: 'i' format requires -2147483648 <= number <= 2147483647 | Hi ! Can you try setting `datasets.config.MAX_TABLE_NBYTES_FOR_PICKLING` to a smaller value than `4 << 30` (4GiB), for example `500 << 20` (500MiB) ? It should reduce the maximum size of the arrow table being pickled during multiprocessing.
If it fixes the issue, we can consider lowering the default value for everyone. | ## Describe the bug
A clear and concise description of what the bug is.
using latest datasets=datasets-1.16.1-py3-none-any.whl
process my own multilingual dataset by following codes, and the number of rows in all dataset is 306000, the max_length of each sentence is 256:
then I get this error:
I have seen the issue in #2134 and #2150, so I don't understand why latest repo still can't deal with big dataset.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: linux docker
- Python version: 3.6
| 51 | struct.error: 'i' format requires -2147483648 <= number <= 2147483647
## Describe the bug
A clear and concise description of what the bug is.
using latest datasets=datasets-1.16.1-py3-none-any.whl
process my own multilingual dataset by following codes, and the number of rows in all dataset is 306000, the max_length of each sentence is 256:
then I get this error:
I have seen the issue in #2134 and #2150, so I don't understand why latest repo still can't deal with big dataset.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: linux docker
- Python version: 3.6
Hi ! Can you try setting `datasets.config.MAX_TABLE_NBYTES_FOR_PICKLING` to a smaller value than `4 << 30` (4GiB), for example `500 << 20` (500MiB) ? It should reduce the maximum size of the arrow table being pickled during multiprocessing.
If it fixes the issue, we can consider lowering the default value for everyone. |
https://github.com/huggingface/datasets/issues/3457 | Add CMU Graphics Lab Motion Capture dataset | This dataset has files in ASF/AMC format. [ The skeleton file is the ASF file (Acclaim Skeleton File). The motion file is the AMC file (Acclaim Motion Capture data). ]
Some questions :
1. How do we go about representing these features using datasets.Features and generate examples ?
2. The dataset download link for ASF/AMC files does not have metadata information, for eg : category and subcategory information. We will need to crawl the website for this information. The authors mention "Please don't crawl this database for all motions." Can we mail the authors for this information ?
The dataset structure is as follows :
```
subjects
- 01
- 01_01.amc
- 01_02.amc
.
.
.
- 01.asf
- 02
- 02_01.amc
- 02_02.amc
.
.
.
- 02.asf
```
There is no metadata regarding the category, sub-category and motion description.
Need your inputs. @mariosasko / @lhoestq
Thank you.
| ## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md). | 149 | Add CMU Graphics Lab Motion Capture dataset
## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
This dataset has files in ASF/AMC format. [ The skeleton file is the ASF file (Acclaim Skeleton File). The motion file is the AMC file (Acclaim Motion Capture data). ]
Some questions :
1. How do we go about representing these features using datasets.Features and generate examples ?
2. The dataset download link for ASF/AMC files does not have metadata information, for eg : category and subcategory information. We will need to crawl the website for this information. The authors mention "Please don't crawl this database for all motions." Can we mail the authors for this information ?
The dataset structure is as follows :
```
subjects
- 01
- 01_01.amc
- 01_02.amc
.
.
.
- 01.asf
- 02
- 02_01.amc
- 02_02.amc
.
.
.
- 02.asf
```
There is no metadata regarding the category, sub-category and motion description.
Need your inputs. @mariosasko / @lhoestq
Thank you.
|
https://github.com/huggingface/datasets/issues/3457 | Add CMU Graphics Lab Motion Capture dataset | Hi @dnaveenr! Thanks for working on this!
1. We can use the `Sequence(Value("string"))` feature type for the subject's AMC files and `Value("string")` for the subject's ASF file (`Value("string")` represents the file paths) + the types for categories/subcategories and descriptions.
2. We can use this URL to download the motion descriptions: http://mocap.cs.cmu.edu/search.php?subjectnumber=<subject_number>&motion=%%%&maincat=%&subcat=%&subtext=yes where `subject_number` is the number between 1 and 144. And to get categories/subcategories, feel free to contact the authors (they state in the FAQ they are happy to help) and ask them if they can provide the mapping from categories/subcategories to the AMC files to avoid crawling. You can also mention that your goal is to make their dataset more accessible by adding its loading script to the Hub.
The AMC files are also available in the tvd, c3d, mpg and avi formats (the links are in the [FAQ](http://mocap.cs.cmu.edu/faqs.php) section), so it would be nice to have one config for each of these additional formats.
And additionally, we can add a `Data Preprocessing` section to the card where we explain how to load/process the files. I can help with that. | ## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md). | 181 | Add CMU Graphics Lab Motion Capture dataset
## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Hi @dnaveenr! Thanks for working on this!
1. We can use the `Sequence(Value("string"))` feature type for the subject's AMC files and `Value("string")` for the subject's ASF file (`Value("string")` represents the file paths) + the types for categories/subcategories and descriptions.
2. We can use this URL to download the motion descriptions: http://mocap.cs.cmu.edu/search.php?subjectnumber=<subject_number>&motion=%%%&maincat=%&subcat=%&subtext=yes where `subject_number` is the number between 1 and 144. And to get categories/subcategories, feel free to contact the authors (they state in the FAQ they are happy to help) and ask them if they can provide the mapping from categories/subcategories to the AMC files to avoid crawling. You can also mention that your goal is to make their dataset more accessible by adding its loading script to the Hub.
The AMC files are also available in the tvd, c3d, mpg and avi formats (the links are in the [FAQ](http://mocap.cs.cmu.edu/faqs.php) section), so it would be nice to have one config for each of these additional formats.
And additionally, we can add a `Data Preprocessing` section to the card where we explain how to load/process the files. I can help with that. |
https://github.com/huggingface/datasets/issues/3457 | Add CMU Graphics Lab Motion Capture dataset | Hi @mariosasko ,
1. Thanks for this, so we can add the file paths.
2. Yes, I had already mailed the authors a couple of days back actually, asking for the metadata details[ i.e category, sub-category and motion description] . They are yet to respond though, I will wait for a couple of days and try to follow up with them again. :) Else we can use the workaround solution.
Yes. Supporting all the formats would be helpful.
> And additionally, we can add a Data Preprocessing section to the card where we explain how to load/process the files. I can help with that.
Okay. Got it. | ## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md). | 107 | Add CMU Graphics Lab Motion Capture dataset
## Adding a Dataset
- **Name:** CMU Graphics Lab Motion Capture database
- **Description:** The database contains free motions which you can download and use.
- **Data:** http://mocap.cs.cmu.edu/
- **Motivation:** Nice motion capture dataset
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Hi @mariosasko ,
1. Thanks for this, so we can add the file paths.
2. Yes, I had already mailed the authors a couple of days back actually, asking for the metadata details[ i.e category, sub-category and motion description] . They are yet to respond though, I will wait for a couple of days and try to follow up with them again. :) Else we can use the workaround solution.
Yes. Supporting all the formats would be helpful.
> And additionally, we can add a Data Preprocessing section to the card where we explain how to load/process the files. I can help with that.
Okay. Got it. |
https://github.com/huggingface/datasets/issues/3455 | Easier information editing | Hi ! I guess you are talking about the dataset cards that are in this repository on github ?
I think github allows to submit a PR even for 1 line though the `Edit file` button on the page of the dataset card.
Maybe let's mention this in `CONTRIBUTING.md` ? | **Is your feature request related to a problem? Please describe.**
It requires a lot of effort to improve a datasheet.
**Describe the solution you'd like**
UI or at least a link to the place where the code that needs to be edited is (and an easy way to edit this code directly from the site, without cloning, branching, makefile etc.)
