onnxjs support for torchvision FCN models

[Andrew Green]

I've exported a pretrained resnet101-FCN from pytorch using the following:

import torch
import torchvision.models as models
import torch.onnx
import numpy as np

seg_rn = models.segmentation.fcn_resnet101(pretrained=True, progress=True, num_classes=2, aux_loss=None)
seg_rn.eval()

x = torch.randn(1,3,512,512, requires_grad=True)
torch_out = seg_rn(x)

torch.onnx.export(seg_rn, x, "seg_rn.onnx", 
                    export_params=True, 
                    opset_version=12, 
                    do_constant_folding=True, 
                    input_names=['input'], 
                    output_names=['out'], 
                    dynamic_axes={'input': {0:'batch_size'}, 'out':{0:'batch_size'}})

And tried to import it in onnxjs with this:

window.session = new InferenceSession({ backendHint: 'webgl' }); // does that make it global?

const modelURL = "./models/seg_rn.onnx";
await window.session.loadModel(modelURL);

However, I get the following error when the model tries to load:

Uncaught (in promise) TypeError: cannot resolve operator 'Shape' with opsets: ai.onnx v11

Obviously this is related to the Shape operator.

• What exactly is the issue?
• I saw that Shape was updated recently (just after the last release), will that update fix this issue?
• It looks like I'm using different opsets when I export the model and when I try to load it. How do I control the opset used by onnxjs?

Thanks