CODEWITHSUNDEEP
pythoninitializationappendpytorchempty-list
infobox
infobox

Reputation: 13

pytorch: How do I properly initialize Tensor without any entries?

What I am doing right now is this:

In [1]: torch.Tensor([[[] for _ in range(3)] for _ in range(5)])                     
Out[1]: tensor([], size=(5, 3, 0))

This works fine for me, but is there maybe a torch function that does this that I am missing?

Thanks in advance!

Edit: My use case is this: I use this to aggregate Tensors with all dimensions the same and that dont have the empty dimension. I am using torch.cat:

# results start with shape (a,b,0)
results = torch.Tensor([[[] for _ in range(b)] for _ in range(a)]) 

for t in range(time):
    # r has shape (a,b)
    r = model(...) 
    # results now has shape (a,b,t)
    results = torch.cat([results,r.unsqueeze(2)],dim=-1)

Simply appending to a list is impractical for me as I have to do reshaping operations on results on every step (Im doing beam search).

One solution would also be to not initialize results until I have the first returned Tensor, but this feels unpythonic/wrong.

Upvotes: 1

Views: 15071

Answers (2)

trialNerror
trialNerror

Reputation: 3573

You have the torch.empty function :

torch.empty(5,3,0)
>>> tensor([], size=(5, 3, 0))

is a tensor without any entry

Upvotes: 1

Mughees
Mughees

Reputation: 953

This can be another way depending on your usecase.

alpha = torch.tensor([])

In[5]:  alpha[:,None,None,None]
Out[5]: tensor([], size=(0, 1, 1, 1))

Otherways:

torch.tensor([[[[]]]]) #tensor([], size=(1, 1, 1, 0))
torch.tensor([[[[],[]]]]) #tensor([], size=(1, 1, 2, 0))

Upvotes: 1

Related Questions