Performance tips for writing PyTorch code

Where can I find advice for writing performant PyTorch code? Are there things to be careful about like going back and forth between PyTorch and Numpy or always trying to use batching? What do you reckon are main aspects to keep in mind?


If you are using the GPU, then going back and forth between pytorch and numpy will be a performance bottleneck.

I dont think we’ve really established performance tips as such.