commit 56fedafff8bbc10da5cb3f8b293b8508534d6e4e Author: ModelHub XC Date: Fri May 8 11:35:50 2026 +0800 初始化项目，由ModelHub XC社区提供模型 Model: othrif/wav2vec2-large-xlsr-arabic Source: Original Platform diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000..07f0db3 --- /dev/null +++ b/.gitattributes @@ -0,0 +1,16 @@ +*.bin.* filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tar.gz filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md new file mode 100644 index 0000000..b985699 --- /dev/null +++ b/README.md @@ -0,0 +1,127 @@ +--- +language: ar +datasets: +- common_voice +metrics: +- wer +tags: +- audio +- automatic-speech-recognition +- speech +- xlsr-fine-tuning-week +license: apache-2.0 +model-index: +- name: XLSR Wav2Vec2 Arabic by Othmane Rifki + results: + - task: + name: Speech Recognition + type: automatic-speech-recognition + dataset: + name: Common Voice ar + type: common_voice + args: ar + metrics: + - name: Test WER + type: wer + value: 46.77 +--- + +# Wav2Vec2-Large-XLSR-53-Arabic + +Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Arabic using the [Common Voice](https://huggingface.co/datasets/common_voice). +When using this model, make sure that your speech input is sampled at 16kHz. + +## Usage + +The model can be used directly (without a language model) as follows: + +```python +import torch +import torchaudio +from datasets import load_dataset +from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor + +test_dataset = load_dataset("common_voice", "ar", split="test[:2%]") + +processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") +model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") + +resampler = torchaudio.transforms.Resample(48_000, 16_000) + +# Preprocessing the datasets. +# We need to read the audio files as arrays +def speech_file_to_array_fn(batch): + speech_array, sampling_rate = torchaudio.load(batch["path"]) + batch["speech"] = resampler(speech_array).squeeze().numpy() + return batch + +test_dataset = test_dataset.map(speech_file_to_array_fn) +inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True) + +with torch.no_grad(): + logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits + +predicted_ids = torch.argmax(logits, dim=-1) + +print("Prediction:", processor.batch_decode(predicted_ids)) +print("Reference:", test_dataset["sentence"][:2]) +``` + + +## Evaluation + +The model can be evaluated as follows on the Arabic test data of Common Voice. + + +```python +import torch +import torchaudio +from datasets import load_dataset, load_metric +from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor +import re + +test_dataset = load_dataset("common_voice", "ar", split="test") +wer = load_metric("wer") + +processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") +model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") +model.to("cuda") + +chars_to_ignore_regex = '[\\\\\\\\\\\\\\\\؛\\\\\\\\\\\\\\\\—\\\\\\\\\\\\\\\\_get\\\\\\\\\\\\\\\\«\\\\\\\\\\\\\\\\»\\\\\\\\\\\\\\\\ـ\\\\\\\\\\\\\\\\ـ\\\\\\\\\\\\\\\\,\\\\\\\\\\\\\\\\?\\\\\\\\\\\\\\\\.\\\\\\\\\\\\\\\\!\\\\\\\\\\\\\\\\-\\\\\\\\\\\\\\\\;\\\\\\\\\\\\\\\\:\\\\\\\\\\\\\\\\"\\\\\\\\\\\\\\\\“\\\\\\\\\\\\\\\\%\\\\\\\\\\\\\\\\‘\\\\\\\\\\\\\\\\”\\\\\\\\\\\\\\\\�\\\\\\\\\\\\\\\\#\\\\\\\\\\\\\\\\،\\\\\\\\\\\\\\\\☭,\\\\\\\\\\\\\\\\؟]' +resampler = torchaudio.transforms.Resample(48_000, 16_000) + +# Preprocessing the datasets. +# We need to read the audio files as arrays +def speech_file_to_array_fn(batch): + batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + speech_array, sampling_rate = torchaudio.load(batch["path"]) + batch["speech"] = resampler(speech_array).squeeze().numpy() + return batch + +test_dataset = test_dataset.map(speech_file_to_array_fn) + +# Preprocessing the datasets. +# We need to read the audio files as arrays +def evaluate(batch): + inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True) + + with torch.no_grad(): + logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits + + pred_ids = torch.argmax(logits, dim=-1) + batch["pred_strings"] = processor.batch_decode(pred_ids) + return batch + +result = test_dataset.map(evaluate, batched=True, batch_size=8) + +print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"]))) +``` + +**Test Result**: 46.77 + + +## Training + +The Common Voice `train`, `validation` datasets were used for training. + +The script used for training can