**Describe alternatives you've considered**
The current Ux is to have the 8 steps for contribution while One just wishes to change a line a type etc.
| 50 | Easier information editing
**Is your feature request related to a problem? Please describe.**
It requires a lot of effort to improve a datasheet.
**Describe the solution you'd like**
UI or at least a link to the place where the code that needs to be edited is (and an easy way to edit this code directly from the site, without cloning, branching, makefile etc.)
**Describe alternatives you've considered**
The current Ux is to have the 8 steps for contribution while One just wishes to change a line a type etc.
Hi ! I guess you are talking about the dataset cards that are in this repository on github ?
I think github allows to submit a PR even for 1 line though the `Edit file` button on the page of the dataset card.
Maybe let's mention this in `CONTRIBUTING.md` ? |
https://github.com/huggingface/datasets/issues/3455 | Easier information editing | We now host all the datasets on the HF Hub, where you can easily edit them through UI (for single file changes) or Git workflow (for single/multiple file changes) | **Is your feature request related to a problem? Please describe.**
It requires a lot of effort to improve a datasheet.
**Describe the solution you'd like**
UI or at least a link to the place where the code that needs to be edited is (and an easy way to edit this code directly from the site, without cloning, branching, makefile etc.)
**Describe alternatives you've considered**
The current Ux is to have the 8 steps for contribution while One just wishes to change a line a type etc.
| 29 | Easier information editing
**Is your feature request related to a problem? Please describe.**
It requires a lot of effort to improve a datasheet.
**Describe the solution you'd like**
UI or at least a link to the place where the code that needs to be edited is (and an easy way to edit this code directly from the site, without cloning, branching, makefile etc.)
**Describe alternatives you've considered**
The current Ux is to have the 8 steps for contribution while One just wishes to change a line a type etc.
We now host all the datasets on the HF Hub, where you can easily edit them through UI (for single file changes) or Git workflow (for single/multiple file changes) |
https://github.com/huggingface/datasets/issues/3452 | why the stratify option is omitted from test_train_split function? | Hi ! It's simply not added yet :)
If someone wants to contribute to add the `stratify` parameter I'd be happy to give some pointers.
In the meantime, I guess you can use `sklearn` or other tools to do a stratified train/test split over the **indices** of your dataset and then do
```
train_dataset = dataset.select(train_indices)
test_dataset = dataset.select(test_indices)
``` | why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset. | 60 | why the stratify option is omitted from test_train_split function?
why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset.
Hi ! It's simply not added yet :)
If someone wants to contribute to add the `stratify` parameter I'd be happy to give some pointers.
In the meantime, I guess you can use `sklearn` or other tools to do a stratified train/test split over the **indices** of your dataset and then do
```
train_dataset = dataset.select(train_indices)
test_dataset = dataset.select(test_indices)
``` |
https://github.com/huggingface/datasets/issues/3452 | why the stratify option is omitted from test_train_split function? | Hi @lhoestq I would like to add `stratify` parameter, can you give me some pointers for adding the same ? | why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset. | 20 | why the stratify option is omitted from test_train_split function?
why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset.
Hi @lhoestq I would like to add `stratify` parameter, can you give me some pointers for adding the same ? |
https://github.com/huggingface/datasets/issues/3452 | why the stratify option is omitted from test_train_split function? | Hi ! Sure :)
The `train_test_split` method is defined here:
https://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3253-L3253
and inside `train_test_split ` we need to create the right `train_indices` and `test_indices` that are passed here to `.select()`:
https://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3450-L3464
For example if your dataset is like
| | label |
|---:|--------:|
| 0 | 1 |
| 1 | 1 |
| 2 | 0 |
| 3 | 0 |
and the user passes `stratify=dataset["label"]`, then you should get indices that look like this
```
train_indices = [0, 2]
test_indices = [1, 3]
```
these indices will be passed to `.select` to return the stratified train and test splits :)
Feel free to îng me if you have any question ! | why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset. | 114 | why the stratify option is omitted from test_train_split function?
why the stratify option is omitted from test_train_split function?
is there any other way implement the stratify option while splitting the dataset? as it is important point to be considered while splitting the dataset.
Hi ! Sure :)
The `train_test_split` method is defined here:
https://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3253-L3253
and inside `train_test_split ` we need to create the right `train_indices` and `test_indices` that are passed here to `.select()`:
https://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3450-L3464
For example if your dataset is like
| | label |
|---:|--------:|
| 0 | 1 |
| 1 | 1 |
| 2 | 0 |
| 3 | 0 |
and the user passes `stratify=dataset["label"]`, then you should get indices that look like this
```
train_indices = [0, 2]
test_indices = [1, 3]
```
these indices will be passed to `.select` to return the stratified train and test splits :)
Feel free to îng me if you have any question ! |
https://github.com/huggingface/datasets/issues/3450 | Unexpected behavior doing Split + Filter | Hi ! This is an issue with `datasets` 1.12. Sorry for the inconvenience. Can you update to `>=1.13` ?
see https://github.com/huggingface/datasets/issues/3190
Maybe we should also backport the bug fix to `1.12` (in a new version `1.12.2`) | ## Describe the bug
I observed unexpected behavior when applying 'train_test_split' followed by 'filter' on dataset. Elements of the training dataset eventually end up in the test dataset (after applying the 'filter')
## Steps to reproduce the bug
```
from datasets import Dataset
import pandas as pd
dic = {'x': [1,2,3,4,5,6,7,8,9], 'y':['q','w','e','r','t','y','u','i','o']}
df = pd.DataFrame.from_dict(dic)
dataset = Dataset.from_pandas(df)
split_dataset = dataset.train_test_split(test_size=0.5, shuffle=False, seed=42)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]
eval_dataset_2 = eval_dataset.filter(lambda example: example['x'] % 2 == 0)
print( eval_dataset['x'])
print(eval_dataset_2['x'])
```
One observes that elements in eval_dataset2 are actually coming from the training dataset...
## Expected results
The expected results would be that the filtered eval dataset would only contain elements from the original eval dataset.
## Actual results
Specify the actual results or traceback.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.12.1
- Platform: Windows 10
- Python version: 3.7
- PyArrow version: 5.0.0
| 36 | Unexpected behavior doing Split + Filter
## Describe the bug
I observed unexpected behavior when applying 'train_test_split' followed by 'filter' on dataset. Elements of the training dataset eventually end up in the test dataset (after applying the 'filter')
## Steps to reproduce the bug
```
from datasets import Dataset
import pandas as pd
dic = {'x': [1,2,3,4,5,6,7,8,9], 'y':['q','w','e','r','t','y','u','i','o']}
df = pd.DataFrame.from_dict(dic)
dataset = Dataset.from_pandas(df)
split_dataset = dataset.train_test_split(test_size=0.5, shuffle=False, seed=42)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]
eval_dataset_2 = eval_dataset.filter(lambda example: example['x'] % 2 == 0)
print( eval_dataset['x'])
print(eval_dataset_2['x'])
```
One observes that elements in eval_dataset2 are actually coming from the training dataset...
## Expected results
The expected results would be that the filtered eval dataset would only contain elements from the original eval dataset.
## Actual results
Specify the actual results or traceback.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.12.1
- Platform: Windows 10
- Python version: 3.7
- PyArrow version: 5.0.0
Hi ! This is an issue with `datasets` 1.12. Sorry for the inconvenience. Can you update to `>=1.13` ?
see https://github.com/huggingface/datasets/issues/3190
Maybe we should also backport the bug fix to `1.12` (in a new version `1.12.2`) |
https://github.com/huggingface/datasets/issues/3449 | Add `__add__()`, `__iadd__()` and similar to `Dataset` class | I was going through the codebase, and I believe the implementation of __add__() and __iadd__() will be similar to concatenate_datasets() after the elimination of code for arguments other than the list of datasets (info, split, axis).
(Assuming elimination of axis means concatenating over axis 1.) | **Is your feature request related to a problem? Please describe.**
No.
