- trainable_mask.train_color_filter=True
- trainable_mask.train_mask_vals=True
- files.downsample=2
- training.batch_size=4