Add batch to training process

[TanguyJeanneau]

So far, the training is made:
per epoch (all N videos in the datasetdataset) //
per video (all frames or a subselection of M frames) //
per couple (2 subsequent frames of the video) //
process the 2 frames //
then optimize (Adam)

We should train by batch, so
per epoch (all N videos in the datasetdataset) //
per video (all frames or a subselection of M frames) //
process all frames at once in a batch //
then optimize (Adam)

OR

per epoch (all N videos in the datasetdataset) //
per video batch (all frames or a subselection of M frames, for P videos) //
process all frames at once in a batch //
then optimize (Adam)