search

kennethwdk / CID

Code for "Contextual Instance Decoupling for Robust Multi-Person Pose Estimation", CVPR 2022 Oral
MIT License
50 stars 11 forks source link

Reasoning is slow #1

Closed james13138 closed 2 years ago

james13138 commented 2 years ago

Thank you for sharing your work. I tried, but failed to reproduce the efficiency you showed in the paper. CID in the paper is much faster than dekr, but I tested two models, and even CID is slower than dekr. The following is a screenshot of my GPU model and output reasoning time. If you have time, please help me. Thank you, and thank you again for sharing your work GPU： GPU CID： 时间输出代码截图 时间输出 DEKR： 代码截图 输出时间

james13138 commented 2 years ago

Thank you for your feedback email. Thank you for helping me answer my questions and provide help. Thank you for sharing your work

james13138 commented 2 years ago

''' Hello,

All FPS in Table 4 only measure the model feedforward time without flip test, not including the data loading time (read image and tranform), which is also adopted in most object detection framework like Mask RCNN; It is recommended to use torch.cuda.synchronize() before calling time.time() to calculate the gpu process. For example, to measure the runtime of a = model(b), you can use following code. import timeit torch.cuda.synchronize() start_time = timeit.default_timer()

/ test code, e.g., a = model(b) /

torch.cuda.synchronize() total_time = timeit.default_timer() - start_time

The FPS of DEKR is measued using the early version code, I notice that the authors have modified code recently, so if you use the newest DEKR code, the results may not be the same as the paper.

'''