Deep-All results on PACS about randconv

[wildphoton]

@wangtingwei1993 Thanks for your interest! I think the major source difference is that we are using the official data splits for train/val/test while JiGen made their own splits. JiGen also uses random cropping/flip augmentation while ours does not. I think we are all using the PyTorch pretrain weights. Optimizer, LR decay and batch size are probably not the exactly same. But I won't expect them to make a huge difference. I believe the data splits (maybe + extra augmentations) made the JiGen results not the same. I will update our PACS scripts soon.