Saving train state of step 25
Browse files- .ipynb_checkpoints/run_distillation-checkpoint.py +683 -678
- __pycache__/evaluate.cpython-39.pyc +0 -0
- distil-whisper/events.out.tfevents.1715057787.server02.1349950.0 +3 -0
- distil-whisper/events.out.tfevents.1715063050.server02.1368197.0 +3 -0
- distil-whisper/events.out.tfevents.1715063266.server02.1369570.0 +3 -0
- distil-whisper/events.out.tfevents.1715063402.server02.1370564.0 +3 -0
- distil-whisper/events.out.tfevents.1715063677.server02.1372191.0 +3 -0
- distil-whisper/events.out.tfevents.1715063742.server02.1372871.0 +3 -0
- distil-whisper/events.out.tfevents.1715064564.server02.1376229.0 +3 -0
- distil-whisper/events.out.tfevents.1715065478.server02.1379863.0 +3 -0
- run_distillation.py +9 -4
- run_evaluate.py +0 -0
- tokenizer.json +21 -0
.ipynb_checkpoints/run_distillation-checkpoint.py
CHANGED
@@ -750,11 +750,14 @@ def main():
|
|
750 |
else:
|
751 |
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
752 |
|
|
|
|
|
753 |
# 2. Initialize the accelerator
|
754 |
# We will let the accelerator handle device placement for us in this example
|
755 |
# We simply have to specify the training precision and any trackers being used
|
756 |
# We'll use the same dtype arguments as our JAX/Flax training script and convert
|
757 |
# it to accelerate format
|
|
|
758 |
if training_args.dtype == "float16":
|
759 |
mixed_precision = "fp16"
|
760 |
teacher_dtype = torch.float16
|
@@ -1007,686 +1010,688 @@ def main():
|
|
1007 |
)
|
1008 |
else:
|
1009 |
is_multilingual = False
|
|
|
|
|
1010 |
|
1011 |
# 8. Create a single speech processor - make sure all processes wait until data is saved
|
1012 |
-
if accelerator.is_main_process:
|
1013 |
-
|
1014 |
-
|
1015 |
-
|
1016 |
-
|
1017 |
-
|
1018 |
-
|
1019 |
-
accelerator.wait_for_everyone()
|
1020 |
-
processor = WhisperProcessor.from_pretrained(training_args.output_dir)
|
1021 |
-
|
1022 |
-
# 9. Resample speech dataset: `datasets` takes care of automatically loading and resampling the audio,
|
1023 |
-
# so we just need to set the correct target sampling rate.
|
1024 |
-
sampling_rate = feature_extractor.sampling_rate
|
1025 |
-
raw_datasets = raw_datasets.cast_column(
|
1026 |
-
|
1027 |
-
|
1028 |
-
)
|
1029 |
-
|
1030 |
-
# 10. Preprocessing the datasets: we need to read the audio files as arrays and tokenize the targets.
|
1031 |
-
# 10.1: Define the pre-processing constants
|
1032 |
-
max_input_length = int(data_args.max_duration_in_seconds * sampling_rate)
|
1033 |
-
min_input_length = int(data_args.min_duration_in_seconds * sampling_rate)
|
1034 |
-
max_label_length = (
|
1035 |
-
|
1036 |
-
)
|
1037 |
-
|
1038 |
-
timestamp_probability = data_args.timestamp_probability
|
1039 |
-
condition_on_prev_probability = data_args.condition_on_prev_probability
|
1040 |
-
return_timestamps = data_args.return_timestamps if timestamp_probability > 0 else False
|
1041 |
-
|
1042 |
-
timestamp_ids = tokenizer.timestamp_ids()
|
1043 |
-
timestamp_begin = tokenizer.all_special_ids[-1]
|
1044 |
-
timestamp_position = 3 if is_multilingual else 1
|
1045 |
-
|
1046 |
-
decoder_start_token_id = student_model.config.decoder_start_token_id # <|startoftranscript|>
|
1047 |
-
decoder_prev_token_id = tokenizer.all_special_ids[-3] # <|startofprev|>
|
1048 |
-
prompt_cutoff_length = max_label_length // 2
|
1049 |
-
|
1050 |
-
num_workers = data_args.preprocessing_num_workers
|
1051 |
-
dataloader_num_workers = training_args.dataloader_num_workers
|
1052 |
-
prefetch_factor = training_args.dataloader_prefetch_factor
|
1053 |
-
|
1054 |
-
metric = evaluate.load("wer")
|
1055 |
-
normalizer = (
|
1056 |
-
|
1057 |
-
|
1058 |
-
|
1059 |
-
)
|
1060 |
-
wer_threshold = data_args.wer_threshold
|
1061 |
-
use_pseudo_labels = data_args.use_pseudo_labels
|
1062 |
-
train_text_column_name = "whisper_transcript" if use_pseudo_labels else "text"
|
1063 |
-
|
1064 |
-
# 10.2: filter based on maximum number of training/evaluation samples
|
1065 |
-
if training_args.do_train and data_args.max_train_samples is not None:
|
1066 |
-
|
1067 |
-
|
1068 |
-
|
1069 |
-
|
1070 |
-
|
1071 |
-
|
1072 |
-
if training_args.do_eval and data_args.max_eval_samples is not None:
|
1073 |
-
|
1074 |
-
|
1075 |
-
|
1076 |
-
|
1077 |
-
|
1078 |
-
|
1079 |
-
|
1080 |
-
# 10.3: filter training data based on WER threshold -> this is KEY to good distillation performance
|
1081 |
-
def is_wer_in_range(ground_truth, whisper_transcript):
|
1082 |
-
|
1083 |
-
|
1084 |
-
|
1085 |
-
|
1086 |
-
|
1087 |
-
|
1088 |
-
|
1089 |
-
|
1090 |
-
|
1091 |
-
|
1092 |
-
|
1093 |
-
|
1094 |
-
filter_by_wer_threshold = partial(
|
1095 |
-
|
1096 |
-
|
1097 |
-
|
1098 |
-
)
|
1099 |
-
|
1100 |
-
if wer_threshold is not None and use_pseudo_labels:
|
1101 |
-
|
1102 |
-
|
1103 |
-
|
1104 |
-
|
1105 |
-
|
1106 |
-
|
1107 |
-
|
1108 |
-
# 10.4: pre-process training/evaluation datasets
|
1109 |
-
def prepare_train_dataset(batch):
|
1110 |
-
|
1111 |
-
|
1112 |
-
|
1113 |
-
|
1114 |
-
|
1115 |
-
|
1116 |
-
|
1117 |
-
|
1118 |
-
|
1119 |
-
|
1120 |
-
|
1121 |
-
|
1122 |
-
|
1123 |
-
|
1124 |
-
|
1125 |
-
|
1126 |
-
|
1127 |
-
|
1128 |
-
|
1129 |
-
|
1130 |
-
|
1131 |
-
|
1132 |
-
|
1133 |
-
|
1134 |
-
|
1135 |
-
|
1136 |
-
|
1137 |
-
|
1138 |
-
|
1139 |
-
|
1140 |
-
|
1141 |
-
|
1142 |
-
|
1143 |
-
|
1144 |
-
|
1145 |
-
|
1146 |
-
|
1147 |
-
|
1148 |
-
|
1149 |
-
|
1150 |
-
|
1151 |
-
|
1152 |
-
|
1153 |
-
|
1154 |
-
|
1155 |
-
|
1156 |
-
|
1157 |
-
|
1158 |
-
|
1159 |
-
|
1160 |
-
|
1161 |
-
|
1162 |
-
|
1163 |
-
|
1164 |
-
|
1165 |
-
|
1166 |
-
|
1167 |
-
|
1168 |
-
|
1169 |
-
|
1170 |
-
|
1171 |
-
|
1172 |
-
|
1173 |
-
def prepare_eval_dataset(batch):
|
1174 |
-
|
1175 |
-
|
1176 |
-
|
1177 |
-
|
1178 |
-
|
1179 |
-
|
1180 |
-
|
1181 |
-
|
1182 |
-
|
1183 |
-
|
1184 |
-
|
1185 |
-
vectorized_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
|
1186 |
-
if training_args.do_train:
|
1187 |
-
|
1188 |
-
|
1189 |
-
|
1190 |
-
|
1191 |
-
|
1192 |
-
|
1193 |
-
|
1194 |
-
|
1195 |
-
|
1196 |
-
|
1197 |
-
|
1198 |
-
|
1199 |
-
|
1200 |
-
|
1201 |
-
|
1202 |
-
|
1203 |
-
if training_args.do_eval:
|
1204 |
-
|
1205 |
-
|
1206 |
-
|
1207 |
-
|
1208 |
-
|
1209 |
-
|
1210 |
-
|
1211 |
-
|
1212 |
-
|
1213 |
-
|
1214 |
-
|
1215 |
-
|
1216 |
-
# 10.5: Filter training data with inputs longer than `max_input_length`
|
1217 |
-
def is_audio_in_length_range(length):
|
1218 |
-
|
1219 |
-
|
1220 |
-
filter_by_audio_fn = partial(
|
1221 |
-
|
1222 |
-
)
|
1223 |
-
with accelerator.main_process_first():
|
1224 |
-
|
1225 |
-
|
1226 |
-
|
1227 |
-
|
1228 |
-
|
1229 |
-
|
1230 |
-
# 10.6: Filter training data with labels longer than `max_label_length`
|
1231 |
-
def is_labels_in_length_range(labels):
|
1232 |
-
|
1233 |
-
|
1234 |
-
filter_by_labels_fn = partial(
|
1235 |
-
|
1236 |
-
)
|
1237 |
-
with accelerator.main_process_first():
