This fork adds support for ResNet18 models conversion to PyTorch. It also adds TF & PyTorch converted checkpoints with SimCLR models pretrained on CIFAR10 and CIFAR100.

The implementation was tested as follows:

• CIFAR10 & CIFAR100 SimCLR models were trained using the code & commands provided in the Google's simclr repo as of this commit;
• Only unsupervised pretraining phase was executed; no linear head fine-tuning was performed;
• The last checkpoint was converted to PyTorch (see convert_resnet18.ipynb);
• The converted models were loaded into PyTorch, all layers apart from the last block and an FC head were frozen, and using supervised fine-tuning I was able to achieve 95% and 74.4% validation accuracy on CIFAR10 and CIFAR100 respectively;
• The weights of pretrained backbones (ignore FC weights) can be found in cifar10-ckpts and cifar100-ckpts folders;
• Note: I cannot guarantee that the code in this repository still works with ResNet50/101, I didn't test it after my modifications.

Pretrained SimCLRv2 models in Pytorch.

python download.py r152_3x_sk1
python convert.py r152_3x_sk1/model.ckpt-250228 [--ema]
python verify.py r152_3x_sk1.pth