Can not train model. Tensor mismatch

[Vadim2S]

I try run training. Just test run yet. One file as train and same file as eval.

And I am get error in file train_ddp.py line 154 or so.
Error: RuntimeError: The size of tensor a (10) must match the size of tensor b (9) at non-singleton dimension 0
Code: mse = torch.mean(torch.square(inputs - video_recon), dim=(1, 2, 3))

I am investigate and see what (before rearrange):
inputs.shape = [1, 3, 10, 256, 256]
video_recon.shape = [1, 3, 9, 256, 256]

[qqingzheng]

May I check your training sh script? it seems that you set a wrong number of frames.

Cuz we employ temporal causal convolution instead of regular convolution, the number of frames must be 4x+1 (if temporal compression rate is 4).

[Vadim2S]

What exactly meaning of --num_frames and --eval_num_frames? At first glance I am thinking about something like batch_size or train\val subset length.
Thanks! approach "4*n+1" works! Num_frames parameters must be from row 13, 17, 21, 25, 29, 33, etc Any other values leads to tensor mismatch.