3 PyTorch Tips For Better Performance!

Key Takeaways

here are three simple yet very effective pie church tips for better performance let's go number one if you want to construct a tensor from another data like a numper array don't use tensor but instead use s tensor or from numpy because this will create a full copy and this avoids copying now we only have to be aware that if we change the original data then it also affects the tensor number two try to avoid gpu to cpu transfers or vice versa for example try to avoid these functions cpu item or numpy because each of these functions transfers data between the devices instead use tensor.detach this will create a new tensor that is detached from the computational graph number three create tensors directly on the gpu for example don't create a tensor like this and then call dot cuda because this creates it on the cpu and then transfers it instead create it directly on the gpu by specifying the device in the constructor like so