Editing forward functions of MobileNetV2 or other backboned models

[zeynepozdemir]

I want to use mobilenetv2, one of the backboned models defined in mmpretrain, as a backboned model in mmfewshot.

I explained the problem I had there in the link below.
open-mmlab/mmfewshot#129

1: https://github.com/open-mmlab/mmpretrain/blob/main/mmpretrain/models/backbones/mobilenet_v2.py

2: https://github.com/open-mmlab/mmfewshot/blob/main/mmfewshot/classification/models/backbones/resnet12.py

In short, the forward part of the code in 1 returns {tuple :1}. It contains {Tensor: (25,1280,3,3).

2 gives direct output as {Tensor:(25,1600)}.

After that, I want to use Ptotypical Networks, so I need to get output as in 2. How can I do it without breaking the model structure?

It is very important for me to solve this critical point, can you help me?

[Ezra-Yu]

the mmpretrain.model.backbone.forward return a tuple, and mmfewshot.classification.backbone.forward just return a tensor.
Try to modify the code of mmfewshot in https://github.com/open-mmlab/mmfewshot/blob/af4fad50b0e00946467244dc8bce4d232d2b79c2/mmfewshot/classification/models/classifiers/base.py#L73

From :

        x = self.backbone(img)

To:

        x = self.backbone(img)[-1]

[zeynepozdemir]

I got the following error when I applied the technique you said.

I solved my problem by updating the forward function like this.. Thank You :)