|
1238 |
-
|
1239 |
-
|
1240 |
-
|
1241 |
-
|
1242 |
-
|
1243 |
-
|
1244 |
-
# Pre-processing complete!
|
1245 |
-
# For large datasets it is advised to run the preprocessing on a
|
1246 |
-
# single machine first with `--preprocessing_only` since there will mostly likely
|
1247 |
-
# be a timeout when running the script in distributed mode.
|
1248 |
-
# In a second step, `--preprocessing_only` can then be set to `False` to load the
|
1249 |
-
# cached dataset
|
1250 |
-
if data_args.preprocessing_only:
|
1251 |
-
|
1252 |
-
|
1253 |
-
|
1254 |
-
|
1255 |
-
|
1256 |
-
|
1257 |
-
|
1258 |
-
|
1259 |
-
|
1260 |
-
|
1261 |
-
# 11. Define Evaluation Metrics
|
1262 |
-
def compute_metrics(preds, labels):
|
1263 |
-
|
1264 |
-
|
1265 |
-
|
1266 |
-
|
1267 |
-
|
1268 |
-
|
1269 |
-
|
1270 |
-
|
1271 |
-
|
1272 |
-
|
1273 |
-
|
1274 |
-
|
1275 |
-
|
1276 |
-
|
1277 |
-
|
1278 |
-
|
1279 |
-
|
1280 |
-
|
1281 |
-
|
1282 |
-
|
1283 |
-
|
1284 |
-
|
1285 |
-
# 12. Define Training Schedule
|
1286 |
-
# Store some constants
|
1287 |
-
per_device_train_batch_size = int(training_args.per_device_train_batch_size)
|
1288 |
-
train_batch_size = per_device_train_batch_size * accelerator.num_processes
|
1289 |
-
gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
|
1290 |
-
per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
|
1291 |
-
|
1292 |
-
if not data_args.streaming and training_args.max_steps < 0:
|
1293 |
-
|
1294 |
-
|
1295 |
-
|
1296 |
-
elif training_args.max_steps > 0:
|
1297 |
-
|
1298 |
-
|
1299 |
-
|
1300 |
-
|
1301 |
-
|
1302 |
-
|
1303 |
-
|
1304 |
-
|
1305 |
-
|
1306 |
-
else:
|
1307 |
-
|
1308 |
-
|
1309 |
-
if training_args.eval_steps is None:
|
1310 |
-
|
1311 |
-
|
1312 |
-
|
1313 |
-
|
1314 |
-
else:
|
1315 |
-
|
1316 |
-
|
1317 |
-
# 13. Define optimizer, LR scheduler, collator
|
1318 |
-
decay_parameters = get_parameter_names(
|
1319 |
-
|
1320 |
-
|
1321 |
-
|
1322 |
-
)
|
1323 |
-
decay_parameters = [name for name in decay_parameters if "bias" not in name]
|
1324 |
-
optimizer_grouped_parameters = [
|
1325 |
-
|
1326 |
-
|
1327 |
-
|
1328 |
-
|
1329 |
-
|
1330 |
-
|
1331 |
-
|
1332 |
-
|
1333 |
-
]
|
1334 |
-
optimizer = torch.optim.AdamW(
|
1335 |
-
|
1336 |
-
|
1337 |
-
|
1338 |
-
|
1339 |
-
)
|
1340 |
-
|
1341 |
-
# LR scheduler gets stepped by `num_processes` each time -> account for this in warmup / total steps
|
1342 |
-
lr_scheduler = get_scheduler(
|
1343 |
-
|
1344 |
-
|
1345 |
-
|
1346 |
-
|
1347 |
-
)
|
1348 |
-
|
1349 |
-
data_collator = DataCollatorSpeechSeq2SeqWithPadding(
|
1350 |
-
|
1351 |
-
|
1352 |
-
|
1353 |
-
|
1354 |
-
|
1355 |
-
|
1356 |
-
)
|
1357 |
-
|
1358 |
-
# 14. Define generation arguments - we need to do this before we wrap the models in DDP
|
1359 |
-
# so that we can still access the configs
|
1360 |
-
num_beams = (
|
1361 |
-
|
1362 |
-
|
1363 |
-
|
1364 |
-
)
|
1365 |
-
|
1366 |
-
gen_kwargs = {
|
1367 |
-
|
1368 |
-
|
1369 |
-
|
1370 |
-
}
|
1371 |
-
if is_multilingual:
|
1372 |
-
|
1373 |
-
|
1374 |
-
|
1375 |
-
|
1376 |
-
|
1377 |
-
|
1378 |
-
|
1379 |
-
|
1380 |
-
# 15. Prepare everything with accelerate
|
1381 |
-
student_model, teacher_model, optimizer, lr_scheduler = accelerator.prepare(
|
1382 |
-
|
1383 |
-
)
|
1384 |
-
|
1385 |
-
def kl_divergence(target_distribution, log_predicted_distribution, labels):
|
1386 |
-
|
1387 |
-
|
1388 |
-
|
1389 |
-
|
1390 |
-
|
1391 |
-
|
1392 |
-
|
1393 |
-
|
1394 |
-
|
1395 |
-
|
1396 |
-
# Define gradient update step fn
|
1397 |
-
def train_step(
|
1398 |
-
|
1399 |
-
|
1400 |
-
):
|
1401 |
-
|
1402 |
-
|
1403 |
-
|
1404 |
-
|
1405 |
-
|
1406 |
-
|
1407 |
-
|
1408 |
-
|
1409 |
-
|
1410 |
-
|
1411 |
-
|
1412 |
-
|
1413 |
-
|
1414 |
-
|
1415 |
-
|
1416 |
-
|
1417 |
-
|
1418 |
-
|
1419 |
-
|
1420 |
-
|
1421 |
-
|
1422 |
-
|
1423 |
-
|
1424 |
-
|
1425 |
-
|
1426 |
-
|
1427 |
-
|
1428 |
-
|
1429 |
-
# Define eval fn
|
1430 |
-
def eval_step(batch):
|
1431 |
-
|
1432 |
-
|
1433 |
-
|
1434 |
-
|
1435 |
-
|
1436 |
-
|
1437 |
-
|
1438 |
-
|
1439 |
-
|
1440 |
-
|
1441 |
-
|
1442 |
-
|
1443 |
-
|
1444 |
-
|
1445 |
-
|
1446 |
-
|
1447 |
-
|
1448 |
-
|
1449 |
-
|
1450 |
-
|
1451 |
-
|
1452 |
-
|
1453 |
-
|
1454 |
-
|
1455 |
-
|
1456 |