**Describe the solution you'd like**
I would like to be able to concatenate datasets as follows:
```python
>>> dataset["train"] += dataset["validation"]
```
... instead of using `concatenate_datasets()`:
```python
>>> raw_datasets["train"] = concatenate_datasets([raw_datasets["train"], raw_datasets["validation"]])
>>> del raw_datasets["validation"]
```
**Describe alternatives you've considered**
Well, I have considered `concatenate_datasets()` 😀
**Additional context**
N.a.
| 45 | Add `__add__()`, `__iadd__()` and similar to `Dataset` class
**Is your feature request related to a problem? Please describe.**
No.
**Describe the solution you'd like**
I would like to be able to concatenate datasets as follows:
```python
>>> dataset["train"] += dataset["validation"]
```
... instead of using `concatenate_datasets()`:
```python
>>> raw_datasets["train"] = concatenate_datasets([raw_datasets["train"], raw_datasets["validation"]])
>>> del raw_datasets["validation"]
```
**Describe alternatives you've considered**
Well, I have considered `concatenate_datasets()` 😀
**Additional context**
N.a.
I was going through the codebase, and I believe the implementation of __add__() and __iadd__() will be similar to concatenate_datasets() after the elimination of code for arguments other than the list of datasets (info, split, axis).
(Assuming elimination of axis means concatenating over axis 1.) |
https://github.com/huggingface/datasets/issues/3449 | Add `__add__()`, `__iadd__()` and similar to `Dataset` class | Most data frame libraries (Polars, Pandas, ...) override `__add__` to perform (mathematical) summation, so having different behavior here is a bad idea. | **Is your feature request related to a problem? Please describe.**
No.
**Describe the solution you'd like**
I would like to be able to concatenate datasets as follows:
```python
>>> dataset["train"] += dataset["validation"]
```
... instead of using `concatenate_datasets()`:
```python
>>> raw_datasets["train"] = concatenate_datasets([raw_datasets["train"], raw_datasets["validation"]])
>>> del raw_datasets["validation"]
```
**Describe alternatives you've considered**
Well, I have considered `concatenate_datasets()` 😀
**Additional context**
N.a.
| 22 | Add `__add__()`, `__iadd__()` and similar to `Dataset` class
**Is your feature request related to a problem? Please describe.**
No.
**Describe the solution you'd like**
I would like to be able to concatenate datasets as follows:
```python
>>> dataset["train"] += dataset["validation"]
```
... instead of using `concatenate_datasets()`:
```python
>>> raw_datasets["train"] = concatenate_datasets([raw_datasets["train"], raw_datasets["validation"]])
>>> del raw_datasets["validation"]
```
**Describe alternatives you've considered**
Well, I have considered `concatenate_datasets()` 😀
**Additional context**
N.a.
Most data frame libraries (Polars, Pandas, ...) override `__add__` to perform (mathematical) summation, so having different behavior here is a bad idea. |
https://github.com/huggingface/datasets/issues/3448 | JSONDecodeError with HuggingFace dataset viewer | Hi ! I think the issue comes from the dataset_infos.json file: it has the "flat" field twice.
Can you try deleting this file and regenerating it please ? | ## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
| 28 | JSONDecodeError with HuggingFace dataset viewer
## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
Hi ! I think the issue comes from the dataset_infos.json file: it has the "flat" field twice.
Can you try deleting this file and regenerating it please ? |
https://github.com/huggingface/datasets/issues/3448 | JSONDecodeError with HuggingFace dataset viewer | Thanks! That fixed that, but now I am getting:
Server Error
Status code: 400
Exception: KeyError
Message: 'feature'
I checked the dataset_infos.json and pubmed_neg.py script, I don't use 'feature' anywhere as a key. Is the dataset viewer expecting that I do? | ## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
| 41 | JSONDecodeError with HuggingFace dataset viewer
## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
Thanks! That fixed that, but now I am getting:
Server Error
Status code: 400
Exception: KeyError
Message: 'feature'
I checked the dataset_infos.json and pubmed_neg.py script, I don't use 'feature' anywhere as a key. Is the dataset viewer expecting that I do? |
https://github.com/huggingface/datasets/issues/3448 | JSONDecodeError with HuggingFace dataset viewer | It seems that the `feature` key is missing from some feature type definition in your dataset_infos.json:
```json
"tokens": {
"dtype": "list",
"id": null,
"_type": "Sequence"
},
"tags": {
"dtype": "list",
"id": null,
"_type": "Sequence"
}
```
They should be
```json
"tokens": {
"dtype": "list",
"id": null,
"_type": "Sequence"
"feature": {"dtype": "string", "id": null, "_type": "Value"}
},
"tags": {
"dtype": "list",
"id": null,
"_type": "Sequence",
"feature": {"num_classes": 5, "names": ["-", "S", "H", "N", "C"], "names_file": null, "id": null, "_type": "ClassLabel"}
}
```
Note that you can generate the dataset_infos.json automatically to avoid mistakes:
```bash
datasets-cli test ./path/to/dataset --save_infos
``` | ## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
| 98 | JSONDecodeError with HuggingFace dataset viewer
## Dataset viewer issue for 'pubmed_neg'
**Link:** https://huggingface.co/datasets/IGESML/pubmed_neg
I am getting the error:
Status code: 400
Exception: JSONDecodeError
Message: Expecting property name enclosed in double quotes: line 61 column 2 (char 1202)
I have checked all files - I am not using single quotes anywhere. Not sure what is causing this issue.
Am I the one who added this dataset ? Yes
It seems that the `feature` key is missing from some feature type definition in your dataset_infos.json:
```json
"tokens": {
"dtype": "list",
"id": null,
"_type": "Sequence"
},
"tags": {
"dtype": "list",
"id": null,
"_type": "Sequence"
}
```
They should be
```json
"tokens": {
"dtype": "list",
"id": null,
"_type": "Sequence"
"feature": {"dtype": "string", "id": null, "_type": "Value"}
},
"tags": {
"dtype": "list",
"id": null,
"_type": "Sequence",
"feature": {"num_classes": 5, "names": ["-", "S", "H", "N", "C"], "names_file": null, "id": null, "_type": "ClassLabel"}
}
```
Note that you can generate the dataset_infos.json automatically to avoid mistakes:
```bash
datasets-cli test ./path/to/dataset --save_infos
``` |
https://github.com/huggingface/datasets/issues/3447 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading | Hi ! Indeed it says "downloading and preparing" but in your case it didn't need to download anything since you used local files (it would have thrown an error otherwise). I think we can improve the logging to make it clearer in this case | ## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
| 44 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading
## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
Hi ! Indeed it says "downloading and preparing" but in your case it didn't need to download anything since you used local files (it would have thrown an error otherwise). I think we can improve the logging to make it clearer in this case |
https://github.com/huggingface/datasets/issues/3447 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading | @lhoestq Thank you for explaining. I am sorry but I was not clear about my intention. I didn't want to kill internet traffic; I wanted to kill all write activity. In other words, you can imagine that my storage has only read access but crashes on write.
When run_clm.py is invoked with the same parameters, the hash in the cache directory "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/..." doesn't change, and my job can load cached data properly. This is great.
Unfortunately, when params change (which happens sometimes), the hash changes and the old cache is invalid. datasets builder would create a new cache directory with the new hash and create JSON builder there, even though every JSON builder is the same. I didn't find a way to avoid such behavior.
This problem can be resolved when using datasets.map() for tokenizing and grouping text. This function allows me to specify output filenames with --cache_file_names, so that the cached files are always valid.
This is the code that I used to freeze cache filenames for tokenization. I wish I could do the same to datasets.load_dataset()
```
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
cache_file_names={k: os.path.join(model_args.cache_dir, f'{k}-tokenized') for k in raw_datasets},
)
``` | ## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
| 201 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading
## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
@lhoestq Thank you for explaining. I am sorry but I was not clear about my intention. I didn't want to kill internet traffic; I wanted to kill all write activity. In other words, you can imagine that my storage has only read access but crashes on write.
