1 理解
1-1 视频看这个，但是要资料的话其公众号说了免费，加了一堆问题就是不发。
https://www.bilibili.com/video/BV1sW4y1J7cL/?p=23&spm_id_from=333.880.my_history.page.click&vd_source=f88ed35500cb30c7be9bbe418a5998ca
1-2 代码
文本讲解+实际代码清晰
https://zhuanlan.zhihu.com/p/403433120
https://zhuanlan.zhihu.com/p/403433120
区分
1 不用于RNN使用隐藏态递推，使用位置编码代替了信息。
2 在1的基础上，单个词使用q k v直接并行计算，加速作用。
多个词也并行计算，但是为了妨碍编码阶段后续输入泄露信息，使用掩码遮蔽掉t时刻后续的信息。
、
首先逐个词语预测，然后第一轮预测结果当做第二次的输入，全部进去重新预测一边。
第一次预测阶段，逐个输出，当前位置以后的词都是0，mask掉后续未预测词的因素影响
第二次预测，自第一次预测基础上，一次性全部考虑前后所有词相互影响。
数据预处理datasets.py
# Starting Reference: http://nlp.seas.harvard.edu/2018/04/03/attention.html#greedy-decoding
import torch
import torch.utils.data as Data
# Encoder_input 训练和预测 Decoder_input 训练 Decoder_output 真值用于统计准确度
sentences = [['我 是 学 生 P', 'S I am a student', 'I am a student E'], # S: 开始符号
['我 喜 欢 学 习', 'S I like learning P', 'I like learning P E'], # E: 结束符号
['我 是 男 生 P', 'S I am a boy', 'I am a boy E']] # P: 占位符号，如果当前句子不足固定长度用P占位
src_vocab = {'P': 0, '我': 1, '是': 2, '学': 3, '生': 4, '喜': 5, '欢': 6, '习': 7, '男': 8} # 词源字典 字：索引
src_idx2word = {src_vocab[key]: key for key in src_vocab}
src_vocab_size = len(src_vocab) # 字典字的个数
tgt_vocab = {'P': 0, 'S': 1, 'E': 2, 'I': 3, 'am': 4, 'a': 5, 'student': 6, 'like': 7, 'learning': 8, 'boy': 9}
idx2word = {tgt_vocab[key]: key for key in tgt_vocab} # 把目标字典转换成 索引：字的形式
tgt_vocab_size = len(tgt_vocab) # 目标字典尺寸
src_len = len(sentences[0][0].split(" ")) # Encoder输入的最大长度
tgt_len = len(sentences[0][1].split(" ")) # Decoder输入输出最大长度
# 把sentences 转换成字典索引
def make_data():
enc_inputs, dec_inputs, dec_outputs = [], [], []
for i in range(len(sentences)): # 句子数目
enc_input = [[src_vocab[n] for n in sentences[i][0].split()]]
dec_input = [[tgt_vocab[n] for n in sentences[i][1].split()]]
dec_output = [[tgt_vocab[n] for n in sentences[i][2].split()]]
enc_inputs.extend(enc_input)
dec_inputs.extend(dec_input)
dec_outputs.extend(dec_output)
return torch.LongTensor(enc_inputs), torch.LongTensor(dec_inputs), torch.LongTensor(dec_outputs)
# 自定义数据集函数
class MyDataSet(Data.Dataset):
def __init__(self, enc_inputs, dec_inputs, dec_outputs):
super(MyDataSet, self).__init__()
self.enc_inputs = enc_inputs
self.dec_inputs = dec_inputs
self.dec_outputs = dec_outputs
def __len__(self):
return self.enc_inputs.shape[0]
def __getitem__(self, idx):
return self.enc_inputs[idx], self.dec_inputs[idx], self.dec_outputs[idx]
训练main.py
# Starting Reference: http://nlp.seas.harvard.edu/2018/04/03/attention.html#greedy-decoding
import torch.nn as nn
import torch.optim as optim
from datasets import *
from transformer import Transformer
if __name__ == "__main__":
enc_inputs, dec_inputs, dec_outputs = make_data()
print('enc_inputs \n',enc_inputs.shape,"\n",enc_inputs)
print('dec_inputs \n',dec_inputs.shape,"\n",dec_inputs)
print('dec_outputs \n',dec_outputs.shape,"\n",dec_outputs)
loader = Data.DataLoader(MyDataSet(enc_inputs, dec_inputs, dec_outputs), 2, True)
# 批次2
model = Transformer().cuda()
criterion = nn.CrossEntropyLoss(ignore_index=0) # 忽略 占位符 索引为0.
