File size: 16,797 Bytes
256a159 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 |
# flake8: noqa
# yapf: disable
import getpass
import os.path as osp
from datetime import datetime
from typing import List, Optional
import mmengine
import pytz
import tabulate
from mmengine import ConfigDict
from opencompass.utils import (LarkReporter, dataset_abbr_from_cfg,
get_infer_output_path, get_logger,
model_abbr_from_cfg)
from opencompass.utils.prompt import get_prompt_hash
METRIC_WHITELIST = ['score', 'auc_score', 'accuracy', 'humaneval_pass@1', 'rouge1', 'avg_toxicity_score', 'bleurt_diff', 'matthews_correlation', 'truth']
METRIC_BLACKLIST = ['bp', 'sys_len', 'ref_len']
class PretrainSummarizer:
""""""
def __init__(self, config: ConfigDict, dataset_abbrs: Optional[List[str]] = None, summary_groups: List = [], prompt_db = None) -> None:
self.tasks = []
self.cfg = config
self.logger = get_logger()
# Enable lark bot if lark_url is presented
self.lark_reporter = None
if self.cfg.get('lark_bot_url', None):
self.lark_reporter = LarkReporter(self.cfg['lark_bot_url'])
def summarize(
self,
output_path: str = None,
time_str: str = datetime.now().strftime('%Y%m%d_%H%M%S')): # noqa
model_cfgs = self.cfg['models']
dataset_cfgs = self.cfg['datasets']
summarizer_cfg = self.cfg.get('summarizer', {})
work_dir = self.cfg['work_dir']
# pick up results
raw_results = {}
parsed_results = {}
dataset_metrics = {}
model_abbrs = [model_abbr_from_cfg(model) for model in model_cfgs]
for model in model_cfgs:
model_abbr = model_abbr_from_cfg(model)
parsed_results[model_abbr] = {}
raw_results[model_abbr] = {}
for dataset in dataset_cfgs:
dataset_abbr = dataset_abbr_from_cfg(dataset)
filepath = get_infer_output_path(model, dataset, osp.join(work_dir, 'results'))
if not osp.exists(filepath):
continue
result = mmengine.load(filepath)
raw_results[model_abbr][dataset_abbr] = result
if 'error' in result:
self.debug(f'error in {model_abbr} {dataset_abbr} {result["error"]}')
continue
else:
parsed_results[model_abbr][dataset_abbr] = []
dataset_metrics[dataset_abbr] = []
for metric, score in result.items():
if metric not in METRIC_BLACKLIST and isinstance(score, (int, float)):
parsed_results[model_abbr][dataset_abbr].append(score)
dataset_metrics[dataset_abbr].append(metric)
else:
continue
if len(parsed_results[model_abbr][dataset_abbr]) == 0:
self.logger.warning(f'unknown result format: {result}, continue')
del parsed_results[model_abbr][dataset_abbr]
del dataset_metrics[dataset_abbr]
continue
indice = sorted(
list(range(len(dataset_metrics[dataset_abbr]))),
key=lambda i: (
METRIC_WHITELIST.index(dataset_metrics[dataset_abbr][i])
if dataset_metrics[dataset_abbr][i] in METRIC_WHITELIST
else len(METRIC_WHITELIST)
)
)
parsed_results[model_abbr][dataset_abbr] = [parsed_results[model_abbr][dataset_abbr][i] for i in indice]
dataset_metrics[dataset_abbr] = [dataset_metrics[dataset_abbr][i] for i in indice]
# parse eval mode
dataset_eval_mode = {}
for dataset in dataset_cfgs:
inferencer = dataset.get('infer_cfg', {}).get('inferencer', {}).get('type', '')
inferencer = inferencer if isinstance(inferencer, str) else inferencer.__name__
dataset_abbr = dataset_abbr_from_cfg(dataset)
if 'GenInferencer' in inferencer:
dataset_eval_mode[dataset_abbr] = 'gen'
elif 'PPLInferencer' in inferencer:
dataset_eval_mode[dataset_abbr] = 'ppl'
else:
dataset_eval_mode[dataset_abbr] = 'unknown'
self.logger.warning(f'unknown inferencer: {inferencer} - {dataset_abbr}')
# calculate group metrics
summary_groups = summarizer_cfg.get('summary_groups', [])
for sg in summary_groups:
for model_abbr in model_abbrs:
results = {}
eval_modes = []
for dataset_abbr in sg['subsets']:
if dataset_abbr in parsed_results[model_abbr]:
results[dataset_abbr] = parsed_results[model_abbr][dataset_abbr][0]
eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
if len(results) == len(sg['subsets']):
if 'weights' in sg:
numerator = sum(results[k] * sg['weights'][k] for k in sg['weights'])
denominator = sum(sg['weights'].values())
metric = 'weighted_average'
else:
numerator = sum(results[k] for k in results)
denominator = len(results)
