[BUG] MobileNetV3 doesn't work from timm version 0.6..

[JeonghwaYoo-R]

Describe the bug When I update timm from 0.5.4 to latest version, I encountered error. I only used forward_features from timm, then append some custom module.

The number of class is 8.

When I roll back the version to 0.5.4, no error occured.

I used mobilenetv3_large_100_miil_in21k

To Reproduce Steps to reproduce the behavior:

1. Update the timm to 0.6.xx
2. Use MobileNet model (mobilenetv3_large_100_miil_in21k)

Expected behavior A clear and concise description of what you expected to happen.

Screenshots If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: Window 10
• This repository version [e.g. pip 0.3.1 or commit ref]
• PyTorch version w/ CUDA/cuDNN [e.g. from conda list, 1.7.0 py3.8_cuda11.0.221_cudnn8.0.3_0]

Additional context Add any other context about the problem here.

[rwightman]

@JeonghwaYoo-R in 0.6 I altered forward_features, added forward_head to be more consistent across all model types, forward_features now breaks before final head layers and global pooling. MobileVitV3 has an odd head in that there is another conv layer after the global pool.

https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/mobilenetv3.py#L192-L211

you can see forward_features returns features earlier than before as that last conv is part of the 'head'. If you want old behaviour you can

model.reset_classifier(num_classes=0) or create_model('name', num_classes=0), reset the classifier part only instead of calling forward features, then pass the output ofmodel(x)` to your custom module

instead of forward_features only, call

x = model.forward_features(x)
x = model.forward_head(x, pre_logits=True)

[JeonghwaYoo-R]

@rwightman

Thank you! I solved this problem by creating a model through num_class=0.