-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
156 lines (131 loc) · 6.03 KB
/
model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm
from dataset import Dataset, EOS, SOS
import random
import math
import argparse
class Encoder(nn.Module):
def __init__(self, dataset, device, hidden_size = 64, char_size = 64):
super(Encoder, self).__init__()
self.dataset = dataset
self.device = device
self.char_embeddings = nn.Embedding(len(dataset.char_to_ix), char_size)
self.hidden_size = hidden_size // 2
self.lstm = nn.LSTM(input_size=char_size,
hidden_size=hidden_size // 2,
num_layers=1,
batch_first = True,
bidirectional=True,
dropout=0.0)
def forward(self, x):
x = self.char_embeddings(x)
outs, (h, c) = self.lstm(x)
#collapse forward and backwards dimensions to a single hidden state
final_hidden = torch.cat((outs[:, -1, :self.hidden_size], outs[:, 0, self.hidden_size:]), dim=1)
return outs, final_hidden.unsqueeze(dim=0)
class Decoder(nn.Module):
def __init__(self, dataset, device, hidden_size):
super(Decoder, self).__init__()
self.dataset = dataset
self.device = device
self.out_embeddings = nn.Embedding(len(dataset.out_to_ix), hidden_size)
#this should technically be an LSTMCell
#since it's only used one timestep at a time
self.decoder_lstm = nn.LSTM(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
batch_first=True,
dropout=0.0)
self.linear = nn.Linear(hidden_size, len(dataset.out_to_ix))
print("Decoder")
print(self)
def forward(self, x, hidden, source_hiddens):
hidden_state, cell_state = self.decode(x, hidden)
out_scores = self.linear(hidden_state)
return (hidden_state, cell_state), out_scores
def decode(self, x, hidden):
x = self.out_embeddings(x)
_, (hidden_state, cell_state) = self.decoder_lstm(x, hidden)
return (hidden_state, cell_state)
class DecoderAttention(nn.Module):
def __init__(self, dataset, device, hidden_size):
super(DecoderAttention, self).__init__()
self.dataset = dataset
self.device = device
self.out_embeddings = nn.Embedding(len(dataset.out_to_ix), hidden_size)
#this should technically be an LSTMCell
#since it's only used one timestep at a time
self.decoder_lstm = nn.LSTM(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
batch_first=True,
dropout=0.0)
self.W_C = nn.Linear(hidden_size * 2, hidden_size)
self.W_S = nn.Linear(hidden_size, len(dataset.out_to_ix))
print("Decoder")
print(self)
def forward(self, x, hidden, source_hiddens):
(hidden_state, cell_state), context_vectors = self.decode(x, hidden, source_hiddens)
contexts = torch.cat([context_vectors, hidden_state], dim=2)
out_scores = self.W_S(F.tanh(self.W_C(contexts)))
return (hidden_state, cell_state), out_scores
def decode(self, x, hidden, source_hiddens):
x = self.out_embeddings(x)
_, (hidden_state, cell_state) = self.decoder_lstm(x, hidden)
# hidden_state: (1, batch_size, hidden)
# source_hiddens: (batch_size, seq_len, hidden)
scores = F.softmax(torch.einsum("dbh,bsh->bs", (hidden_state, source_hiddens)))
# scores: (batch_size, seq_len)
context_vectors = torch.einsum("bsh,bs->bh", (source_hiddens, scores)).unsqueeze(dim=0)
# context_vectors: (batch_size, hidden)
return (hidden_state, cell_state), context_vectors
class Seq2Seq(nn.Module):
def __init__(self, dataset, device, hidden_size = 64, character_size = 64, decoder_type='base'):
super(Seq2Seq, self).__init__()
self.dataset = dataset
self.device = device
self.encoder = Encoder(dataset, device, hidden_size, character_size)
self.decoder_type = decoder_type
if decoder_type == 'base':
self.decoder = Decoder(dataset, device, hidden_size)
elif decoder_type == 'attn':
self.decoder = DecoderAttention(dataset, device, hidden_size)
self.to(device)
def _decoder_step(self, x, hidden_inputs, source_hiddens):
(new_hidden), scores = self.decoder(x, hidden_inputs, source_hiddens)
scores = scores.squeeze(dim=0)
#argmax
_, x = torch.topk(scores, 1)
return new_hidden, scores, x
def train_step(self, word, target, criterion):
source_hiddens, final_hidden = self.encoder(word)
current_hidden = (final_hidden, final_hidden)
x = target[:, 0:1]
batch_loss = 0
for x_ix in range(len(target[0]) - 1):
#decode a timestep
current_hidden, scores, x = self._decoder_step(x, current_hidden, source_hiddens)
x_target = target[:, x_ix+1]
batch_loss += criterion(scores, x_target)
return batch_loss
def free_run(self, word, return_attentions=False):
source_hiddens, final_hidden = self.encoder(word)
current_hidden = (final_hidden, final_hidden)
x = torch.tensor([[self.dataset.out_to_ix[SOS]]], device=self.device)
chars = []
out_char = None
EOS_ix = self.dataset.out_to_ix[EOS]
if self.decoder_type == 'attn':
attentions = []
while (out_char != EOS_ix) and len(chars) < 32:
#decode a timestep
current_hidden, scores, x = self._decoder_step(x, current_hidden, source_hiddens)
if self.decoder_type == 'attn':
attentions.append(self.decoder.attention_cache)
out_char = x[0][0].item()
chars.append(out_char)
if return_attentions and self.decoder_type == 'attn':
return chars, attentions
return chars