So I\'m trying to train a SimCLR network with a custom lightweight ConvNet backbone (tried it with a ResNet already) on a dataset containing first 5 letters of the alphabet