CODEWITHSUNDEEP
pythonnumpypytorchgradienttensor
tstseby
tstseby

Reputation: 1319

PyTorch: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead

Calling tensor.numpy() gives the error:

RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

tensor.cpu().detach().numpy() gives the same error.

Upvotes: 57

Views: 128703

Answers (5)

Deepanshu Mehta
Deepanshu Mehta

Reputation: 1780

Best solution is to use torch.no_grad(): the context manager which disables the tracking of the gradient locally.

Just write your code inside this contact manager like:

with torch.no_grad():
   graph_x = some_list_of_numbers
   graph_y = some_list_of_tensors

   plt.plot(graph_x, graph_y)
   plt.show()

Upvotes: 4

Greg Davey
Greg Davey

Reputation: 91

I just ran into this problem when running through epochs, I recorded the loss to a list

final_losses.append(loss)

Once I ran through all the epochs I wanted to graph the output

plt.plot(range(epochs), final_loss)
plt.ylabel('RMSE Loss')
plt.xlabel('Epoch');

I was running this on my Mac, with no problem, but, I needed to run this on a windows PC and it produced the error noted above. So, I checked the type of each variable.

Type(range(epochs)), type(final_losses)

range, list

Seems like it should be OK.

It took a little bit of fidgeting to realize that the final_losses list was a list of tensors. I then converted them to an actual list with a new list variable fi_los.

fi_los = [fl.item() for fl in final_losses ]
plt.plot(range(epochs), fi_los)
plt.ylabel('RMSE Loss')
plt.xlabel('Epoch');

Success!

Upvotes: 7

Rickantonais
Rickantonais

Reputation: 496

I had the same error message but it was for drawing a scatter plot on matplotlib.

There is 2 steps I could get out of this error message :

  1. import the fastai.basics library with : from fastai.basics import *

  2. If you only use the torch library, remember to take off the requires_grad with :

    with torch.no_grad():
    (your code)

Upvotes: 31

Nannigalaxy
Nannigalaxy

Reputation: 627

For existing tensor

from torch.autograd import Variable

type(y)  # <class 'torch.Tensor'>

y = Variable(y, requires_grad=True)
y = y.detach().numpy()

type(y)  #<class 'numpy.ndarray'>

Upvotes: 3

Blupon
Blupon

Reputation: 1039

 Error reproduced

import torch

tensor1 = torch.tensor([1.0,2.0],requires_grad=True)

print(tensor1)
print(type(tensor1))

tensor1 = tensor1.numpy()

print(tensor1)
print(type(tensor1))

which leads to the exact same error for the line tensor1 = tensor1.numpy():

tensor([1., 2.], requires_grad=True)
<class 'torch.Tensor'>
Traceback (most recent call last):
  File "/home/badScript.py", line 8, in <module>
    tensor1 = tensor1.numpy()
RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

Process finished with exit code 1

Generic solution

this was suggested to you in your error message, just replace var with your variable name

import torch

tensor1 = torch.tensor([1.0,2.0],requires_grad=True)

print(tensor1)
print(type(tensor1))

tensor1 = tensor1.detach().numpy()

print(tensor1)
print(type(tensor1))

which returns as expected

tensor([1., 2.], requires_grad=True)
<class 'torch.Tensor'>
[1. 2.]
<class 'numpy.ndarray'>

Process finished with exit code 0

Some explanation

You need to convert your tensor to another tensor that isn't requiring a gradient in addition to its actual value definition. This other tensor can be converted to a numpy array. Cf. this discuss.pytorch post. (I think, more precisely, that one needs to do that in order to get the actual tensor out of its pytorch Variable wrapper, cf. this other discuss.pytorch post).

Upvotes: 55

Related Questions