Convert HRNet mmpose model to ONNX

[leonardts]

Hello and thank you very much for this awesome repo!

I tried to export an HRNet model to ONNX using pytorch2onnx.py. I used the config file "configs/top_down/hrnet/coco/hrnet_w32_coco_256x192.py" and the checkpoint "hrnet_w32_coco_256x192-b9e0b3ab_20200708.pth" and I used the default values for the other parameters. Unfortunately I got the two following issues:

• mmpose - WARNING - The model and loaded state dict do not match exactly

• Traceback (most recent call last):

  File "tools/pytorch2onnx.py", line 152, in <module>
    verify=args.verify)
  File "tools/pytorch2onnx.py", line 74, in pytorch2onnx
    opset_version=opset_version)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/__init__.py", line 230, in export
    custom_opsets, enable_onnx_checker, use_external_data_format)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/utils.py", line 91, in export
    use_external_data_format=use_external_data_format)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/utils.py", line 639, in _export
    dynamic_axes=dynamic_axes)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/utils.py", line 421, in _model_to_graph
    dynamic_axes=dynamic_axes, input_names=input_names)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/utils.py", line 203, in _optimize_graph
    graph = torch._C._jit_pass_onnx(graph, operator_export_type)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/__init__.py", line 263, in _run_symbolic_function
    return utils._run_symbolic_function(*args, **kwargs)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/torch/onnx/utils.py", line 934, in _run_symbolic_function
    return symbolic_fn(g, *inputs, **attrs)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/onnx/symbolic.py", line 17, in symbolic_fn
    g, interpolate_mode, args)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/onnx/onnx_utils/symbolic_helper.py", line 254, in _get_interpolate_attributes
    scales = _interpolate_get_scales_if_available(g, scales)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/onnx/onnx_utils/symbolic_helper.py", line 229, in _interpolate_get_scales_if_available
    available_scales = _maybe_get_const(scales[0], 'f') != -1 and not _is_none(
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/onnx/onnx_utils/symbolic_helper.py", line 64, in _maybe_get_const
    return _parse_arg(value, desc)
  File "/home/leonard/anaconda3/envs/open-mmlab/lib/python3.7/site-packages/mmcv/onnx/onnx_utils/symbolic_helper.py", line 33, in _parse_arg
    return float(tval)
ValueError: only one element tensors can be converted to Python scalars

Please note that the second issue is only related to HRNet. When I tried to export the simplebaseline model I got only the warning "mmpose - WARNING - The model and loaded state dict do not match exactly".
I have pytorch 1.7.1 and cuda 9.2.

[astahlman]

I'm seeing the same thing on master.

I found a similar issue in pytorch/pytorch#22674 in which someone reported the same error when exporting a mobilenet_v2 model to onnx, but that issue was closed as of pytorch v1.2

[RichardTurner-dev]

Hello there! I'm having the exact same issue as @leonardts. I have pytorch 1.7.1 and cuda 10.1.

[innerlee]

@RunningLeon could you please offer some help?

[RunningLeon]

@leonardts Please try with PyTorch==1.6.0. This tool is not fully tested on the other version of PyTorch.

[leonardts]

Hi!
Thank you for you answer! With pytorch 1.6.0 I have the following result:

2021-02-02 13:56:31,899 - mmpose - WARNING - The model and loaded state dict do not match exactly

