python
复制代码
import math
import torch
from torch import nn
from d2l import torch as d2l
def masked_softmax(X, valid_lens):
"""通过在最后一个轴上掩蔽元素来执行softmax操作"""
# X:3D张量，valid_lens:1D或2D张量
if valid_lens is None:
return nn.functional.softmax(X, dim=-1)
else:
shape = X.shape
if valid_lens.dim() == 1:
valid_lens = torch.repeat_interleave(valid_lens, shape[1])
else:
valid_lens = valid_lens.reshape(-1)
# 最后一轴上被掩蔽的元素使用一个非常大的负值替换，从而其softmax输出为0
X = d2l.sequence_mask(X.reshape(-1, shape[-1]), valid_lens,
value=-1e6)
return nn.functional.softmax(X.reshape(shape), dim=-1)
class AdditiveAttention(nn.Module):
def __init__(self,key_size,query_size,num_hiddens,dropout,**kwargs):
super(AdditiveAttention,self).__init__(**kwargs)
self.w_k=nn.Linear(key_size,num_hiddens,bias=False)
self.w_q=nn.Linear(query_size,num_hiddens,bias=False)
self.w_v=nn.Linear(num_hiddens,1,bias=False)
self.dropout=nn.Dropout(dropout)
def forward(self,queries,keys,values,valid_lens):
queries,keys=self.w_q(queries),self.w_k(keys)
#print(queries.shape)
# q=queries.unsqueeze(2)
# k=keys.unsqueeze(1)
#利用广播机制进行相加:
#queries.unsqueeze(2):[B,sq_len,1,hiddens]
#keys.unsqueeze(1):[B,1,sq_len,hiddens]
features=queries.unsqueeze(2)+keys.unsqueeze(1)
#print(features.shape)
features=torch.tanh(features)
scores=self.w_v(features).squeeze(-1)
self.attention_weights=masked_softmax(scores,valid_lens)
#批量矩阵乘乘法:bmm
return torch.bmm(self.dropout(self.attention_weights),values)
#输入测试:
queries, keys = torch.normal(0, 1, (2, 1, 20)), torch.ones((2, 10, 2))
values = torch.arange(40, dtype=torch.float32).reshape(1, 10, 4).repeat(2, 1, 1)
valid_lens = torch.tensor([2, 6])
attention = AdditiveAttention(key_size=2, query_size=20, num_hiddens=8,dropout=0.1)
attention.eval()
print(attention(queries, keys, values, valid_lens))
d2l.show_heatmaps(attention.attention_weights.reshape((1, 1, 2, 10)),
xlabel='Keys', ylabel='Queries')
python
复制代码
import math
import torch
from torch import nn
from d2l import torch as d2l
def sequence_mask(X, valid_lens, value=float("-inf")):
"""简单实现的掩蔽功能"""
batch_size, seq_len = X.shape
mask = torch.arange(seq_len, device=X.device).expand(batch_size, seq_len) < valid_lens.unsqueeze(1)
X[~mask] = value
return X
def masked_softmax(X, valid_lens):
"""通过在最后一个轴上掩蔽元素来执行softmax操作"""
if valid_lens is None:
return nn.functional.softmax(X, dim=-1)
else:
shape = X.shape
if valid_lens.dim() == 1:
valid_lens = valid_lens.unsqueeze(1).repeat(1, shape[1])
X = X.reshape(-1, shape[-1])
X = sequence_mask(X, valid_lens.reshape(-1), value=-1e6)
return nn.functional.softmax(X.reshape(shape), dim=-1)
class DotProductAttention(nn.Module):
"""缩放点积注意力"""
def __init__(self, dropout, **kwargs):
super(DotProductAttention, self).__init__(**kwargs)
self.dropout = nn.Dropout(dropout)
# queries的形状：(batch_size，查询的个数，d)
# keys的形状：(batch_size，"键－值"对的个数，d)
# values的形状：(batch_size，"键－值"对的个数，值的维度)
# valid_lens的形状:(batch_size，)或者(batch_size，查询的个数)
def forward(self, queries, keys, values, valid_lens=None):
d = queries.shape[-1]
# 设置transpose_b=True为了交换keys的最后两个维度
scores = torch.bmm(queries, keys.transpose(1,2)) / math.sqrt(d)
self.attention_weights = masked_softmax(scores, valid_lens)
return torch.bmm(self.dropout(self.attention_weights), values)
#输入测试:
queries = torch.normal(0, 1, (2, 3, 4))
keys = torch.normal(0, 1, (2,5, 4))
values = torch.normal(0, 1, (2, 5, 6))
valid_lens = torch.tensor([3, 4])
attention = DotProductAttention(dropout=0.5)
attention.eval()
attention(queries, keys, values, valid_lens)
print(attention.attention_weights.shape)
d2l.show_heatmaps(attention.attention_weights[1].reshape((1, 1, 3, 5)),
xlabel='Keys', ylabel='Queries')
