File size: 24,411 Bytes
09b13b3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 |
#!/home/haroon/python_virtual_envs/whisper_fine_tuning/bin/python
#!/usr/bin/env python
# coding: utf-8
# # Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers
# In this Colab, we present a step-by-step guide on how to fine-tune Whisper
# for any multilingual ASR dataset using Hugging Face 🤗 Transformers. This is a
# more "hands-on" version of the accompanying [blog post](https://huggingface.co/blog/fine-tune-whisper).
# For a more in-depth explanation of Whisper, the Common Voice dataset and the theory behind fine-tuning, the reader is advised to refer to the blog post.
# ## Introduction
# Whisper is a pre-trained model for automatic speech recognition (ASR)
# published in [September 2022](https://openai.com/blog/whisper/) by the authors
# Alec Radford et al. from OpenAI. Unlike many of its predecessors, such as
# [Wav2Vec 2.0](https://arxiv.org/abs/2006.11477), which are pre-trained
# on un-labelled audio data, Whisper is pre-trained on a vast quantity of
# **labelled** audio-transcription data, 680,000 hours to be precise.
# This is an order of magnitude more data than the un-labelled audio data used
# to train Wav2Vec 2.0 (60,000 hours). What is more, 117,000 hours of this
# pre-training data is multilingual ASR data. This results in checkpoints
# that can be applied to over 96 languages, many of which are considered
# _low-resource_.
#
# When scaled to 680,000 hours of labelled pre-training data, Whisper models
# demonstrate a strong ability to generalise to many datasets and domains.
# The pre-trained checkpoints achieve competitive results to state-of-the-art
# ASR systems, with near 3% word error rate (WER) on the test-clean subset of
# LibriSpeech ASR and a new state-of-the-art on TED-LIUM with 4.7% WER (_c.f._
# Table 8 of the [Whisper paper](https://cdn.openai.com/papers/whisper.pdf)).
# The extensive multilingual ASR knowledge acquired by Whisper during pre-training
# can be leveraged for other low-resource languages; through fine-tuning, the
# pre-trained checkpoints can be adapted for specific datasets and languages
# to further improve upon these results. We'll show just how Whisper can be fine-tuned
# for low-resource languages in this Colab.
# <figure>
# <img src="https://raw.githubusercontent.com/sanchit-gandhi/notebooks/main/whisper_architecture.svg" alt="Trulli" style="width:100%">
# <figcaption align = "center"><b>Figure 1:</b> Whisper model. The architecture
# follows the standard Transformer-based encoder-decoder model. A
# log-Mel spectrogram is input to the encoder. The last encoder
# hidden states are input to the decoder via cross-attention mechanisms. The
# decoder autoregressively predicts text tokens, jointly conditional on the
# encoder hidden states and previously predicted tokens. Figure source:
# <a href="https://openai.com/blog/whisper/">OpenAI Whisper Blog</a>.</figcaption>
# </figure>
# The Whisper checkpoints come in five configurations of varying model sizes.
# The smallest four are trained on either English-only or multilingual data.
# The largest checkpoint is multilingual only. All nine of the pre-trained checkpoints
# are available on the [Hugging Face Hub](https://huggingface.co/models?search=openai/whisper). The
# checkpoints are summarised in the following table with links to the models on the Hub:
#
# | Size | Layers | Width | Heads | Parameters | English-only | Multilingual |
# |--------|--------|-------|-------|------------|------------------------------------------------------|---------------------------------------------------|
# | tiny | 4 | 384 | 6 | 39 M | [✓](https://huggingface.co/openai/whisper-tiny.en) | [✓](https://huggingface.co/openai/whisper-tiny.) |
# | base | 6 | 512 | 8 | 74 M | [✓](https://huggingface.co/openai/whisper-base.en) | [✓](https://huggingface.co/openai/whisper-base) |
# | small | 12 | 768 | 12 | 244 M | [✓](https://huggingface.co/openai/whisper-small.en) | [✓](https://huggingface.co/openai/whisper-small) |
# | medium | 24 | 1024 | 16 | 769 M | [✓](https://huggingface.co/openai/whisper-medium.en) | [✓](https://huggingface.co/openai/whisper-medium) |
