-
Notifications
You must be signed in to change notification settings - Fork 2
/
transformer.py
220 lines (176 loc) · 8.27 KB
/
transformer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import torch
import torch.nn as nn
#N= batch size, ie the no. of sentence pairs in a batch
#src_len= length of the src sentences, padding= 0
#query_len, key_len, value_len= src/trg sentence length depending on wheather used in encoder or decoder
class MultiHeadAttention(nn.Module):
def __init__(self, d_emb, heads):
'''
d_emb= embedding dim (same as d_model in the paper)
heads= no. of diff heads in the multi- attention layer
'''
super(MultiHeadAttention, self).__init__()
self.d_emb= d_emb
self.heads= heads
assert d_emb % heads== 0
#dim of each head, to be concatenated later
self.head_dim= d_emb// heads
self.fc_q= nn.Linear(self.d_emb, self.d_emb)
self.fc_k= nn.Linear(self.d_emb, self.d_emb)
self.fc_v= nn.Linear(self.d_emb, self.d_emb)
self.fc_o= nn.Linear(d_emb, d_emb)
def forward(self, query, key, value, mask= None):
#query shape= N x query_len x d_emb
N= query.shape[0]
query_len, key_len, value_len= query.shape[1], key.shape[1], value.shape[1]
#linearly project-> reshape
query= self.fc_q(query)
key= self.fc_q(key)
value= self.fc_q(value)
query= query.reshape(N, query_len, self.heads, self.head_dim)
key= key.reshape(N, key_len, self.heads, self.head_dim)
value= value.reshape(N, value_len, self.heads, self.head_dim)
#query shape= N x query_len x heads x heads_dim
energy= torch.einsum('nqhd, nkhd-> nhqk', [query, key])
#energy shape= N x heads x query_len x key_len
if mask is not None:
energy= energy.masked_fill(mask== 0, float('-inf'))
#apply softmax along the key dim
attention_weights= torch.softmax((energy/(self.head_dim)**(1/2)), dim= 3)
#attention_weights shape= N x heads x query_len x key_len
#value shape= N x value_len x heads x heads_dim
#out shape= N x query_len x heads x heads_dim
#out shape= N x query_len x d_emb (concatinating heads)
out= torch.einsum('nhqk, nvhd-> nqhd', [attention_weights, value]).reshape(N, query_len, -1)
out= self.fc_o(out)
return out
class PositionwiseFeedForwardLayer(nn.Module):
def __init__(self, d_emb, d_ff):
'''
d_ff= dim of the inner layer
'''
super(PositionwiseFeedForwardLayer, self).__init__()
self.features= nn.Sequential(nn.Linear(d_emb, d_ff),
nn.ReLU(),
nn.Linear(d_ff, d_emb))
def forward(self, x):
out= self.features(x)
return out
class EncoderBlock(nn.Module):
def __init__(self, d_emb, heads, d_ff, dropout):
super(EncoderBlock, self).__init__()
self.self_attention= MultiHeadAttention(d_emb, heads)
self.norm1= nn.LayerNorm(d_emb)
self.positionwise_ffn= PositionwiseFeedForwardLayer(d_emb, d_ff)
self.norm2= nn.LayerNorm(d_emb)
self.dropout= nn.Dropout(p= dropout)
def forward(self, src, src_mask):
'''
src_mask= prevents encoder from attending to <pad> tokens
'''
out_sa= self.self_attention(src, src, src, src_mask)
out_sa= self.norm1(src+ self.dropout(out_sa))
out= self.positionwise_ffn(out_sa)
out= self.norm2(out_sa+ self.dropout(out))
return out
class Encoder(nn.Module):
def __init__(self, src_vocab_size, d_emb, max_length, heads, d_ff, dropout, n_blocks):
super(Encoder, self).__init__()
self.token_emb= nn.Embedding(src_vocab_size, d_emb)
#learned positinal embeddings as opposed to the static ones used in the paper
self.positional_emb= nn.Embedding(max_length, d_emb)
self.encoder_blocks= nn.ModuleList([EncoderBlock(d_emb, heads, d_ff, dropout) for _ in range(n_blocks)])
self.dropout= nn.Dropout(p= dropout)
