yaoyao-liu/meta-transfer-learning

resnet model not consistant between pytorch and tensorflow

leon-liangwu opened this issue · 4 comments

leon-liangwu commented

I find that the ResNet models of PyTorch version and TensorFlow version are different.
Is that right? What is the performance of the PyTorch version?

yaoyao-liu commented

Thanks for your interest in our work.
We use the TensorFlow version for the experiments in the paper.
PyTorch version is a re-implementation based on this repo. So some details are different from the original TensorFlow version.

For the performance of the PyTorch version, you may refer to this issue.

leon-liangwu commented

Thanks for your reply very much.
Have you ever tested the same architecture from your TensorFlow version in PyTorch?
What is that performance?

yaoyao-liu commented

Currently, we don't have the PyTorch results for other backbones.

If you hope to re-implement the architecture (ResNet-12), you may follow this repo.

We plan to update other backbones (ResNet-12, WRN-28-10) later.

leon-liangwu commented

OK. Thanks for your information.