PyTorch binary classifier with ResNet and Face dataset

The code uses the pretained weights of ResNet18, replaces the last fc layer with output size 1 or 2 for my binary classifier.

• If output size of 1 is used, sigmoid function is used on the output to give a value between 0 and 1, it’s then rounded to 0 or 1 for the class labels.
  • BCELoss() is used
• If output size of 2 is used, the index of the max value in the output tensor is used as the class label
  • CrossEntropyLoss() is used

• With learning rate = 1e-3, the loss and accuracy fluctuate and do not converge. Reducing learning rate to 1e-5 solves the problem

• Freezing all layers except the last fc layer result in very slow convergence and lower accuracy in the end

• CrossEntropyLoss() results in slightly better accuracy than BCELoss() for my dataset

• Name: Modified LFW Face Dataset
• https://medium.com/hackernoon/binary-face-classifier-using-pytorch-2d835ccb7816