Tensorflow's ResNetV2 as timm model in Pytorch?

[lsecerovic]

Does the ResNetV2 model, as defined in tf, exist as a timm model in Pytorch?

TF implementation references Identity Mappings in Deep Residual Networks paper.

I've found resnetv2_50 in timm, but it references some other papers(this, this and this), so it doesn't look like the same thing as in TF.

[rwightman]

@lsecerovic it is (or rather can be) the same thing, the timm resnetv2 was added to support the BiT weights (uses weight standardization) and also support the 'hybrid vit' models that used the same arch but in a non preactivation style.

However, the resnetv2_50 model you point to should match the older TF resnetv2 definitions, with padding being a possible issue (depending on if the TF version uses 'SAME' or if it uses a fixed padding style that's more compatible with PyTorch). There's also a resnetv2_101 I've trained. Weights could be converted from TF but don't think those ones are great?