RuntimeError: downsample feature map of size torch.Size([1, 64, 135, 240])

[wolf943134497]

Hi @ducha-aiki @jatentaki Thanks for your great works! I met some errors: Traceback (most recent call last): File "detect.py", line 249, in described_samples = extract(dataset, args.h5_path) File "detect.py", line 201, in extract batched_features = extract(bitmaps) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch_dimcheck/dimcheck.py", line 110, in wrapped result = func(*args, kwargs) File "/home/disk/disk/model/disk.py", line 57, in features descriptors, heatmaps = self._split(self.unet(images)) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in call result = self.forward(*input, *kwargs) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch_dimcheck/dimcheck.py", line 110, in wrapped result = func(args, kwargs) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/unets/unet.py", line 100, in forward features.append(layer(features[-1])) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in call result = self.forward(*input, *kwargs) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch/nn/modules/container.py", line 92, in forward input = module(input) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in call result = self.forward(input, **kwargs) File "/home/anaconda3/envs/pytorch1/lib/python3.6/site-packages/unets/ops.py", line 42, in forward raise RuntimeError(msg) RuntimeError: Trying to downsample feature map of size torch.Size([1, 64, 135, 240])

which version of pytorch? any suggestions? Thanks!

[ducha-aiki]

I am not the author, please don't assign to me a credit for creating DISK.

Regarding your question --- set resize values to be power of 2, e.g. 1024x768. Altough I would also prefer to hear about more robust solution from @jatentaki

[jatentaki]

Haha, @ducha-aiki is not a maintainer but apparently responds even faster than I do.

More precisely, it has to be not only power of 2, but multiples of 16 (since we have 4 downsampling steps, each reducing the resolution by 2 --> 2^4 = 16). In practice, please set the --height and --width flags to a size which is closest to your input size and divisible by 16 (don't worry if it doesn't exactly preserve the aspect ratio, it will internally maintain it with padding).

[jatentaki]

I added more informative exception messages around this issue. @wolf943134497 is the problem resolved?

[wolf943134497]

@jatentaki the problem is solved! Thanks for your suggestions and great works! Sorry to disturb you, ducha-aiki. Thanks for your help!