Tuple to Tensor conversion doesn't work

data
1

Hello, lately I tried to train a neural network using Sequential and DataLoader. Data loader returns a tuple with length=4 in which every element is an image as torch.Tensor. Because I can’t pass tuple but Tensor as an input to the neural network I tried converting it by firstly making it a numpy array and then converting it to Tensor. However, when trying to convert to numpy array I get a following error: ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (4,) + inhomogeneous part. Each tensor has the same shape as shown in the image below.

Screenshot from 2024-02-28 19-04-58
Screenshot from 2024-02-28 19-04-582314×750 91.4 KB

Is this expected behaviour with tensors as tuple elements? What is interesting is that the same code works just fine in google colab. A workaround I found is to firstly convert every element of tuple to numpy array and then convert whole thing to tensor.

2

You could use torch.stack to transform the tuple into a single batch:

batch = tuple([torch.randn(3, 480, 640) for _ in range(4)])

for b in batch:
    print(b.shape)
# torch.Size([3, 480, 640])
# torch.Size([3, 480, 640])
# torch.Size([3, 480, 640])
# torch.Size([3, 480, 640])

batch = torch.stack(batch)
print(batch.shape)
# torch.Size([4, 3, 480, 640])