-
def generate_step(batch):
|
1457 |
-
|
1458 |
-
|
1459 |
-
|
1460 |
-
|
1461 |
-
|
1462 |
-
logger.info("***** Running training *****")
|
1463 |
-
logger.info(f" Num examples = {total_train_steps * train_batch_size * gradient_accumulation_steps}")
|
1464 |
-
if not data_args.streaming:
|
1465 |
-
|
1466 |
-
logger.info(" Instantaneous batch size per device =" f" {training_args.per_device_train_batch_size}")
|
1467 |
-
logger.info(" Gradient accumulation steps =" f" {gradient_accumulation_steps}")
|
1468 |
-
logger.info(
|
1469 |
-
|
1470 |
-
)
|
1471 |
-
logger.info(f" Total optimization steps = {total_train_steps}")
|
1472 |
-
|
1473 |
-
# ======================== Training ================================
|
1474 |
-
train_time = 0
|
1475 |
-
train_start = time.time()
|
1476 |
-
steps_trained_progress_bar = tqdm(
|
1477 |
-
|
1478 |
-
)
|
1479 |
-
continue_training = True
|
1480 |
-
epochs_trained = 0
|
1481 |
-
cur_step = 0
|
1482 |
-
|
1483 |
-
checkpoint = None
|
1484 |
-
if training_args.resume_from_checkpoint is not None:
|
1485 |
-
|
1486 |
-
elif last_checkpoint is not None:
|
1487 |
-
|
1488 |
-
|
1489 |
-
if checkpoint is not None:
|
1490 |
-
|
1491 |
-
|
1492 |
-
|
1493 |
-
|
1494 |
-
|
1495 |
-
|
1496 |
-
|
1497 |
-
|
1498 |
-
|
1499 |
-
|
1500 |
-
|
1501 |
-
|
1502 |
-
|
1503 |
-
|
1504 |
-
|
1505 |
-
|
1506 |
-
|
1507 |
-
|
1508 |
-
|
1509 |
-
|
1510 |
-
|
1511 |
-
|
1512 |
-
|
1513 |
-
|
1514 |
-
|
1515 |
-
else:
|
1516 |
-
|
1517 |
-
|
1518 |
-
for epoch in range(epochs_trained, num_epochs):
|
1519 |
-
|
1520 |
-
|
1521 |
-
|
1522 |
-
|
1523 |
-
|
1524 |
-
|
1525 |
-
|
1526 |
-
|
1527 |
-
|
1528 |
-
|
1529 |
-
|
1530 |
-
|
1531 |
-
|
1532 |
-
|
1533 |
-
|
1534 |
-
|
1535 |
-
|
1536 |
-
|
1537 |
-
|
1538 |
-
|
1539 |
-
|
1540 |
-
|
1541 |
-
|
1542 |
-
|
1543 |
-
|
1544 |
-
|
1545 |
-
|
1546 |
-
|
1547 |
-
|
1548 |
-
|
1549 |
-
|
1550 |
-
|
1551 |
-
|
1552 |
-
|
1553 |
-
|
1554 |
-
|
1555 |
-
|
1556 |
-
|
1557 |
-
|
1558 |
-
|
1559 |
-
|
1560 |
-
|
1561 |
-
|
1562 |
-
|
1563 |
-
|
1564 |
-
|
1565 |
-
|
1566 |
-
|
1567 |
-
|
1568 |
-
|
1569 |
-
|
1570 |
-
|
1571 |
-
|
1572 |
-
|
1573 |
-
|
1574 |
-
|
1575 |
-
|
1576 |
-
|
1577 |
-
|
1578 |
-
|
1579 |
-
|
1580 |
-
|
1581 |
-
|
1582 |
-
|
1583 |
-
|
1584 |
-
|
1585 |
-
|
1586 |
-
|
1587 |
-
|
1588 |
-
|
1589 |
-
|
1590 |
-
|
1591 |
-
|
1592 |
-
|
1593 |
-
|
1594 |
-
|
1595 |
-
|
1596 |
-
|
1597 |
-
|
1598 |
-
|
1599 |
-
|
1600 |
-
|
1601 |
-
|
1602 |
-
|
1603 |
-
|
1604 |
-
|
1605 |
-
|
1606 |
-
|
1607 |
-
|
1608 |
-
|
1609 |
-
|
1610 |
-
|
1611 |
-
|
1612 |
-
|
1613 |
-
|
1614 |
-
|
1615 |
-
|
1616 |
-
|
1617 |
-
|
1618 |
-
|
1619 |
-
|
1620 |
-
|
1621 |
-
|
1622 |
-
|
1623 |
-
|
1624 |
-
|
1625 |
-
|
1626 |
-
|
1627 |
-
|
1628 |
-
|
1629 |
-
|
1630 |
-
|
1631 |
-
|
1632 |
-
|
1633 |
-
|
1634 |
-
|
1635 |
-
|
1636 |
-
|
1637 |
-
|
1638 |
-
|
1639 |
-
|
1640 |
-
|
1641 |
-
|
1642 |
-
|
1643 |
-
|
1644 |
-
|
1645 |
-
|
1646 |
-
|
1647 |
-
|
1648 |
-
|
1649 |
-
|
1650 |
-
|
1651 |
-
|
1652 |
-
|
1653 |
-
|
1654 |
-
|
1655 |
-
|
1656 |
-
|
1657 |
-
|
1658 |
-
|
1659 |
-
|
1660 |
-
|
1661 |
-
|
1662 |
-
|
1663 |
-
|
1664 |
-
|
1665 |
-
|
1666 |
-
|
1667 |
-
|
1668 |
-
|
1669 |
-
|
1670 |
-
|
1671 |
-
|
1672 |
-
|
1673 |
-
|
1674 |
-
|
1675 |
-
|
1676 |
-
|
1677 |
-
|
1678 |
-
|
1679 |
-
|
1680 |
-
|
1681 |
-
|
1682 |
-
|
1683 |
-
|
1684 |
-
|
1685 |
-
|
1686 |
-
|
1687 |
-
|
1688 |
-
|
1689 |
-
accelerator.end_training()
|
1690 |
|
1691 |
|
1692 |
if __name__ == "__main__":
|
|
|
750 |
else:
|
751 |
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
752 |
|
753 |
+
|
754 |
+
|
755 |
# 2. Initialize the accelerator
|
756 |
# We will let the accelerator handle device placement for us in this example
|
757 |
# We simply have to specify the training precision and any trackers being used
|
758 |
# We'll use the same dtype arguments as our JAX/Flax training script and convert
|
759 |
# it to accelerate format
|
760 |
+
|
761 |
if training_args.dtype == "float16":
|
762 |
mixed_precision = "fp16"
|
763 |
teacher_dtype = torch.float16
|
|
|
1010 |
)
|
1011 |
else:
|
1012 |
is_multilingual = False
|
1013 |
+
|
1014 |
+
print(f" is_multilingual : {is_multilingual}")
