UpResNet10 Model Conversion

[tcyrus]

When trying to convert the UpResNet10 model to CoreML, I get a similar error to the CUnet model conversion.

Unfortunately, I didn't log the output at the time.

[tcyrus]

I was able to convert the UpResNet10 models for Chainer (model source) to ONNX.

The Chainer models look slightly different from the Caffe models when viewed in Netron.

1. Model is Horizontally Mirrored
   This is probably ok, but it's confusing

2. Unusual Input and Output Labels
   Again, probably ok, but it's still confusing

3. ReLU is sometimes replaced with LeakyReLU
   This is because Caffe uses the same layer type (ReLU) for both ReLU and LeakyReLU (docs)

4. Split layers are merged with the input layer.
   This should be fine as the Split doesn't have any specific attributes (docs)

5. CropCenter is replaced with Slice
   This is mainly confusing me because I don't know the difference between the two

6. Modified step
   This is probably the trickiest part to explain since I know little to nothing about ML.
   Based on my limited understanding, parts of the UpResNet10 algorithm can be broken up into smaller groups of layers which I'm going to call "steps" since I don't know what the real name for it is. These steps probably have different weight values attached to them, but they're the same in terms of structure.

   Step for Caffe Model:
   

   Step for Chainer Model:
   

   This can be broken down into smaller differences:
   a. Pooling is replaced with ReduceMean (😱)
   This is really confusing for me since I believe the proper operation should be GlobalAveragePool
   b. InnerProduct is replaced with Gemm
   This is a documented conversion for CoreML (link) if that helps
   c. Axpy is replaced with a combination of Unsqueeze, Expand, Mul, and Add (😱)

7. Eltwise is replaced with Add at the end
   found some code doing this conversion here

[tcyrus]

The steps to resolution depend on the reason for the differences:

• If the conversion is 1 to 1, then ignore it
• If the conversion is un-optimal or has different behavior (ReduceMean), then fix it
• If the conversion is many to one (Axpy), make note of it as a potential ONNX Backend Optimisation

[tcyrus]

Found confirmation that the correct conversion is to GlobalAveragePool (link)

I need to correct the model via editing the onnx file once confirming the input / output dimensions.

[tcyrus]

After looking through more details. the Mul + Add seems to be the proper replacement to Axpy (link). This could also be the reason for the mirroring of the model.

I am still unclear on the reason behind Unsqueeze + Expand, but this could be related to the input or output shape of Gemm.

[tcyrus]

I took a break from this for a while. I'll probably try passing the resulting model through onnxsim and/or onnxoptimizer to try and see if it'll simplify the graph by replacing redundant graph operations.

[tcyrus]

If I'm reading this commit correctly, the current UpResNet10 models for waifu2x-caffe are generated using the models from waifu2x-chainer. Based on this knowledge (also, I don't understand any of this), I have decided to use the resulting model from the conversion.
I'll close this issue once I upload the models, but I do not recommend using them unless you know what you're doing.