Enrico Fini

Separate section for unsupervised continual learning

1

Hey, I'll just leave it here as an idea, it would be nice to have a section for "Unsupervised Continual Learning" or something like that. It is a hot topic...

Implement multi-crop (SwAV, DCv2, DINO, MoCo)

1

and improve SimCLR's implementation

On the use of Apex AMP and hybrid stages

6

Is there a specific reason why you used Apex AMP instead of the native AMP provided by PyTorch? Have you tried native AMP? I tried to train `poolformer_s12` and `poolformer_s24`...

gradient checkpointing with timm

1

looks very easy, we can just call this function: https://github.com/rwightman/pytorch-image-models/blob/8ff45e41f7a6aba4d5fdadee7dc3b7f2733df045/timm/models/resnet.py#L697