|
1015 |
|
1016 |
# 8. Create a single speech processor - make sure all processes wait until data is saved
|
1017 |
+
# if accelerator.is_main_process:
|
1018 |
+
# feature_extractor.save_pretrained(training_args.output_dir)
|
1019 |
+
# tokenizer.save_pretrained(training_args.output_dir)
|
1020 |
+
# # save the config and generation config as well
|
1021 |
+
# config.save_pretrained(training_args.output_dir)
|
1022 |
+
# student_model.generation_config.save_pretrained(training_args.output_dir)
|
1023 |
+
|
1024 |
+
# accelerator.wait_for_everyone()
|
1025 |
+
# processor = WhisperProcessor.from_pretrained(training_args.output_dir)
|
1026 |
+
|
1027 |
+
# # 9. Resample speech dataset: `datasets` takes care of automatically loading and resampling the audio,
|
1028 |
+
# # so we just need to set the correct target sampling rate.
|
1029 |
+
# sampling_rate = feature_extractor.sampling_rate
|
1030 |
+
# raw_datasets = raw_datasets.cast_column(
|
1031 |
+
# data_args.audio_column_name,
|
1032 |
+
# datasets.features.Audio(sampling_rate=sampling_rate),
|
1033 |
+
# )
|
1034 |
+
|
1035 |
+
# # 10. Preprocessing the datasets: we need to read the audio files as arrays and tokenize the targets.
|
1036 |
+
# # 10.1: Define the pre-processing constants
|
1037 |
+
# max_input_length = int(data_args.max_duration_in_seconds * sampling_rate)
|
1038 |
+
# min_input_length = int(data_args.min_duration_in_seconds * sampling_rate)
|
1039 |
+
# max_label_length = (
|
1040 |
+
# data_args.max_label_length if data_args.max_label_length is not None else student_model.config.max_length
|
1041 |
+
# )
|
1042 |
+
|
1043 |
+
# timestamp_probability = data_args.timestamp_probability
|
1044 |
+
# condition_on_prev_probability = data_args.condition_on_prev_probability
|
1045 |
+
# return_timestamps = data_args.return_timestamps if timestamp_probability > 0 else False
|
1046 |
+
|
1047 |
+
# timestamp_ids = tokenizer.timestamp_ids()
|
1048 |
+
# timestamp_begin = tokenizer.all_special_ids[-1]
|
1049 |
+
# timestamp_position = 3 if is_multilingual else 1
|
1050 |
+
|
1051 |
+
# decoder_start_token_id = student_model.config.decoder_start_token_id # <|startoftranscript|>
|
1052 |
+
# decoder_prev_token_id = tokenizer.all_special_ids[-3] # <|startofprev|>
|
1053 |
+
# prompt_cutoff_length = max_label_length // 2
|
1054 |
+
|
1055 |
+
# num_workers = data_args.preprocessing_num_workers
|
1056 |
+
# dataloader_num_workers = training_args.dataloader_num_workers
|
1057 |
+
# prefetch_factor = training_args.dataloader_prefetch_factor
|
1058 |
+
|
1059 |
+
# metric = evaluate.load("wer")
|
1060 |
+
# normalizer = (
|
1061 |
+
# BasicTextNormalizer()
|
1062 |
+
# if data_args.language is not None
|
1063 |
+
# else EnglishTextNormalizer(tokenizer.english_spelling_normalizer)
|
1064 |
+
# )
|
1065 |
+
# wer_threshold = data_args.wer_threshold
|
1066 |
+
# use_pseudo_labels = data_args.use_pseudo_labels
|
1067 |
+
# train_text_column_name = "whisper_transcript" if use_pseudo_labels else "text"
|
1068 |
+
|
1069 |
+
# # 10.2: filter based on maximum number of training/evaluation samples
|
1070 |
+
# if training_args.do_train and data_args.max_train_samples is not None:
|
1071 |
+
# raw_datasets["train"] = (
|
1072 |
+
# raw_datasets["train"].take(data_args.max_train_samples)
|
1073 |
+
# if data_args.streaming
|
1074 |
+
# else raw_datasets["train"].select(range(data_args.max_train_samples))
|
1075 |
+
# )
|
1076 |
+
|
1077 |
+
# if training_args.do_eval and data_args.max_eval_samples is not None:
|
1078 |
+
# for eval_split in all_eval_splits:
|
1079 |
+
# raw_datasets[eval_split] = (
|
1080 |
+
# raw_datasets[eval_split].take(data_args.max_eval_samples)
|
1081 |
+
# if data_args.streaming
|
1082 |
+
# else raw_datasets[eval_split].select(range(data_args.max_eval_samples))
|
1083 |
+
# )
|
1084 |
+
|
1085 |
+
# # 10.3: filter training data based on WER threshold -> this is KEY to good distillation performance
|
1086 |
+
# def is_wer_in_range(ground_truth, whisper_transcript):
|
1087 |
+
# norm_ground_truth = normalizer(ground_truth)
|
1088 |
+
# if whisper_transcript is not None and whisper_transcript.upper() == whisper_transcript:
|
1089 |
+
# # filter entirely upper-case transcriptions: these are erroneous generations from large-v3
|
1090 |
+
# return False
|
1091 |
+
# elif len(norm_ground_truth) > 0 and whisper_transcript is not None:
|
1092 |
+
# norm_whisper_transcript = normalizer(whisper_transcript)
|
1093 |
+
# wer = 100 * metric.compute(predictions=[norm_whisper_transcript], references=[norm_ground_truth])
|
1094 |
+
# return wer < wer_threshold
|
1095 |
+
# else:
|
1096 |
+
# # filter automatically since we can't know the WER
|
1097 |
+
# return False
|
1098 |
+
|
1099 |
+
# filter_by_wer_threshold = partial(
|
1100 |
+
# raw_datasets["train"].filter,
|
1101 |
+
# function=is_wer_in_range,
|
1102 |
+
# input_columns=["text", "whisper_transcript"],
|
1103 |
+
# )
|
1104 |
+
|
1105 |
+
# if wer_threshold is not None and use_pseudo_labels:
|
1106 |
+
# with accelerator.main_process_first():
|
1107 |
+
# raw_datasets["train"] = (
|
1108 |
+
# filter_by_wer_threshold(num_proc=num_workers, desc="filtering train dataset by wer")
|
1109 |
+
# if not data_args.streaming
|
1110 |
+
# else filter_by_wer_threshold()
|
1111 |
+
# )
|
1112 |
+
|
1113 |
+
# # 10.4: pre-process training/evaluation datasets
|
1114 |
+
# def prepare_train_dataset(batch):
|
1115 |
+
# """
|
1116 |
+
# Pre-process the raw dataset in a three stage process:
|
1117 |
+
# 1. Convert the audio arrays to log-mel spectrogram inputs
|
1118 |
+
# 2. Possibly filter the timestamp tokens from the token ids (depending on the timestamp probability)
|
1119 |
+
# 3. Possibly add prompt tokens if conditioning on previous text (depending on the conditioning probability)
|
1120 |
+
# """
|
1121 |
+
# # process audio input
|
1122 |
+
# audio = [sample["array"] for sample in batch["audio"]]
|
1123 |
+
# inputs = feature_extractor(audio, sampling_rate=sampling_rate)
|
1124 |
+
# batch["input_features"] = inputs.input_features
|
1125 |
+
# batch["input_length"] = [len(sample) for sample in audio]
|
1126 |
+
|
1127 |
+
# # process text targets - for training these are the Whisper-generated pseudo-labels
|
1128 |
+
# input_str_batched = batch[train_text_column_name]
|
1129 |
+
# condition_on_prev_batched = batch.get("condition_on_prev", len(input_str_batched) * [None])
|
1130 |
+
|
1131 |
+
# all_token_ids = []
|
1132 |
+
# all_token_ids_unprompted = []
|
1133 |
+
# for prev_ids, input_str in zip(condition_on_prev_batched, input_str_batched):
|
1134 |
+
# token_ids = tokenizer(input_str, add_special_tokens=not use_pseudo_labels).input_ids
|
1135 |
+
|
1136 |
+
# # check whether we have timestamps in the PLs and filter if required
|
1137 |
+
# has_timestamps = len(set(token_ids) & set(timestamp_ids)) > 0
|
1138 |
+
# if has_timestamps:
|
1139 |
+
# # sample from binomial distribution to get probability of training on timestamps
|
1140 |
+
# predict_timestamps = bool(np.random.binomial(1, timestamp_probability))
|
1141 |
+
# if not predict_timestamps:
|
1142 |
+
# # filter timestamps and insert the <|notimestamps|> task token
|
1143 |
+
# token_ids = [token for token in token_ids if token < timestamp_begin]
|
1144 |
+
# token_ids.insert(timestamp_position, timestamp_begin)
|
1145 |
+
|
1146 |
+
# all_token_ids_unprompted.append(token_ids)
|
1147 |
+
# # check whether to condition on previous text - we do this with probability condition_on_prev_probability
|
1148 |
+
# condition_on_prev = bool(np.random.binomial(1, condition_on_prev_probability))
|
1149 |
+
# if not condition_on_prev:
|
1150 |
+
# prev_ids = None
|
1151 |
+
# elif "condition_on_prev" not in batch and len(all_token_ids_unprompted) > 1:
|
1152 |
+
# # prompt ids are the penultimate token ids in the batch
|
1153 |
+
# prev_ids = all_token_ids_unprompted[-2]
|
1154 |
+
|
1155 |
+
# if prev_ids is not None:
|
1156 |
+
# if has_timestamps and not predict_timestamps:
|
1157 |
+
# # filter timestamp ids from prompt when not predicting timestamps
|
1158 |
+
# prev_ids = [token for token in prev_ids if token < timestamp_begin]
|
