About motion_cnn.py,why data loading needs so much time ?it seems to be close to the batch time. about two-stream-action-recognition

[jeffreyyihuang]

Hi,
For the first question, the reason why the training process will temporally stuck is that the model is still processing on CPU(including I/O and resizing image to the size of model input). If you raise the number of batch size, the time of preprocessing will also increase. Therefore, it is normal that the time on CPU will increase, due to the hardware limitation.

For the second question, to make it more precise, batch time means the time consumption on GPU, and the data time means the time consumption on CPU.

If you would like to this verify this question, you can test the code in here and figure out the reason why it takes so much time on CPU.

I hope my clarification has been helpful.
Jeffrey