Only supports ResNet-50 for now. Since the augmentation in PyTorch will be slightly different from the orignal, expect some differences in accuracy. I am not entirely sure about why the crop_only is 5 point worse. Original weights from BYOL. This is a basic script which should generally work with most versions of PyTorch and Torchvision, but it's written with PyTorch (1.4) and Torchvision (0.5).
# convert the weights
python convert.py pretrain_res50x1.pkl pretrain_res50x1.pth.tar
# validate the weights
python validate.py pretrain_res50x1.pth.tar /datasets/imagenet/val
| Name | Original Acc | Converted Acc |
|---|---|---|
| pretrain_res50x1 | 74.4 | 74.6 |
| res50x1_batchsize_2048 | 72.4 | 72.3 |
| res50x1_batchsize_1024 | 72.2 | 72.3 |
| res50x1_batchsize_512 | 72.2 | 72.1 |
| res50x1_batchsize_256 | 71.8 | 71.9 |
| res50x1_batchsize_128 | 69.6 (+- 0.5) | 69.7 |
| res50x1_batchsize_64 | 59.7 (+- 1.5) | 58.2 |
| res50x1_crop_and_blur_only | 61.1 (+- 0.3) | 62.9 |
| res50x1_crop_and_color_only | 70.7 | 69.1 |
| res50x1_crop_only | 59.4 (+- 0.3) | 55.3 |
| res50x1_no_color | 63.4 (+- 0.7) | 63.8 |
| res50x1_no_grayscale | 70.3 | 70.5 |