一、随机初始化的embedding
import torch.nn as nn
import torch
user_embedding = nn.Embedding(5, 4)
weight = user_embedding.weight.data.numpy()
print(weight)
input = torch.tensor([1,2])
print(user_embedding(input))
# 输出
[[-1.2546824 0.15995328 -1.3933309 0.78180116]
[-1.160267 0.15576266 -1.2538452 0.3032509 ]
[-0.2289746 -0.86487883 -2.103254 -0.12548219]
[ 0.26746908 -0.60910237 -0.2590135 0.97958463]
[ 0.44616193 0.7447642 0.98776644 1.2315478 ]]
tensor([[-1.1603, 0.1558, -1.2538, 0.3033],
[-0.2290, -0.8649, -2.1033, -0.1255]], grad_fn=<EmbeddingBackward>)
二、自定义embedding权重
方式一:
import torch.nn as nn
import torch
t = torch.tensor([[0.8245, 0.1198, 0.6111],
[0.8917, 0.0699, 0.3550],
[0.0691, 0.2732, 0.7184],
[0.0359, 0.9635, 0.2277],
[0.8692, 0.9012, 0.7116]])
user_embedding = nn.Embedding(5, 3, _weight=nn.Parameter(t))
weight = user_embedding.weight.data.numpy()
print(weight)
input = torch.tensor([1,2])
print(user_embedding(input))
# 输出
[[0.8245 0.1198 0.6111]
[0.8917 0.0699 0.355 ]
[0.0691 0.2732 0.7184]
[0.0359 0.9635 0.2277]
[0.8692 0.9012 0.7116]]
tensor([[0.8917, 0.0699, 0.3550],
[0.0691, 0.2732, 0.7184]], grad_fn=<EmbeddingBackward>)
注意:如果仅仅使用类embedding操作,而不需要更新参数
设置 nn.Embedding(5, 3, _weight=nn.Parameter(t, requires_grad=False
))
方式二:
import torch.nn.functional as F
import torch
input = torch.tensor([[1,2,4,5],[4,3,2,9]])
embedding_matrix = torch.rand(10, 3)
print(embedding_matrix)
print(F.embedding(input, embedding_matrix))
# 输出
tensor([[0.8245, 0.1198, 0.6111],
[0.8917, 0.0699, 0.3550],
[0.0691, 0.2732, 0.7184],
[0.0359, 0.9635, 0.2277],
[0.8692, 0.9012, 0.7116],
[0.4018, 0.3034, 0.3883],
[0.2971, 0.2333, 0.3194],
[0.1596, 0.5504, 0.8969],
[0.4099, 0.5024, 0.0058],
[0.1426, 0.4245, 0.7905]])
tensor([[[0.8917, 0.0699, 0.3550],
[0.0691, 0.2732, 0.7184],
[0.8692, 0.9012, 0.7116],
[0.4018, 0.3034, 0.3883]],
[[0.8692, 0.9012, 0.7116],
[0.0359, 0.9635, 0.2277],
[0.0691, 0.2732, 0.7184],
[0.1426, 0.4245, 0.7905]]])