Training with a single GPU

[cinjon]

Heads up that validate doesn't do anything in train_recognizer.py if it's not distributed training. The keyword arg isn't used in the function and trying to use it in the config will fail because there aren't enough data loaders.

[cinjon]

Actually, this is really a problem with workflow. When it is has more than one array item like in https://github.com/open-mmlab/mmaction/blob/master/configs/TSN/tsn_kinetics400_2d_rgb_r50_seg3_f1s1.py#L110, it fails because we aren't building the dataloaders for the later ones.

I see that there is a commented out one below that with just one dataloader (for 'train'). Is this meant to be the approach?