# | large | 32 | 1280 | 20 | 1550 M | x | [✓](https://huggingface.co/openai/whisper-large) |
#
# For demonstration purposes, we'll fine-tune the multilingual version of the
# [`"small"`](https://huggingface.co/openai/whisper-small) checkpoint with 244M params (~= 1GB).
# As for our data, we'll train and evaluate our system on a low-resource language
# taken from the [Common Voice](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0)
# dataset. We'll show that with as little as 8 hours of fine-tuning data, we can achieve
# strong performance in this language.
# ------------------------------------------------------------------------
#
# \\({}^1\\) The name Whisper follows from the acronym “WSPSR”, which stands for “Web-scale Supervised Pre-training for Speech Recognition”.
# ## Load Dataset
# Using 🤗 Datasets, downloading and preparing data is extremely simple.
# We can download and prepare the Common Voice splits in just one line of code.
#
# First, ensure you have accepted the terms of use on the Hugging Face Hub: [mozilla-foundation/common_voice_11_0](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0). Once you have accepted the terms, you will have full access to the dataset and be able to download the data locally.
#
# Since Hindi is very low-resource, we'll combine the `train` and `validation`
# splits to give approximately 8 hours of training data. We'll use the 4 hours
# of `test` data as our held-out test set:
# In[1]:
from datasets import load_dataset, DatasetDict
common_voice = DatasetDict()
common_voice["train"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="train+validation", token=True)
common_voice["test"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="test", token=True)
print(common_voice)
# Most ASR datasets only provide input audio samples (`audio`) and the
# corresponding transcribed text (`sentence`). Common Voice contains additional
# metadata information, such as `accent` and `locale`, which we can disregard for ASR.
# Keeping the notebook as general as possible, we only consider the input audio and
# transcribed text for fine-tuning, discarding the additional metadata information:
# In[2]:
common_voice = common_voice.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "path", "segment", "up_votes"])
print(common_voice)
# ## Prepare Feature Extractor, Tokenizer and Data
# The ASR pipeline can be de-composed into three stages:
# 1) A feature extractor which pre-processes the raw audio-inputs
# 2) The model which performs the sequence-to-sequence mapping
# 3) A tokenizer which post-processes the model outputs to text format
#
# In 🤗 Transformers, the Whisper model has an associated feature extractor and tokenizer,
# called [WhisperFeatureExtractor](https://huggingface.co/docs/transformers/main/model_doc/whisper#transformers.WhisperFeatureExtractor)
# and [WhisperTokenizer](https://huggingface.co/docs/transformers/main/model_doc/whisper#transformers.WhisperTokenizer)
# respectively.
#
# We'll go through details for setting-up the feature extractor and tokenizer one-by-one!
# ### Load WhisperFeatureExtractor
# The Whisper feature extractor performs two operations:
# 1. Pads / truncates the audio inputs to 30s: any audio inputs shorter than 30s are padded to 30s with silence (zeros), and those longer that 30s are truncated to 30s
# 2. Converts the audio inputs to _log-Mel spectrogram_ input features, a visual representation of the audio and the form of the input expected by the Whisper model
# <figure>
# <img src="https://raw.githubusercontent.com/sanchit-gandhi/notebooks/main/spectrogram.jpg" alt="Trulli" style="width:100%">
