-
Notifications
You must be signed in to change notification settings - Fork 0
/
build_greedy_subsets.py
138 lines (113 loc) · 6.29 KB
/
build_greedy_subsets.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
import os
import argparse
import pandas as pd
import numpy as np
from datasets import load_dataset
import logging
from tqdm import tqdm
from itertools import chain
import json
import collections
tqdm.pandas()
def read_dir_to_np(dir_results, df_oie):
'''
:param args:
:return: np array with 5 column : [scores, index_sum, index_doc, topic_id, story_id]
'''
npys = os.listdir(dir_results)
list_index_doc = sorted([file for file in npys if 'index_doc' in file])
list_index_sum = sorted([file for file in npys if 'index_sum' in file])
list_topic_id = sorted([file for file in npys if 'topic_id' in file])
list_story_id = sorted([file for file in npys if 'story_id' in file])
list_scores = sorted([file for file in npys if 'scores' in file])
scores = np.concatenate(
[np.load(os.path.join(dir_results, score_npy)) for score_npy in list_scores])
index_sum = np.concatenate(
[np.load(os.path.join(dir_results, index_sum_npy)) for index_sum_npy in list_index_sum])
index_doc_npy = np.concatenate(
[np.load(os.path.join(dir_results, index_doc_npy)) for index_doc_npy in list_index_doc])
topic_id = np.concatenate(
[np.load(os.path.join(dir_results, topic_npy)) for topic_npy in list_topic_id])
# stories_id = np.concatenate(
# [np.load(os.path.join(args.dir_results, story_npy)) for story_npy in list_story_id])
stories_id = np.array(df_oie.iloc[index_doc_npy]['story_id'])
result = np.column_stack((scores, index_sum, index_doc_npy, topic_id, stories_id))
logger.info(f'number of pairs: {result.shape}')
return result
def flatten(lst):
return set(chain.from_iterable(lst))
if __name__=='__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', type=str)
parser.add_argument('--threshold', type=float, default=0.5)
args = parser.parse_args()
root_logger = logging.getLogger()
logger = root_logger.getChild(__name__)
handler = logging.StreamHandler()
formatter = logging.Formatter(fmt="%(asctime)s %(name)s.%(lineno)d %(levelname)s : %(message)s", datefmt="%H:%M:%S")
handler.setFormatter(formatter)
root_logger.addHandler(handler)
root_logger.setLevel(logging.INFO)
with open(os.path.join(args.data_dir, "num_of_pairs.txt"), "r") as f:
num_of_pairs = int(f.readlines()[0])
logger.info("Loading data..")
df_oie = pd.read_csv(os.path.join(args.data_dir, "df_oie.csv"))
if "df_pairs.parquet" in os.listdir(args.data_dir):
pairs = pd.read_parquet(os.path.join(args.data_dir, "df_pairs.parquet"))
else:
scores = read_dir_to_np(os.path.join(args.data_dir, "result_npy"), df_oie)
pairs = pd.DataFrame(scores, columns=["score", "index_sum", "index_doc", "topic_id", "story_id"])
pairs[['index_sum', 'index_doc', 'topic_id', 'story_id']] = pairs[['index_sum', 'index_doc', 'topic_id', 'story_id']].astype(int)
pairs.to_parquet(os.path.join(args.data_dir, "df_pairs.parquet"))
pairs['aligned'] = pairs['score'] > args.threshold
logger.info(f"num of computed pairs: {len(pairs)}")
logger.info(f"num of original pairs: {num_of_pairs}")
assert len(pairs) == num_of_pairs
if "df_alignments.parquet" in os.listdir(args.data_dir):
alignments = pd.read_parquet(os.path.join(args.data_dir, "df_alignments.parquet"))
else:
logger.info("create dataframe of alignments")
data = []
for (topic, story_id), row in tqdm(pairs.groupby(['topic_id', 'story_id'])):
oie_sum_group = row.groupby('index_sum')[['aligned']].apply(sum) > 0
data.append((topic, story_id, oie_sum_group[oie_sum_group['aligned'] == True].index.values.tolist()))
alignments = pd.DataFrame(data, columns=['topic', 'story', 'alignment'])
alignments.to_parquet(os.path.join(args.data_dir, "df_alignments.parquet"))
logger.info("create topic df..")
all_oie = pd.DataFrame(pairs.groupby('topic_id')['index_sum'].progress_apply(set))
total_coverage = pd.DataFrame(alignments.groupby('topic')['alignment'].progress_apply(flatten))
topics = pd.concat([all_oie, total_coverage], axis=1)
topics.rename(columns={'index_sum': 'all_oie'}, inplace=True)
topics['non_aligned'] = topics['all_oie'] - topics['alignment']
topics['total_coverage'] = topics['alignment'].apply(len) / topics['all_oie'].apply(len)
topics['hallucinated'] = topics['non_aligned'].apply(len) / topics['all_oie'].apply(len)
alignments['length'] = alignments['alignment'].apply(len)
max_idx_per_topic = alignments.groupby('topic')[['length']].idxmax()
topics['dominant'] = alignments.loc[max_idx_per_topic['length']]['alignment'].apply(set).tolist()
topics["relative_dominant_score"] = topics["dominant"].apply(len) / topics["alignment"].apply(len)
topics["absolute_dominant_score"] = topics["dominant"].apply(len) / topics["all_oie"].apply(len)
logger.info("Create df of all alignments")
all_coverage_data = collections.defaultdict(list)
data_stories = collections.defaultdict(list)
alignments["alignment"] = alignments["alignment"].apply(set)
for topic_id, group in tqdm(alignments.groupby("topic")[["story", "alignment"]]):
topic_indices = topics.loc[topic_id]["all_oie"].copy()
total_aligned = set()
while len(group) > 0:
group["diff"] = topic_indices - group["alignment"]
group["diff_len"] = group["diff"].apply(len)
idx = group["diff_len"].argmin()
row_id = group.index[idx]
to_add = group.loc[row_id]["alignment"]
total_aligned.update(to_add)
all_coverage_data[topic_id].append(total_aligned.copy())
data_stories[topic_id].append(group.loc[row_id]["story"])
group.drop(row_id, inplace=True)
topic_indices -= total_aligned
max_length = max(len(v) for v in all_coverage_data.values())
for topic_id, coverage in all_coverage_data.items():
while len(coverage) < max_length:
all_coverage_data[topic_id].append(coverage[-1])
df_coverage = pd.DataFrame(all_coverage_data).T
df_coverage = df_coverage.merge(topics[["all_oie", "alignment"]], how="left", left_index=True, right_index=True)
df_coverage.to_json(os.path.join(args.data_dir, "source_coverage.jsonl"), orient="records", lines=True)