Closed zhuimenghx closed 3 years ago
Thanks for your interest in the paper. 1, You can use any number of GPUs for training models on Kinetics. Remember to adjust your batch size and learning rate accordingly. 2, We use 32 frames which give better results on UCF101 and HMDB51.
Thanks for your interest in the paper. 1, You can use any number of GPUs for training models on Kinetics. Remember to adjust your batch size and learning rate accordingly. 2, We use 32 frames which give better results on UCF101 and HMDB51.
In addition, how many clips are used for average prediction for ucf101 and hmdb51?
Hello, I would like to ask some related issues about R(2+1)d paper ①Did you use 64 GPUs when training K400? What is the specific number and model of GPU? ② Is the clip_length of the pre-training model used by UCF101 and HMDB51 8 frames or 16 frames?