ProGamerGov

The invalid device ordinal is error is normally given when you specify a non existent GPU ID. The GPU value in the script should be set to the PyTorch GPU...

Basically this is what neural-style-pt does with GPU IDs (example with the Python Interpreter): ``` import torch a = torch.randn(3) a.to('cpu') # Puts tensor 'a' on the CPU if it...

Are you running multiscale generation for each frame before adding them together, or adding the frames together for each step?

Nice work on translating my old code! I think that `loss.backward()` shouldn't require `retain_graph=True`, as I don't think that we need to save the intermediate values. The intermediate values will...

PyTorch has a few differences with Torch7, due to the Autograd feature, and because of Python. I suspect that the issue lies with something in the MaskedStyleLoss function. This example...

@genekogan Looks good! I'm not sure about the licensing issue with translated code with respect to the segmentation code from Lua. That could conflict with neural-style-pt license, and as such...

@genekogan Are making sure that that TV weight is set to 0 in your experiments?

@genekogan I was actually recently looking into histogram loss myself, after seeing the results from: https://arxiv.org/pdf/1701.08893.pdf. It was used in [deep-painterly-harmonization](https://github.com/luanfujun/deep-painterly-harmonization/blob/master/neural_paint.lua#L602), and it seems like a better idea than performing...

Also, on a bit of an unrelated note, have you tried to get gradient normalization from the original Lua/Torch7 code working in PyTorch? I did figure out that it's more...

@genekogan I do recall seeing some issues with gradients when using masks that had very small/thin regions surrounded by other regions. Maybe something like that could be being exaggerated by...