-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_tracincp.py
261 lines (239 loc) · 14.5 KB
/
test_tracincp.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
from dataset.simfluence_dataset import SimfluenceDataset
from dataset.tracincp_dataset import TracInCPDataset
import torch
import numpy as np
import matplotlib.pyplot as plt
import os
import sys
import json
from tqdm import tqdm
import fire
from scipy import stats
# 导入Simulator
from model.Simulator import Simulator
from model.XlmrSimulator import XlmrSimulator
from model.VectorSimulator import VectorSimulator
from utils.eval import eval_simulator, eval_tracincp_simulator, eval_tracincp_self_influence
import logging
logging.basicConfig(
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
level=os.environ.get("LOGLEVEL", "INFO").upper(),
stream=sys.stdout,
)
logger = logging.getLogger("Simfuence.test")
data_paths = [
# boolq
]
data_paths_dict = {
'boolq': [
'./runs/boolq/output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-28/',
'./runs/boolq/output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-29/',
'./runs/boolq/output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-30/',
'./runs/boolq/output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-31/',
'./runs/boolq/output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-32/',
],
'sst2': [
'./runs/sst2/output_sst2_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-last-token_seed-28',
'./runs/sst2/output_sst2_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-last-token_seed-29',
'./runs/sst2/output_sst2_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-last-token_seed-30',
'./runs/sst2/output_sst2_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-last-token_seed-31',
'./runs/sst2/output_sst2_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-last-token_seed-32',
],
'rte': [
'./runs/rte/output_rte_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-28',
'./runs/rte/output_rte_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-29',
'./runs/rte/output_rte_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-30',
'./runs/rte/output_rte_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-31',
'./runs/rte/output_rte_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-32',
],
'webnlg': [
'./runs/webnlg/output_webnlg_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-28',
'./runs/webnlg/output_webnlg_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-29',
'./runs/webnlg/output_webnlg_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-30',
'./runs/webnlg/output_webnlg_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-31',
'./runs/webnlg/output_webnlg_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-32',
],
'wmt16_de_en': [
'runs/wmt16_de_en/output_wmt16_de_en_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-5',
'runs/wmt16_de_en/output_wmt16_de_en_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-6',
'runs/wmt16_de_en/output_wmt16_de_en_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-7',
'runs/wmt16_de_en/output_wmt16_de_en_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-8',
'runs/wmt16_de_en/output_wmt16_de_en_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-9',
],
'webnlg_ood_train': [
'./runs/webnlg/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-1',
'./runs/webnlg/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-2',
'./runs/webnlg/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-3',
],
'rte_ood_train': [
'./runs/rte/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-1',
'./runs/rte/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-2',
'./runs/rte/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-3',
],
'flan': [
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-5/',
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-6/',
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-7/',
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-8/',
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-9/',
],
'flan_pythia-1b': [
"runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-5/",
"runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-6/",
"runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-7/",
"runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-8/",
"runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-9/",
],
'flan_rte_ood': [
'runs/flan_ood/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-1/',
'runs/flan_ood/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-2/',
'runs/flan_ood/output_rte-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_seed-3/',
],
'flan_webnlg_ood': [
'runs/flan_ood/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-1',
'runs/flan_ood/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-2',
'runs/flan_ood/output_webnlg-ood_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-1e-6_weight-decay-0.001_epoch-3_loss-output-token_seed-3',
],
'debug': [
'runs/rte/output_rte_bs-4_shot-200_sample-128_lr-2e-6_weight-decay-0.001_epoch-3_seed-1',
'runs/rte/output_rte_bs-4_shot-200_sample-128_lr-2e-6_weight-decay-0.001_epoch-3_seed-1',
'runs/rte/output_rte_bs-4_shot-200_sample-128_lr-2e-6_weight-decay-0.001_epoch-3_seed-1',
],
'flan_tracincp': [
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-410m-deduped_lr-2e-7_weight-decay-0.001_epoch-3_loss-output-token_seed-5/'