# <figcaption align = "center"><b>Figure 2:</b> Conversion of sampled audio array to log-Mel spectrogram.
# Left: sampled 1-dimensional audio signal. Right: corresponding log-Mel spectrogram. Figure source:
# <a href="https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html">Google SpecAugment Blog</a>.
# </figcaption>
# We'll load the feature extractor from the pre-trained checkpoint with the default values:
# In[3]:
from transformers import WhisperFeatureExtractor
feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-small")
# ### Load WhisperTokenizer
# The Whisper model outputs a sequence of _token ids_. The tokenizer maps each of these token ids to their corresponding text string. For Hindi, we can load the pre-trained tokenizer and use it for fine-tuning without any further modifications. We simply have to
# specify the target language and the task. These arguments inform the
# tokenizer to prefix the language and task tokens to the start of encoded
# label sequences:
# In[4]:
from transformers import WhisperTokenizer
tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")
# ### Combine To Create A WhisperProcessor
# To simplify using the feature extractor and tokenizer, we can _wrap_
# both into a single `WhisperProcessor` class. This processor object
# inherits from the `WhisperFeatureExtractor` and `WhisperProcessor`,
# and can be used on the audio inputs and model predictions as required.
# In doing so, we only need to keep track of two objects during training:
# the `processor` and the `model`:
# In[5]:
from transformers import WhisperProcessor
processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")
# ### Prepare Data
# Let's print the first example of the Common Voice dataset to see
# what form the data is in:
# In[6]:
print(common_voice["train"][0])
# Since
# our input audio is sampled at 48kHz, we need to _downsample_ it to
# 16kHz prior to passing it to the Whisper feature extractor, 16kHz being the sampling rate expected by the Whisper model.
#
# We'll set the audio inputs to the correct sampling rate using dataset's
# [`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=cast_column#datasets.DatasetDict.cast_column)
# method. This operation does not change the audio in-place,
# but rather signals to `datasets` to resample audio samples _on the fly_ the
# first time that they are loaded:
# In[7]:
from datasets import Audio
common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16000))
# Re-loading the first audio sample in the Common Voice dataset will resample
# it to the desired sampling rate:
# In[8]:
print(common_voice["train"][0])
# We'll define our pre-processing strategy. We advise that you **do not** lower-case the transcriptions or remove punctuation unless mixing different datasets. This will enable you to fine-tune Whisper models that can predict punctuation and casing. Later, you will see how we can evaluate the predictions without punctuation or casing, so that the models benefit from the WER improvement obtained by normalising the transcriptions while still predicting fully formatted transcriptions.
# In[9]:
from transformers.models.whisper.english_normalizer import BasicTextNormalizer
do_lower_case = False
do_remove_punctuation = False
normalizer = BasicTextNormalizer()
# Now we can write a function to prepare our data ready for the model:
# 1. We load and resample the audio data by calling `batch["audio"]`. As explained above, 🤗 Datasets performs any necessary resampling operations on the fly.
# 2. We use the feature extractor to compute the log-Mel spectrogram input features from our 1-dimensional audio array.
# 3. We perform any optional pre-processing (lower-case or remove punctuation).
# 4. We encode the transcriptions to label ids through the use of the tokenizer.
# In[10]:
def prepare_dataset(batch):
# load and (possibly) resample audio data to 16kHz
audio = batch["audio"]
# compute log-Mel input features from input audio array
batch["input_features"] = processor.feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
# compute input length of audio sample in seconds
batch["input_length"] = len(audio["array"]) / audio["sampling_rate"]
# optional pre-processing steps
transcription = batch["sentence"]
if do_lower_case:
transcription = transcription.lower()
if do_remove_punctuation:
transcription = normalizer(transcription).strip()
# encode target text to label ids
batch["labels"] = processor.tokenizer(transcription).input_ids
return batch
# We can apply the data preparation function to all of our training examples using dataset's `.map` method. The argument `num_proc` specifies how many CPU cores to use. Setting `num_proc` > 1 will enable multiprocessing. If the `.map` method hangs with multiprocessing, set `num_proc=1` and process the dataset sequentially.
# In[11]:
common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names["train"], num_proc=2)
# Finally, we filter any training data with audio samples longer than 30s. These samples would otherwise be truncated by the Whisper feature-extractor which could affect the stability of training. We define a function that returns `True` for samples that are less than 30s, and `False` for those that are longer:
# In[12]:
max_input_length = 30.0
def is_audio_in_length_range(length):
return length < max_input_length
# We apply our filter function to all samples of our training dataset through 🤗 Datasets' `.filter` method:
# In[13]:
common_voice["train"] = common_voice["train"].filter(
is_audio_in_length_range,
input_columns=["input_length"],
)
# ## Training and Evaluation
# Now that we've prepared our data, we're ready to dive into the training pipeline.
# The [🤗 Trainer](https://huggingface.co/transformers/master/main_classes/trainer.html?highlight=trainer)
# will do much of the heavy lifting for us. All we have to do is:
#
# - Define a data collator: the data collator takes our pre-processed data and prepares PyTorch tensors ready for the model.
