ValueError: Don't know how to translate op Unsqueeze when running converted PyTorch Model

Question

I'm running into problems trying to use a PyTorch model exported as an ONNX model with Caffe2. Here is my export code

the_model = torchvision.models.densenet121(pretrained=True)
garbage, model_inputs = preprocessing("test.jpg")
torch_out = torch.onnx._export(the_model,             
                           model_inputs,                       
                           "model_weights/chexnet-py.onnx",
                           export_params=True)

Now here is my testing code

model = onnx.load("model_weights/chexnet-py.onnx")
garbage, model_inputs = preprocessing("text.jpg")
prepared_backend = onnx_caffe2.backend.prepare(model)
W = {model.graph.input[0].name: model_inputs.numpy()}
c2_out = prepared_backend.run(W)[0]

This is returning the following error

ValueError: Don't know how to translate op Unsqueeze when running    converted PyTorch Model

Additional information pytorch version 1.0.0a0+6f664d3 Caffe2 is latest version (attempted building from source, pip, and conda). All gave same result.

Answer 1

Try looking into this, if you have to edit package called onnx-caffe2 to add the mapping b/w Unsqueeze to ExpandDims https://github.com/onnx/onnx/issues/1481

Look for the answer:

I found that the Caffe2 equivalence for Unsqueeze in ONNX is ExpandDims, and there is a special mapping in onnx_caffe2/backend.py around line 121 for those operators that are different only in their names and attribute names, but somehow Unsqueeze isn't presented there (have no idea why). So I manually added the mapping rules for it in the _renamed_operators and _per_op_renamed_attrs dicts and the code would look like:

_renamed_operators = {
    'Caffe2ConvTranspose':   'ConvTranspose',
    'GlobalMaxPool':         'MaxPool',
    'GlobalAveragePool':     'AveragePool',
    'Pad':                   'PadImage',
    'Neg':                   'Negative',
    'BatchNormalization':    'SpatialBN',
    'InstanceNormalization': 'InstanceNorm',
    'MatMul':                'BatchMatMul',
    'Upsample':              'ResizeNearest',
    'Equal':                 'EQ',
    'Unsqueeze':             'ExpandDims',  # add this line
}

_global_renamed_attrs = {'kernel_shape': 'kernels'}
_per_op_renamed_attrs = {
    'Squeeze':              {'axes': 'dims'},
    'Transpose':            {'perm': 'axes'},
    'Upsample':             {'mode': ''},
    'Unsqueeze':            {'axes': 'dims'},  # add this line
}

And everything works as expected.

I am not the OP, thanks to OP though.