How to efficiently normalize a batch of tensor to [0, 1]

[Red-Eyed]

The performance drop in is in min() function
I read that it calls item(), and it is slow.

Speed of torch.max

Thanks for clarification! Seems like it is synchronization thing resulting inaccurate timing of .item() operation. Since torch.max() return a int number, does that mean it will call .item() in the end and that is the reason why torch.max seems “slow”?