1159 |
+
|
1160 |
+
# # check that the length of the prompt does not exceed more than half the max label length (224)
|
1161 |
+
# if len(prev_ids) > prompt_cutoff_length:
|
1162 |
+
# prev_ids = prev_ids[-prompt_cutoff_length + 1 :]
|
1163 |
+
# prev_ids = [decoder_prev_token_id] + prev_ids
|
1164 |
+
|
1165 |
+
# # and that the total length of the labels does not exceed the max label length (448)
|
1166 |
+
# if len(prev_ids + token_ids) > max_label_length:
|
1167 |
+
# trim_length = len(prev_ids + token_ids) - max_label_length + 1
|
1168 |
+
# prev_ids = prev_ids[trim_length:]
|
1169 |
+
# prev_ids = [decoder_prev_token_id] + prev_ids
|
1170 |
+
|
1171 |
+
# token_ids = prev_ids + token_ids
|
1172 |
+
|
1173 |
+
# all_token_ids.append(token_ids)
|
1174 |
+
|
1175 |
+
# batch["labels"] = all_token_ids
|
1176 |
+
# return batch
|
1177 |
+
|
1178 |
+
# def prepare_eval_dataset(batch):
|
1179 |
+
# # process audio input
|
1180 |
+
# sample = batch["audio"]
|
1181 |
+
# inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
|
1182 |
+
# batch["input_features"] = inputs.input_features[0]
|
1183 |
+
# batch["input_length"] = len(sample["array"])
|
1184 |
+
|
1185 |
+
# # process targets - for evaluation these are the ground-truth transcriptions
|
1186 |
+
# input_str = batch["text"]
|
1187 |
+
# batch["labels"] = tokenizer(input_str).input_ids
|
1188 |
+
# return batch
|
1189 |
+
|
1190 |
+
# vectorized_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
|
1191 |
+
# if training_args.do_train:
|
1192 |
+
# # with streaming mode we can only have 1 worker, whereas with non-streaming
|
1193 |
+
# # we can use `num_workers` (which is much faster)
|
1194 |
+
# # We gate the pre-processing function accordingly
|
1195 |
+
# map_fn_train = partial(
|
1196 |
+
# raw_datasets["train"].map,
|
1197 |
+
# function=prepare_train_dataset,
|
1198 |
+
# remove_columns=raw_datasets_train_features,
|
1199 |
+
# batched=True,
|
1200 |
+
# batch_size=data_args.preprocessing_batch_size,
|
1201 |
+
# )
|
1202 |
+
# with accelerator.main_process_first():
|
1203 |
+
# vectorized_datasets["train"] = (
|
1204 |
+
# map_fn_train(num_proc=num_workers, desc="preprocess train dataset")
|
1205 |
+
# if not data_args.streaming
|
1206 |
+
# else map_fn_train()
|
1207 |
+
# )
|
1208 |
+
# if training_args.do_eval:
|
1209 |
+
# for eval_split in all_eval_splits:
|
1210 |
+
# raw_datasets_eval_features = list(raw_datasets[eval_split].features.keys())
|
1211 |
+
# map_fn_eval = partial(
|
1212 |
+
# raw_datasets[eval_split].map, function=prepare_eval_dataset, remove_columns=raw_datasets_eval_features
|
1213 |
+
# )
|
1214 |
+
# with accelerator.main_process_first():
|
1215 |
+
# vectorized_datasets[eval_split] = (
|
1216 |
+
# map_fn_eval(num_proc=num_workers, desc="preprocess eval dataset")
|
1217 |
+
# if not data_args.streaming
|
1218 |
+
# else map_fn_eval()
|
1219 |
+
# )
|
1220 |
+
|
1221 |
+
# # 10.5: Filter training data with inputs longer than `max_input_length`
|
1222 |
+
# def is_audio_in_length_range(length):
|
1223 |
+
# return min_input_length < length < max_input_length
|
1224 |
+
|
1225 |
+
# filter_by_audio_fn = partial(
|
1226 |
+
# vectorized_datasets.filter, function=is_audio_in_length_range, input_columns=["input_length"]
|
1227 |
+
# )
|
1228 |
+
# with accelerator.main_process_first():
|
1229 |
+
# vectorized_datasets = (
|
1230 |
+
# filter_by_audio_fn(num_proc=num_workers, desc="filtering train dataset by audio length")
|
1231 |
+
# if not data_args.streaming
|
1232 |
+
# else filter_by_audio_fn()
|
1233 |
+
# )
|
1234 |
+
|
1235 |
+
# # 10.6: Filter training data with labels longer than `max_label_length`
|
1236 |
+
# def is_labels_in_length_range(labels):
|
1237 |
+
# return 0 < len(labels) <= max_label_length
|
1238 |
+
|
1239 |
+
# filter_by_labels_fn = partial(
|
1240 |
+
# vectorized_datasets.filter, function=is_labels_in_length_range, input_columns=["labels"]
|
1241 |
+
# )
|
1242 |
+
# with accelerator.main_process_first():
|
1243 |
+
# vectorized_datasets = (
|
1244 |
+
# filter_by_labels_fn(num_proc=num_workers, desc="filtering train dataset")
|
1245 |
+
# if not data_args.streaming
|
1246 |
+
# else filter_by_labels_fn()
|
1247 |
+
# )
|
1248 |
+
|
1249 |
+
# # Pre-processing complete!
|
1250 |
+
# # For large datasets it is advised to run the preprocessing on a
|
1251 |
+
# # single machine first with `--preprocessing_only` since there will mostly likely
|
1252 |
+
# # be a timeout when running the script in distributed mode.
|
1253 |
+
# # In a second step, `--preprocessing_only` can then be set to `False` to load the
|
1254 |
+
# # cached dataset
|
1255 |
+
# if data_args.preprocessing_only:
|
1256 |
+
# if data_args.streaming:
|
1257 |
+
# raise ValueError(
|
1258 |
+
# "When using streaming mode, dataset pre-processing is performed on the fly, hence there is no notion"
|
1259 |
+
# "of a cached pre-processed dataset. Remove the argument `--preprocessing_only` to run pre-processing "
|
1260 |
+
# "on the fly with streaming mode."
|
1261 |
+
# )
|
1262 |
+
# cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
|
1263 |
+
# logger.info(f"Data preprocessing finished. Files cached at {cache}.")
|
1264 |
+
# return
|
1265 |
+
|
1266 |
+
# # 11. Define Evaluation Metrics
|
1267 |
+
# def compute_metrics(preds, labels):
|
1268 |
+
# # replace padded labels by the padding token
|
1269 |
+
# for idx in range(len(labels)):
|
1270 |
+
# labels[idx][labels[idx] == -100] = tokenizer.pad_token_id
|
1271 |
+
|
1272 |
+
# pred_str = tokenizer.batch_decode(preds, skip_special_tokens=True, decode_with_timestamps=return_timestamps)
|
1273 |
+
# # we do not want to group tokens when computing the metrics
|
1274 |
+
# label_str = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
1275 |
+
# wer_ortho = 100 * metric.compute(predictions=pred_str, references=label_str)
|
1276 |
+
|
1277 |
+
# # normalize everything and re-compute the WER
|
1278 |
+
# norm_pred_str = [normalizer(pred) for pred in pred_str]
|
1279 |
+
# norm_label_str = [normalizer(label) for label in label_str]
|
1280 |
+
# # for logging, we need the pred/labels to match the norm_pred/norm_labels, so discard any filtered samples here
|
1281 |
+
# pred_str = [pred_str[i] for i in range(len(norm_pred_str)) if len(norm_label_str[i]) > 0]
|
1282 |
+
# label_str = [label_str[i] for i in range(len(norm_label_str)) if len(norm_label_str[i]) > 0]
|
1283 |
+
# # filtering step to only evaluate the samples that correspond to non-zero normalized references:
|
1284 |
+
# norm_pred_str = [norm_pred_str[i] for i in range(len(norm_pred_str)) if len(norm_label_str[i]) > 0]
|
1285 |
+
# norm_label_str = [norm_label_str[i] for i in range(len(norm_label_str)) if len(norm_label_str[i]) > 0]
|
1286 |
+
|
1287 |
+
# wer = 100 * metric.compute(predictions=norm_pred_str, references=norm_label_str)
|
1288 |
+
# return {"wer": wer, "wer_ortho": wer_ortho}, pred_str, label_str, norm_pred_str, norm_label_str
|
1289 |
+
|
1290 |
+
# # 12. Define Training Schedule
|
1291 |
+
# # Store some constants
|
1292 |
+
# per_device_train_batch_size = int(training_args.per_device_train_batch_size)
|
1293 |
+
# train_batch_size = per_device_train_batch_size * accelerator.num_processes
|
1294 |
+
# gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
|
1295 |
+
# per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
|
1296 |
+
|
1297 |
+
# if not data_args.streaming and training_args.max_steps < 0:
|
1298 |
+
# num_epochs = int(training_args.num_train_epochs)
|
1299 |
+
# steps_per_epoch = len(vectorized_datasets["train"]) // (train_batch_size * gradient_accumulation_steps)
|
1300 |
+
# total_train_steps = steps_per_epoch * num_epochs
|
1301 |
+
# elif training_args.max_steps > 0:
|
1302 |
+
# logger.info("max_steps is given, it will override any value given in num_train_epochs")
|
1303 |
+
# total_train_steps = int(training_args.max_steps)
|
1304 |
+
# if not data_args.streaming:
|
1305 |
+
# steps_per_epoch = len(vectorized_datasets["train"]) // (train_batch_size * gradient_accumulation_steps)
|
1306 |
+
