i found a bug: if using bceloss in `train.py`, then `unet3plus.py` should not apply a sigmoid activation # ❌ before # return F.sigmoid(d1) # ✅ after return d1 Otherwise you're applying sigmoid twice and the loss may not decrease
i found a bug: if using bceloss in
train.py, thenunet3plus.pyshould not apply a sigmoid activation❌ before
return F.sigmoid(d1)
✅ after
return d1
Otherwise you're applying sigmoid twice and the loss may not decrease