Permute() or view()

nlp
1

I want to change the input form(batch, timesteps, features) into (timesteps, batch, features), but i don’t understand the difference between permute() and view(), and which one to use.

2

You should use permute, as otherwise you’ll just get a new view on the same data using different strides.
Here is a small example:

batch_size = 2
time_steps = 3
features = 4

x = torch.arange(batch_size*time_steps*features).view(batch_size, time_steps, features)
print(x[0])
y = x.permute(1, 0, 2)
print(y[:, 0])  # Same as x[0]
z = x.view(time_steps, batch_size, features)
print(z[:, 0])  # NOT the same!
3

Thank you, it helps a lot.