Which activation function should I use in the last layer output while using Resnet50?

vision
1

Hi, I hope you are great.
I have been creating my own cnn model with pytorch. But I am not sure which activation function should I use in resnet50 for the last layer (output) and what is the difference between them. Do you have any idea about that?

2

What is your task? Generally for image classification its softmax.

1 Like
3

Keep in mind that commonly used loss functions for classification tasks require logits, i.e. no final activation function, if nn.CrossEntropyLoss is used, or log-probabilities, i.e. F.log_softmax as the final activation function, if nn.NLLLoss is used.

2 Likes
4

I got glass images. I want to predict which glass image is good (I mean quality) which image is not good (1 or 0).

5

Your problem statement sounds like its a binary classification. Usually Sigmoid is preferred.
While inference any value greater than equal to 0.5 is considered as 1 and lesser than 0.5 considered as 0.

For loss you can use Binary Cross Entropy
loss = F.binary_cross_entropy(torch.sigmoid(input), target)

1 Like