optimizer = optim.SGD(model.parameters(), lr=1e-3, momentum=0.99)
for epoch in range(10):
for enc_inputs, dec_inputs, dec_outputs in loader: # enc_inputs : [batch_size, src_len]
# dec_inputs : [batch_size, tgt_len]
# dec_outputs: [batch_size, tgt_len]
enc_inputs, dec_inputs, dec_outputs = enc_inputs.cuda(), dec_inputs.cuda(), dec_outputs.cuda()
print('======================')
print('enc_inputs \n',enc_inputs)
print('dec_inputs \n',dec_inputs)
'''
enc_inputs tensor([[1, 2, 3, 4, 0]])
dec_inputs tensor([[1, 3, 4, 5, 6]])
'''
outputs, enc_self_attns, dec_self_attns, dec_enc_attns = model(enc_inputs, dec_inputs)
# outputs: [batch_size * tgt_len, tgt_vocab_size]
loss = criterion(outputs, dec_outputs.view(-1))
print('Epoch:', '%04d' % (epoch + 1), 'loss =', '{:.6f}'.format(loss))
optimizer.zero_grad()
loss.backward()
optimizer.step()
torch.save(model, 'model.pth')
print("保存模型")
测试预测
from datasets import *
def get_attn_pad_mask(seq_q, seq_k): # seq_q: [batch_size, seq_len] ,seq_k: [batch_size, seq_len]
batch_size, len_q = seq_q.size()
batch_size, len_k = seq_k.size()
pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # 判断 输入那些含有P(=0),用1标记 ,[batch_size, 1, len_k]
return pad_attn_mask.expand(batch_size, len_q, len_k) # 扩展成多维度
def test(model, enc_input, start_symbol):
# Starting Reference: http://nlp.seas.harvard.edu/2018/04/03/attention.html#greedy-decoding
enc_outputs, enc_self_attns = model.Encoder(enc_input)
dec_input = torch.zeros(1, tgt_len).type_as(enc_input.data)
next_symbol = start_symbol
for i in range(0, tgt_len):
print('dec_input',dec_input)
dec_input[0][i] = next_symbol
dec_outputs, _, _ = model.Decoder(dec_input, enc_input, enc_outputs)
projected = model.projection(dec_outputs)
prob = projected.squeeze(0).max(dim=-1, keepdim=False)[1]
next_word = prob.data[i]
next_symbol = next_word.item()
return dec_input
enc_inputs, dec_inputs, dec_outputs = make_data()
loader = Data.DataLoader(MyDataSet(enc_inputs, dec_inputs, dec_outputs), 2, True)
enc_inputs, _, _ = next(iter(loader))
model = torch.load('model.pth')
predict_dec_input = test(model, enc_inputs[0].view(1, -1).cuda(), start_symbol=tgt_vocab["S"])
predict=predict_dec_input.data[0]
print('解码器输入',[src_idx2word[int(i)] for i in enc_inputs[0]], '->',
[idx2word[n.item()] for n in predict.squeeze()])
#解码器输入 ['我', '是', '学', '生', 'P'] -> ['S', 'I', 'am', 'a', 'student']
print("===============第二次================")
print("===============第二次================")
print("===============第二次================")
predict, _, _, _ = model(enc_inputs[0].view(1, -1).cuda(), predict_dec_input)
predict = predict.data.max(1, keepdim=True)[1]
print("解码器输出",[src_idx2word[int(i)] for i in enc_inputs[0]], '->',
[idx2word[n.item()] for n in predict.squeeze()])
#解码器输出 ['我', '是', '学', '生', 'P'] -> ['I', 'am', 'a', 'student', 'E']
模型transformer.py
import numpy as np
import torch.nn as nn
from datasets import *
d_model = 512 # 字 Embedding 的维度
d_ff = 2048 # 前向传播隐藏层维度
d_k = d_v = 64 # K(=Q), V的维度
n_layers = 6 # 有多少个encoder和decoder
n_heads = 8 # Multi-Head Attention设置为8
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pos_table = np.array([
[pos / np.power(10000, 2 * i / d_model) for i in range(d_model)]
