I was checking grad value of bias w.r.t the last conv-filter.
Here is my code.
myLoss = abs(pred - answer).mean() # 1x1x256x256 -> scalar value
print( myLoss.item() )
print( last_layers_bias.grad )
I guess myLoss.item() and last_layers_bais.grad has same value.
But two values are so different, and I geuss this is a bug?