ResNet strikes back: An improved training procedure in timm

[loukasilias]

Hi everyone,
I would like to ask you how I could load the model after importing the timm library.
The same question has to do with other models as well (swin transformers etc.).

Thank you.

[talhaanwarch]

I think its ecaresnet50t

[rwightman]

Weights related to experiments from the paper are all here https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-rsb-weights

Anything with an a1_0 , a2_0, a3_0 are the results for that recipe for seed 0 as per the paper tables. I have a few other weights for resnet50 specifically that demo an a1h (a1 based recipe with more aug), and the b , c1 / c2 , d recipes in the appendix. The default weights for resnet50 now use a1_0. But you can download those files manually and load into model using standard pytorch mechanisms (or the checkpoint / initial-checkpoint args in include val / train scripts)