When an imported module tries to use numpy, I get an error from pytorch

Asked Mar 11 '23 at 21:44

Active Mar 11 '23 at 21:44

Viewed 56 times

I am trying to graph the loss of my neural network with pyplot. When I try to pass in the arguments to the function like so:

dataframe = dict (
    Epoch = epochsArray,
    Loss = lossesArray
)
figure = px.line(dataframe, x='Epoch', y='Loss', title='Loss Per Epoch') #This line is where the error is

I get an error from PyTorch: RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead. How can I call tensor.detach().numpy() inside a module that was imported?

asked Mar 11 '23 at 21:44

Crossbow4000

what version of numpy and pytorch you have ? – zaki98 Mar 11 '23 at 21:56
numpy version: 1.23.3 pytorch version: 1.13.1 I updated both of them a few days ago – Crossbow4000 Mar 11 '23 at 21:59
1

would using torch.no_grad() work ? https://stackoverflow.com/a/56023903/9908000 – zaki98 Mar 11 '23 at 22:13
I get a different error now `RuntimeError: Only Tensors created explicitly by the user (graph leaves) support the deepcopy protocol at the moment` – Crossbow4000 Mar 11 '23 at 22:29
1

this is why u should store loss using `loss.item()`. You can even use `torch.gather` with `no_grad()`. However `no_grad()` should work without having to call `deepcopy`. why is `deepcopy` being called? – Nelson aka SpOOKY Mar 12 '23 at 01:21
Using loss.item() and no_grad() fixed my issues, thank you – Crossbow4000 Mar 19 '23 at 16:58

When an imported module tries to use numpy, I get an error from pytorch

0 Answers0