How to convert a list of cuda tensors to a list of cpu tensors?

[Suraj]

So I have a big list of cuda tensors containing loss values of each iteration but I can’t even scatter plot it. How can I convert it to a list of cpu tensors or even numpy array without iterating over the indices? Thank You!

[Nikronic (Nikan Doosti)]

Hi,

In this post I have written a code to transfer a list of CPU tensors to GPU which all have the same shape (in your case all have shape of (1, )).

Input a list of tensors to a model without the need to manually transfer each item to cuda

Hi, I’d like to input a list of tensors to my model. However, pytorch doesn’t automatically transfer each item in this list to cuda and I need to manually do that each time by .cuda() . Is there a better more elegant way of dealing with this?

Bests

[Suraj]

Thank you so much @Nikronic .

Just creating a new tensor with torch.tensor() worked. Then simply plotted the scatter plot on torch tensor (with device = cpu).

new_tensor = torch.tensor(list_of_cuda_tensors, device = 'cpu')

[chethanjjj (Chethan)]

But, what if you want to keep it as a list of tensors after the transfer from gpu to cpu

[Thilani (Thilani Delpagoda)]

Thanks! This helped me with my code which ran on cpu but running a plot in gpu.

[nlgranger (Nicolas Granger)]

In a context where performance is a concern, you’d be better off stacking the scalar tensors first then moving to cpu: torch.stack(list_of_losses).cpu(). Otherwise [t.item() for t in list_of_losses] seems more idiomatic to me if you want the result as a list of floats.