About converting model to .onnx-> ncnn?

[zelenooki87]

I am very impressed with your project and upscale models. Tried project in google colab and results are fascinanting, even better than commercial products. Cause I have AMD radeon GPU, I would like to convert FeMaSR models to ncnn, (.bin and .params) and use it in OSource programs like cupscale or quality scaler. Do you have recommendations for conversion? Thank you very, very much in advance!

[chaofengc]

Thanks for interest. Since I have no experience in the conversion, I am not able to give detailed instructions. You may try some official tutorials, for example:

• https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html
• https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx

[zelenooki87]

I think it is not so hard. https://github.com/Tencent/ncnn/tree/master/tools/pnnx Could you specify input shape of the pretrained models? tnx

[chaofengc]

The input width and length should be multiplier of 32(16) for x2(x4) model

[zelenooki87]

please suggest values to type when converting torch summary of 4x SR .pth model https://pastebin.com/4JSUFB0f tried with [1,3,224,224] standard values but without success maybe to try 1,4,256,256 for 4x model? without input shape conversion is impossible. Again thank you for great work!

[chaofengc]

224×224 should be ok. You may try 256×256, which should definitely work.