When run_clm.py is invoked with the same parameters, the hash in the cache directory "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/..." doesn't change, and my job can load cached data properly. This is great.
Unfortunately, when params change (which happens sometimes), the hash changes and the old cache is invalid. datasets builder would create a new cache directory with the new hash and create JSON builder there, even though every JSON builder is the same. I didn't find a way to avoid such behavior.
This problem can be resolved when using datasets.map() for tokenizing and grouping text. This function allows me to specify output filenames with --cache_file_names, so that the cached files are always valid.
This is the code that I used to freeze cache filenames for tokenization. I wish I could do the same to datasets.load_dataset()
```
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
cache_file_names={k: os.path.join(model_args.cache_dir, f'{k}-tokenized') for k in raw_datasets},
)
``` |
https://github.com/huggingface/datasets/issues/3447 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading | Hi ! `load_dataset` may re-generate your dataset if some parameters changed indeed. If you want to freeze a dataset loaded with `load_dataset`, I think the best solution is just to save it somewhere on your disk with `.save_to_disk(my_dataset_dir)` and reload it with `load_from_disk(my_dataset_dir)`. This way you will be able to reload the dataset without having to run `load_dataset` | ## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
| 58 | HF_DATASETS_OFFLINE=1 didn't stop datasets.builder from downloading
## Describe the bug
According to https://huggingface.co/docs/datasets/loading_datasets.html#loading-a-dataset-builder, setting HF_DATASETS_OFFLINE to 1 should make datasets to "run in full offline mode". It didn't work for me. At the very beginning, datasets still tried to download "custom data configuration" for JSON, despite I have run the program once and cached all data into the same --cache_dir.
"Downloading" is not an issue when running with local disk, but crashes often with cloud storage because (1) multiply GPU processes try to access the same file, AND (2) FileLocker fails to synchronize all processes, due to storage throttling. 99% of times, when the main process releases FileLocker, the file is not actually ready for access in cloud storage and thus triggers "FileNotFound" errors for all other processes. Well, another way to resolve the problem is to investigate super reliable cloud storage, but that's out of scope here.
## Steps to reproduce the bug
```
export HF_DATASETS_OFFLINE=1
python run_clm.py --model_name_or_path=models/gpt-j-6B --train_file=trainpy.v2.train.json --validation_file=trainpy.v2.eval.json --cache_dir=datacache/trainpy.v2
```
## Expected results
datasets should stop all "downloading" behavior but reuse the cached JSON configuration. I think the problem here is part of the cache directory path, "default-471372bed4b51b53", is randomly generated, and it could change if some parameters changed. And I didn't find a way to use a fixed path to ensure datasets to reuse cached data every time.
## Actual results
The logging shows datasets are still downloading into "datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426".
```
12/16/2021 10:25:59 - WARNING - datasets.builder - Using custom data configuration default-471372bed4b51b53
12/16/2021 10:25:59 - INFO - datasets.builder - Generating dataset json (datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426)
Downloading and preparing dataset json/default to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426...
100%|██████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 17623.13it/s]
12/16/2021 10:25:59 - INFO - datasets.utils.download_manager - Downloading took 0.0 min
12/16/2021 10:26:00 - INFO - datasets.utils.download_manager - Checksum Computation took 0.0 min
100%|███████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 1206.99it/s]
12/16/2021 10:26:00 - INFO - datasets.utils.info_utils - Unable to verify checksums.
12/16/2021 10:26:00 - INFO - datasets.builder - Generating split train
12/16/2021 10:26:01 - INFO - datasets.builder - Generating split validation
12/16/2021 10:26:02 - INFO - datasets.utils.info_utils - Unable to verify splits sizes.
Dataset json downloaded and prepared to datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/c2d554c3377ea79c7664b93dc65d0803b45e3279000f993c7bfd18937fd7f426. Subsequent calls will reuse this data.
100%|█████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 53.54it/s]
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.16.1
- Platform: Linux
- Python version: 3.8.10
- PyArrow version: 6.0.1
Hi ! `load_dataset` may re-generate your dataset if some parameters changed indeed. If you want to freeze a dataset loaded with `load_dataset`, I think the best solution is just to save it somewhere on your disk with `.save_to_disk(my_dataset_dir)` and reload it with `load_from_disk(my_dataset_dir)`. This way you will be able to reload the dataset without having to run `load_dataset` |
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | Yes I agree, these should be as aligned as possible. Maybe we can also check the feedback in the survey at http://hf.co/oss-survey and see if people mentioned related things on the API (in particular if we go the breaking change way, it would be good to be sure we are taking the right direction for the community). | ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 57 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
Yes I agree, these should be as aligned as possible. Maybe we can also check the feedback in the survey at http://hf.co/oss-survey and see if people mentioned related things on the API (in particular if we go the breaking change way, it would be good to be sure we are taking the right direction for the community). |
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | I like this proposal.
> There is also an important difference in terms of behavior:
Dataset.map adds new columns (with dict.update)
BUT
IterableDataset discards previous columns (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.
> The main breaking change would be the change of behavior of IterableDataset.map, because currently it discards all the previous columns instead of keeping them.
Yes, this behavior of `IterableDataset.map` was surprising to me the first time I used it because I was expecting the same behavior as `Dataset.map`, so I'm OK with the breaking change here.
> IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
\+ it's also missing the actual formatting code (we return unformatted tensors)
> We could have a completely aligned map method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
> For information, TFDS does lazy map by default, and has an additional .cache() method.
If I understand this part correctly, the idea would be for `Dataset.map` to behave similarly to `Dataset.with_transform` (lazy processing) and to have an option to cache processed data (with `.cache()`). This idea is really nice because it can also be applied to `IterableDataset` to fix https://github.com/huggingface/datasets/issues/3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?)
> If the two APIs are more aligned it would be awesome for the examples in transformers, and it would create a satisfactory experience for users that want to switch from one mode to the other.
Yes, it would be amazing to have an option to easily switch between these two modes.
I agree with the rest.
| ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 322 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
I like this proposal.
> There is also an important difference in terms of behavior:
Dataset.map adds new columns (with dict.update)
BUT
IterableDataset discards previous columns (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.
> The main breaking change would be the change of behavior of IterableDataset.map, because currently it discards all the previous columns instead of keeping them.
Yes, this behavior of `IterableDataset.map` was surprising to me the first time I used it because I was expecting the same behavior as `Dataset.map`, so I'm OK with the breaking change here.
> IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
\+ it's also missing the actual formatting code (we return unformatted tensors)
> We could have a completely aligned map method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
> For information, TFDS does lazy map by default, and has an additional .cache() method.
If I understand this part correctly, the idea would be for `Dataset.map` to behave similarly to `Dataset.with_transform` (lazy processing) and to have an option to cache processed data (with `.cache()`). This idea is really nice because it can also be applied to `IterableDataset` to fix https://github.com/huggingface/datasets/issues/3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?)
> If the two APIs are more aligned it would be awesome for the examples in transformers, and it would create a satisfactory experience for users that want to switch from one mode to the other.
Yes, it would be amazing to have an option to easily switch between these two modes.
I agree with the rest.
|
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | > If I understand this part correctly, the idea would be for Dataset.map to behave similarly to Dataset.with_transform (lazy processing) and to have an option to cache processed data (with .cache()). This idea is really nice because it can also be applied to IterableDataset to fix #3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?)
Yea this is too big of a change in my opinion. Anyway it's fine as it is right now with streaming=lazy and regular=eager. | ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 105 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
> If I understand this part correctly, the idea would be for Dataset.map to behave similarly to Dataset.with_transform (lazy processing) and to have an option to cache processed data (with .cache()). This idea is really nice because it can also be applied to IterableDataset to fix #3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?)