# num_epochs = int(np.ceil(total_train_steps / steps_per_epoch))
|
1307 |
+
# else:
|
1308 |
+
# # Setting a very large number of epochs so we go as many times as necessary over the iterator.
|
1309 |
+
# num_epochs = sys.maxsize
|
1310 |
+
# steps_per_epoch = total_train_steps
|
1311 |
+
# else:
|
1312 |
+
# raise ValueError("max_steps must be specified when training with a streaming (iterable) dataset")
|
1313 |
+
|
1314 |
+
# if training_args.eval_steps is None:
|
1315 |
+
# logger.info(
|
1316 |
+
# f"eval_steps is not set, evaluating at the end of {'each epoch' if not data_args.streaming else 'training'}"
|
1317 |
+
# )
|
1318 |
+
# eval_steps = steps_per_epoch
|
1319 |
+
# else:
|
1320 |
+
# eval_steps = training_args.eval_steps
|
1321 |
+
|
1322 |
+
# # 13. Define optimizer, LR scheduler, collator
|
1323 |
+
# decay_parameters = get_parameter_names(
|
1324 |
+
# student_model,
|
1325 |
+
# [nn.LayerNorm],
|
1326 |
+
# forbidden_module=[student_model.model.encoder] if training_args.freeze_encoder else None,
|
1327 |
+
# )
|
1328 |
+
# decay_parameters = [name for name in decay_parameters if "bias" not in name]
|
1329 |
+
# optimizer_grouped_parameters = [
|
1330 |
+
# {
|
1331 |
+
# "params": [param for name, param in student_model.named_parameters() if name in decay_parameters],
|
1332 |
+
# "weight_decay": training_args.weight_decay,
|
1333 |
+
# },
|
1334 |
+
# {
|
1335 |
+
# "params": [param for name, param in student_model.named_parameters() if name not in decay_parameters],
|
1336 |
+
# "weight_decay": 0.0,
|
1337 |
+
# },
|
1338 |
+
# ]
|
1339 |
+
# optimizer = torch.optim.AdamW(
|
1340 |
+
# params=optimizer_grouped_parameters,
|
1341 |
+
# lr=training_args.learning_rate,
|
1342 |
+
# betas=(training_args.adam_beta1, training_args.adam_beta2),
|
1343 |
+
# eps=training_args.adam_epsilon,
|
1344 |
+
# )
|
1345 |
+
|
1346 |
+
# # LR scheduler gets stepped by `num_processes` each time -> account for this in warmup / total steps
|
1347 |
+
# lr_scheduler = get_scheduler(
|
1348 |
+
# name=training_args.lr_scheduler_type,
|
1349 |
+
# optimizer=optimizer,
|
1350 |
+
# num_warmup_steps=training_args.warmup_steps * accelerator.num_processes,
|
1351 |
+
# num_training_steps=total_train_steps * accelerator.num_processes,
|
1352 |
+
# )
|
1353 |
+
|
1354 |
+
# data_collator = DataCollatorSpeechSeq2SeqWithPadding(
|
1355 |
+
# processor=processor,
|
1356 |
+
# decoder_start_token_id=decoder_start_token_id,
|
1357 |
+
# decoder_prev_token_id=decoder_prev_token_id,
|
1358 |
+
# input_padding="longest",
|
1359 |
+
# target_padding="max_length",
|
1360 |
+
# max_target_length=max_label_length,
|
1361 |
+
# )
|
1362 |
+
|
1363 |
+
# # 14. Define generation arguments - we need to do this before we wrap the models in DDP
|
1364 |
+
# # so that we can still access the configs
|
1365 |
+
# num_beams = (
|
1366 |
+
# training_args.generation_num_beams
|
1367 |
+
# if training_args.generation_num_beams is not None
|
1368 |
+
# else getattr(student_model.generation_config, "num_beams", 1)
|
1369 |
+
# )
|
1370 |
+
|
1371 |
+
# gen_kwargs = {
|
1372 |
+
# "max_length": max_label_length,
|
1373 |
+
# "num_beams": num_beams,
|
1374 |
+
# "return_timestamps": return_timestamps,
|
1375 |
+
# }
|
1376 |
+
# if is_multilingual:
|
1377 |
+
# # forcing the language and task tokens helps multilingual models in their generations
|
1378 |
+
# gen_kwargs.update(
|
1379 |
+
# {
|
1380 |
+
# "language": data_args.language,
|
1381 |
+
# "task": data_args.task,
|
1382 |
+
# }
|
1383 |
+
# )
|
1384 |
+
|
1385 |
+
# # 15. Prepare everything with accelerate
|
1386 |
+
# student_model, teacher_model, optimizer, lr_scheduler = accelerator.prepare(
|
1387 |
+
# student_model, teacher_model, optimizer, lr_scheduler
|
1388 |
+
# )
|
1389 |
+
|
1390 |
+
# def kl_divergence(target_distribution, log_predicted_distribution, labels):
|
1391 |
+
# kl_loss = nn.KLDivLoss(reduction="none")
|
1392 |
+
# divergence = kl_loss(log_predicted_distribution, target_distribution)
|
1393 |
+
# # ignore padded tokens from divergence, i.e. where labels are not set to -100
|
1394 |
+
# padding_mask = labels >= 0
|
1395 |
+
# padding_mask = padding_mask.unsqueeze(-1)
|
1396 |
+
# divergence = divergence * padding_mask
|
1397 |
+
# # take the average over the mini-batch
|
1398 |
+
# divergence = divergence.sum() / padding_mask.sum()
|
1399 |
+
# return divergence
|
1400 |
+
|
1401 |
+
# # Define gradient update step fn
|
1402 |
+
# def train_step(
|
1403 |
+
# batch,
|
1404 |
+
# temperature=2.0,
|
1405 |
+
# ):
|
1406 |
+
# student_model.train()
|
1407 |
+
# teacher_model.eval()
|
1408 |
+
|
1409 |
+
# student_outputs = student_model(**batch)
|
1410 |
+
# with torch.no_grad():
|
1411 |
+
# if share_hidden_states:
|
1412 |
+
# # if the student and teacher share the same frozen encoder then we don't have to recompute the
|
1413 |
+
# # encoder hidden-states for the teacher model, we can just re-use from the student
|
1414 |
+
# encoder_outputs = BaseModelOutput(student_outputs.encoder_last_hidden_state.to(dtype=teacher_dtype))
|
1415 |
+
# teacher_outputs = teacher_model(encoder_outputs=encoder_outputs, labels=batch["labels"])
|
1416 |
+
# else:
|
1417 |
+
# # do the full forward pass for the teacher model (encoder + decoder)
|
1418 |
+
# teacher_outputs = teacher_model(**batch)
|
1419 |
+
|
1420 |
+
# # CE (data) loss
|
1421 |
+
# ce_loss = student_outputs.loss
|
1422 |
+
# # rescale distribution by temperature to ensure gradients scale correctly
|
1423 |
+
# teacher_distribution = nn.functional.softmax(teacher_outputs.logits / temperature, dim=-1)
|
1424 |
+
# # log softmax of student predictions for numerical stability
|
1425 |
+
# student_distribution = nn.functional.log_softmax(student_outputs.logits / temperature, dim=-1)
|
1426 |
+
# # KL-divergence loss (scaled by temperature)
|
1427 |
+
# kl_loss = kl_divergence(teacher_distribution, student_distribution, batch["labels"]) * temperature**2
|
1428 |
+
|
1429 |
+
# # use Distil-Whisper formulation (fix weight of CE loss and tune KL weight)
|
1430 |
+
# loss = 0.8 * ce_loss + training_args.kl_weight * kl_loss
|
1431 |
+
# metrics = {"loss": loss, "ce_loss": ce_loss, "kl_loss": kl_loss}
|
1432 |
+
# return loss, metrics
|
1433 |
+
|
1434 |
+
# # Define eval fn
|
1435 |
+
# def eval_step(batch):
|
1436 |
+
# student_model.eval()
|
1437 |
+
# teacher_model.eval()
|
1438 |
+
|
1439 |
+
# with torch.no_grad():
|
1440 |
+
# student_outputs = student_model(**batch)
|
1441 |
+
# if share_hidden_states:
|
1442 |
+
# encoder_outputs = BaseModelOutput(student_outputs.encoder_last_hidden_state.to(dtype=teacher_dtype))
|
1443 |
+
# teacher_outputs = teacher_model(encoder_outputs=encoder_outputs, labels=batch["labels"])
|
1444 |
+
# else:
|
1445 |
+
# teacher_outputs = teacher_model(**batch)
|
1446 |
+
|
1447 |
+
# # CE (data) loss
|
1448 |
+
# ce_loss = student_outputs.loss
|
1449 |
+
|
1450 |
+
# # log softmax / softmax for numerical stability
|
1451 |
+
# student_distribution = nn.functional.log_softmax(student_outputs.logits, dim=-1)
|
1452 |
+
# teacher_distribution = nn.functional.softmax(teacher_outputs.logits, dim=-1)
|
1453 |
+
# # temperature is always 1 for eval
|
1454 |
+
# kl_loss = kl_divergence(teacher_distribution, student_distribution, batch["labels"])
|
1455 |
+
|
1456 |
+
# # use Distil-Whisper formulation (fix weight of CE loss and tune KL weight)
|
1457 |
+
# loss = 0.8 * ce_loss + training_args.kl_weight * kl_loss
|
1458 |
+
# metrics = {"loss": loss, "ce_loss": ce_loss, "kl_loss": kl_loss}
|
1459 |
+
# return metrics
|
1460 |
+
|
1461 |
+
# def generate_step(batch):
|
1462 |
+
# student_model.eval()
|
1463 |
+
# output_ids = accelerator.unwrap_model(student_model).generate(batch["input_features"], **gen_kwargs)
|
1464 |
+
# output_ids = accelerator.pad_across_processes(output_ids, dim=1, pad_index=tokenizer.pad_token_id)
|
1465 |
+
# return output_ids
|
1466 |
+
|
1467 |
+
# logger.info("***** Running training *****")
|
1468 |
+
# logger.info(f" Num examples = {total_train_steps * train_batch_size * gradient_accumulation_steps}")
|
1469 |
+
