forked from dbiir/UER-py
-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_c3.py
201 lines (153 loc) · 6.92 KB
/
run_c3.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
"""
This script provides an exmaple to wrap UER-py for C3 (a multiple choice dataset).
"""
import argparse
import json
import random
import torch
import torch.nn as nn
from uer.layers import *
from uer.encoders import *
from uer.utils.constants import *
from uer.utils import *
from uer.utils.optimizers import *
from uer.utils.config import load_hyperparam
from uer.utils.seed import set_seed
from uer.model_saver import save_model
from run_classifier import build_optimizer, load_or_initialize_parameters, train_model, batch_loader, evaluate
from uer.opts import finetune_opts
class MultipleChoice(nn.Module):
def __init__(self, args):
super(MultipleChoice, self).__init__()
self.embedding = str2embedding[args.embedding](args, len(args.tokenizer.vocab))
self.encoder = str2encoder[args.encoder](args)
self.dropout = nn.Dropout(args.dropout)
self.output_layer = nn.Linear(args.hidden_size, 1)
def forward(self, src, tgt, seg, soft_tgt=None):
"""
Args:
src: [batch_size x choices_num x seq_length]
tgt: [batch_size]
seg: [batch_size x choices_num x seq_length]
"""
choices_num = src.shape[1]
src = src.view(-1, src.size(-1))
seg = seg.view(-1, seg.size(-1))
# Embedding.
emb = self.embedding(src, seg)
# Encoder.
output = self.encoder(emb, seg)
output = self.dropout(output)
logits = self.output_layer(output[:, 0, :])
reshaped_logits = logits.view(-1, choices_num)
if tgt is not None:
loss = nn.NLLLoss()(nn.LogSoftmax(dim=-1)(reshaped_logits), tgt.view(-1))
return loss, reshaped_logits
else:
return None, reshaped_logits
def read_dataset(args, path):
with open(path, mode="r", encoding="utf-8") as f:
data = json.load(f)
examples = []
for i in range(len(data)):
for j in range(len(data[i][1])):
example = ["\n".join(data[i][0]).lower(), data[i][1][j]["question"].lower()]
for k in range(len(data[i][1][j]["choice"])):
example += [data[i][1][j]["choice"][k].lower()]
for k in range(len(data[i][1][j]["choice"]), args.max_choices_num):
example += ["No Answer"]
example += [data[i][1][j].get("answer", "").lower()]
examples += [example]
dataset = []
for i, example in enumerate(examples):
tgt = 0
for k in range(args.max_choices_num):
if example[2 + k] == example[6]:
tgt = k
dataset.append(([], tgt, []))
for k in range(args.max_choices_num):
src_a = args.tokenizer.convert_tokens_to_ids([CLS_TOKEN] + args.tokenizer.tokenize(example[k + 2]) + [SEP_TOKEN])
src_b = args.tokenizer.convert_tokens_to_ids(args.tokenizer.tokenize(example[1]) + [SEP_TOKEN])
src_c = args.tokenizer.convert_tokens_to_ids(args.tokenizer.tokenize(example[0]) + [SEP_TOKEN])
src = src_a + src_b + src_c
seg = [1] * (len(src_a) + len(src_b)) + [2] * len(src_c)
if len(src) > args.seq_length:
src = src[: args.seq_length]
seg = seg[: args.seq_length]
while len(src) < args.seq_length:
src.append(0)
seg.append(0)
dataset[-1][0].append(src)
dataset[-1][2].append(seg)
return dataset
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument("--max_choices_num", default=4, type=int,
help="The maximum number of cadicate answer, shorter than this will be padded.")
parser.add_argument("--tokenizer", choices=["bert", "char", "space"], default="bert",
help="Specify the tokenizer."
"Original Google BERT uses bert tokenizer on Chinese corpus."
"Char tokenizer segments sentences into characters."
"Space tokenizer segments sentences into words according to space."
)
args = parser.parse_args()
args.labels_num = args.max_choices_num
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build multiple choice model.
model = MultipleChoice(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch, _) in enumerate(batch_loader(batch_size, src, tgt, seg)):
loss = train_model(args, model, optimizer, scheduler, src_batch, tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
model.module.load_state_dict(torch.load(args.output_model_path))
else:
model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))
if __name__ == "__main__":
main()