Yea this is too big of a change in my opinion. Anyway it's fine as it is right now with streaming=lazy and regular=eager. |
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | Yes indeed, thanks. I added it to the list of methods to align in the first post | ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 17 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
Yes indeed, thanks. I added it to the list of methods to align in the first post |
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | I just encountered the problem of the missing `fn_kwargs` parameter in the `map` method. I am commenting to give a workaround in case someone has the same problem and does not find a solution.
You can wrap your function call inside a class that contains the other parameters needed by the function called by map, like this:
```python
def my_func(x, y, z):
# Do things
class MyFuncWrapper:
def __init__(self, y, z):
self.y = y
self.z = z
def __call__(self, x):
return my_func(x, self.y, self.z)
```
Then, give an instance of the `MyFuncWrapper` to the map function. | ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 96 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
I just encountered the problem of the missing `fn_kwargs` parameter in the `map` method. I am commenting to give a workaround in case someone has the same problem and does not find a solution.
You can wrap your function call inside a class that contains the other parameters needed by the function called by map, like this:
```python
def my_func(x, y, z):
# Do things
class MyFuncWrapper:
def __init__(self, y, z):
self.y = y
self.z = z
def __call__(self, x):
return my_func(x, self.y, self.z)
```
Then, give an instance of the `MyFuncWrapper` to the map function. |
https://github.com/huggingface/datasets/issues/3444 | Align the Dataset and IterableDataset processing API | The main differences have been addressed (map, formatting) but there are still a few things to implement like Dataset.take, Dataset.skip, IterableDataset.set_format, IterableDataset.formatted_as, IterableDataset.reset_format.
The rest cannot be implemented for the general case. E.g. train_test_split and select can only work on an iterable dataset if the underlying dataset format allows it (we need to know the number of rows and have some sort of random access) | ## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger | 65 | Align the Dataset and IterableDataset processing API
## Intro
items marked like <s>this</s> are done already :)
Currently the two classes have two distinct API for processing:
### The `.map()` method
Both have those parameters in common: function, batched, batch_size
- IterableDataset is missing those parameters:
<s>with_indices</s>, with_rank, <s>input_columns</s>, <s>drop_last_batch</s>, <s>remove_columns</s>, features, disable_nullable, fn_kwargs, num_proc
- Dataset also has additional parameters that are exclusive, due to caching:
keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, suffix_template, new_fingerprint
- <s>There is also an important difference in terms of behavior:
**Dataset.map adds new columns** (with dict.update)
BUT
**IterableDataset discards previous columns** (it overwrites the dict)
IMO the two methods should have the same behavior. This would be an important breaking change though.</s>
- Dataset.map is eager while IterableDataset.map is lazy
### The `.shuffle()` method
- <s>Both have an optional seed parameter, but IterableDataset requires a mandatory parameter buffer_size to control the size of the local buffer used for approximate shuffling.</s>
- <s>IterableDataset is missing the parameter generator</s>
- Also Dataset has exclusive parameters due to caching: keep_in_memory, load_from_cache_file, indices_cache_file_name, writer_batch_size, new_fingerprint
### The `.with_format()` method
- IterableDataset only supports "torch" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs
- other methods like `set_format`, `reset_format` or `formatted_as` are also missing
### Other methods
- Both have the same `remove_columns` method
- IterableDataset is missing: <s>cast</s>, <s>cast_column</s>, <s>filter</s>, <s>rename_column</s>, <s>rename_columns</s>, class_encode_column, flatten, prepare_for_task, train_test_split, shard
- Some other methods are missing but we can discuss them: set_transform, formatted_as, with_transform
- And others don't really make sense for an iterable dataset: select, sort, add_column, add_item
- Dataset is missing skip and take, that IterableDataset implements.
## Questions
I think it would be nice to be able to switch between streaming and regular dataset easily, without changing the processing code significantly.
1. What should be aligned and what shouldn't between those two APIs ?
IMO the minimum is to align the main processing methods.
It would mean aligning breaking the current `Iterable.map` to have the same behavior as `Dataset.map` (add columns with dict.update), and add multiprocessing as well as the missing parameters. DONE ✅
It would also mean implementing the missing methods: cast, cast_column, filter, rename_column, rename_columns, class_encode_column, flatten, prepare_for_task, train_test_split, shard. WIP 🟠
2. What are the breaking changes for IterableDataset ?
The main breaking change would be the change of behavior of `IterableDataset.map`, because currently it discards all the previous columns instead of keeping them. DONE ✅
3. Shall we also do some changes for regular datasets ?
I agree the simplest would be to have the exact same methods for both Dataset and IterableDataset. However this is probably not a good idea because it would prevent users from using the best benefits of them. That's why we can keep some aspects of regular datasets as they are:
- keep the eager Dataset.map with caching
- keep the with_transform method for lazy processing
- keep Dataset.select (it could also be added to IterableDataset even though it's not recommended)
We could have a completely aligned `map` method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.
For information, TFDS does lazy map by default, and has an additional `.cache()` method.
## Opinions ?
I'd love to gather some opinions about this here. If the two APIs are more aligned it would be awesome for the examples in `transformers`, and it would create a satisfactory experience for users that want to switch from one mode to the other.
cc @mariosasko @albertvillanova @thomwolf @patrickvonplaten @sgugger
The main differences have been addressed (map, formatting) but there are still a few things to implement like Dataset.take, Dataset.skip, IterableDataset.set_format, IterableDataset.formatted_as, IterableDataset.reset_format.
