-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathlm_setup.py
193 lines (152 loc) · 6.76 KB
/
lm_setup.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
import gc
import multiprocessing
import glob
import re
import logging
import os
import numpy as np
from collections import namedtuple
from base.utils import print_title, os_exec, verify_dir_exists
from lm_model import WrappedLm1bModel, encode_paragraphs, LM1B_MODEL_HIDDEN_DIM
LM1B_URL_PREFIX = 'http://download.tensorflow.org/models/LM_LSTM_CNN/'
LM1B_DATA_DIR = 'data/lm1b/'
LM1B_VOCAB_FILENAME = 'vocab-2016-09-10.txt'
LM1B_VOCAB_URL = LM1B_URL_PREFIX + LM1B_VOCAB_FILENAME
LM1B_VOCAB_PATH = LM1B_DATA_DIR + LM1B_VOCAB_FILENAME
LM1B_CKPT_BASE_URL = LM1B_URL_PREFIX + 'all_shards-2016-09-10/ckpt-base'
LM1B_CKPT_CHAR_EMB_URL = LM1B_URL_PREFIX + 'all_shards-2016-09-10/ckpt-char-embedding'
LM1B_CKPT_LSTM_URL = LM1B_URL_PREFIX + 'all_shards-2016-09-10/ckpt-lstm'
LM1B_GRAPH_DEF_FILENAME = 'altered-graph-2017-10-06.pbtxt'
LM1B_GRAPH_DEF_PATH = LM1B_DATA_DIR + LM1B_GRAPH_DEF_FILENAME
LmDataShardConfig = namedtuple('LmDataShardConfig', [
'dataset', # 'train' or 'dev'
'sequences', # 'contexts' or 'questions'
'payload', # 'EMB' or 'L1' or 'L2'
'num_shards', # int
'shard' # int
])
LmDatasetEncodings = namedtuple('LmDatasetEncodings', [
'h_starts',
'lens',
'hs'
])
LmData = namedtuple('LmData', [
'trn_ctxs', # LmDatasetEncodings
'trn_qtns', # LmDatasetEncodings
'dev_ctxs', # LmDatasetEncodings
'dev_qtns', # LmDatasetEncodings
'tst_ctxs', # LmDatasetEncodings
'tst_qtns' # LmDatasetEncodings
])
def _get_shard_seq_idxs(num_examples, num_shards, current_shard):
assert 0 < current_shard and current_shard <= num_shards
shard_size = num_examples // num_shards
if num_examples % num_shards:
shard_size += 1
first_seq_idx = (current_shard - 1) * shard_size
after_last_seq_idx = min(first_seq_idx + shard_size, num_examples)
return first_seq_idx, after_last_seq_idx
def _write_lm_data_shard_process(lm_data_shard_cfg, device, seqs_originals, seqs_sent_lens):
if device == 'cpu':
cuda_visible_devices = ''
else:
assert device.startswith('gpu')
cuda_visible_devices = device[3:]
os.environ['CUDA_VISIBLE_DEVICES'] = cuda_visible_devices
lm_model = WrappedLm1bModel(
LM1B_GRAPH_DEF_PATH,
LM1B_DATA_DIR + 'ckpt-*',
LM1B_VOCAB_PATH)
hs, lens = encode_paragraphs(lm_model, lm_data_shard_cfg, seqs_originals, seqs_sent_lens)
shard_path = _get_shard_path(lm_data_shard_cfg)
logging.getLogger().info('Saving to {:s} ...'.format(shard_path))
verify_dir_exists(shard_path)
np.savez(shard_path, hs=hs, lens=lens)
def _get_shard_path_prefix(payload, dataset, sequences):
return LM1B_DATA_DIR + 'lm_encodings_{:s}_{:s}_{:s}'.format(
payload, dataset, sequences)
def _get_shard_path(lm_data_shard_cfg):
path_prefix = _get_shard_path_prefix(
lm_data_shard_cfg.payload,
lm_data_shard_cfg.dataset,
lm_data_shard_cfg.sequences)
shard_path = path_prefix + '_{:03d}_of_{:03d}.npz'.format(
lm_data_shard_cfg.shard,
lm_data_shard_cfg.num_shards)
return shard_path
def _load_lm_dataset_encodings(dataset, sequences, payload):
logger = logging.getLogger()
path_prefix = _get_shard_path_prefix(payload, dataset, sequences)
paths = glob.glob(path_prefix + '*')
assert paths, 'Did not find expected LM data shards ' + path_prefix
last_path = sorted(paths)[-1]
p = re.compile('_of_([0-9]*?)\.npz$')
