How to extract stage feature maps from the model?

[ma7555]

Hello wangermeng,

I want to implement this as a backbone in a different larger model. I need to be able to extract the 4 feature maps from different stages.
Upon checking model.summary() I do not find any layers with output equal to the paper reference. (H/32 W/32, C1) as an example for F1.

I have found out that after every attention block the output is:

(None, 3136, 64) >> (None, 56, 56, 64)
(None, 784, 128) >> (None, 28, 28, 128)
(None, 196, 320) >> (None, 14, 14, 64)
(None, 50, 512) >> (None, 7, 7, 512) ????????????????????????

The last layer should be (None, 49, 512)?

The line responsible for this change is where a 1 is added in the last stage

PVT-tensorflow2/model/PVT.py

Line 123 in 0fdc147

num_patch = feat_size_list[i][0] * feat_size_list[i][1] + 1