TypeError: unrecognized operator 'Upsample'

[christophstach]

When using a style transfer model compiled to onnx, I get the following Error in onnxjs 0.1.5:

core.js:15724 ERROR Error: Uncaught (in promise): TypeError: unrecognized operator 'Upsample'
TypeError: unrecognized operator 'Upsample'

I am using the pytorch function

x = F.interpolate(x, mode='nearest', scale_factor=self.scale_factor)

which is translated to the onnx Upsample operator, wich is not supported in onnxjs yet.

[gnsmrky]

@christophstach You will have to implement Upsample() by Concat() ops. Here is a working ONNX.js Fast Neural Style. Still buggy but should work ok on desktop browsers.

I am still working on putting up the modified PyTorch code to GitHub. Should be done soon.

[fs-eire]

Thanks to @gnsmrky for the explaination. According to latest ONNX spec, Unsample is deprecated in opset v10. Is there an equivalent alternative operator does the same thing?

[gnsmrky]

@fs-eire I believe the equivalent op would be Resize. Still need PyTorch ONNX export to change accordingly, though.

[gnsmrky]

@christophstach Take a look at my PyTorch fast-neural-style (FNS) for Web. I've put up all the details to make the PyTorch FNS work with ONNX.js. Some network model tweaking methods for running in browsers are mentioned. Should give you a good reference to make it run with ONNX.js.

More technical details for op replacements are at Making the PyTorch fast-neural-style (FNS) to ONNX.js conversion work in browsers. It talks about how to make the network compatible with ONNX.js by modifying the PyTorch codes.

Hope this helps you and everyone here to know how PyTorch --> ONNX.js works better.

:)