01_pytorch_workflow_exercises question

[sharkweekshane]

Hi folks,
Been loving the course so far (Cue fire animation)

I'm working through the second set of exercises (01) and I keep getting an error when I try to plot the predictions

Plot the predictions (these may need to be on a specific device)

plot_predictions(predictions = y_preds.cpu())

I get error: RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

I see a similar question was asked about this, but I can't wrap my mind around it. Can someone help?

Edit: I also tried plot_predictions(predictions = y_preds.cpu().detach().numpy()) as suggested by the error, and I got another error that I also don't understand

[DavS123]

Hi,

y_preds.cpu().detach().numpy() should we replaces by y_preds.detach().numpy().cpu()

should do the trick.

best wishes

[sharkweekshane]

Hi Dav,
Thanks for the reply. I tried doing it your way and ended up with another error (sigh)

input: plot_predictions(predictions = y_preds.detach().numpy().cpu())
output: AttributeError: 'numpy.ndarray' object has no attribute 'cpu'

I'm still a novice coder, so I'm not sure how to tackle this error. My inclination tells me that I should try converting the tensor to an array before I try ploting.

[DavS123]

Hi
here is a good reference.

https://www.tutorialspoint.com/how-to-convert-a-pytorch-tensor-with-gradient-to-a-numpy-array

Apprently y_preds.detach().numpy() does the trick.

best wishes

[mrdbourke]

Hi @sharkweekshane,

I've faced that error plenty of times too. And it can be a bit hard to troubleshoot.

I just went through the solutions notebook and it seems to work fine end-to-end: https://github.com/mrdbourke/pytorch-deep-learning/blob/main/extras/solutions/01_pytorch_workflow_exercise_solutions.ipynb

I'd say a potential issue you may be facing:

• Data already on cpu() (this happens if you haven't set the device to GPU)/sent to GPU

You can check what device your PyTorch tensor is on with tensor.device:

y_preds.device

Out:

device(type='cuda', index=0)

y_preds.cpu().device

Out:

device(type='cpu')

However, tensor.device will error if your tensor is already NumPy.

As for this line:

tensor.cpu().detach().numpy()

This is saying "send tensor to the cpu(), detach() the gradient tracking (this is automatically set for learning in PyTorch) and then turn the PyTorch tesnor to numpy() datatype".

Hopefully this clears things up!