File size: 11,707 Bytes
6e13c6f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 |
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# fine-tune wav2vec BERT v2\n",
"\n",
"[](https://colab.research.google.com/github/phineas-pta/fine-tune-whisper-vi/blob/main/train/w2v-bert-v2.ipynb)\n",
"\n",
"on colab: mount gdrive using GUI before training\n",
"\n",
"on kaggle: select kaggle free T4×2 for auto double batch size"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"from huggingface_hub import notebook_login\n",
"notebook_login()\n",
"# !huggingface-cli login --token=███"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"scrolled": true,
"trusted": true
},
"outputs": [],
"source": [
"# workaround for a bug in `datasets` package\n",
"%pip uninstall -y cudf dask-cuda dask-cudf\n",
"%pip install -q cudf-cu12 --extra-index-url=https://pypi.nvidia.com\n",
"%pip install -qU 'datasets[audio]' accelerate transformers jiwer bitsandbytes\n",
"# install then `import evaluate` throw error on kaggle"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"import torch\n",
"from dataclasses import dataclass\n",
"import datasets as hugDS\n",
"from transformers import Wav2Vec2BertForCTC, SeamlessM4TFeatureExtractor, Wav2Vec2CTCTokenizer, TrainingArguments, Trainer\n",
"import jiwer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"SAMPLING_RATE = 16_000\n",
"def load_my_data(mode, **kwargs):\n",
"\ttmp = hugDS.load_dataset(**kwargs, trust_remote_code=True, streaming=True).cast_column(\"audio\", hugDS.Audio(sampling_rate=SAMPLING_RATE))\n",
"\tmatch mode:\n",
"\t\tcase 0:\n",
"\t\t\treturn tmp\n",
"\t\tcase 1:\n",
"\t\t\treturn tmp.select_columns([\"audio\", \"transcription\"])\n",
"\t\tcase 2:\n",
"\t\t\treturn tmp.select_columns([\"audio\", \"sentence\"]).rename_column(\"sentence\", \"transcription\")\n",
"\t\tcase _:\n",
"\t\t\traise ValueError(\"oh no!\")\n",
"\n",
"MY_DATA = hugDS.IterableDatasetDict()\n",
"\n",
"MY_DATA[\"train\"] = hugDS.concatenate_datasets([ # total: 1.5M samples\n",
"\tload_my_data(path=\"google/fleurs\", name=\"vi_vn\", split=\"train\", mode=1), # 3k\n",
"\tload_my_data(path=\"vivos\", split=\"train\", mode=2), # 11.7k\n",
"\tload_my_data(path=\"doof-ferb/fpt_fosd\", split=\"train\", mode=0), # 25.9k\n",
"\tload_my_data(path=\"doof-ferb/infore1_25hours\", split=\"train\", mode=0), # 14.9k\n",
"\tload_my_data(path=\"doof-ferb/vlsp2020_vinai_100h\", split=\"train\", mode=0), # 56.4k\n",
"\tload_my_data(path=\"doof-ferb/LSVSC\", split=\"train\", mode=1), # 45k\n",
"\tload_my_data(path=\"quocanh34/viet_vlsp\", split=\"train\", mode=0), # 171k\n",
"\tload_my_data(path=\"linhtran92/viet_youtube_asr_corpus_v2\", split=\"train\", mode=1), # 195k\n",
"\tload_my_data(path=\"doof-ferb/infore2_audiobooks\", split=\"train\", mode=0), # 315k\n",
"\tload_my_data(path=\"linhtran92/viet_bud500\", split=\"train\", mode=0), # 634k\n",
"])\n",
"\n",
"MY_DATA[\"test\"] = hugDS.concatenate_datasets([ # total: 15k samples\n",
"\tload_my_data(path=\"mozilla-foundation/common_voice_16_1\", name=\"vi\", split=\"test\", mode=2), # 1.3k\n",
"\t# remove FLEURS because error when running in batch\n",
"\tload_my_data(path=\"vivos\", split=\"test\", mode=2), # .7k\n",
"])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"modelID = \"trick4kid/w2v-bert-2.0-vietnamese-CV16.0\"\n",
"FEATURE_EXTRACTOR = SeamlessM4TFeatureExtractor.from_pretrained(modelID)\n",
"TOKENIZER = Wav2Vec2CTCTokenizer.from_pretrained(modelID)\n",
"MODEL = Wav2Vec2BertForCTC.from_pretrained(\n",
"\tmodelID, ctc_loss_reduction=\"mean\", add_adapter=True, mask_time_prob=0.,\n",
"\tlayerdrop=0., pad_token_id=TOKENIZER.pad_token_id, vocab_size=len(TOKENIZER)\n",
")\n",
"\n",
"DUMMY_TOKEN = -100"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"def prepare_dataset(batch):\n",
"\taudio = batch[\"audio\"]\n",
"\tbatch[\"input_features\"] = FEATURE_EXTRACTOR(audio[\"array\"], sampling_rate=SAMPLING_RATE).input_features[0] # compute log-Mel input features\n",
"\tbatch[\"labels\"] = TOKENIZER(batch[\"transcription\"]).input_ids # encode target text to label ids\n",
"\treturn batch\n",
"MY_DATA = MY_DATA.map(prepare_dataset) # no `num_proc` coz streaming"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"@dataclass\n",
"class DataCollatorCTCWithPadding:\n",
"\tdef __call__(self, features):\n",
"\t\t# split inputs and labels since they have to be of different lengths and need different padding methods\n",
