hitachinsk
About setting parameter batch_size=1, num_frames=5
In the second stage of the FGT network training, I found that batch_size was only set to 1 and only 5 frames per video were selected for training. Thus, the size of the input tensor is (b, t, c, h, w)=>(1, 5, c, h, w). I would like to know why batch_size is set so small.
In my experiments, I set batch size to 2, not 1. The batch size is the mini-batch per GPU, therefore, if you adopt 4 GPUs, and the batch size set to 2, the overall batch is 8 (2*4).
If you have more GPU memory, you can select more frames in a video for training, which may lead to better performance. I only selected 5 frames because of the limitation of GPU memory.
I have corrected the configuration files, thanks for pointing out this error.
Thank you for providing the precious details.
Furthermore, I find some duplicate definitions. In the second stage FGT network training, I find some duplicate definitions betwen configuration file 'train.yaml' and the 'inputs.py' file. Also, when run code in train.py, it load another configuration file 'flowCheckPoint/config.yaml', which also contains some duplicate definitions.