training set Size and Duration for video on UCF-101

[flying0712]

Hi, when training on the UCF-101 trainlist01 with 9000 videos each taking 17 frames, it takes 1 hour per epoch and 11 days for 270 epochs on a single machine with 8x V100 GPUs(batch size=1, larger causes GPU memory overflow), which is quite time-consuming. So, could you please share how you set up the the training set size and how long the training duration is( I noticed the paper used 32x GPUs with a batch size of 256 ), and whether you used all video frames or just the first 17? Thank you!

[RobertLuo1]

Hi, thanks for your attention on our work. As indicated in the ucf101_lfqgan_128_L.yaml, we use 64x NPU for training and the global batch size is 128. The total epoch is 2000, which is adopted from the original magvit-v1 paper. The training duration is about 3 days. During training, the 17 video frames are randomly sampled from a given video.