There seems to be some kind of incompatibility between PyTorch’s Vulkan build and the ESRGAN/Real-ESRGAN neural net architecture. I have an AMD card that isn’t supported by ROCm (RX 5700), so is there any way for me to run this?
I’m hoping there’s a fix that doesn’t require me to like, re-invent the ESRGAN architecture.
Error Vulkan::convolution not available! Reason: The provided (weight, bias, stride, padding, dilation, groups, transposed, output_padding, output_min, output_max) parameters are either invalid individually or their combination is not supported by Vulkan impl.
Running ‘classic’ ESRGAN has a similar-seeming issue:
...
File "/home/travis/development/ESRGAN/RRDBNet_arch.py", line 34, in forward
return x5 * 0.2 + x
RuntimeError: Vulkan binary elementwise ops require channel dimension to be equal!
I know that this project can be compiled with NCNN to run with Vulkan - what’s the problem with the regular PyTorch implementation?