More information regarding this work can be found in this link.

ResNet50SimCLR.pt contains a pretrained model that can be used directly for inference as shown in UC4aDemonstration.ipynb.

More encoders, pretrained in a self-supervised learning fashion, can be found here.

Example for loading the pre-trained encoders:

backbone = torchvision.models.resnet50(pretrained=False)
backbone.fc = nn.Sequential(nn.Linear(2048, 2048), nn.ReLU(), backbone.fc)
backbone = torch.nn.parallel.DataParallel(backbone,device_ids=[0,1])
backbone.load_state_dict(torch.load('ResNet50_Simclr_500_Epochs.pt'))
backbone.module.fc = nn.Identity()
backbone = backbone.module

main.py handles both the self-supervised learning pretraining as well as the supervised training of the linear classifier.

Examples of the data and the class activation mappings can be found in UC4aDemonstration.ipynb.