Huge difference in evaluation between model.train() and model.eval() on torchvision resnet

Question

I\'m evaluating performance of my deep CNN model - running k-nearest neighbors classification using embeddings generated by model. I noticed that when I forgot to switch to