Lower accuracy when converted to fp32 trt model

[poornimajd]

Hello,Great work! I have a pytorch segmentation model,which has been trained on a custom dataset.When I convert this model to the tensorrt version,I get a decrease in miou score from 89 to 66 ,though it is just fp32 version.I wonder why does the accuracy decrease so much.I used the following dependencies Pytorch- 1.7.1 Cuda- 10.2 TensorRT-7.1.3 I also tried passing the input as x= x.contiguous() , as referred in the previous issues. My model has a bilinear interpolation layer ,but since it is a higher version of tensorRT ,I thought that this cannot bring down the accuracy so much. What could be the problem here? Any suggestion is greatly appreciated. Thank you

[purse1996]

Do you solve the problem? I meet a similar problem, the miou decreases from 82 to 78. Thank you very much.

[ryandok]

Do you solve the problem? I meet a similar problem too. Thank you very much.

[poornimajd]

no not yet.

[YOONAHLEE]

I also had the same issue. Converting a torch model with the same precision (fp32) changed the model output.

[Shubham-Omprakash-Patil]

Having the same issue. Original PyTorch model gave excellent results. However, PyTorch --> TRT conversion led to decrease in score and increase in errors.