This is a converter to convert TensorFlow checkpoints provided in SimCLR repo to PyTorch format, to facilitate related research.
-
Firstly, download and unzip the checkpoints from SimCLR repo, you will get 3 folders:
ResNet50_1x,ResNet50_2x, andResNet50_4x. -
Run the following commands to convert the 3 checkpoints:
python convert.py ResNet50_1x/model.ckpt-225206 resnet50-1x.pth python convert.py ResNet50_2x/model.ckpt-225206 resnet50-2x.pth python convert.py ResNet50_4x/model.ckpt-225206 resnet50-4x.pth
You will get 3 PyTorch checkpoints,
resnet50-1x.pth,resnet50-2x.pth,resnet50-4x.pth. The model definition is inresent_wider.py.
To validate the correctness of the conversion, I tested the performance of the models using PyTorch standard augmentation on ImageNet (but without normalization, as the original TF models were not trained with normalization), using commands:
python eval.py /path/to/imagenet -a resnet50-1x/resnet50-2x/resnet50-4x
The performance is:
| Model | TensorFlow Top-1 | PyTorch Top-1 |
|---|---|---|
| ResNet-50 (1x) | 69.1 | 68.9 |
| ResNet-50 (2x) | 74.2 | 74.1 |
| ResNet-50 (4x) | 76.6 | 76.4 |
There is a slight degradation, which should be due to the difference in data pre-processing (e.g., resize) in two frameworks.