Open yellofi opened 3 months ago
Hello, I have evaluated our method using Tensor-RT (FP16), and the mIoU values are consistent with those measured in PyTorch, without any degradation. Unfortunately, I don't have the specific numerical values at hand right now. However, I can confirm that on Cityscapes, the accuracy of TRT-FP16 aligns with the reported accuracy in the paper for Torch-FP32. Additionally, the accuracy of TRT-INT8 is approximately 0.7 to 1.0 mIOU lower than Torch-32.
First of all, I appreciate your amazing work.
The model inference time via Tensor-RT is much faster than the torch in your research paper. I wonder if you evaluated models via both torch and Tensor-RT (FP16), or only torch.
Could the degradation occur on Tensor-RT (FP16)?