I train a model with log_softmax activation in the last layer,
then while evaluating the model, I should only print the model(x) to get the real probs, right?
so if it’s that why Keras only takes softmax instead of log softmax and it totally works?
log_softmax instead of first applying
softmax and later
log for numerical stability as described in the LogSumExp trick.
If you want to print the probabilities, you could just use
torch.exp on the output.
thank you for the reply
so it means that if i use log_softmax, then i should call exp on the output - exp(out)
otherwise the argmax will return incorrect output for classifying the data ?
If you just want the argmax you can keep the log_softmax and use argmax after that, but if you want the correct softmax probabilities you should use exp(out)
thank you for your answer