The rest cannot be implemented for the general case. E.g. train_test_split and select can only work on an iterable dataset if the underlying dataset format allows it (we need to know the number of rows and have some sort of random access) |
https://github.com/huggingface/datasets/issues/3440 | datasets keeps reading from cached files, although I disabled it | Hi ! What version of `datasets` are you using ? Can you also provide the logs you get before it raises the error ? | ## Describe the bug
Hi,
I am trying to avoid dataset library using cached files, I get the following bug when this tried to read the cached files. I tried to do the followings:
```
from datasets import set_caching_enabled
set_caching_enabled(False)
```
also force redownlaod:
```
download_mode='force_redownload'
```
but none worked so far, this is on a cluster and on some of the machines this reads from the cached files, I really appreciate any idea on how to fully remove caching @lhoestq
many thanks
```
File "run_clm.py", line 496, in <module>
main()
File "run_clm.py", line 419, in main
train_result = trainer.train(resume_from_checkpoint=checkpoint)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 943, in train
self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/transformers/trainer.py", line 1445, in _maybe_log_save_evaluate
metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 172, in evaluate
output = self.eval_loop(
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 241, in eval_loop
metrics = self.compute_pet_metrics(eval_datasets, model, self.extra_info[metric_key_prefix], task=task)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 268, in compute_pet_metrics
centroids = self._compute_per_token_train_centroids(model, task=task)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 353, in _compute_per_token_train_centroids
data = get_label_samples(self.get_per_task_train_dataset(task), label)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 350, in get_label_samples
return dataset.filter(lambda example: int(example['labels']) == label)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 470, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/fingerprint.py", line 406, in wrapper
out = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2519, in filter
indices = self.map(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2036, in map
return self._map_single(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 503, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 470, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/fingerprint.py", line 406, in wrapper
out = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2248, in _map_single
return Dataset.from_file(cache_file_name, info=info, split=self.split)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 654, in from_file
return cls(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 593, in __init__
self.info.features = self.info.features.reorder_fields_as(inferred_features)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/features/features.py", line 1092, in reorder_fields_as
return Features(recursive_reorder(self, other))
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/features/features.py", line 1081, in recursive_reorder
raise ValueError(f"Keys mismatch: between {source} and {target}" + stack_position)
ValueError: Keys mismatch: between {'indices': Value(dtype='uint64', id=None)} and {'candidates_ids': Sequence(feature=Value(dtype='null', id=None), length=-1, id=None), 'labels': Value(dtype='int64', id=None), 'attention_mask': Sequence(feature=Value(dtype='int8', id=None), length=-1, id=None), 'input_ids': Sequence(feature=Value(dtype='int32', id=None), length=-1, id=None), 'extra_fields': {}, 'task': Value(dtype='string', id=None)}
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: linux
- Python version: 3.8.12
- PyArrow version: 6.0.1
| 24 | datasets keeps reading from cached files, although I disabled it
## Describe the bug
Hi,
I am trying to avoid dataset library using cached files, I get the following bug when this tried to read the cached files. I tried to do the followings:
```
from datasets import set_caching_enabled
set_caching_enabled(False)
```
also force redownlaod:
```
download_mode='force_redownload'
```
but none worked so far, this is on a cluster and on some of the machines this reads from the cached files, I really appreciate any idea on how to fully remove caching @lhoestq
many thanks
```
File "run_clm.py", line 496, in <module>
main()
File "run_clm.py", line 419, in main
train_result = trainer.train(resume_from_checkpoint=checkpoint)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 943, in train
self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/transformers/trainer.py", line 1445, in _maybe_log_save_evaluate
metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 172, in evaluate
output = self.eval_loop(
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 241, in eval_loop
metrics = self.compute_pet_metrics(eval_datasets, model, self.extra_info[metric_key_prefix], task=task)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 268, in compute_pet_metrics
centroids = self._compute_per_token_train_centroids(model, task=task)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 353, in _compute_per_token_train_centroids
data = get_label_samples(self.get_per_task_train_dataset(task), label)
File "/users/dara/codes/fewshot/debug/fewshot/third_party/trainers/trainer.py", line 350, in get_label_samples
return dataset.filter(lambda example: int(example['labels']) == label)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 470, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/fingerprint.py", line 406, in wrapper
out = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2519, in filter
indices = self.map(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2036, in map
return self._map_single(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 503, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 470, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/fingerprint.py", line 406, in wrapper
out = func(self, *args, **kwargs)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2248, in _map_single
return Dataset.from_file(cache_file_name, info=info, split=self.split)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 654, in from_file
return cls(
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 593, in __init__
self.info.features = self.info.features.reorder_fields_as(inferred_features)
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/features/features.py", line 1092, in reorder_fields_as
return Features(recursive_reorder(self, other))
File "/users/dara/conda/envs/multisuccess/lib/python3.8/site-packages/datasets/features/features.py", line 1081, in recursive_reorder
raise ValueError(f"Keys mismatch: between {source} and {target}" + stack_position)
ValueError: Keys mismatch: between {'indices': Value(dtype='uint64', id=None)} and {'candidates_ids': Sequence(feature=Value(dtype='null', id=None), length=-1, id=None), 'labels': Value(dtype='int64', id=None), 'attention_mask': Sequence(feature=Value(dtype='int8', id=None), length=-1, id=None), 'input_ids': Sequence(feature=Value(dtype='int32', id=None), length=-1, id=None), 'extra_fields': {}, 'task': Value(dtype='string', id=None)}
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: linux
- Python version: 3.8.12
- PyArrow version: 6.0.1
Hi ! What version of `datasets` are you using ? Can you also provide the logs you get before it raises the error ? |
https://github.com/huggingface/datasets/issues/3431 | Unable to resolve any data file after loading once | Hi ! `load_dataset` accepts as input either a local dataset directory or a dataset name from the Hugging Face Hub.
So here you are getting this error the second time because it tries to load the local `wiki_dpr` directory, instead of `wiki_dpr` from the Hub. It doesn't work since it's a **cache** directory, not a **dataset** directory in itself.
To fix that you can use another cache directory like `cache_dir="/data2/whr/lzy/open_domain_data/retrieval/cache"` | when I rerun my program, it occurs this error
" Unable to resolve any data file that matches '['**train*']' at /data2/whr/lzy/open_domain_data/retrieval/wiki_dpr with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'zip']", so how could i deal with this problem?
thx.
And below is my code .
| 70 | Unable to resolve any data file after loading once
when I rerun my program, it occurs this error
" Unable to resolve any data file that matches '['**train*']' at /data2/whr/lzy/open_domain_data/retrieval/wiki_dpr with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'zip']", so how could i deal with this problem?
thx.
And below is my code .
Hi ! `load_dataset` accepts as input either a local dataset directory or a dataset name from the Hugging Face Hub.
So here you are getting this error the second time because it tries to load the local `wiki_dpr` directory, instead of `wiki_dpr` from the Hub. It doesn't work since it's a **cache** directory, not a **dataset** directory in itself.
To fix that you can use another cache directory like `cache_dir="/data2/whr/lzy/open_domain_data/retrieval/cache"` |
https://github.com/huggingface/datasets/issues/3425 | Getting configs names takes too long | It looks like it's currently calling `HfFileSystem.ls()` ~8 times at the root and for each subdirectory:
- ""
- "en.noblocklist"
- "en.noclean"
- "en"
- "multilingual"
- "realnewslike"
Currently `ls` is slow because it iterates on all the files inside the repository.
An easy optimization would be to cache the result of each call to `ls`.
We can also optimize `ls` by using a tree structure per directory instead of a list of all the files.
|
## Steps to reproduce the bug
```python
from datasets import get_dataset_config_names
get_dataset_config_names("allenai/c4")
```
## Expected results
I would expect to get the answer quickly, at least in less than 10s
## Actual results
It takes about 45s on my environment
## Environment info
- `datasets` version: 1.16.1
- Platform: Linux-5.11.0-1022-aws-x86_64-with-glibc2.31
- Python version: 3.9.6
- PyArrow version: 4.0.1 | 76 | Getting configs names takes too long
## Steps to reproduce the bug
```python
from datasets import get_dataset_config_names
get_dataset_config_names("allenai/c4")
```
## Expected results
I would expect to get the answer quickly, at least in less than 10s
## Actual results
It takes about 45s on my environment
## Environment info
- `datasets` version: 1.16.1
- Platform: Linux-5.11.0-1022-aws-x86_64-with-glibc2.31
- Python version: 3.9.6
- PyArrow version: 4.0.1
It looks like it's currently calling `HfFileSystem.ls()` ~8 times at the root and for each subdirectory:
- ""
- "en.noblocklist"
- "en.noclean"
- "en"
- "multilingual"
- "realnewslike"
Currently `ls` is slow because it iterates on all the files inside the repository.
An easy optimization would be to cache the result of each call to `ls`.
We can also optimize `ls` by using a tree structure per directory instead of a list of all the files.
|
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | Hi ! Thanks for reporting :)
When using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.
We can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens. | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 55 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
Hi ! Thanks for reporting :)
When using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.
We can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens. |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | > Hi ! Thanks for reporting :)
>
> When using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.
>
> We can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens.
Hi ! Thanks for reply
Do u have some plans to fix the problem?
| ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 74 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
> Hi ! Thanks for reporting :)
>
> When using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.
>
> We can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens.
Hi ! Thanks for reply
Do u have some plans to fix the problem?
|
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | Isn’t that somehow a bug on PyTorch side? (Just asking because this behavior seems quite general and maybe not what would be intended) | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 23 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
Isn’t that somehow a bug on PyTorch side? (Just asking because this behavior seems quite general and maybe not what would be intended) |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | From PyTorch's documentation [here](https://pytorch.org/docs/stable/data.html#dataset-types):
> When using an IterableDataset with multi-process data loading. The same dataset object is replicated on each worker process, and thus the replicas must be configured differently to avoid duplicated data. See [IterableDataset](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset) documentations for how to achieve this.