num_shards = int(p.search(last_path).group(1))
shard_paths = []
lens_list = []
for shard in range(1, num_shards+1):
shard_cfg = LmDataShardConfig(
dataset, sequences, payload, num_shards, shard)
shard_path = _get_shard_path(shard_cfg)
shard_paths.append(shard_path)
shard_lens = np.load(shard_path)['lens']
logger.info('Loaded shard lengths from {:<60s}: {:s} {:s}'.format(shard_path, shard_lens.dtype, shard_lens.shape))
lens_list.append(shard_lens)
lens = np.concatenate(lens_list, axis=0)
h_starts = np.insert(lens[:-1], 0, [0]).cumsum(dtype=np.int32)
num_vectors = lens.sum()
hs = np.zeros((num_vectors, LM1B_MODEL_HIDDEN_DIM), dtype=np.float32)
pos = 0
for shard_path in shard_paths:
shard_hs = np.load(shard_path)['hs']
hs[pos:pos+shard_hs.shape[0], :] = shard_hs
pos += shard_hs.shape[0]
logger.info('Loaded shard hs from {:<60s}: {:s} {:s}'.format(shard_path, shard_hs.dtype, shard_hs.shape))
del shard_hs
gc.collect()
assert pos == num_vectors
dataset_str = '{:s} {:s} {:s}'.format(payload, dataset, sequences)
logger.info('Joined shard lengths for {:<60s}: {:s} {:s}'.format(
dataset_str, str(lens.dtype), str(lens.shape)))
logger.info('Joined shard hs for {:<60s}: {:s} {:s}'.format(
dataset_str, str(hs.dtype), str(hs.shape)))
gc.collect()
return LmDatasetEncodings(h_starts, lens, hs)
##########################################
# Interface
##########################################
def download_lm_model():
print_title('Downloading LM1B model')
for url in [LM1B_CKPT_BASE_URL, LM1B_CKPT_CHAR_EMB_URL, LM1B_CKPT_LSTM_URL, LM1B_VOCAB_URL]:
wget_cmd = 'wget {} -P {}'.format(url, LM1B_DATA_DIR)
os_exec(wget_cmd)
def write_lm_data_shard(data, lm_data_shard_cfg, device):
logger = logging.getLogger()
logger.info('Writing LM data for lm_data_shard_cfg {:}'.format(lm_data_shard_cfg))
ds_mapping = {'train': data.trn.tabular, 'dev': data.dev.tabular}
tab_ds = ds_mapping[lm_data_shard_cfg.dataset]
seqs_mapping = {'contexts': tab_ds.ctxs, 'questions': tab_ds.qtns}
seqs = seqs_mapping[lm_data_shard_cfg.sequences]
seqs_originals = [seq.tokenized.originals for seq in seqs]
seqs_sent_lens = [seq.tokenized.sent_lens for seq in seqs]
for originals, sent_lens in zip(seqs_originals, seqs_sent_lens):
assert sum(sent_lens) == len(originals)
first_seq_idx, after_last_seq_idx = _get_shard_seq_idxs(
len(seqs_originals), lm_data_shard_cfg.num_shards, lm_data_shard_cfg.shard)
logger.info('There are a total of {:d} {:s} in {:s} dataset, shard to be written covers indices {:08d}-{:08d}'.format(
len(seqs_originals), lm_data_shard_cfg.sequences, lm_data_shard_cfg.dataset, first_seq_idx, after_last_seq_idx-1))
seqs_originals = seqs_originals[first_seq_idx:after_last_seq_idx]
seqs_sent_lens = seqs_sent_lens[first_seq_idx:after_last_seq_idx]
job = multiprocessing.Process(
target = _write_lm_data_shard_process,
args = (lm_data_shard_cfg, device, seqs_originals, seqs_sent_lens))
job.start()
job.join()
logger.info('Done writing LM data for lm_data_shard_cfg {:}'.format(lm_data_shard_cfg))
def get_lm_data(payload):
trn_ctxs = _load_lm_dataset_encodings('train', 'contexts', payload)
trn_qtns = _load_lm_dataset_encodings('train', 'questions', payload)
dev_ctxs = _load_lm_dataset_encodings('dev', 'contexts', payload)
dev_qtns = _load_lm_dataset_encodings('dev', 'questions', payload)
return LmData(
trn_ctxs, trn_qtns, dev_ctxs, dev_qtns, tst_ctxs=None, tst_qtns=None)