what does num_frame_per_clib mean in train.py and test.py

[LindseyScharpman]

I noticed that the values of num_frame_per_clib parameters in training and prediction are different. The example is RGB flow: why num_frame_per_clib is 16 when predicted, and when training is 64. So can num_frame_per_clib be set at 64 when predict?

Another problem is that num_frame_per_clib value is 64 during training, but if the number of RGB frames extracted from a video is more than 64, for example, the number of RGB frames of the video is more than 128 during training, does that mean that the num_frame_per_clib value can be set to 128?

[huangchaoxing]

@LossNAN I am also confused about the difference of 16 and 64. It seems for 3D CNN，if we train the model at 64 frames length then the test length should also be 64. Or does I3D allows varied-length (I really doubt)?

@wannibar @LossNAN I suppose the num_frame_per_clib value should be fixed at 64. But we might sample different chunk everytime when we come across this video during training. Is it controlled by the s_index for sampling ?

[huangchaoxing]

@HyunWoo-Nick I think the single test length should be 64 or 16, but you do multiple 64 or 16 to cover all the frames and do the average during testing. You might not be able to directly feed the whole video in one time into the network, because it will cause tensor size mismatch with the final fully connected layers.