metric = 'naive_average'
results[metric] = numerator / denominator
eval_modes = list(set(eval_modes))
eval_mode = eval_modes[0] if len(eval_modes) == 1 else 'mixed'
# add to global results
raw_results[model_abbr][sg['name']] = results
parsed_results[model_abbr][sg['name']] = [numerator / denominator]
dataset_metrics[sg['name']] = [metric]
dataset_eval_mode[sg['name']] = eval_mode
elif len(results) == 0:
continue
else:
raw_results[model_abbr][sg['name']] = {'error': 'missing datasets: {}'.format(set(sg['subsets']) - set(results.keys()))}
prompt_version = {dataset_abbr_from_cfg(d): get_prompt_hash(d)[:6] for d in dataset_cfgs}
# format table
summarizer_dataset_abbrs = []
if summarizer_cfg.get('dataset_abbrs') is None:
for dataset in dataset_cfgs:
dataset_abbr = dataset_abbr_from_cfg(dataset)
if dataset_abbr in dataset_metrics:
for metric in dataset_metrics[dataset_abbr]:
summarizer_dataset_abbrs.append((dataset_abbr, metric))
else:
summarizer_dataset_abbrs.append((dataset_abbr, None))
for dataset_abbr in dataset_metrics:
for metric in dataset_metrics[dataset_abbr]:
if (dataset_abbr, metric) not in summarizer_dataset_abbrs:
summarizer_dataset_abbrs.append((dataset_abbr, metric))
else:
for item in summarizer_cfg['dataset_abbrs']:
if isinstance(item, str):
summarizer_dataset_abbrs.append((item, None))
elif isinstance(item, (list, tuple)):
summarizer_dataset_abbrs.append((item[0], item[1]))
table = []
checkpoints = [model_abbr.rsplit('_', 1)[1] if '_' in model_abbr else model_abbr for model_abbr in model_abbrs]
# model_abbrs = [model_abbr.rsplit("_", 1)[0] for model_abbr in model_abbrs]
header = ['dataset', 'version', 'metric', 'mode'] + model_abbrs
time_zone = pytz.timezone('Asia/Shanghai')
now = datetime.now(time_zone)
time = now.strftime('%m/%d %H:%M')
times = [time] * len(model_abbrs)
table.append(header)
table.append(['time', 'version', 'metric', 'mode'] + times)
table.append(['checkpoint', 'version', 'metric', 'mode']+ checkpoints)
# check long bench
max_seq_lens = [str(model_cfg.max_seq_len) for model_cfg in model_cfgs]
table.append(['max_seq_len', 'version', 'metric', 'mode']+ max_seq_lens)
dataset_score = [0]* len(model_abbrs)
dataset_num = [0] * len(model_abbrs)
for dataset_abbr, metric in summarizer_dataset_abbrs:
# if dataset_abbr not in dataset_metrics:
# table.append([dataset_abbr, '-', '-', '-'] + ['-'] * len(model_abbrs))
# continue
if metric is None and dataset_abbr in dataset_metrics:
index = 0
metric = dataset_metrics[dataset_abbr][0]
elif dataset_abbr in dataset_metrics and metric in dataset_metrics[dataset_abbr]:
index = dataset_metrics[dataset_abbr].index(metric)
elif not dataset_abbr.startswith('---'):
table.append([dataset_abbr, '-', '-', '-'] + ['-'] * len(model_abbrs))
continue
if dataset_abbr.startswith('---'):
row = [dataset_abbr,'-','-','-']
else:
row = [dataset_abbr, prompt_version.get(dataset_abbr, '-'), metric, dataset_eval_mode.get(dataset_abbr, '-')]
for i, model_abbr in enumerate(model_abbrs):
if dataset_abbr in parsed_results[model_abbr]:
row.append('{:.02f}'.format(parsed_results[model_abbr][dataset_abbr][index]))
dataset_score[i] += parsed_results[model_abbr][dataset_abbr][index]
dataset_num[i] += 1
else:
if dataset_abbr.startswith('---') and dataset_num[i] != 0:
row.append('{:.02f}'.format(dataset_score[i] / dataset_num[i]))
dataset_score[i] = 0
dataset_num[i] = 0
else:
row.append('-')
table.append(row)
# format raw txt
raw_dataset_abbrs = []
for model_abbr in model_abbrs:
for dataset_abbr in raw_results[model_abbr]:
if dataset_abbr not in raw_dataset_abbrs:
raw_dataset_abbrs.append(dataset_abbr)
raw_txts = []
for model_abbr in model_abbrs:
raw_txts.append('-------------------------------')
raw_txts.append(f'Model: {model_abbr}')
for dataset_abbr in raw_dataset_abbrs:
result = raw_results[model_abbr].get(dataset_abbr, '{}')
raw_txts.append(f'{dataset_abbr}: {result}')
raw_txts = '\n'.join(raw_txts)
# output to screean
print(tabulate.tabulate(table, headers='firstrow'))
# output to file
if output_path is None:
output_path = osp.join(work_dir, 'summary', f'summary_{time_str}.txt')
