forked from minggg/squad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
args.py
244 lines (219 loc) · 9.69 KB
/
args.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
"""Command-line arguments for setup.py, train.py, test.py.
Author:
Chris Chute (chute@stanford.edu)
"""
import argparse
def get_setup_args():
"""Get arguments needed in setup.py."""
parser = argparse.ArgumentParser('Download and pre-process SQuAD')
add_common_args(parser)
parser.add_argument('--train_url',
type=str,
default='https://github.com/chrischute/squad/data/train-v2.0.json')
parser.add_argument('--dev_url',
type=str,
default='https://github.com/chrischute/squad/data/dev-v2.0.json')
parser.add_argument('--test_url',
type=str,
default='https://github.com/chrischute/squad/data/test-v2.0.json')
parser.add_argument('--glove_url',
type=str,
default='http://nlp.stanford.edu/data/glove.840B.300d.zip')
parser.add_argument('--dev_meta_file',
type=str,
default='./data/dev_meta.json')
parser.add_argument('--test_meta_file',
type=str,
default='./data/test_meta.json')
parser.add_argument('--word2idx_file',
type=str,
default='./data/word2idx.json')
parser.add_argument('--char2idx_file',
type=str,
default='./data/char2idx.json')
parser.add_argument('--answer_file',
type=str,
default='./data/answer.json')
parser.add_argument('--para_limit',
type=int,
default=400,
help='Max number of words in a paragraph')
parser.add_argument('--ques_limit',
type=int,
default=50,
help='Max number of words to keep from a question')
parser.add_argument('--test_para_limit',
type=int,
default=1000,
help='Max number of words in a paragraph at test time')
parser.add_argument('--test_ques_limit',
type=int,
default=100,
help='Max number of words in a question at test time')
parser.add_argument('--char_dim',
type=int,
default=64,
help='Size of char vectors (char-level embeddings)')
parser.add_argument('--glove_dim',
type=int,
default=300,
help='Size of GloVe word vectors to use')
parser.add_argument('--glove_num_vecs',
type=int,
default=2196017,
help='Number of GloVe vectors')
parser.add_argument('--ans_limit',
type=int,
default=30,
help='Max number of words in a training example answer')
parser.add_argument('--char_limit',
type=int,
default=16,
help='Max number of chars to keep from a word')
parser.add_argument('--include_test_examples',
type=lambda s: s.lower().startswith('t'),
default=True,
help='Process examples from the test set')
args = parser.parse_args()
return args
def get_train_args():
"""Get arguments needed in train.py."""
parser = argparse.ArgumentParser('Train a model on SQuAD')
add_common_args(parser)
add_train_test_args(parser)
parser.add_argument('--eval_steps',
type=int,
default=50000,
help='Number of steps between successive evaluations.')
parser.add_argument('--lr',
type=float,
default=0.5,
help='Learning rate.')
parser.add_argument('--l2_wd',
type=float,
default=0,
help='L2 weight decay.')
parser.add_argument('--num_epochs',
type=int,
default=30,
help='Number of epochs for which to train. Negative means forever.')
parser.add_argument('--drop_prob',
type=float,
default=0.2,
help='Probability of zeroing an activation in dropout layers.')
parser.add_argument('--metric_name',
type=str,
default='F1',
choices=('NLL', 'EM', 'F1'),
help='Name of dev metric to determine best checkpoint.')
parser.add_argument('--max_checkpoints',
type=int,
default=5,
help='Maximum number of checkpoints to keep on disk.')
parser.add_argument('--max_grad_norm',
type=float,
default=5.0,
help='Maximum gradient norm for gradient clipping.')
parser.add_argument('--seed',
type=int,
default=224,
help='Random seed for reproducibility.')
parser.add_argument('--ema_decay',
type=float,
default=0.999,
help='Decay rate for exponential moving average of parameters.')
args = parser.parse_args()
if args.metric_name == 'NLL':
# Best checkpoint is the one that minimizes negative log-likelihood
args.maximize_metric = False
elif args.metric_name in ('EM', 'F1'):
# Best checkpoint is the one that maximizes EM or F1
args.maximize_metric = True
else:
raise ValueError(f'Unrecognized metric name: "{args.metric_name}"')
return args
def get_test_args():
"""Get arguments needed in test.py."""
parser = argparse.ArgumentParser('Test a trained model on SQuAD')
add_common_args(parser)
add_train_test_args(parser)
parser.add_argument('--split',
type=str,
default='dev',
choices=('train', 'dev', 'test'),
help='Split to use for testing.')
parser.add_argument('--sub_file',
type=str,
default='submission.csv',
help='Name for submission file.')
# Require load_path for test.py
args = parser.parse_args()
if not args.load_path:
raise argparse.ArgumentError('Missing required argument --load_path')
return args
def add_common_args(parser):
"""Add arguments common to all 3 scripts: setup.py, train.py, test.py"""
parser.add_argument('--train_record_file',
type=str,
default='./data/train.npz')
parser.add_argument('--dev_record_file',
type=str,
default='./data/dev.npz')
parser.add_argument('--test_record_file',
type=str,
default='./data/test.npz')
parser.add_argument('--word_emb_file',
type=str,
default='./data/word_emb.json')
parser.add_argument('--char_emb_file',
type=str,
default='./data/char_emb.json')
parser.add_argument('--train_eval_file',
type=str,
default='./data/train_eval.json')
parser.add_argument('--dev_eval_file',
type=str,
default='./data/dev_eval.json')
parser.add_argument('--test_eval_file',
type=str,
default='./data/test_eval.json')
def add_train_test_args(parser):
"""Add arguments common to train.py and test.py"""
parser.add_argument('--name',
'-n',
type=str,
required=True,
help='Name to identify training or test run.')
parser.add_argument('--max_ans_len',
type=int,
default=15,
help='Maximum length of a predicted answer.')
parser.add_argument('--num_workers',
type=int,
default=4,
help='Number of sub-processes to use per data loader.')
parser.add_argument('--save_dir',
type=str,
default='./save/',
help='Base directory for saving information.')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch size per GPU. Scales automatically when \
multiple GPUs are available.')
parser.add_argument('--use_squad_v2',
type=lambda s: s.lower().startswith('t'),
default=True,
help='Whether to use SQuAD 2.0 (unanswerable) questions.')
parser.add_argument('--hidden_size',
type=int,
default=100,
help='Number of features in encoder hidden layers.')
parser.add_argument('--num_visuals',
type=int,
default=10,
help='Number of examples to visualize in TensorBoard.')
parser.add_argument('--load_path',
type=str,
default=None,
help='Path to load as a model checkpoint.')