A model that concats two inputs is not converted correctly.

❓Question

I want to convert an inpainting model with two inputs.

https://github.com/researchmm/AOT-GAN-for-Inpainting

One input is the masked original image (x) and the other input is the mask image (mask). When converted to torchscript, the model concats the two inputs as intended, as shown in the attached image.

x = torch.rand(1, 3, 512, 512)
mask = torch.rand(1, 1, 512, 512)

ts = torch.jit.trace(model, (x,mask))

However, when I convert to coreml model with the code below, the conversion itself completes correctly, but the converted mlmodel double joins one input (mask) and ignores the other input (x).

import coremltools as ct
mlmodel = ct.convert(ts, inputs=[ct.ImageType(shape=x.shape),ct.ImageType(shape=mask.shape)])
mlmodel.save("inpainting.mlmodel")

concat receives x:1 and mask. The output of the ignored input (x) is x:0, and the mask input (mask) outputs x:1 and mask and is concat.

Is there a way to concat the two inputs as intended by the original model?

Thank you for your work. Best,

apple / coremltools

A model that concats two inputs is not converted correctly. #1725

❓Question