unexpected key in source state_dict: head.0.0.0.conv1.weight, head.0.0.0.bn1.weight, head.0.0.0.bn1.bias, head.0.0.0.bn1.running_mean, head.0.0.0.bn1.running_var, head.0.0.0.bn1.num_batches_tracked, head.0.0.0.conv2.weight, head.0.0.0.bn2.weight, head.0.0.0.bn2.bias, head.0.0.0.bn2.running_mean, head.0.0.0.bn2.running_var, head.0.0.0.bn2.num_batches_tracked, head.0.0.0.conv3.weight, head.0.0.0.bn3.weight, head.0.0.0.bn3.bias, head.0.0.0.bn3.running_mean, head.0.0.0.bn3.running_var, head.0.0.0.bn3.num_batches_tracked, head.0.0.0.downsample.0.weight, head.0.0.0.downsample.1.weight, head.0.0.0.downsample.1.bias, head.0.0.0.downsample.1.running_mean, head.0.0.0.downsample.1.running_var, head.0.0.0.downsample.1.num_batches_tracked, head.0.1.0.conv1.weight, head.0.1.0.bn1.weight, head.0.1.0.bn1.bias, head.0.1.0.bn1.running_mean, head.0.1.0.bn1.running_var, head.0.1.0.bn1.num_batches_tracked, head.0.1.0.conv2.weight, head.0.1.0.bn2.weight, head.0.1.0.bn2.bias, head.0.1.0.bn2.running_mean, head.0.1.0.bn2.running_var, head.0.1.0.bn2.num_batches_tracked, head.0.1.0.conv3.weight, head.0.1.0.bn3.weight, head.0.1.0.bn3.bias, head.0.1.0.bn3.running_mean, head.0.1.0.bn3.running_var, head.0.1.0.bn3.num_batches_tracked, head.0.1.0.downsample.0.weight, head.0.1.0.downsample.1.weight, head.0.1.0.downsample.1.bias, head.0.1.0.downsample.1.running_mean, head.0.1.0.downsample.1.running_var, head.0.1.0.downsample.1.num_batches_tracked, head.0.2.0.conv1.weight, head.0.2.0.bn1.weight, head.0.2.0.bn1.bias, head.0.2.0.bn1.running_mean, head.0.2.0.bn1.running_var, head.0.2.0.bn1.num_batches_tracked, head.0.2.0.conv2.weight, head.0.2.0.bn2.weight, head.0.2.0.bn2.bias, head.0.2.0.bn2.running_mean, head.0.2.0.bn2.running_var, head.0.2.0.bn2.num_batches_tracked, head.0.2.0.conv3.weight, head.0.2.0.bn3.weight, head.0.2.0.bn3.bias, head.0.2.0.bn3.running_mean, head.0.2.0.bn3.running_var, head.0.2.0.bn3.num_batches_tracked, head.0.2.0.downsample.0.weight, head.0.2.0.downsample.1.weight, head.0.2.0.downsample.1.bias, head.0.2.0.downsample.1.running_mean, head.0.2.0.downsample.1.running_var, head.0.2.0.downsample.1.num_batches_tracked, head.1.0.0.conv1.weight, head.1.0.0.bn1.weight, head.1.0.0.bn1.bias, head.1.0.0.bn1.running_mean, head.1.0.0.bn1.running_var, head.1.0.0.bn1.num_batches_tracked, head.1.0.0.conv2.weight, head.1.0.0.bn2.weight, head.1.0.0.bn2.bias, head.1.0.0.bn2.running_mean, head.1.0.0.bn2.running_var, head.1.0.0.bn2.num_batches_tracked, head.1.0.0.conv3.weight, head.1.0.0.bn3.weight, head.1.0.0.bn3.bias, head.1.0.0.bn3.running_mean, head.1.0.0.bn3.running_var, head.1.0.0.bn3.num_batches_tracked, head.1.0.0.downsample.0.weight, head.1.0.0.downsample.1.weight, head.1.0.0.downsample.1.bias, head.1.0.0.downsample.1.running_mean, head.1.0.0.downsample.1.running_var, head.1.0.0.downsample.1.num_batches_tracked, head.1.1.0.conv1.weight, head.1.1.0.bn1.weight, head.1.1.0.bn1.bias, head.1.1.0.bn1.running_mean, head.1.1.0.bn1.running_var, head.1.1.0.bn1.num_batches_tracked, head.1.1.0.conv2.weight, head.1.1.0.bn2.weight, head.1.1.0.bn2.bias, head.1.1.0.bn2.running_mean, head.1.1.0.bn2.running_var, head.1.1.0.bn2.num_batches_tracked, head.1.1.0.conv3.weight, head.1.1.0.bn3.weight, head.1.1.0.bn3.bias, head.1.1.0.bn3.running_mean, head.1.1.0.bn3.running_var, head.1.1.0.bn3.num_batches_tracked, head.1.1.0.downsample.0.weight, head.1.1.0.downsample.1.weight, head.1.1.0.downsample.1.bias, head.1.1.0.downsample.1.running_mean, head.1.1.0.downsample.1.running_var, head.1.1.0.downsample.1.num_batches_tracked, head.2.0.0.conv1.weight, head.2.0.0.bn1.weight, head.2.0.0.bn1.bias, head.2.0.0.bn1.running_mean, head.2.0.0.bn1.running_var, head.2.0.0.bn1.num_batches_tracked, head.2.0.0.conv2.weight, head.2.0.0.bn2.weight, head.2.0.0.bn2.bias, head.2.0.0.bn2.running_mean, head.2.0.0.bn2.running_var, head.2.0.0.bn2.num_batches_tracked, head.2.0.0.conv3.weight, head.2.0.0.bn3.weight, head.2.0.0.bn3.bias, head.2.0.0.bn3.running_mean, head.2.0.0.bn3.running_var, head.2.0.0.bn3.num_batches_tracked, head.2.0.0.downsample.0.weight, head.2.0.0.downsample.1.weight, head.2.0.0.downsample.1.bias, head.2.0.0.downsample.1.running_mean, head.2.0.0.downsample.1.running_var, head.2.0.0.downsample.1.num_batches_tracked, head.3.0.0.conv1.weight, head.3.0.0.bn1.weight, head.3.0.0.bn1.bias, head.3.0.0.bn1.running_mean, head.3.0.0.bn1.running_var, head.3.0.0.bn1.num_batches_tracked, head.3.0.0.conv2.weight, head.3.0.0.bn2.weight, head.3.0.0.bn2.bias, head.3.0.0.bn2.running_mean, head.3.0.0.bn2.running_var, head.3.0.0.bn2.num_batches_tracked, head.3.0.0.conv3.weight, head.3.0.0.bn3.weight, head.3.0.0.bn3.bias, head.3.0.0.bn3.running_mean, head.3.0.0.bn3.running_var, head.3.0.0.bn3.num_batches_tracked, head.3.0.0.downsample.0.weight, head.3.0.0.downsample.1.weight, head.3.0.0.downsample.1.bias, head.3.0.0.downsample.1.running_mean, head.3.0.0.downsample.1.running_var, head.3.0.0.downsample.1.num_batches_tracked, fc.weight, fc.bias, stage4.2.fuse_layers.1.0.0.0.weight, stage4.2.fuse_layers.1.0.0.1.weight, stage4.2.fuse_layers.1.0.0.1.bias, stage4.2.fuse_layers.1.0.0.1.running_mean, stage4.2.fuse_layers.1.0.0.1.running_var, stage4.2.fuse_layers.1.0.0.1.num_batches_tracked, stage4.2.fuse_layers.1.2.0.weight, stage4.2.fuse_layers.1.2.1.weight, stage4.2.fuse_layers.1.2.1.bias, stage4.2.fuse_layers.1.2.1.running_mean, stage4.2.fuse_layers.1.2.1.running_var, stage4.2.fuse_layers.1.2.1.num_batches_tracked, stage4.2.fuse_layers.1.3.0.weight, stage4.2.fuse_layers.1.3.1.weight, stage4.2.fuse_layers.1.3.1.bias, stage4.2.fuse_layers.1.3.1.running_mean, stage4.2.fuse_layers.1.3.1.running_var, stage4.2.fuse_layers.1.3.1.num_batches_tracked, stage4.2.fuse_layers.2.0.0.0.weight, stage4.2.fuse_layers.2.0.0.1.weight, stage4.2.fuse_layers.2.0.0.1.bias, stage4.2.fuse_layers.2.0.0.1.running_mean, stage4.2.fuse_layers.2.0.0.1.running_var, stage4.2.fuse_layers.2.0.0.1.num_batches_tracked, stage4.2.fuse_layers.2.0.1.0.weight, stage4.2.fuse_layers.2.0.