#
# - Evaluation metrics: during evaluation, we want to evaluate the model using the [word error rate (WER)](https://huggingface.co/metrics/wer) metric. We need to define a `compute_metrics` function that handles this computation.
#
# - Load a pre-trained checkpoint: we need to load a pre-trained checkpoint and configure it correctly for training.
#
# - Define the training configuration: this will be used by the 🤗 Trainer to define the training schedule.
#
# Once we've fine-tuned the model, we will evaluate it on the test data to verify that we have correctly trained it
# to transcribe speech in Hindi.
# ### Define a Data Collator
# The data collator for a sequence-to-sequence speech model is unique in the sense that it
# treats the `input_features` and `labels` independently: the `input_features` must be
# handled by the feature extractor and the `labels` by the tokenizer.
#
# The `input_features` are already padded to 30s and converted to a log-Mel spectrogram
# of fixed dimension by action of the feature extractor, so all we have to do is convert the `input_features`
# to batched PyTorch tensors. We do this using the feature extractor's `.pad` method with `return_tensors=pt`.
#
# The `labels` on the other hand are un-padded. We first pad the sequences
# to the maximum length in the batch using the tokenizer's `.pad` method. The padding tokens
# are then replaced by `-100` so that these tokens are **not** taken into account when
# computing the loss. We then cut the BOS token from the start of the label sequence as we
# append it later during training.
#
# We can leverage the `WhisperProcessor` we defined earlier to perform both the
# feature extractor and the tokenizer operations:
# In[14]:
import torch
from dataclasses import dataclass
from typing import Any, Dict, List, Union
@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
processor: Any
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# split inputs and labels since they have to be of different lengths and need different padding methods
# first treat the audio inputs by simply returning torch tensors
input_features = [{"input_features": feature["input_features"]} for feature in features]
batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
# get the tokenized label sequences
label_features = [{"input_ids": feature["labels"]} for feature in features]
# pad the labels to max length
labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
# replace padding with -100 to ignore loss correctly
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
# if bos token is appended in previous tokenization step,
# cut bos token here as it's append later anyways
if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().cpu().item():
labels = labels[:, 1:]
batch["labels"] = labels
return batch
# Let's initialise the data collator we've just defined:
# In[15]:
data_collator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)
# ### Evaluation Metrics
# We'll use the word error rate (WER) metric, the 'de-facto' metric for assessing
# ASR systems. For more information, refer to the WER [docs](https://huggingface.co/metrics/wer). We'll load the WER metric from 🤗 Evaluate:
# In[16]:
import evaluate
metric = evaluate.load("wer")
# We then simply have to define a function that takes our model
# predictions and returns the WER metric. This function, called
# `compute_metrics`, first replaces `-100` with the `pad_token_id`
# in the `label_ids` (undoing the step we applied in the
# data collator to ignore padded tokens correctly in the loss).
# It then decodes the predicted and label ids to strings. Finally,
# it computes the WER between the predictions and reference labels.
# Here, we have the option of evaluating with the 'normalised' transcriptions
# and predictions. We recommend you set this to `True` to benefit from the WER
# improvement obtained by normalising the transcriptions.
# In[17]:
# evaluate with the 'normalised' WER
do_normalize_eval = True
def compute_metrics(pred):
pred_ids = pred.predictions
label_ids = pred.label_ids
# replace -100 with the pad_token_id
label_ids[label_ids == -100] = processor.tokenizer.pad_token_id
# we do not want to group tokens when computing the metrics
pred_str = processor.tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
label_str = processor.tokenizer.batch_decode(label_ids, skip_special_tokens=True)
if do_normalize_eval:
pred_str = [normalizer(pred) for pred in pred_str]
label_str = [normalizer(label) for label in label_str]
