-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathgenerate_cars.py
190 lines (166 loc) · 7.13 KB
/
generate_cars.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
import functools
import json
import gzip
from datetime import datetime
from collections import defaultdict, Counter
import pandas as pd
import os
from pathlib import Path
import contextlib
import itertools
import more_itertools
from hashlib import md5
import argparse
import ir_datasets
from lz4.frame import decompress, LZ4FrameFile
_logger = ir_datasets.log.easy()
class Sessionizer:
def __init__(self, glove_model='glove-wiki-gigaword-50', threshold=0.5):
import gensim.downloader
self.prev_prefix = None
self.prev_repr = None
self.prev_session_id = None
self.prev_qid = None
self.user_session_counter = Counter()
with _logger.duration(f'loading {glove_model}'):
self.glove = gensim.downloader.load(glove_model)
self.threshold = threshold
def next_session_id(self, qid, query, prefix):
sim, query_repr = self.get_sim_repr(qid, query, prefix)
if sim > self.threshold:
session_id = self.prev_session_id
else:
session_id = f'{prefix}_{self.user_session_counter[prefix]}'
self.user_session_counter[prefix] += 1
self.prev_user_id = prefix
self.prev_repr = query_repr
self.prev_session_id = session_id
self.prev_qid = qid
self.prev_prefix = prefix
return session_id
def get_sim_repr(self, qid, query, prefix):
np = ir_datasets.lazy_libs.numpy()
if self.prev_prefix != prefix:
return 0., None
if self.prev_qid == qid:
return 1., self.prev_repr
vecs = [self.glove.get_vector(t.lower()) for t in query.split() if t.lower() in self.glove]
if vecs:
query_repr = np.stack(vecs).mean(axis=0)
if self.prev_repr is not None:
return np.dot(query_repr, self.prev_repr)/(np.linalg.norm(query_repr)*np.linalg.norm(self.prev_repr)), query_repr
return 0., query_repr
return 0., None
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--out_dir', type=Path, required=True)
parser.add_argument('--dataset', default='aol-ia')
parser.add_argument('--splits', nargs='+', default='train:2006-04-08:2006-05-17:5 dev:2006-05-17:2006-05-24:5 test:2006-05-24:2007-01-01:50'.split())
parser.add_argument('--run', required=True)
parser.add_argument('--trial', action='store_true')
parser.add_argument('--title_only', action='store_true')
args = parser.parse_args()
splits = {}
for split in args.splits:
cols = split.split(':')
if len(cols) == 3:
name, start, end, context = *cols, 50
elif len(cols) == 4:
name, start, end, context = cols
splits[name] = (datetime.fromisoformat(start), datetime.fromisoformat(end), int(context))
dataset = ir_datasets.load(args.dataset)
queries_todo = {}
logs_by_id = defaultdict(list)
qlogs_cls = dataset.qlogs_cls()
field = next(f for f in ['session_id', 'user_id'] if f in qlogs_cls._fields)
field = qlogs_cls._fields.index(field)
qlogs = dataset.qlogs
if args.trial:
qlogs = itertools.islice(qlogs, 100_000)
for qlog in _logger.pbar(qlogs, desc='grouping logs'):
logs_by_id[qlog[field]].append(qlog)
sessionizer = Sessionizer()
sessions = defaultdict(list)
for prefix, logs in _logger.pbar(logs_by_id.items(), desc='building sessions'):
logs = sorted(logs, key=lambda x: x.time)
for log in logs:
session_id = sessionizer.next_session_id(log.query_id, log.query, prefix)
sessions[session_id].append(log)
del logs_by_id
with _logger.duration('filtering sessions'):
sessions = {sid: logs for sid, logs in sessions.items() if len(set(l.query_id for l in logs)) > 1}
if not args.out_dir.exists():
args.out_dir.mkdir(exist_ok=True, parents=True)
ds = dataset.docs_store()
runs_by_did = defaultdict(dict)
with _logger.duration(f'reading run {args.run}'), gzip.open(args.run, 'rb') as fin:
for line in fin:
qid, _, did, rank, _, _ = line.decode().strip().split()
runs_by_did[str(qid)][str(did)] = int(rank)
with contextlib.ExitStack() as stack:
split2file = {split: stack.enter_context(open(args.out_dir/f'{split}.json', 'wt')) for split in splits}
for session_id, logs in _logger.pbar(sorted(sessions.items()), desc='creating records'):
# make sure it appears in a split
split = [s for s, (start, end, _) in splits.items() if start <= logs[0].time < end]
if len(split) == 0:
continue
split = split[0]
context = splits[split][2]
record = {
'session_id': session_id,
'query': []
}
for log in logs:
# matching_query = [q for q in record['query'] if q['text'] == log.query]
clicked_dids = {i.doc_id for i in log.items if i.clicked}
if args.run:
dids = get_dids_from_run(runs_by_did[log.query_id], clicked_dids, context)
else:
dids = [i.doc_id for i in log.items]
docs = ds.get_many(dids)
query = {
'id': str(len(record['query'])),
'text': log.query,
'tokens': log.query.split(),
'candidates': []
}
record['query'].append(query)
for doc_id in dids:
if doc_id not in docs:
continue
doc = docs[doc_id]
query['candidates'].append({
'id': doc_id,
'title': doc.title,
'content': '' if args.title_only else doc.text,
'url': doc.url,
'label': doc_id in clicked_dids,
})
# ensure each query has at least one clicked and one non-clicked document
record['query'] = [q for q in record['query'] if any(c['label'] for c in q['candidates']) and any(not c['label'] for c in q['candidates'])]
# make sure the session has more than 1 query
if len(record['query']) > 1:
split2file[split].write(json.dumps(record) + '\n')
def get_dids_from_run(run_by_did, clicked_dids, context):
run_by_rank = {rank: did for did, rank in run_by_did.items()}
target_dids = set()
for did in clicked_dids:
if did in run_by_did:
rank = run_by_did[did]
start, stop = rank - context // 2, rank + context // 2
if stop - start != context: # if odd
stop += 1
while start not in run_by_rank:
start += 1
stop += 1
while stop - 1 not in run_by_rank:
start -= 1
stop -= 1
while start not in run_by_rank:
start += 1
for rank in range(start, stop):
if rank in run_by_rank:
target_dids.add(run_by_rank[rank])
return sorted(target_dids)
if __name__ == "__main__":
main()