It looks like an intended behavior from PyTorch
As suggested in the [docstring of the IterableDataset class](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset), we could pass a `worker_init_fn` to the DataLoader to fix this. It could be called `streaming_worker_init_fn` for example.
However, while this solution works, I'm worried that many users simply don't know about this parameter and just start their training with duplicate data without knowing it. That's why I'm more in favor of integrating the check on the worker id directly in `datasets` in our implementation of `IterableDataset.__iter__`. | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 127 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
From PyTorch's documentation [here](https://pytorch.org/docs/stable/data.html#dataset-types):
> When using an IterableDataset with multi-process data loading. The same dataset object is replicated on each worker process, and thus the replicas must be configured differently to avoid duplicated data. See [IterableDataset](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset) documentations for how to achieve this.
It looks like an intended behavior from PyTorch
As suggested in the [docstring of the IterableDataset class](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset), we could pass a `worker_init_fn` to the DataLoader to fix this. It could be called `streaming_worker_init_fn` for example.
However, while this solution works, I'm worried that many users simply don't know about this parameter and just start their training with duplicate data without knowing it. That's why I'm more in favor of integrating the check on the worker id directly in `datasets` in our implementation of `IterableDataset.__iter__`. |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | Hi there @lhoestq @cloudyuyuyu
I met that problem recently, and #4375 is really useful because I finally found out I am training with duplicate data.
However, in multi-GPU training, I'm using DDP mode and IterableDataset, which still yields duplicate data for each progress. And this is dangerous because users maybe not realize this behavior. | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 54 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
Hi there @lhoestq @cloudyuyuyu
I met that problem recently, and #4375 is really useful because I finally found out I am training with duplicate data.
However, in multi-GPU training, I'm using DDP mode and IterableDataset, which still yields duplicate data for each progress. And this is dangerous because users maybe not realize this behavior. |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | If the worker_info.id is unique per process it should work fine, could you check that they're unique ?
The code to get the worker_info in each worker is `torch.utils.data.get_worker_info()` | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 29 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
If the worker_info.id is unique per process it should work fine, could you check that they're unique ?
The code to get the worker_info in each worker is `torch.utils.data.get_worker_info()` |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | test.py
```python
import json
import os
import torch
from torch.utils.data import IterableDataset, DataLoader
from transformers import PreTrainedTokenizer, TrainingArguments
from common.arguments import DataTrainingArguments, ModelArguments
class MyIterableDataset(IterableDataset):
def __iter__(self):
worker_info = torch.utils.data.get_worker_info()
print(worker_info)
return iter(range(3))
if __name__ == '__main__':
dataset = MyIterableDataset()
dataloader = DataLoader(dataset, num_workers=1)
for i in dataloader:
print(i)
```
```sh
$ python3 -m torch.distributed.launch \
--nproc_per_node=2 test.py
WorkerInfo(id=0, num_workers=1, seed=5545685212307804959, dataset=<__main__.MyIterableDataset object at 0x7f92648cf6a0>)
WorkerInfo(id=0, num_workers=1, seed=3174108029709729025, dataset=<__main__.MyIterableDataset object at 0x7f19ab961670>)
tensor([0])
tensor([1])
tensor([2])
tensor([0])
tensor([1])
tensor([2])
```
@lhoestq they are not unique | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 84 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
test.py
```python
import json
import os
import torch
from torch.utils.data import IterableDataset, DataLoader
from transformers import PreTrainedTokenizer, TrainingArguments
from common.arguments import DataTrainingArguments, ModelArguments
class MyIterableDataset(IterableDataset):
def __iter__(self):
worker_info = torch.utils.data.get_worker_info()
print(worker_info)
return iter(range(3))
if __name__ == '__main__':
dataset = MyIterableDataset()
dataloader = DataLoader(dataset, num_workers=1)
for i in dataloader:
print(i)
```
```sh
$ python3 -m torch.distributed.launch \
--nproc_per_node=2 test.py
WorkerInfo(id=0, num_workers=1, seed=5545685212307804959, dataset=<__main__.MyIterableDataset object at 0x7f92648cf6a0>)
WorkerInfo(id=0, num_workers=1, seed=3174108029709729025, dataset=<__main__.MyIterableDataset object at 0x7f19ab961670>)
tensor([0])
tensor([1])
tensor([2])
tensor([0])
tensor([1])
tensor([2])
```
@lhoestq they are not unique |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?
I guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 47 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?
I guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | > It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?
>
> I guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is
Maybe we should document it? | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 55 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
> It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?
>
> I guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is
Maybe we should document it? |
https://github.com/huggingface/datasets/issues/3423 | data duplicate when setting num_works > 1 with streaming data | hmm actually let me open a new issue on DDP - original post was for single node | ## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
| 17 | data duplicate when setting num_works > 1 with streaming data
## Describe the bug
The data is repeated num_works times when we load_dataset with streaming and set num_works > 1 when construct dataloader
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import pandas as pd
import numpy as np
import os
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import shutil
NUM_OF_USER = 1000000
NUM_OF_ACTION = 50000
NUM_OF_SEQUENCE = 10000
NUM_OF_FILES = 32
NUM_OF_WORKERS = 16
if __name__ == "__main__":
shutil.rmtree("./dataset")
for i in range(NUM_OF_FILES):
sequence_data = pd.DataFrame(
{
"imei": np.random.randint(1, NUM_OF_USER, size=NUM_OF_SEQUENCE),
"sequence": np.random.randint(1, NUM_OF_ACTION, size=NUM_OF_SEQUENCE)
}
)
if not os.path.exists("./dataset"):
os.makedirs("./dataset")
sequence_data.to_csv(f"./dataset/sequence_data_{i}.csv",
index=False)
dataset = load_dataset("csv",
data_files=[os.path.join("./dataset",file) for file in os.listdir("./dataset") if file.endswith(".csv")],
split="train",
streaming=True).with_format("torch")
data_loader = DataLoader(dataset,
batch_size=1024,
num_workers=NUM_OF_WORKERS)
result = pd.DataFrame()
for i, batch in tqdm(enumerate(data_loader)):
result = pd.concat([result,
pd.DataFrame(batch)],
axis=0)
result.to_csv(f"num_work_{NUM_OF_WORKERS}.csv", index=False)
```
## Expected results
data do not duplicate
## Actual results
data duplicate NUM_OF_WORKERS = 16
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:datasets==1.14.0
- Platform:transformers==4.11.3
- Python version:3.8
- PyArrow version:
hmm actually let me open a new issue on DDP - original post was for single node |
https://github.com/huggingface/datasets/issues/3422 | Error about load_metric | Hi ! I wasn't able to reproduce your error.
Can you try to clear your cache at `~/.cache/huggingface/modules` and try again ? | ## Describe the bug
File "/opt/conda/lib/python3.8/site-packages/datasets/load.py", line 1371, in load_metric
metric = metric_cls(
TypeError: 'NoneType' object is not callable
## Steps to reproduce the bug
```python
metric = load_metric("glue", "sst2")
```
## Environment info
- `datasets` version: 1.16.1
- Platform: Linux-4.15.0-161-generic-x86_64-with-glibc2.10
- Python version: 3.8.3
- PyArrow version: 6.0.1
| 22 | Error about load_metric
## Describe the bug
File "/opt/conda/lib/python3.8/site-packages/datasets/load.py", line 1371, in load_metric
metric = metric_cls(
TypeError: 'NoneType' object is not callable
## Steps to reproduce the bug
```python
metric = load_metric("glue", "sst2")
```
## Environment info
- `datasets` version: 1.16.1
- Platform: Linux-4.15.0-161-generic-x86_64-with-glibc2.10
- Python version: 3.8.3
- PyArrow version: 6.0.1
Hi ! I wasn't able to reproduce your error.
