Added data augmentation methods and staged learning rate

[KaiyangZhou]

Updates:

1. Random erasing data augmentation can be enabled by adding argument --augdata-re.
2. You can use staged learning rate, i.e. different learning rate for different layers. This can be managed by three arguments: (a) --staged-lr: called to use staged learning rate; (b) --new-layers: list of layer names (strings) indicating which layers use the default learning rate and the rest use a scaled learning rate; (c) --base-lr-mult: learning rate multiplier for base layers. For example, when you train resnet50, if you want the randomly initialized self.classifier to have --lr 0.1 and the rest with learning rate scaled by 0.1, you can add --staged-lr --new-layers classifier --base-lr-mult 0.1 to the argument list. See here for more details.

[KaiyangZhou]

In addition, when you do --load-weights path_to_pth, load_pretrained_weights() can handle keys with module., i.e. weights previously saved with nn.DataParallel. Check the code for more details.

[KaiyangZhou]

p.s. feel free to suggest any features that would be useful to the pipeline.

[KaiyangZhou]

Flag to call random erasing is changed to --random-erase.

[KaiyangZhou]

New features in transforms.py:

--color-jiter: randomly change the brightness, contrast and saturation

--color-aug: randomly alter the intensities of RGB channels