Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -26,15 +26,42 @@ Each document exists as a collection of a pdf, a tiff image with the same conten
|
|
26 |
</center>
|
27 |
|
28 |
|
29 |
-
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
```python
|
32 |
-
from datasets import load_dataset
|
33 |
dataset = load_dataset('pixparse/IDL-wds', streaming=True)
|
34 |
print(next(iter(dataset['train'])).keys())
|
35 |
>> dict_keys(['__key__', '__url__', 'json', 'ocr', 'pdf', 'tif'])
|
36 |
```
|
37 |
|
|
|
38 |
For faster download, you can directly use the `huggingface_hub` library. Make sure `hf_transfer` is installed prior to downloading and mind that you have enough space locally.
|
39 |
|
40 |
```python
|
@@ -50,38 +77,6 @@ For faster download, you can directly use the `huggingface_hub` library. Make su
|
|
50 |
|
51 |
```
|
52 |
|
53 |
-
|
54 |
-
|
55 |
-
# TODO The following example uses fitz which is AGPL. We should also recommend the same with pypdf.
|
56 |
-
|
57 |
-
```python
|
58 |
-
from chug import create_wds_loader, create_doc_anno_pipe
|
59 |
-
|
60 |
-
# TODO import image transforms and text transforms from pixparse.data
|
61 |
-
|
62 |
-
|
63 |
-
decoder = create_decoder_pipe(
|
64 |
-
image_preprocess=image_preprocess, # callable of transforms to image tensors
|
65 |
-
anno_preprocess=anno_preprocess, # callable of transforms to text into tensors of labels and targets
|
66 |
-
image_key="pdf",
|
67 |
-
image_fmt="RGB",
|
68 |
-
)
|
69 |
-
|
70 |
-
loader = create_wds_loader(
|
71 |
-
"/my_data/idl-train-*.tar",
|
72 |
-
decoder,
|
73 |
-
is_train=True,
|
74 |
-
resampled=False,
|
75 |
-
start_interval=0,
|
76 |
-
num_samples=2159432,
|
77 |
-
workers=8,
|
78 |
-
batch_size=32, # adjust to your architecture capacity
|
79 |
-
seed=seed, # set a seed
|
80 |
-
world_size=world_size, # get world_size from your training environment
|
81 |
-
)
|
82 |
-
|
83 |
-
```
|
84 |
-
|
85 |
Further, a metadata file `_pdfa-english-train-info-minimal.json` contains the list of samples per shard, with same basename and `.json` or `.pdf` extension,
|
86 |
as well as the count of files per shard.
|
87 |
|
@@ -152,28 +147,7 @@ That way, columnar documents can be better separated. This is a basic heuristic
|
|
152 |
|
153 |
For each pdf document, we store statistics on number of pages per shard, number of valid samples per shard. A valid sample is a sample that can be encoded then decoded, which we did for each sample.
|
154 |
|
155 |
-
This instance of IDL is in [webdataset](https://github.com/webdataset/webdataset/commits/main) .tar format. It can be used with webdataset library or current releases of Hugging Face `datasets`.
|
156 |
-
Here is an example using the "streaming" parameter. We do recommend downloading the dataset to save bandwidth.
|
157 |
-
|
158 |
-
```python
|
159 |
-
dataset = load_dataset('pixparse/IDL-wds', streaming=True)
|
160 |
-
print(next(iter(dataset['train'])).keys())
|
161 |
-
>> dict_keys(['__key__', '__url__', 'json', 'ocr', 'pdf', 'tif'])
|
162 |
-
```
|
163 |
-
|
164 |
-
For faster download, you can use directly the `huggingface_hub` library. Make sure `hf_transfer` is install prior to downloading and mind that you have enough space locally.
|
165 |
|
166 |
-
```python
|
167 |
-
import os
|
168 |
-
|
169 |
-
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
|
170 |
-
|
171 |
-
from huggingface_hub import HfApi, logging
|
172 |
-
|
173 |
-
#logging.set_verbosity_debug()
|
174 |
-
hf = HfApi()
|
175 |
-
hf.snapshot_download("pixparse/pdfa-english-train", repo_type="dataset", local_dir_use_symlinks=False)
|
176 |
-
```
|
177 |
### Data, metadata and statistics.
|
178 |
<center>
|
179 |
<img src="https://huggingface.co/datasets/pixparse/IDL-wds/resolve/main/doc_images/idl_page_example.png" alt="An addendum from an internal legal document" width="600" height="300">
|
|
|
26 |
</center>
|
27 |
|
28 |
|
29 |
+
This instance of IDL is in [webdataset](https://github.com/webdataset/webdataset/commits/main) .tar format.
|
30 |
+
|
31 |
+
### Usage with `chug`
|
32 |
+
Check out [chug](https://github.com/huggingface/chug), our optimized library for sharded dataset loading!
|
33 |
+
|
34 |
+
```python
|
35 |
+
import chug
|
36 |
+
task_cfg = chug.DataTaskDocReadCfg(page_sampling='all')
|
37 |
+
data_cfg = chug.DataCfg(
|
38 |
+
source='pixparse/IDL-wds',
|
39 |
+
split='train',
|
40 |
+
batch_size=None,
|
41 |
+
format='hfids',
|
42 |
+
num_workers=0,
|
43 |
+
)
|
44 |
+
data_loader = chug.create_loader(
|
45 |
+
data_cfg,
|
46 |
+
task_cfg,
|
47 |
+
)
|
48 |
+
sample = next(iter(data_loader))
|
49 |
+
```
|
50 |
+
|
51 |
+
|
52 |
+
|
53 |
+
### Usage with datasets
|
54 |
+
|
55 |
+
This dataset can also be used with webdataset library or current releases of Hugging Face `datasets`.
|
56 |
+
Here is an example using the "streaming" parameter. We do recommend downloading the dataset to save bandwidth.
|
57 |
+
|
58 |
```python
|
|
|
59 |
dataset = load_dataset('pixparse/IDL-wds', streaming=True)
|
60 |
print(next(iter(dataset['train'])).keys())
|
61 |
>> dict_keys(['__key__', '__url__', 'json', 'ocr', 'pdf', 'tif'])
|
62 |
```
|
63 |
|
64 |
+
|
65 |
For faster download, you can directly use the `huggingface_hub` library. Make sure `hf_transfer` is installed prior to downloading and mind that you have enough space locally.
|
66 |
|
67 |
```python
|
|
|
77 |
|
78 |
```
|
79 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
Further, a metadata file `_pdfa-english-train-info-minimal.json` contains the list of samples per shard, with same basename and `.json` or `.pdf` extension,
|
81 |
as well as the count of files per shard.
|
82 |
|
|
|
147 |
|
148 |
For each pdf document, we store statistics on number of pages per shard, number of valid samples per shard. A valid sample is a sample that can be encoded then decoded, which we did for each sample.
|
149 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
### Data, metadata and statistics.
|
152 |
<center>
|
153 |
<img src="https://huggingface.co/datasets/pixparse/IDL-wds/resolve/main/doc_images/idl_page_example.png" alt="An addendum from an internal legal document" width="600" height="300">
|