# if not data_args.streaming:
|
1470 |
+
# logger.info(f" Num epochs = {num_epochs}")
|
1471 |
+
# logger.info(" Instantaneous batch size per device =" f" {training_args.per_device_train_batch_size}")
|
1472 |
+
# logger.info(" Gradient accumulation steps =" f" {gradient_accumulation_steps}")
|
1473 |
+
# logger.info(
|
1474 |
+
# f" Total train batch size (w. parallel & distributed) = {train_batch_size * gradient_accumulation_steps}"
|
1475 |
+
# )
|
1476 |
+
# logger.info(f" Total optimization steps = {total_train_steps}")
|
1477 |
+
|
1478 |
+
# # ======================== Training ================================
|
1479 |
+
# train_time = 0
|
1480 |
+
# train_start = time.time()
|
1481 |
+
# steps_trained_progress_bar = tqdm(
|
1482 |
+
# range(total_train_steps), desc="Train steps ... ", position=0, disable=not accelerator.is_local_main_process
|
1483 |
+
# )
|
1484 |
+
# continue_training = True
|
1485 |
+
# epochs_trained = 0
|
1486 |
+
# cur_step = 0
|
1487 |
+
|
1488 |
+
# checkpoint = None
|
1489 |
+
# if training_args.resume_from_checkpoint is not None:
|
1490 |
+
# checkpoint = training_args.resume_from_checkpoint
|
1491 |
+
# elif last_checkpoint is not None:
|
1492 |
+
# checkpoint = last_checkpoint
|
1493 |
+
|
1494 |
+
# if checkpoint is not None:
|
1495 |
+
# accelerator.load_state(checkpoint)
|
1496 |
+
# # Find num steps and epoch from saved state string pattern
|
1497 |
+
# pattern = r"checkpoint-(\d+)-epoch-(\d+)"
|
1498 |
+
# match = re.search(pattern, checkpoint)
|
1499 |
+
# cur_step = int(match.group(1))
|
1500 |
+
# epochs_trained = int(match.group(2))
|
1501 |
+
|
1502 |
+
# logger.info(" Continuing training from checkpoint, will skip to saved global_step")
|
1503 |
+
# logger.info(f" Continuing training from epoch {epochs_trained}")
|
1504 |
+
# logger.info(f" Continuing training from global step {cur_step}")
|
1505 |
+
|
1506 |
+
# steps_trained_progress_bar.update(cur_step)
|
1507 |
+
|
1508 |
+
# for epoch in range(0, epochs_trained):
|
1509 |
+
# vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
|
1510 |
+
|
1511 |
+
# if not data_args.streaming and training_args.max_steps < 0:
|
1512 |
+
# # we know exactly the number of steps per epoch, so can skip through the required number of batches
|
1513 |
+
# resume_step = (cur_step - epochs_trained * steps_per_epoch) * gradient_accumulation_steps
|
1514 |
+
# else:
|
1515 |
+
# # Currently we don't know how many steps we've taken in the current epoch
|
1516 |
+
# # So we just shuffle the dataset one extra time and start from a fresh epoch
|
1517 |
+
# # This is "good enough" for our purposes but not fully correct
|
1518 |
+
# resume_step = None
|
1519 |
+
# vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
|
1520 |
+
# else:
|
1521 |
+
# resume_step = None
|
1522 |
+
|
1523 |
+
# for epoch in range(epochs_trained, num_epochs):
|
1524 |
+
# vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
|
1525 |
+
# train_dataloader = DataLoader(
|
1526 |
+
# vectorized_datasets["train"],
|
1527 |
+
# collate_fn=data_collator,
|
1528 |
+
# batch_size=per_device_train_batch_size,
|
1529 |
+
# num_workers=dataloader_num_workers,
|
1530 |
+
# prefetch_factor=prefetch_factor,
|
1531 |
+
# pin_memory=training_args.dataloader_pin_memory,
|
1532 |
+
# )
|
1533 |
+
# train_dataloader = accelerator.prepare(train_dataloader)
|
1534 |
+
# if hasattr(train_dataloader, "dataset") and isinstance(train_dataloader.dataset, IterableDataset):
|
1535 |
+
# train_dataloader.dataset.set_epoch(epoch)
|
1536 |
+
|
1537 |
+
# if resume_step is not None:
|
1538 |
+
# # Skip the first N batches in the dataloader when resuming from a checkpoint
|
1539 |
+
# train_dataloader = accelerator.skip_first_batches(train_dataloader, resume_step)
|
1540 |
+
# resume_step = None
|
1541 |
+
|
1542 |
+
# for batch in train_dataloader:
|
1543 |
+
# with accelerator.accumulate(student_model):
|
1544 |
+
# loss, train_metric = train_step(batch, temperature=training_args.temperature)
|
1545 |
+
# accelerator.backward(loss)
|
1546 |
+
# if accelerator.sync_gradients:
|
1547 |
+
# accelerator.clip_grad_norm_(student_model.parameters(), training_args.max_grad_norm)
|
1548 |
+
# optimizer.step()
|
1549 |
+
# lr_scheduler.step()
|
1550 |
+
# optimizer.zero_grad()
|
1551 |
+
|
1552 |
+
# # Check if the accelerator has performed an optimization step behind the scenes
|
1553 |
+
# if accelerator.sync_gradients:
|
1554 |
+
# steps_trained_progress_bar.update(1)
|
1555 |
+
# cur_step += 1
|
1556 |
+
|
1557 |
+
# if cur_step % training_args.logging_steps == 0:
|
1558 |
+
# steps_trained_progress_bar.write(
|
1559 |
+
# f"Step... ({cur_step} / {total_train_steps} | Loss:"
|
1560 |
+
# f" {train_metric['loss']}, Learning Rate:"
|
1561 |
+
# f" {lr_scheduler.get_last_lr()[0]})"
|
1562 |
+
# )
|
1563 |
+
# log_metric(
|
1564 |
+
# accelerator,
|
1565 |
+
# metrics=train_metric,
|
1566 |
+
# learning_rate=lr_scheduler.get_last_lr()[0],
|
1567 |
+
# train_time=train_time + time.time() - train_start,
|
1568 |
+
# step=cur_step,
|
1569 |
+
# epoch=epoch,
|
1570 |
+
# prefix="train",
|
1571 |
+
# )
|
1572 |
+
|
1573 |
+
# # save checkpoint and weights after each save_steps and at the end of training
|
1574 |
+
# if (cur_step % training_args.save_steps == 0) or cur_step == total_train_steps:
|
1575 |
+
# intermediate_dir = os.path.join(training_args.output_dir, f"checkpoint-{cur_step}-epoch-{epoch}")
|
1576 |
+
# accelerator.save_state(output_dir=intermediate_dir)
|
1577 |
+
# accelerator.wait_for_everyone()
|
1578 |
+
# if accelerator.is_main_process:
|
1579 |
+
# rotate_checkpoints(training_args.save_total_limit, output_dir=training_args.output_dir)
|
1580 |
+
|
1581 |
+
# if training_args.push_to_hub:
|
1582 |
+
# upload_folder(
|
1583 |
+
# folder_path=training_args.output_dir,
|
1584 |
+
# repo_id=repo_name,
|
1585 |
+
# repo_type="model",
|
1586 |
+
# commit_message=f"Saving train state of step {cur_step}",
|
1587 |
+
# )
|
1588 |
+
|
1589 |
+
# if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps):
|
1590 |
+
# train_time += time.time() - train_start
|
1591 |
+
# student_model.eval()
|
1592 |
+
# # ======================== Evaluating ==============================
|
1593 |
+
# for eval_split in all_eval_splits:
|
1594 |
+
# eval_metrics = []
|
1595 |
+
# eval_preds = []
|
1596 |
+
# eval_labels = []
|
1597 |
+
# eval_start = time.time()
|
1598 |
+
|
1599 |
+
# validation_dataloader = DataLoader(
|
1600 |
+
# vectorized_datasets[eval_split],
|
1601 |
+
# collate_fn=data_collator,
|
1602 |
+
# batch_size=per_device_eval_batch_size,
|
1603 |
+
# drop_last=False,
|
1604 |
+
# num_workers=dataloader_num_workers,
|
1605 |
+
# prefetch_factor=prefetch_factor,
|
1606 |
+
# pin_memory=training_args.dataloader_pin_memory,
|
1607 |
+
# )
|
1608 |
+
# validation_dataloader = accelerator.prepare(validation_dataloader)
|
1609 |
+
|
1610 |
+
# for batch in tqdm(
|
1611 |
+
# validation_dataloader,
|
1612 |
+
# desc=f"Evaluating {eval_split}...",
|
1613 |
+
# position=2,
|
1614 |
+
# disable=not accelerator.is_local_main_process,
|
1615 |
+
# ):
|
1616 |
+
# # Model forward
|
1617 |
+
# eval_metric = eval_step(batch)
|
1618 |
+
# eval_metric = accelerator.gather_for_metrics(eval_metric)
|
1619 |
+
# eval_metrics.append(eval_metric)
|
1620 |
+
|
1621 |
+
# # generation
|
1622 |
+
# if training_args.predict_with_generate:
|
1623 |
+
# generated_ids = generate_step(batch)
|
1624 |
+
# # Gather all predictions and targets
|
1625 |
+
# generated_ids, labels = accelerator.gather_for_metrics(
|
1626 |
+
# (generated_ids, batch["labels"])
|
1627 |
+
# )
|
1628 |
+
# eval_preds.extend(generated_ids)
|
1629 |
+
# eval_labels.extend(labels)
|
1630 |
+
|
1631 |
+
# eval_time = time.time() - eval_start
|
1632 |
+
# # normalize eval metrics
|
1633 |
+
# eval_metrics = {
|
1634 |
+
# key: torch.mean(torch.stack([d[key] for d in eval_metrics])) for key in eval_metrics[0]
|
1635 |
+
# }
|
1636 |
+
|
1637 |
+
# # compute WER metric
|
1638 |
+
# wer_desc = ""
|
1639 |
+
# if training_args.predict_with_generate:
|
1640 |
+
# wer_metric, pred_str, label_str, norm_pred_str, norm_label_str = compute_metrics(
|
