-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathevaluator.py
179 lines (149 loc) · 6.41 KB
/
evaluator.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
from typing import List, Dict
from arithmetics import TaskPrompt, PromptArithmeticsModel
from args import TrainingArguments
from tasks import AutoTask
from transformers import (
Seq2SeqTrainer,
default_data_collator,
PreTrainedTokenizer,
AutoTokenizer,
)
from datasets import Dataset
import wandb
import torch
from peft import PeftModel
import pandas as pd
import numpy as np
class ArithmeticsEvaluator:
task_prompts: List[TaskPrompt] = None
def __init__(
self,
task_prompts: List[TaskPrompt],
model: PeftModel,
test_datasets: Dict[str, Dataset],
eval_datasets: Dict[str, Dataset],
training_args: TrainingArguments,
tokenizer: PreTrainedTokenizer,
origin_weights: torch.Tensor,
):
self.task_prompts = task_prompts
self.model = model
self.test_datasets = test_datasets
self.eval_datasets = eval_datasets
self.training_args = training_args
self.tokenizer = tokenizer
self.origin_weights = origin_weights
self.results = []
self.orig_run_name = self.training_args.run_name
self.scaling_coefs = np.arange(0.0, 1.05, 0.05)
def set_task(self, task_prompt: TaskPrompt, coef: float = 1):
self.model.prompt_encoder.default.embedding.weight = task_prompt.apply(
self.origin_weights, coef
)
def run(self):
for tp in self.task_prompts:
self.training_args.run_name = (
f"{self.orig_run_name}{tp.task_name.replace(' ', '')}"
)
print(f"Evaluating task origin {tp.task_name}")
print(f"Evaluating on scaling coefs: {self.scaling_coefs}")
best_coef = 1
best_acc = 0
if len(tp.tasks) > 1:
for coef in self.scaling_coefs:
eval_em = []
for dataset_name in tp.tasks:
print(dataset_name, coef)
self.set_task(tp, coef=coef)
print(
f"Current PT weights: {self.model.prompt_encoder.default.embedding.weight}"
)
compute_metrics = AutoTask.get(
dataset_name
).get_compute_metrics(self.tokenizer)
trainer = Seq2SeqTrainer(
model=self.model,
tokenizer=self.tokenizer,
args=self.training_args,
data_collator=default_data_collator,
compute_metrics=compute_metrics,
)
eval_res = trainer.evaluate(
eval_dataset=self.eval_datasets[dataset_name],
metric_key_prefix=f"eval_{dataset_name}",
)
print(eval_res)
if f"eval_{dataset_name}_exact_match" in eval_res.keys():
eval_em.append(eval_res[f"eval_{dataset_name}_exact_match"])
elif f"eval_{dataset_name}_pearsonr" in eval_res.keys():
eval_em.append(eval_res[f"eval_{dataset_name}_pearsonr"])
mean_acc = torch.tensor(eval_em).mean()
if mean_acc > best_acc:
print(f"New best mean acc: {mean_acc} with coef: {coef}")
best_acc = mean_acc
best_coef = coef
print(f"Testing with best coef: {best_coef}")
for dataset_name in tp.tasks:
print(dataset_name, best_coef)
self.set_task(tp, coef=best_coef)
print(
f"Current PT weights: {self.model.prompt_encoder.default.embedding.weight}"
)
compute_metrics = AutoTask.get(dataset_name).get_compute_metrics(
self.tokenizer
)
trainer = Seq2SeqTrainer(
model=self.model,
tokenizer=self.tokenizer,
args=self.training_args,
data_collator=default_data_collator,
compute_metrics=compute_metrics,
)
test_res = trainer.evaluate(
eval_dataset=self.test_datasets[dataset_name],
metric_key_prefix=f"test_{dataset_name}",
)
print(test_res)
if f"test_{dataset_name}_exact_match" in test_res.keys():
self.results.append(
{
"tasks": " ".join(tp.tasks),
f"{dataset_name}_exact_match": test_res[
f"test_{dataset_name}_exact_match"
],
"best_coef": best_coef,
}
)
if f"test_{dataset_name}_macro_f1" in test_res.keys():
self.results.append(
{
"tasks": " ".join(tp.tasks),
f"{dataset_name}_macro_f1": test_res[
f"test_{dataset_name}_macro_f1"
],
"best_coef": best_coef,
}
)
elif f"test_{dataset_name}_f1" in test_res.keys():
self.results.append(
{
"tasks": " ".join(tp.tasks),
f"{dataset_name}_f1": test_res[f"test_{dataset_name}_f1"],
"best_coef": best_coef,
}
)
elif f"test_{dataset_name}_pearsonr" in test_res.keys():
self.results.append(
{
"tasks": " ".join(tp.tasks),
f"{dataset_name}_pearsonr": test_res[
f"test_{dataset_name}_pearsonr"
],
"best_coef": best_coef,
}
)
if wandb.run:
wandb.finish()
df_results = pd.DataFrame.from_dict(self.results)
df_results = df_results.groupby(["tasks"], as_index=False).first()
return df_results