Differenece pretrained model for resnet between torchvision.models and timm

[gihyunkim]

would you let me know the datasets used in pretraining for both models? I really need that information for my research.

[hankyul2]

Hi

If I understand your question correctly, they use the same dataset (ImageNet-1K) for pretraining resnet.
The reason for large gap in top-1 accuracy between timm(top-1: 80.3) and torchvision.models(top-1: 76.5) is difference in training hyperparameters.

timm train resnet 600 epochs with optimized training hyperparameters proposed in ResNet Strikes Back, which brings significant improvement in accuracy. You can find training hyperparameter details in here.

torchvision.models train resnet 90 epochs with original training hyperparameter proposed in Deep Residual Learning. You can find training hyperparameter details in here.

I want my answer could help you .🙂

Thank you

Hankyul