output_csv_path = osp.join(work_dir, 'summary', f'summary_{time_str}.csv')
else:
output_csv_path = output_path.replace('.txt', '.csv')
output_dir = osp.split(output_path)[0]
mmengine.mkdir_or_exist(output_dir)
with open(output_path, 'w', encoding='utf-8') as f:
f.write(time_str + '\n')
f.write('tabulate format\n')
f.write('^' * 128 + '\n')
f.write(tabulate.tabulate(table, headers='firstrow') + '\n')
f.write('$' * 128 + '\n')
f.write('\n' + '-' * 128 + ' THIS IS A DIVIDER ' + '-' * 128 + '\n\n')
f.write('csv format\n')
f.write('^' * 128 + '\n')
f.write('\n'.join([','.join(row) for row in table]) + '\n')
f.write('$' * 128 + '\n')
f.write('\n' + '-' * 128 + ' THIS IS A DIVIDER ' + '-' * 128 + '\n\n')
f.write('raw format\n')
f.write('^' * 128 + '\n')
f.write(raw_txts + '\n')
f.write('$' * 128 + '\n')
self.logger.info(f'write summary to {osp.abspath(output_path)}')
if self.lark_reporter:
content = f'{getpass.getuser()} 的'
content += f'详细评测汇总已输出至 {osp.abspath(output_path)}'
self.lark_reporter.post(content)
with open(output_csv_path, 'w', encoding='utf-8') as f:
f.write('\n'.join([','.join(row) for row in table]) + '\n')
self.logger.info(f'write csv to {osp.abspath(output_csv_path)}')
summary_groups = summarizer_cfg.get('summary_groups', [])
for sg in summary_groups:
for model_abbr in model_abbrs:
results = {}
eval_modes = []
for dataset_abbr in sg['subsets']:
if dataset_abbr in parsed_results[model_abbr]:
results[dataset_abbr] = (parsed_results[model_abbr][dataset_abbr][-1],parsed_results[model_abbr][dataset_abbr][-2])
eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
if len(results) == len(sg['subsets']):
numerator1 = sum(results[k][0] for k in results)
numerator2 = sum(results[k][1] for k in results)
denominator = len(results)
metric = 'correct_bpb-incorrect_bpb'
count_ppl = eval_modes.count('ppl')
count_gen = len(eval_modes)-count_ppl
if count_ppl==0:
results[metric] = -1
else:
results[metric] = (numerator1+count_gen) / count_ppl
eval_modes = list(set(eval_modes))
eval_mode = eval_modes[0] if len(eval_modes) == 1 else 'mixed'
# add to global results
raw_results[model_abbr][sg['name']] = results
parsed_results[model_abbr][sg['name']] = [((numerator1+count_gen) / count_ppl) if count_ppl != 0 else -1, ((numerator2+count_gen) / count_ppl) if count_ppl != 0 else -1]
dataset_metrics[sg['name']] = ['incorrect_bpb','correct_bpb']
dataset_eval_mode[sg['name']] = eval_mode
elif len(results) == 0:
continue
else:
raw_results[model_abbr][sg['name']] = {'error': 'missing datasets: {}'.format(set(sg['subsets']) - set(results.keys()))}
table = []
table.append(['', '', '', ''] + [''] * len(model_abbrs))
table.append(['', '', '', ''] + [''] * len(model_abbrs))
table.append(['', '', '', ''] + [''] * len(model_abbrs))
for dataset_abbr, metric in summarizer_dataset_abbrs:
incorrect_bpb = -1
correct_bpb = -1
if dataset_abbr not in dataset_metrics:
table.append([dataset_abbr, '', '', ''] + [''] * len(model_abbrs))
continue
if metric is None:
index = 0
try:
incorrect_bpb = dataset_metrics[dataset_abbr].index('incorrect_bpb')
correct_bpb = dataset_metrics[dataset_abbr].index('correct_bpb')
except ValueError:
try:
incorrect_bpb = dataset_metrics[dataset_abbr].index('wrong_bpb')
correct_bpb = dataset_metrics[dataset_abbr].index('right_bpb')
except ValueError:
incorrect_bpb = -1
correct_bpb = -1
metric = 'correct_bpb-incorrect_bpb'
elif metric in dataset_metrics[dataset_abbr]:
index = dataset_metrics[dataset_abbr].index(metric)
else:
table.append([dataset_abbr, '-', '-', '-'] + ['-'] * len(model_abbrs))
continue
row = [dataset_abbr, prompt_version.get(dataset_abbr, '-'), metric,
dataset_eval_mode.get(dataset_abbr, '-')]
for model_abbr in model_abbrs:
if dataset_abbr in parsed_results[model_abbr]:
if incorrect_bpb != -1 and correct_bpb != -1:
row.append('{:.02f}/{:.02f}'.format(parsed_results[model_abbr][dataset_abbr][correct_bpb],
parsed_results[model_abbr][dataset_abbr][incorrect_bpb]))
else:
row.append('{:.02f}'.format(-1))
else:
row.append('-')
table.append(row)
with open(output_csv_path, 'a', encoding='utf-8') as f:
f.write('\n'.join([','.join(row) for row in table]) + '\n')
|