1641 |
+
# eval_preds, eval_labels
|
1642 |
+
# )
|
1643 |
+
# eval_metrics.update(wer_metric)
|
1644 |
+
# wer_desc = " ".join([f"Eval {key}: {value} |" for key, value in wer_metric.items()])
|
1645 |
+
# log_pred(
|
1646 |
+
# accelerator,
|
1647 |
+
# pred_str,
|
1648 |
+
# label_str,
|
1649 |
+
# norm_pred_str,
|
1650 |
+
# norm_label_str,
|
1651 |
+
# step=cur_step,
|
1652 |
+
# prefix=eval_split,
|
1653 |
+
# )
|
1654 |
+
|
1655 |
+
# # Print metrics and update progress bar
|
1656 |
+
# steps_trained_progress_bar.write(
|
1657 |
+
# f"Eval results for step ({cur_step} / {total_train_steps} | Eval Loss: {eval_metrics['loss']} |"
|
1658 |
+
# f" {wer_desc})"
|
1659 |
+
# )
|
1660 |
+
|
1661 |
+
# log_metric(
|
1662 |
+
# accelerator,
|
1663 |
+
# metrics=eval_metrics,
|
1664 |
+
# train_time=eval_time,
|
1665 |
+
# step=cur_step,
|
1666 |
+
# epoch=epoch,
|
1667 |
+
# prefix=eval_split,
|
1668 |
+
# )
|
1669 |
+
|
1670 |
+
# # flush the train metrics
|
1671 |
+
# train_start = time.time()
|
1672 |
+
|
1673 |
+
# # break condition
|
1674 |
+
# if cur_step == total_train_steps:
|
1675 |
+
|
1676 |
+
# # un-wrap student model for save
|
1677 |
+
# student_model = accelerator.unwrap_model(student_model)
|
1678 |
+
# student_model.save_pretrained(training_args.output_dir)
|
1679 |
+
|
1680 |
+
# if training_args.push_to_hub:
|
1681 |
+
# upload_folder(
|
1682 |
+
# folder_path=training_args.output_dir,
|
1683 |
+
# repo_id=repo_name,
|
1684 |
+
# repo_type="model",
|
1685 |
+
# commit_message=f"Saving final weights of step {cur_step}",
|
1686 |
+
# )
|
1687 |
+
|
1688 |
+
# continue_training = False
|
1689 |
+
# break
|
1690 |
+
|
1691 |
+
# if not continue_training:
|
1692 |
+
# break
|
1693 |
+
|
1694 |
+
# accelerator.end_training()
|
1695 |
|
1696 |
|
1697 |
if __name__ == "__main__":
|
__pycache__/evaluate.cpython-39.pyc
ADDED
Binary file (142 Bytes). View file
|
|
distil-whisper/events.out.tfevents.1715057787.server02.1349950.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6fb8763b003a4a4d0209179e68aac6e43453e4693f8cee09cd3a53b74ae1f1fa
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715063050.server02.1368197.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ad00996d03cdb169f2976796433bb5c0f4b367ecbe8b4ae2c0b22e7472f45793
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715063266.server02.1369570.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cfe9e8c42d8e2d53b9e63b1b235b356699c4986e401482da4f033bee21824cbe
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715063402.server02.1370564.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f6d637cc9f5a873ec11177c4509c741f2d4e5a13099a60dc50e722fb95533961
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715063677.server02.1372191.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:55b02873c18707d53d847d8e5b6cb1df617e81977e42349876eb8abc83573afd
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715063742.server02.1372871.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8d8c2457fceadefc49be2acdfd53edc83a1a778537b28c2b81a86b591dc464f8
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715064564.server02.1376229.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d38fabb5cd9df9bbdb7b8ebea98ed65d62d7dd69515c9b1e730923dc12733a10
|
3 |
+
size 88
|
distil-whisper/events.out.tfevents.1715065478.server02.1379863.0
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4306c5a95d18085ad3d309c137786b0a2bbe4bebda405c86b23a6617a11b10a5
|
3 |
+
size 392
|
run_distillation.py
CHANGED
@@ -1010,8 +1010,10 @@ def main():
|
|
1010 |
)
|
1011 |
else:
|
1012 |
is_multilingual = False
|
|
|
|
|
1013 |
|
1014 |
-
#
|
1015 |
if accelerator.is_main_process:
|
1016 |
feature_extractor.save_pretrained(training_args.output_dir)
|
1017 |
tokenizer.save_pretrained(training_args.output_dir)
|
@@ -1379,8 +1381,8 @@ def main():
|
|
1379 |
"task": data_args.task,
|
1380 |
}
|
1381 |
)
|
1382 |
-
|
1383 |
-
#
|
1384 |
student_model, teacher_model, optimizer, lr_scheduler = accelerator.prepare(
|
1385 |
student_model, teacher_model, optimizer, lr_scheduler
|
1386 |
)
|
@@ -1485,7 +1487,7 @@ def main():
|
|
1485 |
|
1486 |
checkpoint = None
|
1487 |
if training_args.resume_from_checkpoint is not None:
|
1488 |
-
checkpoint = training_args.resume_from_checkpoint
|
1489 |
elif last_checkpoint is not None:
|
1490 |
checkpoint = last_checkpoint
|
1491 |
|
@@ -1694,3 +1696,6 @@ def main():
|
|
1694 |
|
1695 |
if __name__ == "__main__":
|
1696 |
main()
|
|
|
|
|
|
|
|
1010 |
)
|
1011 |
else:
|
1012 |
is_multilingual = False
|
1013 |
+
|
1014 |
+
print(f" is_multilingual : {is_multilingual}")
|
1015 |
|
1016 |
+
#8. Create a single speech processor - make sure all processes wait until data is saved
|
1017 |
if accelerator.is_main_process:
|
1018 |
feature_extractor.save_pretrained(training_args.output_dir)
|
1019 |
tokenizer.save_pretrained(training_args.output_dir)
|
|
|
1381 |
"task": data_args.task,
|
1382 |
}
|
1383 |
)
|
1384 |
+
print(f" gen_kwargs : {gen_kwargs}")
|
1385 |
+
#15. Prepare everything with accelerate
|
1386 |
student_model, teacher_model, optimizer, lr_scheduler = accelerator.prepare(
|
1387 |
student_model, teacher_model, optimizer, lr_scheduler
|
1388 |
)
|
|
|
1487 |
|
1488 |
checkpoint = None
|
1489 |
if training_args.resume_from_checkpoint is not None:
|
1490 |
+
checkpoint = training_args.resume_from_checkpoint
|
1491 |
elif last_checkpoint is not None:
|
1492 |
checkpoint = last_checkpoint
|
1493 |
|
|
|
1696 |
|
1697 |
if __name__ == "__main__":
|
1698 |
main()
|
1699 |
+
'''
|
1700 |
+
accelerate launch --mixed_precision=bf16 run_distillation.py --model_name_or_path "./distil-large-v3-init" --teacher_model_name_or_path "openai/whisper-large-v3" --train_dataset_name "mozilla-foundation/common_voice_15_0" --train_dataset_config_name "de" --train_split_name "train" --text_column_name "sentence" --eval_dataset_name "mozilla-foundation/common_voice_15_0" --eval_dataset_config_name "de" --eval_split_name "validation" --eval_text_column_name "sentence" --eval_steps 500 --save_steps 50 --warmup_steps 500 --learning_rate 1e-4 --lr_scheduler_type "linear" --logging_steps 25 --save_total_limit 1 --max_steps 500 --per_device_train_batch_size 4 --per_device_eval_batch_size 2 --dataloader_num_workers 2 --preprocessing_num_workers 2 --ddp_timeout 7200 --dtype "bfloat16" --output_dir "./" --use_pseudo_labels "false" --condition_on_prev_probability "0.0" --do_train --do_eval --gradient_checkpointing --overwrite_output_dir --predict_with_generate --freeze_encoder --streaming --push_to_hub
|
1701 |
+
'''
|
run_evaluate.py
ADDED
File without changes
|
tokenizer.json
CHANGED
@@ -14501,6 +14501,12 @@
|
|
14501 |
"type_id": 0
|
14502 |
}
|
14503 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
14504 |
{
|
14505 |
"SpecialToken": {
|
14506 |
"id": "<|transcribe|>",
|
@@ -14533,6 +14539,12 @@
|
|
14533 |
"type_id": 0
|
14534 |
}
|
14535 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
14536 |
{
|
14537 |
"SpecialToken": {
|
14538 |
"id": "<|transcribe|>",
|
@@ -14565,6 +14577,15 @@
|
|
14565 |
}
|
14566 |
],
|
14567 |
"special_tokens": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14568 |
"<|endoftext|>": {
|
14569 |
"id": "<|endoftext|>",
|
14570 |
"ids": [
|
|
|
14501 |
"type_id": 0
|
14502 |
}
|
14503 |
},
|
14504 |
+
{
|
14505 |
+
"SpecialToken": {
|
14506 |
+
"id": "<|de|>",
|
14507 |
+
"type_id": 0
|
14508 |
+
}
|
14509 |
+
},
|
14510 |
{
|
14511 |
"SpecialToken": {
|
14512 |
"id": "<|transcribe|>",
|
|
|
14539 |
"type_id": 0
|
14540 |
}
|
14541 |
},
|
14542 |
+
{
|
14543 |
+
"SpecialToken": {
|
14544 |
+
"id": "<|de|>",
|
14545 |
+
"type_id": 0
|
14546 |
+
}
|
14547 |
+
},
|
14548 |
{
|
14549 |
"SpecialToken": {
|
14550 |
"id": "<|transcribe|>",
|
|
|
14577 |
}
|
14578 |
],
|
14579 |
"special_tokens": {
|
14580 |
+
"<|de|>": {
|
14581 |
+
"id": "<|de|>",
|
14582 |
+
"ids": [
|
14583 |
+
50261
|
14584 |
+
],
|
14585 |
+
"tokens": [
|
14586 |
+
"<|de|>"
|
14587 |
+
]
|
14588 |
+
},
|
14589 |
"<|endoftext|>": {
|
14590 |
"id": "<|endoftext|>",
|
14591 |
"ids": [
|