downstream tasks in norm_layer for EfficientFormerV2

[deepkyu]

Hi,

First of all, congrats for this awesome research 🎉

I have a simple question while reading your EfficientFormerV2 codes.

In your backbone codes for detection and segmentation, I found that norm_layers are not applied in forward_token:

EfficientFormer/segmentation/backbonev2.py

Lines 650 to 653 in 2c0e950

	if self.fork_feat and idx in self.out_indices:
	# norm_layer = getattr(self, f'norm{idx}')
	# x_out = norm_layer(x)
	outs.append(x)

However, for your backbone in classifcation, it forwards with the norm_layer:

EfficientFormer/models/efficientformer_v2.py

Lines 622 to 625 in 2c0e950

	if self.fork_feat and idx in self.out_indices:
	norm_layer = getattr(self, f'norm{idx}')
	x_out = norm_layer(x)
	outs.append(x_out)

Actually, it seems that the difference in the above between classification and other tasks does not occur in your EfficientFormer code.
I tried my best to find the detail in both your code and paper, but I couldn't.

So, I kindly ask you if you could explain why this should be different.

Thank you in advance.

+)
To clarify my question, I added the corresponding code lines from EfficientFormer used in segmentation:

EfficientFormer/segmentation/backbone.py

Lines 478 to 483 in 2c0e950

	if self.fork_feat and idx in self.out_indices:
	norm_layer = getattr(self, f'norm{idx}')
	if len(x.size()) != 4:
	x = x.transpose(1, 2).reshape(B, C, H, W)
	x_out = norm_layer(x)
	outs.append(x_out)

Those seems to be the outputs from each norm layer.