Can you try to clear your cache at `~/.cache/huggingface/modules` and try again ? |
https://github.com/huggingface/datasets/issues/3419 | `.to_json` is extremely slow after `.select` | Hi ! It's slower indeed because a datasets on which `select`/`shard`/`train_test_split`/`shuffle` has been called has to do additional steps to retrieve the data of the dataset table in the right order.
Indeed, if you call `dataset.select([0, 5, 10])`, the underlying table of the dataset is not altered to keep the examples at index 0, 5, and 10. Instead, an indices mapping is added on top of the table, that says that the first example is at index 0, the second at index 5 and the last one at index 10.
Therefore accessing the examples of the dataset is slower because of the additional step that uses the indices mapping.
The step that takes the most time is to query the dataset table from a list of indices here:
https://github.com/huggingface/datasets/blob/047dc756ed20fbf06e6bcaf910464aba0e20610a/src/datasets/formatting/formatting.py#L61-L63
In your case it can be made significantly faster by checking if the indices are contiguous. If they're contiguous, we could pass a python `slice` or `range` instead of a list of integers to `_query_table`. This way `_query_table` will do only one lookup to get the queried batch instead of `batch_size` lookups.
Given that calling `select` with contiguous indices is a common use case I'm in favor of implementing such an optimization :)
Let me know what you think | ## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
| 208 | `.to_json` is extremely slow after `.select`
## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
Hi ! It's slower indeed because a datasets on which `select`/`shard`/`train_test_split`/`shuffle` has been called has to do additional steps to retrieve the data of the dataset table in the right order.
Indeed, if you call `dataset.select([0, 5, 10])`, the underlying table of the dataset is not altered to keep the examples at index 0, 5, and 10. Instead, an indices mapping is added on top of the table, that says that the first example is at index 0, the second at index 5 and the last one at index 10.
Therefore accessing the examples of the dataset is slower because of the additional step that uses the indices mapping.
The step that takes the most time is to query the dataset table from a list of indices here:
https://github.com/huggingface/datasets/blob/047dc756ed20fbf06e6bcaf910464aba0e20610a/src/datasets/formatting/formatting.py#L61-L63
In your case it can be made significantly faster by checking if the indices are contiguous. If they're contiguous, we could pass a python `slice` or `range` instead of a list of integers to `_query_table`. This way `_query_table` will do only one lookup to get the queried batch instead of `batch_size` lookups.
Given that calling `select` with contiguous indices is a common use case I'm in favor of implementing such an optimization :)
Let me know what you think |
https://github.com/huggingface/datasets/issues/3419 | `.to_json` is extremely slow after `.select` | Hi, thanks for the response!
I still don't understand why it is so much slower than iterating and saving:
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 99 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 47 seconds
selected_subset3 = original.select([i for i in range(len(original)) if i % 2 == 0])
selected_subset3.to_json("from_select3.json") # Takes 49 seconds
import json
import time
def fast_to_json(dataset, path):
start = time.time()
with open(path, mode="w") as f:
for example in dataset:
f.write(json.dumps(example, separators=(',', ':')) + "\n")
end = time.time()
print(f"Saved {len(dataset)} examples to {path} in {end - start} seconds.")
fast_to_json(original, "from_original_fast.json")
fast_to_json(selected_subset1, "from_select1_fast.json")
fast_to_json(selected_subset2, "from_select2_fast.json")
fast_to_json(selected_subset3, "from_select3_fast.json")
```
```
Saved 87599 examples to from_original_fast.json in 8 seconds.
Saved 87599 examples to from_select1_fast.json in 10 seconds.
Saved 43799 examples to from_select2_fast.json in 6 seconds.
Saved 43800 examples to from_select3_fast.json in 5 seconds.
``` | ## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
| 157 | `.to_json` is extremely slow after `.select`
## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
Hi, thanks for the response!
I still don't understand why it is so much slower than iterating and saving:
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 99 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 47 seconds
selected_subset3 = original.select([i for i in range(len(original)) if i % 2 == 0])
selected_subset3.to_json("from_select3.json") # Takes 49 seconds
import json
import time
def fast_to_json(dataset, path):
start = time.time()
with open(path, mode="w") as f:
for example in dataset:
f.write(json.dumps(example, separators=(',', ':')) + "\n")
end = time.time()
print(f"Saved {len(dataset)} examples to {path} in {end - start} seconds.")
fast_to_json(original, "from_original_fast.json")
fast_to_json(selected_subset1, "from_select1_fast.json")
fast_to_json(selected_subset2, "from_select2_fast.json")
fast_to_json(selected_subset3, "from_select3_fast.json")
```
```
Saved 87599 examples to from_original_fast.json in 8 seconds.
Saved 87599 examples to from_select1_fast.json in 10 seconds.
Saved 43799 examples to from_select2_fast.json in 6 seconds.
Saved 43800 examples to from_select3_fast.json in 5 seconds.
``` |
https://github.com/huggingface/datasets/issues/3419 | `.to_json` is extremely slow after `.select` | There are slight differences between what you're doing and what `to_json` is actually doing.
In particular `to_json` currently converts batches of rows (as an arrow table) to a pandas dataframe, and then to JSON Lines. From your benchmark it looks like it's faster if we don't use pandas.
Thanks for investigating, I think we can optimize `to_json` significantly thanks to your test. | ## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
| 62 | `.to_json` is extremely slow after `.select`
## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
There are slight differences between what you're doing and what `to_json` is actually doing.
In particular `to_json` currently converts batches of rows (as an arrow table) to a pandas dataframe, and then to JSON Lines. From your benchmark it looks like it's faster if we don't use pandas.
Thanks for investigating, I think we can optimize `to_json` significantly thanks to your test. |
https://github.com/huggingface/datasets/issues/3419 | `.to_json` is extremely slow after `.select` | Thanks for your observations, @eladsegal! I spent some time with this and tried different approaches. Turns out that https://github.com/huggingface/datasets/blob/bb13373637b1acc55f8a468a8927a56cf4732230/src/datasets/io/json.py#L100 is giving the problem when we use `to_json` after `select`. This is when `indices` parameter in `query_table` is not `None` (if it is `None` then `to_json` should work as expected)
In order to circumvent this problem, I found out instead of doing Arrow Table -> Pandas-> JSON we can directly go to JSON by using `to_pydict()` which is a little slower than the current approach but at least `select` works properly now. Lmk what you guys think of it @lhoestq, @eladsegal? | ## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
| 100 | `.to_json` is extremely slow after `.select`
## Describe the bug
Saving a dataset to JSON with `to_json` is extremely slow after using `.select` on the original dataset.
## Steps to reproduce the bug
```python
from datasets import load_dataset
original = load_dataset("squad", split="train")
original.to_json("from_original.json") # Takes 0 seconds
selected_subset1 = original.select([i for i in range(len(original))])
selected_subset1.to_json("from_select1.json") # Takes 212 seconds
selected_subset2 = original.select([i for i in range(int(len(original) / 2))])
selected_subset2.to_json("from_select2.json") # Takes 90 seconds
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: master (https://github.com/huggingface/datasets/commit/6090f3cfb5c819f441dd4a4bb635e037c875b044)
- Platform: Linux-4.4.0-19041-Microsoft-x86_64-with-glibc2.27
- Python version: 3.9.7
- PyArrow version: 6.0.0
Thanks for your observations, @eladsegal! I spent some time with this and tried different approaches. Turns out that https://github.com/huggingface/datasets/blob/bb13373637b1acc55f8a468a8927a56cf4732230/src/datasets/io/json.py#L100 is giving the problem when we use `to_json` after `select`. This is when `indices` parameter in `query_table` is not `None` (if it is `None` then `to_json` should work as expected)
In order to circumvent this problem, I found out instead of doing Arrow Table -> Pandas-> JSON we can directly go to JSON by using `to_pydict()` which is a little slower than the current approach but at least `select` works properly now. Lmk what you guys think of it @lhoestq, @eladsegal? |