-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate.py
220 lines (200 loc) · 8.02 KB
/
evaluate.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import os
import csv
import argparse
import json
import random
import pandas as pd
from tqdm import tqdm
from sklearn.metrics import accuracy_score
from utils import Loader, get_model, set_seed, \
clear_gpu_resources, encode_subset, inference
# maximal number of shots in context window
SHOT_MAP = {
"superglue-rte": 12,
"superglue-boolq": 10,
"hellaswag": 18,
"piqa": 38,
"superglue-wic": 32,
"superglue-wsc": 32,
"ai2_arc_challenge": 46,
"ai2_arc_easy": 52,
"openbookqa": 52
}
BASELINES = [
"incontext_influence_positive",
"incontext_influence_negative",
"datamodel_influence",
"oneshot_influence",
"best_set",
"random",
"perplexity",
"dev_avg_roberta-large",
]
def get_top_dict(args, max_seq_len):
"""
Nested dict to get top train points per [task][test]
+ add items in inverse order from worse -> better
"""
task_shot = SHOT_MAP[args.task] // int(2048 / max_seq_len)
ret = {}
# Perplexity
if "perplexity" in args.resource_file:
data = pd.read_csv(args.resource_file)
for task, d in data.groupby("task"):
ret[task] = {}
for model, d2 in d.groupby("model"):
ret[task][model] = []
for _, row in d2[:task_shot].iterrows():
# add from smallest similarity to biggest
ret[task][model].insert(0, row["index"])
for task in ret:
assert (len(ret[task]) == len(MODELS_ALL))
for model in ret[task]:
assert (len(ret[task][model]) == SHOT_MAP[task])
# Influence
elif "influence" in args.resource_file:
with open(args.resource_file, 'r') as fp:
for line in fp:
line = json.loads(line)
if line["method"] not in args.method:
continue
task = line["task"]
model = line["model"]
scores = dict(sorted(line["scores"].items(), key=lambda item: item[1]))
top = list(scores.keys())
if "negative" in args.method:
top = top[:task_shot]
top = top[::-1] # inverse so worst examples come last
elif "neutral" in args.method:
top = top[int((len(top) - task_shot) / 2): int((len(top) + task_shot) / 2)] # select middle k shot
else: # default
top = top[-task_shot:] # select top k shot at the end, don't need to inverse
if task not in ret:
ret[task] = {}
ret[task][model] = top
elif "random" == args.method:
return ret, task_shot
elif "best_set" == args.method:
with open(args.resource_file, 'r') as fp:
for line in fp:
line = json.loads(line)
if line["method"] not in args.method:
continue
task = line["task"]
model = line["model"]
top = line["scores"][-task_shot:]
if task not in ret:
ret[task] = {}
ret[task][model] = top
# Average Validation Distance
elif args.method in ["dev_avg_roberta-large", "dev_avg_mpnet_avg"]:
with open(args.resource_file, 'r') as fp:
for line in fp:
line = json.loads(line)
task = line["task"]
scores = dict(sorted(line["rank_average_dev"].items(), key=lambda item: item[1]))
top = list(scores.keys())
top = top[-task_shot:] # select top k shot at the end, don't need to inverse
if task not in ret:
ret[task] = {}
ret[task] = top
else:
raise NotImplementedError
return ret, task_shot
def check_line_exists(out_csv, data):
with open(out_csv, 'r') as fp:
reader = csv.DictReader(fp)
for row in reader:
if all(str(row[key]) == str(value) for key, value in data.items()):
return True
return False
def main(args):
if args.method not in BASELINES:
print("Method not valid")
return
set_seed(args.seed)
model_abbr = args.model_name_or_path.split('/')[-1]
fieldnames = ["method", "model", "task", "subset_idx", "k_shot", "val_acc"]
out_csv = os.path.join(args.out_dir, "baseline.csv")
# check if out folder exists
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
# check if csv exists
if not os.path.exists(out_csv):
with open(out_csv, 'w') as fp:
writer = csv.DictWriter(fp, fieldnames=fieldnames)
writer.writeheader()
# load tokenizer + model + exemplars
tokenizer, model = get_model(args)
max_seq_len = model.config.max_seq_length
top_dict, task_shot = get_top_dict(args, max_seq_len)
# Loader with original data dir and helpers
loader = Loader(args.data_dir, args.task)
if args.split == "test":
val_examples = loader.get_test_examples()
elif args.split == "dev":
val_examples = loader.get_dev_examples()
else:
raise
if args.method.startswith("incontext_influence") or args.method in ["perplexity", "datamodel_influence", "oneshot_influence", "best_set"]:
top_k = top_dict[args.task][model_abbr]
elif args.method == "random":
train_indices = [dp["index"] for dp in loader.data_template]
top_k = random.choices(train_indices, k=task_shot)
elif args.method in ["dev_avg_roberta-large", "dev_avg_mpnet_avg"]:
top_k = top_dict[args.task]
else:
raise
# 1) Encode subset from right to left (do this once)
subset = {
"data": loader.get_subset_by_indices(top_k),
"subset_idx": top_k,
}
# 2) Evaluate
pred, true = [], []
demonstration_ids = encode_subset(subset, args.task, tokenizer)
for val_ex in tqdm(val_examples):
prediction, subset_actual = inference(demonstration_ids, val_ex, args.task, max_seq_len, model, tokenizer)
true.append(val_ex["output"])
pred.append(prediction)
# Write
acc = accuracy_score(true, pred)
data = {
"method": args.method,
"model": model_abbr,
"task": args.task,
"subset_idx": ",".join(str(x_i) for x_i in top_k),
"k_shot": task_shot,
"val_acc": acc
}
if not check_line_exists(out_csv, data):
with open(out_csv, 'a') as fp:
writer = csv.DictWriter(fp, fieldnames=fieldnames)
writer.writerow(data)
print(f"[{args.method}-{model_abbr}-{args.task}] Accuracy: {acc}")
print(f"Wrote eval results for ({args.method}, {args.task}, {model_abbr}) at '{out_csv}'")
else:
print(f"(Line exists) [{args.method}-{model_abbr}-{args.task}] Accuracy: {acc}")
clear_gpu_resources()
return
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--task", type=str, default="superglue-rte",
help="List of tasks to run counterfactual for")
parser.add_argument("--model_name_or_path", type=str, default="facebook/opt-6.7b",
help="HF model identifier. For LLaMA, please specify directory of model weights")
parser.add_argument("--split", type=str, default="test",
help="Split to do evaluation on")
parser.add_argument("--resource_file", type=str, default="influence_scores.jsonl",
help="Influence score output file")
parser.add_argument("--data_dir", type=str, default="data-train400-dev200",
help="Data directory")
parser.add_argument("--out_dir", type=str, default="baseline",
help="Output directory")
parser.add_argument("--method", type=str, default="incontext_influence_positive",
help="Baseline type")
parser.add_argument("--cache_dir", type=str, default="/scratch/taing",
help="HF cache directory")
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
main(args)