-
Notifications
You must be signed in to change notification settings - Fork 0
/
benchmark.py
172 lines (155 loc) · 7.16 KB
/
benchmark.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
import argparse
import os
import warnings
from pathlib import Path
import pandas as pd
from gensim.models import KeyedVectors
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import settings
from eval_clm import evaluate_generations
from run_clm import gen_completions
from util import set_random_seed
def generate_completions(model_name_or_path,
dataset_paths,
completions_path,
bias_types,
batch_size=32,
over_write_output=True):
"""
Generates completions for each model, dataset, and bias type.
"""
# Ensure the directory exists
completions_path = Path(completions_path)
completions_path.mkdir(parents=True, exist_ok=True)
for model_ in model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = 'left'
model = AutoModelForCausalLM.from_pretrained(model_, device_map='auto', torch_dtype=torch.float16)
print("loading model")
for data_path in dataset_paths:
data_path = Path(data_path)
for bias_type in bias_types:
output_filename = f"{model_.replace('/', '-')}_{bias_type}_{data_path.stem}.json"
output_path = completions_path / output_filename
if not over_write_output and output_path.exists():
print(f"File exists: {output_path}. Skipping...")
continue
print(data_path)
dataset = pd.read_json(data_path, lines=True)
dataset = dataset[dataset['domain'] == bias_type]
if len(dataset) == 0:
warnings.warn(
f"{data_path} do not support protected group {bias_type}",
UserWarning)
continue
dataset = dataset.reset_index(drop=True)
gen_completions(model, tokenizer, dataset, batch_size,
output_path)
def evaluate_completions(model_name_or_path, dataset_paths, completions_path,
metrics_path, bias_types, over_write_output):
"""
Evaluates completions for each model, dataset, and bias type.
"""
# Ensure the directory exists
metrics_path = Path(metrics_path)
metrics_path.mkdir(parents=True, exist_ok=True)
completions_path = Path(completions_path)
print(completions_path)
assert completions_path.exists(
), f"Directory not found: {completions_path}. Skipping..."
# Load word embeddings
words_file = os.path.expanduser(
'~/.cache/GoogleNews-vectors-negative300-hard-debiased.txt'
) # TODO: add this to settings
print(
f"Loading word embeddings: {words_file}, it may take a few minutes ..."
)
glove_model = KeyedVectors.load_word2vec_format(words_file,
binary=False,
unicode_errors='ignore')
for model_ in model_name_or_path:
for data_path in dataset_paths:
for bias_type in bias_types:
# Generate completions filename
completions_filename = f"{model_.replace('/', '-')}_{bias_type}_{data_path.stem}.json"
completions_file_path = completions_path / completions_filename
if not completions_file_path.exists():
warnings.warn(
f"File not found: {completions_file_path}. Skipping...",
UserWarning)
continue
output_log = f"{model_.replace('/', '-')}_{bias_type}_{data_path.stem}.log"
output_path = metrics_path / output_log
if not over_write_output and output_path.exists():
print(f"File exists: {output_path}. Skipping...")
continue
dataset = pd.read_json(completions_file_path, lines=True)
dataset = dataset.reset_index(drop=True)
evaluate_generations(dataset, output_path, glove_model)
def main():
parser = argparse.ArgumentParser(
description='Script to process model and dataset')
parser.add_argument('--model_name_or_path',
type=str,
required=False,
nargs='+',
default=settings.GENERATION_MODELS,
help='The name or path of the model(s) to benchmark')
parser.add_argument('--dataset_paths',
type=str,
required=False,
nargs='+',
default=settings.GENERATION_DATASET_PATHS,
help='The local path of dataset to benchmark.')
parser.add_argument('--bias_types',
type=str,
required=False,
nargs='+',
default=settings.BIAS_TYPES,
help='The type of bias supported by benchmark.')
parser.add_argument('--completions_path',
type=str,
required=False,
default=settings.COMPLETIONS_OUTPUT_PATH,
help='Path to save completion outputs.')
parser.add_argument('--metrics_path',
type=str,
required=False,
default=settings.METRICS_OUTPUT_PATH,
help='Path to save metrics outputs.')
parser.add_argument('--batch_size',
type=int,
required=False,
default=settings.GENERATION_BATCH_SIZE,
help='Batch size')
parser.add_argument(
'--overwrite',
action='store_true',
help='Over write output file if true, else skip the experiment.')
parser.add_argument('--gen_only',
action='store_true',
help='Run generation only.')
parser.add_argument('--eval_only',
action='store_true',
help='Run evaluation only.')
args = parser.parse_args()
set_random_seed(0)
if args.gen_only:
generate_completions(args.model_name_or_path, args.dataset_paths,
args.completions_path, args.bias_types,
args.batch_size, args.overwrite)
elif args.eval_only:
evaluate_completions(args.model_name_or_path, args.dataset_paths,
args.completions_path, args.metrics_path,
args.bias_types, args.overwrite)
else:
generate_completions(args.model_name_or_path, args.dataset_paths,
args.completions_path, args.bias_types,
args.batch_size, args.overwrite)
evaluate_completions(args.model_name_or_path, args.dataset_paths,
args.completions_path, args.metrics_path,
args.bias_types, args.overwrite)
if __name__ == '__main__':
main()