-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
166 lines (142 loc) · 6.48 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
156
157
158
159
160
161
162
163
164
165
166
from itertools import chain
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from data import (get_word2vec, get_train_data, get_dev_data, get_test_data,
get_word_tag_deprel_lists, load_word_tag_deprel_lists,
save_word_tag_deprel_lists, Encoder, CorpusDataset,
MODEL_LISTS_PATH)
class ParserModel(nn.Module):
def __init__(self, encoder=None, word2vec=None, dropout=0.5,
hidden_size=100):
super().__init__()
if not word2vec:
word2vec = get_word2vec()
if not encoder:
encoder = Encoder(*load_word_tag_deprel_lists(MODEL_LISTS_PATH))
n_tag_ids = len(encoder.id2tag) + 1
n_deprel_ids = len(encoder.id2deprel) + 1
n_word_features = Encoder.n_word_features
n_tag_features = Encoder.n_tag_features
n_deprel_features = Encoder.n_deprel_features
n_classes = encoder.n_classes
embed_size = word2vec.vector_size
self.encoder = encoder
# init embeddings
self.word_embedding = self.init_word_embedding(word2vec)
self.tag_embedding = nn.Embedding(n_tag_ids, embed_size)
self.deprel_embedding = nn.Embedding(n_deprel_ids, embed_size)
# init layers
N = n_word_features + n_tag_features + n_deprel_features
self.linear_stack = nn.Sequential(
nn.Linear(N * embed_size, hidden_size),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(hidden_size, n_classes)
# cross entropy loss already applies softmax to output
)
def init_word_embedding(self, word2vec):
# Add root, unk, null vectors
root = np.random.uniform(-.01, .01, word2vec.vector_size)
unk = np.random.uniform(-.01, .01, word2vec.vector_size)
null = np.random.uniform(-.01, .01, word2vec.vector_size)
weights = torch.tensor(np.vstack((root, word2vec.vectors, unk, null)))
return nn.Embedding.from_pretrained(weights, freeze=False)
def get_concat_embedding(self, X_word, X_tag, X_deprel):
X_word_embed = self.word_embedding(X_word)
X_tag_embed = self.tag_embedding(X_tag)
X_deprel_embed = self.deprel_embedding(X_deprel)
X_concat_embed = torch.cat((X_word_embed, X_tag_embed, X_deprel_embed), 1)
B = X_word.shape[0]
return X_concat_embed.reshape((B, -1)).float()
def forward(self, X_word, X_tag, X_deprel):
X_word, X_tag, X_deprel = X_word.int(), X_tag.int(), X_deprel.int()
X = self.get_concat_embedding(X_word, X_tag, X_deprel)
return self.linear_stack(X)
def predict(self, state):
X_word, X_tag, X_deprel = self.encoder.encode_state(*state)
with torch.no_grad():
logits = self(torch.tensor(X_word).reshape(1, -1),
torch.tensor(X_tag).reshape(1, -1),
torch.tensor(X_deprel).reshape(1, -1))
return self.encoder.decode_target(logits)
raise ValueError("Unable to predict transition.")
def train_epoch(dataloader, model, loss_fn, optimizer):
model.train()
n_batches = len(dataloader)
for i, (X_word, X_tag, X_deprel, y_l, y_ul) in enumerate(dataloader):
# compute prediction and get loss
pred = model(X_word, X_tag, X_deprel)
loss = loss_fn(pred, y_l.argmax(1))
# backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 100 == 0:
print("[Batch {}/{}] Loss: {}".format(i+1, n_batches, loss.item()))
def train(train_dataset, dev_dataset, model, batch_size=2048, n_epochs=10,
lr=0.001):
train_dl = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
dev_dl = DataLoader(dev_dataset, batch_size=batch_size)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr)
for i in range(n_epochs):
print("Begin epoch {}/{}...".format(i+1, n_epochs))
train_epoch(train_dl, model, loss_fn, optimizer)
las, uas = evaluate(dev_dl, model)
print("[Epoch {}/{}] LAS: {}, UAS: {}".format(i+1, n_epochs, las, uas))
def evaluate(dataloader, model):
"""
Evaluate model on dataset and return (LAS, UAS).
LAS = Labelled attachment score, i.e. acc of predicting transition + deprel
UAS = Unlabelled attachment score, i.e. acc of predicting transition
"""
size = len(dataloader.dataset)
model.eval()
las, uas = 0, 0
with torch.no_grad():
for X_word, X_tag, X_deprel, y_l, y_ul in dataloader:
pred = model(X_word, X_tag, X_deprel)
pred_onehot = F.one_hot(pred.argmax(1), y_l.shape[1])
las += (pred_onehot * y_l).sum().item()
uas += (pred_onehot * y_ul).sum().item()
las /= size
uas /= size
return las, uas
def main():
if torch.cuda.is_available():
torch.set_default_tensor_type(torch.cuda.FloatTensor)
print('Running on GPU: {}.'.format(torch.cuda.get_device_name()))
else:
print('Running on CPU.')
print("Loading corpus and pretrained word embedding...")
word2vec = get_word2vec()
train_data = get_train_data()
dev_data = get_dev_data()
test_data = get_test_data()
all = chain(train_data, dev_data, test_data)
word_list, tag_list, deprel_list = get_word_tag_deprel_lists(word2vec, all)
print("Corpus and embedding loaded. {} words, {} tags, {} deprels".format(
len(word_list), len(tag_list), len(deprel_list)))
save_word_tag_deprel_lists(
MODEL_LISTS_PATH, word_list, tag_list, deprel_list)
print("Saved word/tag/deprel lists to {}.".format(MODEL_LISTS_PATH))
encoder = Encoder(word_list, tag_list, deprel_list)
print("Generating datasets from corpus...")
train_dataset = CorpusDataset(train_data, encoder)
dev_dataset = CorpusDataset(dev_data, encoder)
test_dataset = CorpusDataset(test_data, encoder)
print("Datasets generated. Train: n={}, Dev: n={}, Test: n={}".format(
len(train_dataset), len(dev_dataset), len(test_dataset)))
model = ParserModel(encoder, word2vec)
print("Begin training...")
train(train_dataset, dev_dataset, model)
print("Training finished.")
test_las, test_uas = evaluate(DataLoader(test_dataset), model)
print("Test LAS: {}, UAS: {}".format(test_las, test_uas))
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch model state to model.pth")
if __name__ == '__main__':
main()