I got a list that contains lists of different length. How can i transform it in a tensor in pytorch without using padding? Is it possible?
[[3, 5, 10, 11], [1, 5, 10]]
1 Answer
It depends on what you want to achieve with the data structure. You can use torch.sparse, for example:
ll = [[3, 5, 10, 11], [1, 5, 10]]
n = len(ll)
m = max(len(l) for l in ll)
ids = [[], []]
values = []
for i, l in enumerate(ll):
length = len(l)
ids[0] += [i] * length # rows
ids[1] += list(range(length)) # cols
values += l
t = torch.sparse_coo_tensor(ids, values, (n, m))
Otherwise, you can try with embedding techniques for corpus of documents, such as bag-of-words (though it will generate still some "padding"), tf-idf, etc.
bag-of-words with possible duplicates in inner lists
corpus = [[3, 5, 10, 11], [1, 5, 10]]
n = len(corpus)
m = max(max(inner) for inner in corpus)
t = torch.zeros(n, m)
for i, doc in enumerate(corpus):
torch.bincount(corpus)
bag-of-words with distinct values in inner lists
corpus = [[3, 5, 10, 11], [1, 5, 10]]
n = len(corpus)
m = max(max(inner) for inner in corpus)
t = torch.zeros(n, m)
for i, doc in enumerate(corpus):
t[i, doc] = 1
Comments
