-
Notifications
You must be signed in to change notification settings - Fork 20
/
generate_dense_embeddings.py
170 lines (127 loc) · 5.97 KB
/
generate_dense_embeddings.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
#!/usr/bin/env python3
# Copyright GC-DPR authors.
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
"""
Command line tool that produces embeddings for a large documents base based on the pretrained ctx & question encoders
Supposed to be used in a 'sharded' way to speed up the process.
"""
import os
import pathlib
import argparse
import csv
import logging
import pickle
from typing import List, Tuple
from tqdm import tqdm
import numpy as np
import torch
from torch import nn
from torch.cuda.amp import autocast
from torch.utils.data import DataLoader, Dataset
from dpr.models import init_biencoder_components
from dpr.options import add_encoder_params, setup_args_gpu, print_args, set_encoder_params_from_state, \
add_tokenizer_params, add_cuda_params
from dpr.utils.data_utils import Tensorizer
from dpr.utils.model_utils import setup_for_distributed_mode, get_model_obj, load_states_from_checkpoint,move_to_device
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if (logger.hasHandlers()):
logger.handlers.clear()
console = logging.StreamHandler()
logger.addHandler(console)
class CtxDataset(Dataset):
def __init__(self, ctx_rows: List[Tuple[object, str, str]], tensorizer: Tensorizer, insert_title: bool = True):
self.rows = ctx_rows
self.tensorizer = tensorizer
self.insert_title = insert_title
def __len__(self):
return len(self.rows)
def __getitem__(self, item):
ctx = self.rows[item]
return self.tensorizer.text_to_tensor(ctx[1], title=ctx[2] if self.insert_title else None)
def no_op_collate(xx: List[object]):
return xx
def gen_ctx_vectors(ctx_rows: List[Tuple[object, str, str]], model: nn.Module, tensorizer: Tensorizer,
insert_title: bool = True, fp16: bool = False) -> List[Tuple[object, np.array]]:
bsz = args.batch_size
total = 0
results = []
dataset = CtxDataset(ctx_rows, tensorizer, insert_title)
loader = DataLoader(
dataset, shuffle=False, num_workers=2, collate_fn=no_op_collate, drop_last=False, batch_size=bsz)
for batch_id, batch_token_tensors in enumerate(tqdm(loader)):
ctx_ids_batch = move_to_device(torch.stack(batch_token_tensors, dim=0),args.device)
ctx_seg_batch = move_to_device(torch.zeros_like(ctx_ids_batch),args.device)
ctx_attn_mask = move_to_device(tensorizer.get_attn_mask(ctx_ids_batch),args.device)
with torch.no_grad():
if fp16:
with autocast():
_, out, _ = model(ctx_ids_batch, ctx_seg_batch, ctx_attn_mask)
else:
_, out, _ = model(ctx_ids_batch, ctx_seg_batch, ctx_attn_mask)
out = out.float().cpu()
batch_start = batch_id*bsz
ctx_ids = [r[0] for r in ctx_rows[batch_start:batch_start + bsz]]
assert len(ctx_ids) == out.size(0)
total += len(ctx_ids)
results.extend([
(ctx_ids[i], out[i].view(-1).numpy())
for i in range(out.size(0))
])
return results
def main(args):
saved_state = load_states_from_checkpoint(args.model_file)
set_encoder_params_from_state(saved_state.encoder_params, args)
print_args(args)
tensorizer, encoder, _ = init_biencoder_components(args.encoder_model_type, args, inference_only=True)
encoder = encoder.ctx_model
encoder, _ = setup_for_distributed_mode(encoder, None, args.device, args.n_gpu,
args.local_rank,
args.fp16,
args.fp16_opt_level)
encoder.eval()
# load weights from the model file
model_to_load = get_model_obj(encoder)
logger.info('Loading saved model state ...')
logger.debug('saved model keys =%s', saved_state.model_dict.keys())
prefix_len = len('ctx_model.')
ctx_state = {key[prefix_len:]: value for (key, value) in saved_state.model_dict.items() if
key.startswith('ctx_model.')}
model_to_load.load_state_dict(ctx_state)
logger.info('reading data from file=%s', args.ctx_file)
rows = []
with open(args.ctx_file) as tsvfile:
reader = csv.reader(tsvfile, delimiter='\t')
# file format: doc_id, doc_text, title
rows.extend([(row[0], row[1], row[2]) for row in reader if row[0] != 'id'])
shard_size = int(len(rows) / args.num_shards)
start_idx = args.shard_id * shard_size
end_idx = start_idx + shard_size
logger.info('Producing encodings for passages range: %d to %d (out of total %d)', start_idx, end_idx, len(rows))
rows = rows[start_idx:end_idx]
data = gen_ctx_vectors(rows, encoder, tensorizer, True, fp16=args.fp16)
file = args.out_file + '_' + str(args.shard_id) + '.pkl'
pathlib.Path(os.path.dirname(file)).mkdir(parents=True, exist_ok=True)
logger.info('Writing results to %s' % file)
with open(file, mode='wb') as f:
pickle.dump(data, f)
logger.info('Total passages processed %d. Written to %s', len(data), file)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
add_encoder_params(parser)
add_tokenizer_params(parser)
add_cuda_params(parser)
parser.add_argument('--ctx_file', type=str, default=None, help='Path to passages set .tsv file')
parser.add_argument('--out_file', required=True, type=str, default=None,
help='output file path to write results to')
parser.add_argument('--shard_id', type=int, default=0, help="Number(0-based) of data shard to process")
parser.add_argument('--num_shards', type=int, default=1, help="Total amount of data shards")
parser.add_argument('--batch_size', type=int, default=32, help="Batch size for the passage encoder forward pass")
args = parser.parse_args()
assert args.model_file, 'Please specify --model_file checkpoint to init model weights'
setup_args_gpu(args)
main(args)