"\t\tinput_features = [{\"input_features\": feature[\"input_features\"]} for feature in features]\n",
"\t\tlabel_features = [{\"input_ids\" : feature[\"labels\"] } for feature in features]\n",
"\n",
"\t\tbatch = FEATURE_EXTRACTOR.pad(input_features, padding=True, return_tensors=\"pt\")\n",
"\t\tlabels_batch = TOKENIZER.pad(label_features, padding=True, return_tensors=\"pt\")\n",
"\t\tlabels = labels_batch[\"input_ids\"].masked_fill(labels_batch.attention_mask.ne(1), DUMMY_TOKEN) # replace padding with -100 to ignore loss correctly\n",
"\n",
"\t\tbatch[\"labels\"] = labels\n",
"\t\treturn batch\n",
"\n",
"DATA_COLLATOR = DataCollatorCTCWithPadding()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"JIWER_TRANS = jiwer.Compose([ # DO NOT use `jiwer.RemoveEmptyStrings` it can cause rows count mismatch\n",
"\tjiwer.ToLowerCase(),\n",
"\tjiwer.RemoveKaldiNonWords(),\n",
"\tjiwer.RemoveMultipleSpaces(),\n",
"\tjiwer.Strip(),\n",
"\tjiwer.RemovePunctuation(),\n",
"\tjiwer.ReduceToListOfListOfWords(),\n",
"])\n",
"\n",
"def compute_metrics(pred):\n",
"\tpred_logits, label_ids = pred.predictions, pred.label_ids\n",
"\tpred_ids = torch.argmax(pred_logits, axis=-1)\n",
"\tlabel_ids[label_ids == DUMMY_TOKEN] = TOKENIZER.pad_token_id # replace -100 with the pad_token_id\n",
"\n",
"\twer = jiwer.wer( # we do not want to group tokens when computing the metrics\n",
"\t\treference=TOKENIZER.batch_decode(label_ids, group_tokens=False)[0],\n",
"\t\thypothesis=TOKENIZER.batch_decode(pred_ids)[0],\n",
"\t\treference_transform=JIWER_TRANS, hypothesis_transform=JIWER_TRANS\n",
"\t)\n",
"\treturn {\"wer\": wer}"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"# mount gdrive using GUI before training\n",
"%cd '/content/drive/My Drive/coder'\n",
"\n",
"# %cd /kaggle/working\n",
"# !rm -rf ./my-w2v-bert"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"SAVE_PATH = \"./my-w2v-bert\"\n",
"BATCH_SIZE = 4 # should be a power of 2\n",
"# kaggle free P100 train faster than colab free T4\n",
"# kaggle free T4×2: no speed up but auto double batch size\n",
"\n",
"# colab free tier can only run for 8-12h max daily\n",
"# kaggle free tier can only run for 30h max weekly but max 12h per session\n",
"\n",
"has_bf16 = torch.cuda.is_bf16_supported() # GPU Ampere or later\n",
"\n",
"TRAINING_ARGS = TrainingArguments(\n",
"\toutput_dir=SAVE_PATH,\n",
"\tper_device_train_batch_size=BATCH_SIZE,\n",
"\tper_device_eval_batch_size=BATCH_SIZE,\n",
"\tfp16=not has_bf16,\n",
"\tbf16=has_bf16, tf32=has_bf16,\n",
"\t# torch_compile=True, # SDPA not support wav2vec yet\n",
"\treport_to=[\"tensorboard\"],\n",
"\n",
"\tmax_steps=1200, # no `num_train_epochs` coz streaming\n",
"\tlogging_steps=25,\n",
"\tsave_steps=50,\n",
"\teval_steps=50,\n",
"\tevaluation_strategy=\"steps\",\n",
"\tsave_total_limit=2,\n",
"\n",
"\toptim=\"adamw_bnb_8bit\", # 8-bit AdamW optimizer: lower vram usage than default AdamW\n",
"\tlearning_rate=5e-5,\n",
"\twarmup_ratio=.05, # keep between 5-15%\n",
"\tgradient_accumulation_steps=1 if BATCH_SIZE >= 8 else 8 // BATCH_SIZE, # keep effective batch size as min 8 per device\n",
"\tgradient_checkpointing=True,\n",
"\tgradient_checkpointing_kwargs={\"use_reentrant\": False},\n",
"\tload_best_model_at_end=True,\n",
"\tmetric_for_best_model=\"wer\",\n",
"\tgreater_is_better=False, # WER is better when lower\n",
")\n",
"\n",
"TRAINER = Trainer(\n",
"\targs=TRAINING_ARGS,\n",
"\tmodel=MODEL,\n",
"\ttrain_dataset=MY_DATA[\"train\"],\n",
"\teval_dataset=MY_DATA[\"test\"],\n",
"\tdata_collator=DATA_COLLATOR,\n",
"\tcompute_metrics=compute_metrics,\n",
"\ttokenizer=FEATURE_EXTRACTOR, # not TOKENIZER\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"scrolled": true,
"trusted": true
},
"outputs": [],
"source": [
"TRAINER.train() # resume_from_checkpoint=True # only if resume"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"jupyter": {
"outputs_hidden": false
},
"trusted": true
},
"outputs": [],
"source": [
"TRAINER.save_model()\n",
"!zip -FSr res.zip ./my-w2v-bert"
]
}
],
"metadata": {
"accelerator": "GPU",
"colab": {
"gpuType": "T4",
"private_outputs": true,
"provenance": []
},
"kaggle": {
"accelerator": "nvidiaTeslaT4",
"dataSources": [],
"isGpuEnabled": true,
"isInternetEnabled": true,
"language": "python",
"sourceType": "notebook"
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"name": "python",
"version": "3.11.7"
}
},
"nbformat": 4,
"nbformat_minor": 0
}
|