I used the pytorch model, turned it into onnx, and got the test result in trt4. When I used trt5, it didn’t get the output. I output the result of tensorrt reasoning, which is completely different from trt4. The network I use is resnet34+fpn, which has multiple output layers. It seems that the arrangement of trt4 and trt5 output layers has also changed. Is there any detailed change in this aspect?
to help us debug, can you share a small repro that contains the onnx conversion code, inference code and model that demonstrates the differences you are seein between trt4 and trt5?
I am analytical conversion and inference based on the onnx model of sample.I sent out the inference results.The model is based on resnet34+fpn and will get 8 output results.
This is the output of the trt4 inference conversion.
the layer name is:Conv_47
the layer name is:Add_25
the layer name is:Reshape_2
the layer name is:Softmax_1
the layer name is:Reshape_3
the layer name is:Conv_48
the layer name is:Add_26
the layer name is:Reshape_4
the layer name is:Softmax_2
the layer name is:Reshape_5
the layer name is:Conv_49
the layer name is:Add_27
the layer name is:Reshape_6
the layer name is:Softmax_3
the layer name is:Reshape_7
Begin building engine…
End building engine…
start onnx deserializeCudaEngine …
the image proces time is:1.145483 ms
Time taken for image proces is 1.14671 ms.
the trt4 output size is:407676*
the trt4 output size is:403838*
the trt4 output size is:401919*
the trt4 output size is:401010*
the trt4 output size is:8107676*
the trt4 output size is:8103838*
the trt4 output size is:8101919*
the trt4 output size is:8101010*
This is the output of the trt５ inference conversion.
the layer name is:(Unnamed Layer* 160) [Convolution]
the layer name is:(Unnamed Layer* 161) [Scale]
the layer name is:(Unnamed Layer* 162) [Shuffle]
the layer name is:(Unnamed Layer* 163) [Shuffle]
the layer name is:(Unnamed Layer* 164) [Shuffle]
the layer name is:(Unnamed Layer* 165) [Softmax]
the layer name is:(Unnamed Layer* 166) [Shuffle]
the layer name is:(Unnamed Layer* 167) [Shuffle]
the layer name is:(Unnamed Layer* 168) [Shuffle]
the layer name is:(Unnamed Layer* 169) [Convolution]
the layer name is:(Unnamed Layer* 170) [Scale]
the layer name is:(Unnamed Layer* 171) [Convolution]
the layer name is:(Unnamed Layer* 172) [Scale]
the layer name is:(Unnamed Layer* 173) [Shuffle]
the layer name is:(Unnamed Layer* 174) [Shuffle]
the layer name is:(Unnamed Layer* 175) [Shuffle]
the layer name is:(Unnamed Layer* 176) [Softmax]
the layer name is:(Unnamed Layer* 177) [Shuffle]
the layer name is:(Unnamed Layer* 178) [Shuffle]
the layer name is:(Unnamed Layer* 179) [Shuffle]
Begin building engine…
End building engine…
start onnx deserializeCudaEngine …
the trt5 output size is:407676*
the trt5 output size is:8107676*
the trt5 output size is:403838*
the trt5 output size is:8103838*
the trt5 output size is:401919*
the trt5 output size is:8101919*
the trt5 output size is:401010*
the trt5 output size is:8101010*
From the above comparison, we can see that the output order of trt4 and trt5 has changed, and I saved the output for comparison. They are completely different.I don’t know where trt has changed.
I can post my project to git, or send it to your email, I look forward to hearing from you.
you can send me a DM with a google drive, dropbox link or git link.
I am sorry to have dragged on for so long, demo I have uploaded to the cloud disk,
https://drive.google.com/drive/folders/1pgRW0NdeRx7vuMbIkri6oZyp0x_kEPrh?usp=sharing
The focus is on the output after reasoning, which looks a bit different.
thank you for your reply