if pos != 0 else np.zeros(d_model) for pos in range(max_len)])
pos_table[1:, 0::2] = np.sin(pos_table[1:, 0::2]) # 字嵌入维度为偶数时
pos_table[1:, 1::2] = np.cos(pos_table[1:, 1::2]) # 字嵌入维度为奇数时
self.pos_table = torch.FloatTensor(pos_table).cuda() # enc_inputs: [seq_len, d_model]
def forward(self, enc_inputs): # enc_inputs: [batch_size, seq_len, d_model]
enc_inputs += self.pos_table[:enc_inputs.size(1), :]
return self.dropout(enc_inputs.cuda())
def get_attn_pad_mask(seq_q, seq_k): # seq_q: [batch_size, seq_len] ,seq_k: [batch_size, seq_len]
batch_size, len_q = seq_q.size()
batch_size, len_k = seq_k.size()
pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # 判断 输入那些含有P(=0),用1标记 ,[batch_size, 1, len_k]
return pad_attn_mask.expand(batch_size, len_q, len_k) # 扩展成多维度
def get_attn_subsequence_mask(seq): # seq: [batch_size, tgt_len]
attn_shape = [seq.size(0), seq.size(1), seq.size(1)]
subsequence_mask = np.triu(np.ones(attn_shape), k=1) # 生成上三角矩阵,[batch_size, tgt_len, tgt_len]
subsequence_mask = torch.from_numpy(subsequence_mask).byte() # [batch_size, tgt_len, tgt_len]
return subsequence_mask
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask): # Q: [batch_size, n_heads, len_q, d_k]
# K: [batch_size, n_heads, len_k, d_k]
# V: [batch_size, n_heads, len_v(=len_k), d_v]
# attn_mask: [batch_size, n_heads, seq_len, seq_len]
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k) # scores : [batch_size, n_heads, len_q, len_k]
scores.masked_fill_(attn_mask, -1e9) # 如果时停用词P就等于 0
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V) # [batch_size, n_heads, len_q, d_v]
return context, attn
class MultiHeadAttention(nn.Module):
def __init__(self):
super(MultiHeadAttention, self).__init__()
self.W_Q = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_K = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_V = nn.Linear(d_model, d_v * n_heads, bias=False)
self.fc = nn.Linear(n_heads * d_v, d_model, bias=False)
def forward(self, input_Q, input_K, input_V, attn_mask): # input_Q: [batch_size, len_q, d_model]
# input_K: [batch_size, len_k, d_model]
# input_V: [batch_size, len_v(=len_k), d_model]
# attn_mask: [batch_size, seq_len, seq_len]
residual, batch_size = input_Q, input_Q.size(0)
Q = self.W_Q(input_Q).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # Q: [batch_size, n_heads, len_q, d_k]
K = self.W_K(input_K).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # K: [batch_size, n_heads, len_k, d_k]
V = self.W_V(input_V).view(batch_size, -1, n_heads, d_v).transpose(1,
2) # V: [batch_size, n_heads, len_v(=len_k), d_v]
attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1,
1) # attn_mask : [batch_size, n_heads, seq_len, seq_len]
context, attn = ScaledDotProductAttention()(Q, K, V, attn_mask) # context: [batch_size, n_heads, len_q, d_v]
# attn: [batch_size, n_heads, len_q, len_k]
context = context.transpose(1, 2).reshape(batch_size, -1,
n_heads * d_v) # context: [batch_size, len_q, n_heads * d_v]
output = self.fc(context) # [batch_size, len_q, d_model]
return nn.LayerNorm(d_model).cuda()(output + residual), attn
class PoswiseFeedForwardNet(nn.Module):
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.fc = nn.Sequential(
