-
Notifications
You must be signed in to change notification settings - Fork 1
/
translator.py
171 lines (123 loc) · 5.6 KB
/
translator.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
import random
import torch
import torch.nn as nn
import torch.optim as optim
import torchtext
import spacy
import random
def translate_string(model, english, german, text, max_size=40):
spacy_ger = spacy.load("de_core_news_sm")
tokens = [tok.text.lower() for tok in spacy_ger(text)]
tokens.insert(0, german.init_token)
tokens.append(german.eos_token)
text_to_index = [german.vocab.stoi[t] for t in tokens]
text_vector = torch.LongTensor(text_to_index).unsqueeze(1)
with torch.no_grad():
hidden, cell = model.encoder(text_vector)
outputs = [english.vocab.stoi["<sos>"]]
for _ in range(max_size):
previous_word = torch.LongTensor([outputs[-1]])
with torch.no_grad():
output, hidden, cell = model.decoder(previous_word, hidden, cell)
best_guess = output.argmax(1).item()
outputs.append(best_guess)
# Model predicts it's the end of the sentence
if output.argmax(1).item() == english.vocab.stoi["<eos>"]:
break
translated_sentence = [english.vocab.itos[idx] for idx in outputs]
# remove start token
return translated_sentence[1:]
eng_text = torchtext.data.Field(sequential=True, use_vocab=True, init_token="<sos>", eos_token="<eos>",
tokenize="spacy", tokenizer_language='en_core_web_sm', lower=True)
ger_text = torchtext.data.Field(sequential=True, use_vocab=True, init_token="<sos>", eos_token="<eos>",
tokenize="spacy", tokenizer_language='de_core_news_sm', lower=True)
train, val, test = torchtext.datasets.Multi30k.splits(exts=(".de", ".en"), fields=(ger_text, eng_text), root=".data")
batch_size = 64
vocab_size = 10000
eng_text.build_vocab(train, max_size=vocab_size)
ger_text.build_vocab(train, max_size=vocab_size)
print(len(eng_text.vocab), len(ger_text.vocab))
train_load = torchtext.data.BucketIterator(dataset=train, batch_size=batch_size,
sort_within_batch=True, sort_key=lambda x: len(x.src))
for i in train_load:
print(i)
break
print(len(ger_text.vocab), len(eng_text.vocab))
class Encoder(nn.Module):
def __init__(self, input_dim, embed_dim, hidden_dim, num_dropout):
super(Encoder, self).__init__()
self.drop = nn.Dropout(num_dropout)
self.embed_layer = nn.Embedding(input_dim, embed_dim)
self.lstm = nn.LSTM(embed_dim, hidden_dim, num_layers=2, dropout=num_dropout)
def forward(self, x):
x = self.drop(self.embed_layer(x))
out, (h, c) = self.lstm(x)
return h, c
class Decoder(nn.Module):
def __init__(self, input_dim, embed_dim, hidden_dim, output_dim, num_dropout):
super(Decoder, self).__init__()
self.drop = nn.Dropout(num_dropout)
self.embed_layer = nn.Embedding(input_dim, embed_dim)
self.lstm = nn.LSTM(embed_dim, hidden_dim, num_layers=2, dropout=num_dropout)
self.fc = nn.Linear(hidden_dim, output_dim)
def forward(self, x, hidden, cell):
x = x.unsqueeze(0)
x = self.drop(self.embed_layer(x))
out, (hidden, cell) = self.lstm(x)
output = self.fc(out)
output = output.squeeze(0)
# print(output.shape)
return output, hidden, cell
class Network(nn.Module):
def __init__(self, encoder, decoder):
super(Network, self).__init__()
self.encoder = encoder
self.decoder = decoder
def forward(self, source, target, ratio=0.5):
target_len = target.shape[0]
batch_size = source.shape[1]
target_vocab_size = len(eng_text.vocab)
hidden, cell = self.encoder(source)
outputs = torch.zeros(target_len, batch_size, target_vocab_size)
x = target[0]
for t in range(1, target_len):
out, hidden, cell = self.decoder(x, hidden, cell)
outputs[t] = out
pred = out.argmax(1)
x = target[t] if random.random() < ratio else pred
# print(outputs.shape)
return outputs
encoder = Encoder(input_dim=len(ger_text.vocab), embed_dim=512, hidden_dim=1024, num_dropout=0.5)
decoder = Decoder(input_dim=len(eng_text.vocab), embed_dim=512, hidden_dim=1024,
num_dropout=0.5, output_dim=len(eng_text.vocab))
model = Network(encoder=encoder, decoder=decoder)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=3e-4)
sentence = "ein boot mit mehreren männern darauf wird von einem großen pferdegespann ans ufer gezogen."
for epoch in range(20):
epoch_loss = 0
model.eval()
translated_sentence = translate_string(
model=model, text=sentence, german=ger_text, english=eng_text
)
print(f"Translated example sentence: \n {translated_sentence}")
model.train()
for i, batch in enumerate(train_load):
if i+1 >= 50 and (i+1) % 50 == 0:
print(f"{i+1}/{len(train_load)}")
score = model(batch.src, batch.trg)
# print("score shape", score.shape)
score = score[1:].reshape(-1, score.shape[2])
batch.trg = batch.trg[1:].reshape(-1)
loss = criterion(score, batch.trg)
epoch_loss += loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f"Loss at {epoch+1} is: {epoch_loss}")
model.eval()
translated_sentence = translate_string(
model=model, text=sentence, german=ger_text, english=eng_text
)
print(f"Translated example sentence: \n {translated_sentence}")
model.train()