Assume I have an high dimension vector v which dimension is (2000,1), which representing a image feature vector. And I want to get an low embedding for v as following:
f=E*v where E is the embedding matrix with size (30, 2000); now the new vector f is (30,1), which is much smaller than v.
My question is how to implement above idea in pytorch, and I want the embedding matrix E can be trained with data with some loss function?
import torch
from torch import nn
v = torch.randn(2000,1)
…
PS: the whole process is that the vector v will go through an Embedding layer and then output the low dimension vector f.