How to convert 2D matrix to 3D by embeddings in numpy?

Asked Oct 12 '21 at 02:03

Active Oct 12 '21 at 12:02

Viewed 26 times

I have 2D matrix with shape [batch_size, seq_len] where values of the 1st dimension are from range [0, num_classes - 1].

I also have another embedding matrix with shape [num_classes, embedding_size].

How can I transform 2D matrix to 3D with shape [batch_size, seq_len, embedding_size] by replacing the class labels (scalar) in 2D matrix with its embedding vector?

I know how to implement it in for-loop way, but how it can be vectorized?

edited Oct 12 '21 at 12:02

asked Oct 12 '21 at 02:03

Kenenbek Arzymatov

8,439
19
58
109

torch.nn.functional.embedding – MWB Oct 12 '21 at 02:17
@bobcat Is there any numpy like method, without gradients? – Kenenbek Arzymatov Oct 12 '21 at 03:05
`with torch.no_grad(): your_code()` – MWB Oct 12 '21 at 03:46
... actually, if your input doesn't `requires_grad`, then that should be enough – MWB Oct 12 '21 at 03:47
@bobcat thank it is good suggestion. and I update this question to be a numpy question. – Kenenbek Arzymatov Oct 12 '21 at 12:04

How to convert 2D matrix to 3D by embeddings in numpy?

0 Answers0