CODEWITHSUNDEEP
numpypytorchnumpy-ndarray
El Dude
El Dude

Reputation: 5618

PyTorch - to NumPy yields unsized object?

Converting a PyTorch tensor to NumPy I get 

print(nn_result.shape)
# (2433, 2)
np_result = torch.argmax(nn_result).numpy()
type(np_result)
# <type 'numpy.ndarray'>
print(len(np_result))
TypeError: len() of unsized object

Why? I thought per documentation the numpy() function would return a proper ndarray, yet it seems to be incomplete somehow?

Upvotes: 5

Views: 1530

Answers (1)

scarecrow
scarecrow

Reputation: 6864

Perhaps you'd want to use torch.argmax(nn_result, dim=1) ? Since dim defaults to 0, it returns just a single number constructed as a tensor. Let me illustrate with the below example:

>>> x = np.array(1)
>>> x.shape
()
>>> len(x)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: len() of unsized object
>>> x = np.array([1])
>>> x.shape
(1,)
>>> len(x)
1

Essentially np.array will take up any object type that you construct with. In the first case, object is not an array because of which you do not see a valid shape. Since it is not an array calling len throws an error.

torch.argmax with dim=0 returns a tensor as seen in the first case of the above example and hence the error.

Upvotes: 2

Related Questions