metadata

task_categories:
  - image-classification
pretty_name: imagenet5GB
size_categories:
  - 1M<n<10M

ImageNet-1k in 5GB

The full ImageNet-1k compressed to less than 5 GB

Compression procedure:

Resize shorter edge to 288 and crop longer edge to a multiple of 32
Analysis transform: DC-AE f32 c32
Quantization: 8 bit float (e4m3)
Entropy coding: TIFF (CMYK) with deflate

Example dataloader for training

import torch
import datasets
from types import SimpleNamespace
from diffusers import AutoencoderDC
from torchvision.transforms.v2 import ToPILImage, PILToTensor, RandomCrop, CenterCrop
from walloc.walloc import pil_to_latent
from IPython.display import display

device = 'cuda'
config = SimpleNamespace()
config.crop_size = 160
config.valid_crop_size = 288

ds = datasets.load_dataset('danjacobellis/imagenet_288_dcae_fp8')
decoder = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers", torch_dtype=torch.float32).decoder.to(device)

rand_crop = RandomCrop((config.crop_size//32,config.crop_size//32))
cent_crop = CenterCrop((config.valid_crop_size//32,config.valid_crop_size//32))

def train_collate_fn(batch):
    B = len(batch)
    x = torch.zeros((B, 32, config.crop_size//32, config.crop_size//32), dtype=torch.torch.float8_e4m3fn)
    y = torch.zeros(B, dtype=torch.long)
    for i_sample, sample in enumerate(batch):
        y[i_sample] = sample['cls']
        z = pil_to_latent([sample['latent']], N=36, n_bits=8, C=4)[:,:32]
        x[i_sample,:,:,:] = rand_crop(z.to(torch.int8).view(torch.float8_e4m3fn))
    return x, y

def valid_collate_fn(batch):
    B = len(batch)
    x = torch.zeros((B, 32, config.valid_crop_size//32, config.valid_crop_size//32), dtype=torch.torch.float8_e4m3fn)
    y = torch.zeros(B, dtype=torch.long)
    for i_sample, sample in enumerate(batch):
        y[i_sample] = sample['cls']
        z = pil_to_latent([sample['latent']], N=36, n_bits=8, C=4)[:,:32]
        x[i_sample,:,:,:] = cent_crop(z.to(torch.int8).view(torch.float8_e4m3fn))
    return x, y

%%time
# warmup batch
x,y = valid_collate_fn(ds['train'].select(range(64)))
with torch.no_grad():
    xh = decoder(x.to(torch.float32).to(device))

CPU times: user 1.68 s, sys: 124 ms, total: 1.8 s
Wall time: 1.47 s

%%time
x,y = valid_collate_fn(ds['train'].select(range(64)))
with torch.no_grad():
    xh = decoder(x.to(torch.float32).to(device))

CPU times: user 282 ms, sys: 2.51 ms, total: 285 ms
Wall time: 29.2 ms

for img in xh[:2]:
    display(ToPILImage()(img.clamp(-0.5,0.5)+0.5))

!jupyter nbconvert --to markdown README.ipynb

[NbConvertApp] Converting notebook README.ipynb to markdown
[NbConvertApp] Support files will be in README_files/
[NbConvertApp] Making directory README_files
[NbConvertApp] Writing 2752 bytes to README.md