Dear author, thanks for this amazing work. Can you please help me with this problem.
I am trying to export a custom model that uses BEVFormer as backbone to ONNX. My problem is that when I export to ONNX. I got the following error:
Traceback (most recent call last):
File "tools/convert_lanesegnet_to_onnx.py", line 183, in <module>
main()
File "tools/convert_lanesegnet_to_onnx.py", line 148, in main
torch.onnx.export(
File "/opt/conda/lib/python3.8/site-packages/torch/onnx/__init__.py", line 316, in export
return utils.export(model, args, f, export_params, verbose, training,
File "/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py", line 107, in export
_export(model, args, f, export_params, verbose, training, input_names, output_names,
File "/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py", line 737, in _export
proto, export_map, val_use_external_data_format = graph._export_onnx(
RuntimeError: ONNX export failed: Couldn't export Python operator MultiScaleDeformableAttnFunction_fp32
Defined at:
/workspace/source/Mapless/plugin/LaneSegNet/bevformer/modules/temporal_self_attention.py(396): forward_trt
/workspace/source/Mapless/plugin/LaneSegNet/bevformer/modules/encoder.py(614): forward_trt
/workspace/source/Mapless/plugin/LaneSegNet/bevformer/modules/encoder.py(263): forward_trt
/workspace/source/Mapless/plugin/LaneSegNet/lanesegnet/models/modules/bevformer_constructer.py(319): forward_trt
/workspace/source/Mapless/plugin/LaneSegNet/lanesegnet/models/detectors/lanesegnet.py(298): forward_trt
/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py(1090): _slow_forward
/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py(1102): _call_impl
/opt/conda/lib/python3.8/site-packages/torch/jit/_trace.py(118): wrapper
/opt/conda/lib/python3.8/site-packages/torch/jit/_trace.py(127): forward
/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py(1102): _call_impl
/opt/conda/lib/python3.8/site-packages/torch/jit/_trace.py(1166): _get_trace_graph
/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py(388): _trace_and_get_graph_from_model
/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py(437): _create_jit_graph
/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py(493): _model_to_graph
/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py(724): _export
/opt/conda/lib/python3.8/site-packages/torch/onnx/utils.py(107): export
/opt/conda/lib/python3.8/site-packages/torch/onnx/__init__.py(316): export
tools/convert_lanesegnet_to_onnx.py(148): main
tools/convert_lanesegnet_to_onnx.py(183): <module>
I understand that I need to write corresponding ONNX operator for MultiScaleDeformableAttnFunction_fp32.
I have read your code and see that class MultiScaleDeformableAttnFunction_fp32 is the same to that of pytorch model. I do not understand how did you implement this class so that it can be sucessfully exported to ONNX. Can you please explain to me.
mmcv 1.5.2
torch 1.9.1
CUDA 11.1
onnx
onnxruntime
The error happens right at these lines:
As far as I understand, the folder TensorRT/plugin contains custom operator you have implemented. But I dont understand how I can implement my custom ONNX operator in a different code base. I can see some relevant files relating to exporting to onnx:
but do not understand where the symbolic function is called since the above temporal self attn uses MultiScaleDeformableAttnFunction_fp32
Thank you. Any help is appreciated.
Dear author, thanks for this amazing work. Can you please help me with this problem.
I am trying to export a custom model that uses BEVFormer as backbone to ONNX. My problem is that when I export to ONNX. I got the following error:
I understand that I need to write corresponding ONNX operator for MultiScaleDeformableAttnFunction_fp32.
I have read your code and see that class MultiScaleDeformableAttnFunction_fp32 is the same to that of pytorch model. I do not understand how did you implement this class so that it can be sucessfully exported to ONNX. Can you please explain to me.
The error happens right at these lines:
As far as I understand, the folder TensorRT/plugin contains custom operator you have implemented. But I dont understand how I can implement my custom ONNX operator in a different code base. I can see some relevant files relating to exporting to onnx:
but do not understand where the symbolic function is called since the above temporal self attn uses MultiScaleDeformableAttnFunction_fp32
Thank you. Any help is appreciated.