lordofgod opened a new issue #20280: URL: https://github.com/apache/incubator-mxnet/issues/20280
## Description Mxnet 模型的Deconvolution操作在未指定“adj”参数时，可能会导致转换成Onnx的ConvTranspose操作时生成错误的output_padding。 这会导致在使用TensorRT调用Onnx模型时提示“不支持3D deconvolution的非对称padding”。 Mxnet.contrib.onnx.export_mode will convert wrong output_padding value from Deconvolution (Mxnet) to ConvTranspose (Onnx), if the param "adj" is not specified. This bug will cause an asymmetrical padding error while using the converted onnx model in TensorRT. ### Error Message TensorRT error message: ERROR: builtin_op_importers.cpp:691 In function importConvTranspose: [8] Assertion failed: begPadding == endPadding && "TensorRT does not support asymmetrical padding for 3D deconvolutions!" ## To Reproduce ### Steps to reproduce 1. Save a 3D network with deconvolution in Mxnet 2. Convert the model of step1 from mxnet to onnx with python command "mxnet.contrib.onnx.export_model(...)" 3. Test the converted model with TensorRT using trtexec, such as "trtexec.exe --onnx=[converted_onnx_model]" ## What have you tried to solve it? 1. Show onnx model with Netro and modify _op_translations.py ## Environment System: win 10 IDE: vs2019 Mxnet: 1.8.0 Cuda: 11.1 Cudnn: 8.2 TensorRT: 7.2 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