],
'flan_pythia-1b_tracincp': [
'runs/flan/output_flan_bs-8_shot-200_sample-128_model-pythia-1b-deduped_lr-2e-7_weight-decay-0.001_epoch-2_loss-output-token_seed-5/'
],
'dataset_debug_sst2_tracincp': [
'runs/dataset_debug_sst2/output_sst2-dataset-debug-is-mislabelled_bs-4_shot-128_sample-128_model-pythia-410m-deduped_lr-1e-06_weight-decay-0.001_epoch-1_loss-output-token_seed-0/'
]
}
def main(
test_example_nums = 200,
train_example_nums = 200,
task = '',
metric = "",
sim_name = "original", # "vec_sim" "original"
dataset_name = "boolq",
save_dir = "/root/paddlejob/workspace/liuqingyi01/code/Simfluence/output/original_task-output_boolq_bs-4_shot-200_sample-128_model-pythia-410m-deduped_lr-5e-7_weight-decay-0.001_epoch-3_lr-0.001_lambda-0.0_bs-128_train-sample-nums-200_test-sample-nums-200_seed-42_step_thres-None",
hyper_parameter = 0.,
test_example_start_id=-1,
test_example_end_id=-1,
step_ckpt_dir=None,
ckpt_steps=None,
self_influence=False,
):
print("task:", task)
print("metric:", metric)
print("dataset:", dataset_name)
print('simulator', sim_name)
print()
from train import (
DATASET,
DATASET_ADDITIONAL_ARGS,
SIMULATORS,
SIMULATR_ADDIONAL_ARGS,
INPUT_ADDITIONAL_KEYS,
)
# 设置设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
simulator_args = SIMULATR_ADDIONAL_ARGS[sim_name]
if sim_name == 'norder_enc_sim':
order_n = simulator_args['order_n']
else:
order_n = -1
if self_influence:
print('目前选择通过tracincp计算self-influence')
test_dataset = SimfluenceDataset(data_paths_dict[task], test_example_nums=test_example_nums, test_example_start_id=test_example_start_id, test_example_end_id=test_example_end_id, is_train=False, step_thres=None, metric=metric, self_influence=self_influence)
# 加载数据集
# dataset = DATASET[dataset_name]
dataset = TracInCPDataset
if dataset is None:
logger.warning("dataset is None, use default dataset")
else:
test_dataset = dataset(
test_dataset,
is_train=False,
**DATASET_ADDITIONAL_ARGS[dataset_name]
)
# 加载simulator
# model = SIMULATORS[sim_name](train_example_nums=train_example_nums, hyper_parameter=hyper_parameter, test_example_nums=test_example_nums, **simulator_args).to(device)
# if sim_name == 'enc_sim' or sim_name == 'norder_enc_sim':
# if simulator_args['use_initial']:
# model._get_initial_embeds(test_dataset, device)
# model.load_state_dict(torch.load(check_point_path))
# model.to(device)
model = SIMULATORS[sim_name]
print("\n开始测试...")
input_kwargs_keys = INPUT_ADDITIONAL_KEYS[sim_name]
if self_influence:
results = eval_tracincp_self_influence(test_dataset, model, device, step_ckpt_dir, input_kwargs_keys, ckpt_steps)
self_influence = results['self_influence']
sorted_self_influence = sorted(self_influence.items(), key=lambda x: x[1], reverse=True)
# 保存结果
print("\n保存结果...")
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with open(os.path.join(save_dir, f'sorted_self_influence.out'), 'w', encoding='utf-8') as f:
for sample_id, self_inf in sorted_self_influence:
print(json.dumps({sample_id: self_inf}), file=f)
else:
results = eval_tracincp_simulator(test_dataset, model, device, step_ckpt_dir, input_kwargs_keys, ckpt_steps)
# print("测试集mse:", results[0][0])
# print("测试集mae:", results[0][1])
print(f'指标: {metric}')
print(f"测试集mse 均值:{results['all_steps_mse_mean']} 标准差:{results['all_steps_mse_std']}")
print(f"测试集mae 均值:{results['all_steps_mae_mean']} 标准差:{results['all_steps_mae_std']}")
# 保存结果
print("\n保存结果...")
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with open(os.path.join(save_dir, f'pred_and_gt_{metric}_trajectories.out'), 'w', encoding='utf-8') as f:
for test_sample_id, r in results['pred_loss_dict'].items():
r = json.dumps(r)
print(r, file=f)
# 计算last-step spearman correlation ################################
from collections import defaultdict
last_step = {
'pred': defaultdict(list),
'gt': defaultdict(list)
}
with open(os.path.join(save_dir, f'pred_and_gt_{metric}_trajectories.out'), 'r', encoding='utf-8') as f:
lines = f.readlines()
for line in lines:
line = json.loads(line)
for runs_id, trajectory in enumerate(line):
last_step['pred'][runs_id].append(trajectory['pred_loss'][-1])
last_step['gt'][runs_id].append(trajectory['gt_loss'][-1])
total_runs_num = len(last_step['pred'])
last_step_spearman_list = []
for i in range(total_runs_num):
last_step_spearman_list.append(
stats.spearmanr(last_step['gt'][i], last_step['pred'][i] ).statistic
)
# import pdb; pdb.set_trace()
last_step_spearman_np = np.array(last_step_spearman_list)
last_step_spearman_mean = last_step_spearman_np.mean()
last_step_spearman_std = last_step_spearman_np.std()
print(f"测试集last step spearman 均值: {last_step_spearman_mean}, 标准差: {last_step_spearman_std}")
# # 画图
# print("\n画图...")
# fig_save_path = os.path.join(save_dir, 'figs')
# if not os.path.exists(fig_save_path):
# os.mkdir(fig_save_path)
# for test_sample_id, r in tqdm(results['pred_loss_dict].items()):
# for i, item in enumerate(r):
# plt.plot(item['step'], item['gt_loss'], label='gt')
# plt.plot(item['step'], item['pred_loss'], label='predict')
# plt.xlabel('step')
# plt.ylabel('loss')
# plt.title('test sample id:' + str(test_sample_id) + '-' + str(i))
# plt.legend()
# plt.grid()
# plt.savefig(os.path.join(fig_save_path, str(test_sample_id) + '-' + str(i) + '.png'))
# plt.clf()
print('done')
if __name__ == "__main__":
fire.Fire(main)