nn.Linear(d_model, d_ff, bias=False),
nn.ReLU(),
nn.Linear(d_ff, d_model, bias=False))
def forward(self, inputs): # inputs: [batch_size, seq_len, d_model]
residual = inputs
output = self.fc(inputs)
return nn.LayerNorm(d_model).cuda()(output + residual) # [batch_size, seq_len, d_model]
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention() # 多头注意力机制
self.pos_ffn = PoswiseFeedForwardNet() # 前馈神经网络
def forward(self, enc_inputs, enc_self_attn_mask): # enc_inputs: [batch_size, src_len, d_model]
# 输入3个enc_inputs分别与W_q、W_k、W_v相乘得到Q、K、V # enc_self_attn_mask: [batch_size, src_len, src_len]
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs,
# enc_outputs: [batch_size, src_len, d_model],
enc_self_attn_mask) # attn: [batch_size, n_heads, src_len, src_len]
enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size, src_len, d_model]
return enc_outputs, attn
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention() # 多头注意力机制
self.pos_ffn = PoswiseFeedForwardNet() # 前馈神经网络
def forward(self, enc_inputs, enc_self_attn_mask): # enc_inputs: [batch_size, src_len, d_model]
# 输入3个enc_inputs分别与W_q、W_k、W_v相乘得到Q、K、V # enc_self_attn_mask: [batch_size, src_len, src_len]
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs,
# enc_outputs: [batch_size, src_len, d_model],
enc_self_attn_mask) # attn: [batch_size, n_heads, src_len, src_len]
enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size, src_len, d_model]
return enc_outputs, attn
class Encoder(nn.Module):
def __init__(self):
super(Encoder, self).__init__()
self.src_emb = nn.Embedding(src_vocab_size, d_model) # 把字转换字向量
self.pos_emb = PositionalEncoding(d_model) # 加入位置信息
self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
def forward(self, enc_inputs): # enc_inputs: [batch_size, src_len]
enc_outputs = self.src_emb(enc_inputs) # enc_outputs: [batch_size, src_len, d_model]
enc_outputs = self.pos_emb(enc_outputs) # enc_outputs: [batch_size, src_len, d_model]
enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs) # enc_self_attn_mask: [batch_size, src_len, src_len]
enc_self_attns = []
for layer in self.layers:
enc_outputs, enc_self_attn = layer(enc_outputs, enc_self_attn_mask) # enc_outputs : [batch_size, src_len, d_model],
# enc_self_attn : [batch_size, n_heads, src_len, src_len]
enc_self_attns.append(enc_self_attn)
return enc_outputs, enc_self_attns
class DecoderLayer(nn.Module):
def __init__(self):
super(DecoderLayer, self).__init__()
self.dec_self_attn = MultiHeadAttention()
self.dec_enc_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, dec_inputs, enc_outputs, dec_self_attn_mask,
dec_enc_attn_mask): # dec_inputs: [batch_size, tgt_len, d_model]
# enc_outputs: [batch_size, src_len, d_model]
# dec_self_attn_mask: [batch_size, tgt_len, tgt_len]
# dec_enc_attn_mask: [batch_size, tgt_len, src_len]
dec_outputs, dec_self_attn = self.dec_self_attn(dec_inputs, dec_inputs,
dec_inputs,
dec_self_attn_mask) # dec_outputs: [batch_size, tgt_len, d_model]
# dec_self_attn: [batch_size, n_heads, tgt_len, tgt_len]
dec_outputs, dec_enc_attn = self.dec_enc_attn(dec_outputs, enc_outputs,
enc_outputs,
dec_enc_attn_mask) # dec_outputs: [batch_size, tgt_len, d_model]
