where you use the layer activate function is relu，but becwithlogitsloss has the sigmoid, that's error, it will train hardly, that's why non convergence.

as usual, the last layer of segment model is just a conv and no act.

in the paper, he conclused ce loss is better than mse.

An error occurred when I was running python models.py

python models.py
Traceback (most recent call last):
File "models.py", line 163, in
ret = snet(img)
File "/home/A305/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "models.py", line 87, in forward
w = F.avg_pool2d(x3, x3.size(2))
RuntimeError: Given input size: (64x88x32). Calculated output size: (64x1x0). Output size is too small

Is there any way to fix it?

Ubuntu version: 20.04
Pytorch version: 1.7
CUDA version: 11.0

segment_net.eval()

decision_net.eval()

hi:
tks for you job, I have a question
I opened the comment and it didn't work very well in test.py, but I didn't understand why.Please give advice or comments