tensor transformation in pytorch?

user_5982232 0
Viewed 235

I have a tensor of shape (size, 1) and I want to convert it into of shape (size, lookback, 1) by shifting its values. A pandas equivalent is below

size = 7
lookback = 3

data = pd.DataFrame(np.arange(size), columns=['out'])  # input
y = np.full((len(data), lookback, 1), np.nan)          # required/output
for j in range(lookback):
    y[:, j, 0] = data['out'].shift(lookback - j - 1).fillna(method="bfill")

How can I acheive similar in pytorch?

Example input:

[0, 1, 2, 3, 4, 5, 6]

Desired output:

[[0. 0. 0.]
 [0. 0. 1.]
 [0. 1. 2.]
 [1. 2. 3.]
 [2. 3. 4.]
 [3. 4. 5.]
 [4. 5. 6.]]
torch tensor pytorch multidimensional-array arrays
1 Answers

You can use Tensor.unfold for this. First though you will need to pad the front of the tensor, for that you could use nn.functional.pad. E.g.

import torch
import torch.nn.functional as F

size = 7
loopback = 3

data = torch.arange(size, dtype=torch.float)

# pad front of data with 2 values
# replicate padding requires 3d, 4d, or 5d tensor, hence the creation of two unitary dimensions before padding
data_padded = F.pad(data[None, None, ...], (loopback - 1, 0), 'replicate')[0, 0, ...]
# unfold with window size of 3 with step size of 1
y = data_padded.unfold(dimension=0, size=loopback, step=1)

which gives output of

tensor([[0., 0., 0.],
        [0., 0., 1.],
        [0., 1., 2.],
        [1., 2., 3.],
        [2., 3., 4.],
        [3., 4., 5.],
        [4., 5., 6.]])
user_27900470
Related