hulianyuyy/AdaBrowse

GFLOPs, Throughput

Closed this issue · 5 comments

atonyo11 commented

Hi author.
Could you provide me with the code for calculating GFLOPs and throughput?

hulianyuyy commented

I calculate the GFLOPs manually. While the MACs of the resnet18 backbone can be reffered to PyTorch-OpCounter, the FLOPs of 1D CNN, LSTM, classifier and others are mannually computed. As for throughput, you could perform inference with batch size of 1, and the process will report the whole time of processing the Dev or Test set, and thus you could calculate the throughput according to the number of videos in the Dev or Test set.

atonyo11 commented

Thanks!

atonyo11 commented

Sorry, I have one more question.
As I understand, to calculate Gflop, we base on the input. But in this work, input is different for each batch. So how to calculate Gflop?

hulianyuyy commented

We calculat the averaged FLOPs per frame, and assume each video has 100 frames to show the results. E.g., the FLOPs per frame is 3.64G, we give 364G as the FLOPs for comparison for the video.

atonyo11 commented

I got it. Thank you!