The link to the stored-in-image imagenet64x64 dataset. And a code in PyTorch with resnet/wrn for it.
The dataset is from imagenet64x64. They downsampled the imagenet to 16x16, 32x32, and 64x64. Amazing work! The authors provides the data in 10 binary files like cifar10/cifar100. And it's not a natural way in PyTorch to load such big files(each binary file is about 1.6 GB and it costs about 20~30 seconds to load a file).
So I decompress the set of 64x64 into png files like the imagenet, and we can use ImageFolder and DataLoader in PyTorch to load the dataset.
imagenet64 baidu notoriously Slow
But my suggestion is to download from the official website then decompress by yourself using the code
I also upload the ipynb code about how to load the batch data and save to images.
上传了notebook代码， 32x32, 16x16的数据就不弄了。
numpy
imageio
The main function is from the authors' code load batch data
Then I use the code from PyTorch offical examples imagenet to train on the dataset.
pip install torch==1.3.1 torchvision==0.4.2 tqdm
python main.py -a resnet18 [imagenet-folder with train and val folders]wideresnet ···bash CUDA_VISIBLE_DEVICES=3 python main_64.py -a wrn -depth 28 --k 2 -drop 0 -b 64 /shared/imagenet64png --epochs 40 --decay_epochs 25 35 ···
The results(use the default configuration) are coming:
| network | GPU:0 | per epoch | epochs | top1 accuracy(%) | top5 accuracy(%) |
|---|---|---|---|---|---|
| resnet 18 | 1.50G | 3 min 10 sec | 90 | 42.96 | 67.72 |
| resnet 50 | 3.20G | 6 min 10 sec | 90 | 51.96 | 75.95 |
| WRN-28-2,drop 0 | 2.71G | 23 min 46 sec | 40 | 60.02 | 83.04 |
TITAN RTX, imagenet32
| network | Memory | epoch | per epoch | speed | total time | top1 accuracy(%) | top5 accuracy(%) |
|---|---|---|---|---|---|---|---|
| resnet 50 | 1717MiB | 90 | 03:52 | 21.57it/s | 6:09:35 | 38.056 | 62.386 |
| WRN-28-2,drop 0 | 2601MiB | 40 | 05:17 | 15.78it/s | 3:49:15 | 46.606 | 71.806 |
Use the default epochs from resnet imagenet and wide resnet imagenet64x64. I found it's is much faster than the authors reported in their paper.
ImageNet32, WRN-28-2-drop0
ImageNet32 resnet 50
The model is from wide resnet which gets more stars than the official repo.
And I changed it a bit following the implementation of the imagenet64x64 authors WRN imagenet