1.1.weight, stage4.2.fuse_layers.2.0.1.1.bias, stage4.2.fuse_layers.2.0.1.1.running_mean, stage4.2.fuse_layers.2.0.1.1.running_var, stage4.2.fuse_layers.2.0.1.1.num_batches_tracked, stage4.2.fuse_layers.2.1.0.0.weight, stage4.2.fuse_layers.2.1.0.1.weight, stage4.2.fuse_layers.2.1.0.1.bias, stage4.2.fuse_layers.2.1.0.1.running_mean, stage4.2.fuse_layers.2.1.0.1.running_var, stage4.2.fuse_layers.2.1.0.1.num_batches_tracked, stage4.2.fuse_layers.2.3.0.weight, stage4.2.fuse_layers.2.3.1.weight, stage4.2.fuse_layers.2.3.1.bias, stage4.2.fuse_layers.2.3.1.running_mean, stage4.2.fuse_layers.2.3.1.running_var, stage4.2.fuse_layers.2.3.1.num_batches_tracked, stage4.2.fuse_layers.3.0.0.0.weight, stage4.2.fuse_layers.3.0.0.1.weight, stage4.2.fuse_layers.3.0.0.1.bias, stage4.2.fuse_layers.3.0.0.1.running_mean, stage4.2.fuse_layers.3.0.0.1.running_var, stage4.2.fuse_layers.3.0.0.1.num_batches_tracked, stage4.2.fuse_layers.3.0.1.0.weight, stage4.2.fuse_layers.3.0.1.1.weight, stage4.2.fuse_layers.3.0.1.1.bias, stage4.2.fuse_layers.3.0.1.1.running_mean, stage4.2.fuse_layers.3.0.1.1.running_var, stage4.2.fuse_layers.3.0.1.1.num_batches_tracked, stage4.2.fuse_layers.3.0.2.0.weight, stage4.2.fuse_layers.3.0.2.1.weight, stage4.2.fuse_layers.3.0.2.1.bias, stage4.2.fuse_layers.3.0.2.1.running_mean, stage4.2.fuse_layers.3.0.2.1.running_var, stage4.2.fuse_layers.3.0.2.1.num_batches_tracked, stage4.2.fuse_layers.3.1.0.0.weight, stage4.2.fuse_layers.3.1.0.1.weight, stage4.2.fuse_layers.3.1.0.1.bias, stage4.2.fuse_layers.3.1.0.1.running_mean, stage4.2.fuse_layers.3.1.0.1.running_var, stage4.2.fuse_layers.3.1.0.1.num_batches_tracked, stage4.2.fuse_layers.3.1.1.0.weight, stage4.2.fuse_layers.3.1.1.1.weight, stage4.2.fuse_layers.3.1.1.1.bias, stage4.2.fuse_layers.3.1.1.1.running_mean, stage4.2.fuse_layers.3.1.1.1.running_var, stage4.2.fuse_layers.3.1.1.1.num_batches_tracked, stage4.2.fuse_layers.3.2.0.0.weight, stage4.2.fuse_layers.3.2.0.1.weight, stage4.2.fuse_layers.3.2.0.1.bias, stage4.2.fuse_layers.3.2.0.1.running_mean, stage4.2.fuse_layers.3.2.0.1.running_var, stage4.2.fuse_layers.3.2.0.1.num_batches_tracked

Successfully exported ONNX model: tmp.onnx

@RunningLeon this "unexpected key" message is a little bit scary. What do you think about it?

[innerlee]

@jin-s13 we can add a version check for pytorch in the onnx script

[jin-s13]

The unexpected key messages are from the imagenet pre-trained models. That‘s fine.

[leonardts]

@jin-s13 thank you :)
Btw if I run the code with --verify I get the message "The numerical values are same between Pytorch and ONNX" so I think everything is okay.
Thank you very much guys!

[innerlee]

Have you guys starred the repo? 😃

[lamhust2008]

@jin-s13 thank you :) Btw if I run the code with --verify I get the message "The numerical values are same between Pytorch and ONNX" so I think everything is okay. Thank you very much guys!

Hey Leonardts, can you tell me how to convert your hrnet model to onnx from scratch. Thanks!