wer = 100 * metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
# ### Load a Pre-Trained Checkpoint
# Now let's load the pre-trained Whisper `small` checkpoint. Again, this
# is trivial through use of 🤗 Transformers!
# In[18]:
from transformers import WhisperForConditionalGeneration
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model.generation_config.language = "hi" # define your language of choice here
# Override generation arguments - no tokens are forced as decoder outputs (see [`forced_decoder_ids`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.forced_decoder_ids)), no tokens are suppressed during generation (see [`suppress_tokens`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.suppress_tokens)). Set `use_cache` to False since we're using gradient checkpointing, and the two are incompatible:
# In[19]:
model.config.forced_decoder_ids = None
model.config.suppress_tokens = []
model.config.use_cache = False
# ### Define the Training Configuration
# In the final step, we define all the parameters related to training. For more detail on the training arguments, refer to the Seq2SeqTrainingArguments [docs](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Seq2SeqTrainingArguments).
# In[20]:
from transformers import Seq2SeqTrainingArguments
training_args = Seq2SeqTrainingArguments(
output_dir="./",
per_device_train_batch_size=8,
gradient_accumulation_steps=8, # increase by 2x for every 2x decrease in batch size
learning_rate=1e-5,
warmup_steps=500,
max_steps=5000,
gradient_checkpointing=True,
fp16=True,
evaluation_strategy="steps",
per_device_eval_batch_size=4,
predict_with_generate=True,
generation_max_length=225,
save_steps=1000,
eval_steps=1000,
logging_steps=25,
report_to=["tensorboard"],
load_best_model_at_end=True,
metric_for_best_model="wer",
greater_is_better=False,
push_to_hub=True,
)
# **Note**: if one does not want to upload the model checkpoints to the Hub,
# set `push_to_hub=False`.
# We can forward the training arguments to the 🤗 Trainer along with our model,
# dataset, data collator and `compute_metrics` function:
# In[21]:
from transformers import Seq2SeqTrainer
trainer = Seq2SeqTrainer(
args=training_args,
model=model,
train_dataset=common_voice["train"],
eval_dataset=common_voice["test"],
data_collator=data_collator,
compute_metrics=compute_metrics,
tokenizer=processor.feature_extractor,
)
# We'll save the processor object once before starting training. Since the processor is not trainable, it won't change over the course of training:
# In[22]:
processor.save_pretrained(training_args.output_dir)
# ### Training
# Training will take approximately 5-10 hours depending on your GPU. The peak GPU memory for the given training configuration is approximately 36GB.
# Depending on your GPU, it is possible that you will encounter a CUDA `"out-of-memory"` error when you launch training.
# In this case, you can reduce the `per_device_train_batch_size` incrementally by factors of 2
# and employ [`gradient_accumulation_steps`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Seq2SeqTrainingArguments.gradient_accumulation_steps)
# to compensate.
#
# To launch training, simply execute:
# In[ ]:
trainer.train()
# We can label our checkpoint with the `whisper-event` tag on push by setting the appropriate key-word arguments (kwargs):
# In[ ]:
kwargs = {
"dataset_tags": "mozilla-foundation/common_voice_11_0",
"dataset": "Common Voice 11.0", # a 'pretty' name for the training dataset
"language": "hi",
"model_name": "Whisper Small Hi - Sanchit Gandhi", # a 'pretty' name for your model
"finetuned_from": "openai/whisper-small",
"tasks": "automatic-speech-recognition",
"tags": "whisper-event",
}
# The training results can now be uploaded to the Hub. To do so, execute the `push_to_hub` command and save the preprocessor object we created:
# In[ ]:
trainer.push_to_hub(**kwargs)
# ## Closing Remarks
# In this blog, we covered a step-by-step guide on fine-tuning Whisper for multilingual ASR
# using 🤗 Datasets, Transformers and the Hugging Face Hub. For more details on the Whisper model, the Common Voice dataset and the theory behind fine-tuning, refere to the accompanying [blog post](https://huggingface.co/blog/fine-tune-whisper). If you're interested in fine-tuning other
# Transformers models, both for English and multilingual ASR, be sure to check out the
# examples scripts at [examples/pytorch/speech-recognition](https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-recognition).
|