Commit
•
eae87d2
1
Parent(s):
57ac608
Update README.md
Browse files
README.md
CHANGED
@@ -3,7 +3,7 @@ size_categories:
|
|
3 |
- 1M<n<10M
|
4 |
viewer: false
|
5 |
---
|
6 |
-
# Imagenet.int8: Entire Imagenet dataset in 5GB
|
7 |
|
8 |
<p align="center">
|
9 |
<img src="contents/vae.png" alt="small" width="800">
|
@@ -36,31 +36,36 @@ So clearly, it doesn't make sense to download entire Imagenet and process with V
|
|
36 |
(BTW If you think you'll need higher precision, you can always further fine-tune your model on higher precision. But I doubt that.)
|
37 |
|
38 |
|
39 |
-
# How
|
40 |
|
41 |
-
First download this.
|
42 |
|
43 |
```bash
|
44 |
-
|
|
|
|
|
|
|
|
|
45 |
```
|
46 |
|
|
|
|
|
47 |
Then, you need to install [streaming dataset](https://github.com/mosaicml/streaming) to use this. The dataset is MDS format.
|
48 |
|
49 |
```bash
|
50 |
pip install mosaicml-streaming
|
51 |
```
|
52 |
|
53 |
-
Then, you can use
|
54 |
|
55 |
-
|
56 |
|
|
|
57 |
from streaming.base.format.mds.encodings import Encoding, _encodings
|
58 |
import numpy as np
|
59 |
from typing import Any
|
60 |
import torch
|
61 |
from streaming import StreamingDataset
|
62 |
-
from diffusers.models import AutoencoderKL
|
63 |
-
from diffusers.image_processor import VaeImageProcessor
|
64 |
|
65 |
class uint8(Encoding):
|
66 |
def encode(self, obj: Any) -> bytes:
|
@@ -72,7 +77,6 @@ class uint8(Encoding):
|
|
72 |
|
73 |
_encodings["uint8"] = uint8
|
74 |
|
75 |
-
|
76 |
remote_train_dir = "./vae_mds" # this is the path you installed this dataset.
|
77 |
local_train_dir = "./local_train_dir"
|
78 |
|
@@ -93,15 +97,18 @@ train_dataloader = torch.utils.data.DataLoader(
|
|
93 |
)
|
94 |
```
|
95 |
|
96 |
-
|
97 |
|
98 |
-
|
99 |
-
|
|
|
|
|
|
|
|
|
100 |
|
101 |
model = "stabilityai/your-stable-diffusion-model"
|
102 |
vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae").to("cuda:0")
|
103 |
|
104 |
-
|
105 |
batch = next(iter(train_dataloader))
|
106 |
|
107 |
i = 5
|
@@ -120,7 +127,9 @@ img.save("5th_image.png")
|
|
120 |
|
121 |
Enjoy!
|
122 |
|
123 |
-
#
|
|
|
|
|
124 |
|
125 |
```bibtex
|
126 |
@misc{imagenet_int8,
|
|
|
3 |
- 1M<n<10M
|
4 |
viewer: false
|
5 |
---
|
6 |
+
# Imagenet.int8: Entire Imagenet dataset in 5GB
|
7 |
|
8 |
<p align="center">
|
9 |
<img src="contents/vae.png" alt="small" width="800">
|
|
|
36 |
(BTW If you think you'll need higher precision, you can always further fine-tune your model on higher precision. But I doubt that.)
|
37 |
|
38 |
|
39 |
+
# How do I use this?
|
40 |
|
41 |
+
First download this. You can use `huggingface-cli` for that.
|
42 |
|
43 |
```bash
|
44 |
+
# Pro tip : use `hf_transfer` to get faster download speed.
|
45 |
+
pip install hf_transfer
|
46 |
+
export HF_HUB_ENABLE_HF_TRANSFER=True
|
47 |
+
# actual download script.
|
48 |
+
huggingface-cli download --repo-type dataset cloneofsimo/imagenet.int8 --local-dir ./vae_mds
|
49 |
```
|
50 |
|
51 |
+
|
52 |
+
|
53 |
Then, you need to install [streaming dataset](https://github.com/mosaicml/streaming) to use this. The dataset is MDS format.
|
54 |
|
55 |
```bash
|
56 |
pip install mosaicml-streaming
|
57 |
```
|
58 |
|
59 |
+
Then, you can very simply use the dataset like this:
|
60 |
|
61 |
+
(for more info on using Mosaic's StreamingDataset and MDS format, [reference here](https://docs.mosaicml.com/projects/streaming/en/stable/index.html))
|
62 |
|
63 |
+
```python
|
64 |
from streaming.base.format.mds.encodings import Encoding, _encodings
|
65 |
import numpy as np
|
66 |
from typing import Any
|
67 |
import torch
|
68 |
from streaming import StreamingDataset
|
|
|
|
|
69 |
|
70 |
class uint8(Encoding):
|
71 |
def encode(self, obj: Any) -> bytes:
|
|
|
77 |
|
78 |
_encodings["uint8"] = uint8
|
79 |
|
|
|
80 |
remote_train_dir = "./vae_mds" # this is the path you installed this dataset.
|
81 |
local_train_dir = "./local_train_dir"
|
82 |
|
|
|
97 |
)
|
98 |
```
|
99 |
|
100 |
+
By default, batch will have three attributes: `vae_output`, `label`, `label_as_text`.
|
101 |
|
102 |
+
Thats the dataloader! Now, below is the example usage. Notice how you have to reshape the data back to `(B, 4, 32, 32)` as they are decoded flattened.
|
103 |
+
|
104 |
+
```python
|
105 |
+
###### Example Usage. Decode back the 5th image. BTW shuffle plz
|
106 |
+
from diffusers.models import AutoencoderKL
|
107 |
+
from diffusers.image_processor import VaeImageProcessor
|
108 |
|
109 |
model = "stabilityai/your-stable-diffusion-model"
|
110 |
vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae").to("cuda:0")
|
111 |
|
|
|
112 |
batch = next(iter(train_dataloader))
|
113 |
|
114 |
i = 5
|
|
|
127 |
|
128 |
Enjoy!
|
129 |
|
130 |
+
# Citations
|
131 |
+
|
132 |
+
If you find this material helpful, consider citation!
|
133 |
|
134 |
```bibtex
|
135 |
@misc{imagenet_int8,
|