39

I'm currently working with the PyTorch framework and trying to understand foreign code. I got an indices issue and wanted to print the shape of a list.
The only way of doing so (as far as Google tells me) is to convert the list into a numpy array and then getting the shape with numpy.ndarray.shape().

But trying to convert my list into an array, I got a ValueError: only one element tensors can be converted to Python scalars.

My List is a converted PyTorch Tensor (list(pytorchTensor)) and looks somewhat like this:

[
tensor([[-0.2781, -0.2567, -0.2353,  ..., -0.9640, -0.9855, -1.0069],  
        [-0.2781, -0.2567, -0.2353,  ..., -1.0069, -1.0283, -1.0927],  
        [-0.2567, -0.2567, -0.2138,  ..., -1.0712, -1.1141, -1.1784],  
        ...,  
        [-0.6640, -0.6425, -0.6211,  ..., -1.0712, -1.1141, -1.0927],  
        [-0.6640, -0.6425, -0.5997,  ..., -0.9426, -0.9640, -0.9640],  
        [-0.6640, -0.6425, -0.5997,  ..., -0.9640, -0.9426, -0.9426]]),

tensor([[-0.0769, -0.0980, -0.0769,  ..., -0.9388, -0.9598, -0.9808],  
        [-0.0559, -0.0769, -0.0980,  ..., -0.9598, -1.0018, -1.0228],    
        [-0.0559, -0.0769, -0.0769,  ..., -1.0228, -1.0439, -1.0859],  
        ...,  
        [-0.4973, -0.4973, -0.4973,  ..., -1.0018, -1.0439, -1.0228],  
        [-0.4973, -0.4973, -0.4973,  ..., -0.8757, -0.9177, -0.9177],  
        [-0.4973, -0.4973, -0.4973,  ..., -0.9177, -0.8967, -0.8967]]),

tensor([[-0.1313, -0.1313, -0.1100,  ..., -0.8115, -0.8328, -0.8753],  
        [-0.1313, -0.1525, -0.1313,  ..., -0.8541, -0.8966, -0.9391],  
        [-0.1100, -0.1313, -0.1100,  ..., -0.9391, -0.9816, -1.0666],  
        ...,  
        [-0.4502, -0.4714, -0.4502,  ..., -0.8966, -0.8966, -0.8966],  
        [-0.4502, -0.4714, -0.4502,  ..., -0.8115, -0.8115, -0.7903],  
        [-0.4502, -0.4714, -0.4502,  ..., -0.8115, -0.7690, -0.7690]]),
]

Is there a way of getting the shape of that list without converting it into a numpy array?

Mateen Ulhaq
  • 24,552
  • 19
  • 101
  • 135
Henning
  • 545
  • 1
  • 8
  • 16
  • 6
    Not sure if you had the same issue but I was trying to convert a nested list of torch tensors to a bigger tensor that respected the nesting by having more indices/dimensions for the final tensor. I found that incrementally growing the list of torch tensors and then to `torch.stack` it helped. I wrote a non recursive example but I think it should be easy to extend it. Hope it helps: https://stackoverflow.com/questions/54307225/whats-the-difference-between-torch-stack-and-torch-cat-functions/66036075#66036075 – Charlie Parker Feb 04 '21 at 17:33
  • @CharlieParker I was trying to do the same thing. it **should** work. but it doesn't!! – user3180 Apr 30 '21 at 04:11

3 Answers3

35

It seems like you have a list of tensors. For each tensor you can see its size() (no need to convert to list/numpy). If you insist, you can convert a tensor to numpy array using numpy():

Return a list of tensor shapes:

>> [t.size() for t in my_list_of_tensors]

Returns a list of numpy arrays:

>> [t.numpy() for t in my_list_of_tensors]

In terms of performance, it is always best to avoid casting of tensors into numpy arrays, as it may incur sync of device/host memory. If you only need to check the shape of a tensor, use size() function.

Mateen Ulhaq
  • 24,552
  • 19
  • 101
  • 135
Shai
  • 111,146
  • 38
  • 238
  • 371
7

The simplest way to convert pytorch tensor to numpy array is:

nparray = tensor.numpy()

Also, for size and shape:

tensor_size = tensor.size()
tensor_shape = tensor.shape()
tensor_size
>>> (1080)
tensor_shape
>>> (32, 3, 128, 128)
Koke Cacao
  • 468
  • 6
  • 8
3

A real-world example, would require to handle torch no grad issue:

with torch.no_grad():
    probs = [t.numpy() for t in my_tensors]

or

probs = [t.detach().numpy() for t in my_tensors]
loretoparisi
  • 15,724
  • 11
  • 102
  • 146