I use resnet to predict EEG response to image data. And the input data is a matrix x_i with size of 1 * 3 * 500 * 500. Then the output is a vector y_i with size of 1 * 1 * 17 * 100. Let’s call the output EEG response. there is a loop like the following:
for i in range(number of input data):
y_i = model(x_i)
if max_i=200, I would to stack all the y_i to create a matrix with size of 200 * 1 * 17 * 100. How to realize that?
Hi Nikiguo (and gphilip)!
For the specific use case you detail, you will prefer torch.cat().
(torch.stack() will create a new dimension along which to stack.)
>>> import torch
>>> torch.__version__
'1.9.0'
>>> y_0 = torch.arange (5).unsqueeze (0)
>>> y_1 = torch.arange (5).unsqueeze (0) + 10
>>> y_2 = torch.arange (5).unsqueeze (0) + 20
>>> y_0.shape
torch.Size([1, 5])
>>> y_list = [y_0, y_1, y_2]
>>> torch.cat (y_list).shape
torch.Size([3, 5])
>>> torch.cat (y_list)
tensor([[ 0, 1, 2, 3, 4],
[10, 11, 12, 13, 14],
[20, 21, 22, 23, 24]])
>>> torch.stack (y_list).shape
torch.Size([3, 1, 5])
>>> torch.stack (y_list)
tensor([[[ 0, 1, 2, 3, 4]],
[[10, 11, 12, 13, 14]],
[[20, 21, 22, 23, 24]]])
(In your case, you could use stack() followed by squeeze() to
eliminate one of the singleton dimensions.)
Best.
K. Frank
Hi thank you for your help, gphilip. It can work but would increase the dimension of the data. But this can be solved by using squeeze().
aha thanks a lot Frank. That is helpful and it works well for me.