# dec_enc_attn: [batch_size, h_heads, tgt_len, src_len]
dec_outputs = self.pos_ffn(dec_outputs) # dec_outputs: [batch_size, tgt_len, d_model]
return dec_outputs, dec_self_attn, dec_enc_attn
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
self.tgt_emb = nn.Embedding(tgt_vocab_size, d_model)
self.pos_emb = PositionalEncoding(d_model)
self.layers = nn.ModuleList([DecoderLayer() for _ in range(n_layers)])
def forward(self, dec_inputs, enc_inputs, enc_outputs): # dec_inputs: [batch_size, tgt_len]
# enc_intpus: [batch_size, src_len]
# enc_outputs: [batsh_size, src_len, d_model]
dec_outputs = self.tgt_emb(dec_inputs) # [batch_size, tgt_len, d_model]
dec_outputs = self.pos_emb(dec_outputs).cuda() # [batch_size, tgt_len, d_model]
# Mask掉句子中的占位符号和输出顺序细
# 去掉无用标记 例如'我 是 学 生 P' ，P对应句子没有实际意义 用1 替换
dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs).cuda() # [batch_size, tgt_len, tgt_len]
# 用来Mask未来输入信息，返回的是一个上三角矩阵。
dec_self_attn_subsequence_mask = get_attn_subsequence_mask(dec_inputs).cuda() # [batch_size, tgt_len, tgt_len]
# 无用词mask+ 未来信息抹除mask
dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask +
dec_self_attn_subsequence_mask), 0).cuda() # [batch_size, tgt_len, tgt_len]
# 出去无用mask
dec_enc_attn_mask = get_attn_pad_mask(dec_inputs, enc_inputs) # [batc_size, tgt_len, src_len]
print("=======================编码器开始========================")
print('enc_inputs',enc_inputs)
print('dec_inputs',dec_inputs)
print('1取出无用 dec_inputs-dec_inputs \n',dec_self_attn_pad_mask)
print('2三角阵 dec_inputs \n',dec_self_attn_subsequence_mask)
print('3无用词mask+ 未来信息抹除mask dec_inputs\n',dec_self_attn_mask)
print('4取出无用 dec_inputs-enc_inputs\n',dec_enc_attn_mask)
print("########################编码器结束###################")
dec_self_attns, dec_enc_attns = [], []
for layer in self.layers: # dec_outputs: [batch_size, tgt_len, d_model]
# dec_self_attn: [batch_size, n_heads, tgt_len, tgt_len]
# dec_enc_attn: [batch_size, h_heads, tgt_len, src_len]
dec_outputs, dec_self_attn, dec_enc_attn = layer(dec_outputs, enc_outputs, dec_self_attn_mask, dec_enc_attn_mask)
dec_self_attns.append(dec_self_attn)
dec_enc_attns.append(dec_enc_attn)
return dec_outputs, dec_self_attns, dec_enc_attns
class Transformer(nn.Module):
def __init__(self):
super(Transformer, self).__init__()
self.Encoder = Encoder().cuda()
self.Decoder = Decoder().cuda()
self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False).cuda()
def forward(self, enc_inputs, dec_inputs): # enc_inputs: [batch_size, src_len]
# dec_inputs: [batch_size, tgt_len]
enc_outputs, enc_self_attns = self.Encoder(enc_inputs) # enc_outputs: [batch_size, src_len, d_model],
# enc_self_attns: [n_layers, batch_size, n_heads, src_len, src_len]
dec_outputs, dec_self_attns, dec_enc_attns = self.Decoder(
dec_inputs, enc_inputs, enc_outputs) # dec_outpus : [batch_size, tgt_len, d_model],
# dec_self_attns: [n_layers, batch_size, n_heads, tgt_len, tgt_len],
# dec_enc_attn : [n_layers, batch_size, tgt_len, src_len]
dec_logits = self.projection(dec_outputs) # dec_logits: [batch_size, tgt_len, tgt_vocab_size]
return dec_logits.view(-1, dec_logits.size(-1)), enc_self_attns, dec_self_attns, dec_enc_attns
浙公网安备 33010602011771号