def forward(self, src, src_mask):
#src shape= N x src_len
#src_mask shape= N x 1 x 1 x src_len
N, src_len= src.shape
pos= torch.arange(0, src_len).expand(N, src_len)
src= self.dropout(self.token_emb(src)+ self.positional_emb(pos))
for block in self.encoder_blocks:
src= block(src, src_mask)
return src
class DecoderBlock(nn.Module):
def __init__(self, d_emb, heads, d_ff, dropout):
super(DecoderBlock, self).__init__()
self.masked_self_attention= MultiHeadAttention(d_emb, heads)
self.norm1= nn.LayerNorm(d_emb)
self.cross_attention= MultiHeadAttention(d_emb, heads)
self.norm2= nn.LayerNorm(d_emb)
self.positionwise_ffn= PositionwiseFeedForwardLayer(d_emb, d_ff)
self.norm3= nn.LayerNorm(d_emb)
self.dropout= nn.Dropout(p= dropout)
def forward(self, trg, trg_mask, encoded_src, src_mask):
out_msa= self.masked_self_attention(trg, trg, trg, trg_mask)
out_msa= self.norm1(trg+ self.dropout(out_msa))
out_ca= self.cross_attention(out_msa, encoded_src, encoded_src, src_mask)
out_ca= self.norm2(out_msa+ self.dropout(out_ca))
out= self.positionwise_ffn(out_ca)
out= self.norm3(out_ca+ self.dropout(out))
return out
class Decoder(nn.Module):
def __init__(self, trg_vocab_size, d_emb, max_length, heads, d_ff, dropout, n_blocks):
super(Decoder, self).__init__()
self.token_emb= nn.Embedding(trg_vocab_size, d_emb)
self.positional_emb= nn.Embedding(max_length, d_emb)
self.decoder_blocks= nn.ModuleList([DecoderBlock(d_emb, heads, d_ff, dropout) for _ in range(n_blocks)])
self.fc_out = nn.Linear(d_emb, trg_vocab_size)
self.dropout= nn.Dropout(p= dropout)
def forward(self, trg, trg_mask, encoded_src, src_mask):
N, trg_len= trg.shape
pos= torch.arange(0, trg_len).expand(N, trg_len)
trg= self.dropout(self.token_emb(trg)+ self.positional_emb(pos))
for block in self.decoder_blocks:
trg= block(trg, trg_mask, encoded_src, src_mask)
out= self.fc_out(trg)
return out
class Transformer(nn.Module):
def __init__(self,
src_vocab_size,
trg_vocab_size,
src_pad_idx,
trg_pad_idx,
d_emb= 512,
heads= 8,
d_ff= 2048,
n_blocks= 6,
dropout= 0.1,
max_length= 100):
super(Transformer, self).__init__()
self.encoder= Encoder(src_vocab_size, d_emb, max_length, heads, d_ff, dropout, n_blocks)
self.decoder= Decoder(trg_vocab_size, d_emb, max_length, heads, d_ff, dropout, n_blocks)
self.src_pad_idx= src_pad_idx
self.trg_pad_idx= trg_pad_idx
def make_src_mask(self, src):
#src shape= N, src_len
#src_mask= 0 if <pad> present, else 1
src_mask= (src!= self.src_pad_idx).unsqueeze(1).unsqueeze(2)
#src_mask shape= N x 1 x 1 x src_len
return src_mask
def make_trg_mask(self, trg):
N, trg_len= trg.shape
#trg_pad_mask shape= N x 1 x 1 x trg_len
trg_pad_mask= (trg!= self.src_pad_idx).unsqueeze(1).unsqueeze(2)
#trg_subseq_mask= trg_len x trg_len
trg_subseq_mask= torch.tril(torch.ones(trg_len, trg_len)).bool()
#trg_mask= N x 1 x trg_len x trg_len
trg_mask= trg_pad_mask & trg_subseq_mask
return trg_mask
def forward(self, src, trg):
src_mask= self.make_src_mask(src)
trg_mask= self.make_trg_mask(trg)
encoded_src= self.encoder(src, src_mask)
out= self.decoder(trg, trg_mask, encoded_src, src_mask)
return out
if __name__ == "__main__":
#depends on the tokenizer
src_vocab_size = 50
trg_vocab_size = 50
x= torch.randint(0, 50, (5, 20))
trg= torch.randint(0, 50, (5, 20))
src_pad_idx = 0
trg_pad_idx = 0
model = Transformer(src_vocab_size, trg_vocab_size, src_pad_idx, trg_pad_idx)
#target offset by 1 in addition to masked self attention to avoid info. leak
out = model(x, trg[:, :-1])
print(out.shape)