chenwydj
commented
4 years ago Hi @kim2002037 ! Thank you for your interest!
This is because TensorRT (v5.1.5) does not support bilinear interpolation. We cannot get the latency of the model uses bilinear interpolation.
Meanwhile, in PyTorch we have the following FPS measurements:
- NN: 88.9
- Bilinear: 83.7
The FPS gap is only 5.2, which is very small considering the FPS gap (on TensorRT) in our Table 4 (our 163.9 v.s. 136.9 of the 2nd runner-up).
kim2002037
Thank you for your great job. I have one question about latency issue.
When I checked your code, I found that you use bilinear interpolation in train/operations.py
but, In the test part you used nearest interpolation in latency/operations.py.
Can you tell me why? (I'm asking you this question because nearest interpolation is faster than bilinear interpolation in pytorch)
you use as below in train/operations.py out = F.interpolate(x, size=(int(x.size(2))//2, int(x.size(3))//2), mode='bilinear', align_corners=True) out = F.interpolate(out, size=(int(x.size(2)), int(x.size(3))), mode='bilinear', align_corners=True)
you use as below in latency/operations.py out = F.interpolate(x, size=(int(x.size(2))//2, int(x.size(3))//2), mode='nearest') out = F.interpolate(out, size=(int(x.size(2)), int(x.size(3))), mode='nearest')