Error when running PSPNet with PascalVOC

[flriancu]

Hello,

When I run python train.py --arch pspnet --dataset pascal --n_epoch 500 --batch_size 8
I get the following error.
Tested with PyTorch 0.4.1 on Linux.

Traceback (most recent call last):
  File "train.py", line 162, in <module>
    train(args)
  File "train.py", line 92, in train
    loss.backward()
  File "/home/flr/miniconda3/lib/python3.6/site-packages/torch/tensor.py", line 93, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/flr/miniconda3/lib/python3.6/site-packages/torch/autograd/__init__.py", line 90, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

I suspect it might be from the cross_entropy2d in ptsemseg/loss.py, but I am not sure. Please let me know if you need more details.

[adam9500370]

Hi, @flriancu .
You can check #96 .

[flriancu]

@adam9500370 Thanks, solved. Any idea if the align_corners being turned off by default in 0.4 has any influence for the results (see comment in #96)?

[adam9500370]

I think you still need to set align_corners=True when using pre-trained models. (see comment in #81)