be found [here](https://huggingface.co/othrif/wav2vec2-large-xlsr-arabic/tree/main) \ No newline at end of file diff --git a/config.json b/config.json new file mode 100644 index 0000000..a9166b3 --- /dev/null +++ b/config.json @@ -0,0 +1,76 @@ +{ + "_name_or_path": "facebook/wav2vec2-large-xlsr-53", + "activation_dropout": 0.055, + "apply_spec_augment": true, + "architectures": [ + "Wav2Vec2ForCTC" + ], + "attention_dropout": 0.094, + "bos_token_id": 1, + "conv_bias": true, + "conv_dim": [ + 512, + 512, + 512, + 512, + 512, + 512, + 512 + ], + "conv_kernel": [ + 10, + 3, + 3, + 3, + 3, + 2, + 2 + ], + "conv_stride": [ + 5, + 2, + 2, + 2, + 2, + 2, + 2 + ], + "ctc_loss_reduction": "mean", + "ctc_zero_infinity": false, + "do_stable_layer_norm": true, + "eos_token_id": 2, + "feat_extract_activation": "gelu", + "feat_extract_dropout": 0.0, + "feat_extract_norm": "layer", + "feat_proj_dropout": 0.04, + "final_dropout": 0.0, + "gradient_checkpointing": true, + "hidden_act": "gelu", + "hidden_dropout": 0.047, + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 4096, + "layer_norm_eps": 1e-05, + "layerdrop": 0.041, + "mask_channel_length": 10, + "mask_channel_min_space": 1, + "mask_channel_other": 0.0, + "mask_channel_prob": 0.0, + "mask_channel_selection": "static", + "mask_feature_length": 10, + "mask_feature_prob": 0.0, + "mask_time_length": 10, + "mask_time_min_space": 1, + "mask_time_other": 0.0, + "mask_time_prob": 0.082, + "mask_time_selection": "static", + "model_type": "wav2vec2", + "num_attention_heads": 16, + "num_conv_pos_embedding_groups": 16, + "num_conv_pos_embeddings": 128, + "num_feat_extract_layers": 7, + "num_hidden_layers": 24, + "pad_token_id": 49, + "transformers_version": "4.4.0", + "vocab_size": 50 +} diff --git a/evaluate.py b/evaluate.py new file mode 100644 index 0000000..0143860 --- /dev/null +++ b/evaluate.py @@ -0,0 +1,46 @@ +import torch +import torchaudio +from datasets import load_dataset, load_metric +from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor +import re +import tnkeeh as tn + +test_dataset = load_dataset("common_voice", "ar", split="test") +wer = load_metric("wer") + +processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") +model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-arabic") +model.to("cuda") + +#chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\'\�]' +chars_to_ignore_regex = '[\؛\—\_get\«\»\ـ\ـ\,\?\.\!\-\;\:\"\“\%\‘\”\�\#\،\☭,\؟\'ۚ\چ\ڨ\ﺃ\ھ\ﻻ\'ۖ]' +resampler = torchaudio.transforms.Resample(48_000, 16_000) + + +# Preprocessing the datasets. +# We need to read the aduio files as arrays +def speech_file_to_array_fn(batch): + batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + speech_array, sampling_rate = torchaudio.load(batch["path"]) + batch["speech"] = resampler(speech_array).squeeze().numpy() + return batch + +test_dataset = test_dataset.map(speech_file_to_array_fn) +# For arabic diacritics +cleander = tn.Tnkeeh(remove_diacritics=True) +test_dataset = cleander.clean_hf_dataset(test_dataset, 'sentence') +# Preprocessing the datasets. +# We need to read the aduio files as arrays +def evaluate(batch): + inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True) + + with torch.no_grad(): + logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits + + pred_ids = torch.argmax(logits, dim=-1) + batch["pred_strings"] = processor.batch_decode(pred_ids) + return batch + +result = test_dataset.map(evaluate, batched=True, batch_size=32) + +print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"]))) \ No newline at end of file diff --git a/finetune.sh b/finetune.sh new file mode 100755 index 0000000..7111c38 --- /dev/null +++ b/finetune.sh @@ -0,0 +1,40 @@ +#!/usr/bin/env bash + +export model_path=$1 +mkdir -p ${model_path} + +python run_common_voice.py \ + --dataloader_num_workers="8" \ + --model_name_or_path="facebook/wav2vec2-large-xlsr-53" \ + --dataset_config_name="ar" \ + --output_dir=${model_path} \ + --num_train_epochs="50" \ + --per_device_train_batch_size="16" \ + --per_device_eval_batch_size="16" \ + --evaluation_strategy="steps" \ + --warmup_steps="500" \ + --fp16 \ + --freeze_feature_extractor \ + --save_steps="400" \ + --eval_steps="400" \ + --logging_steps="400" \ + --save_total_limit="1" \ + --group_by_length \ + --attention_dropout="0.094" \ + --activation_dropout="0.055" \ + --feat_proj_dropout="0.04" \ + --hidden_dropout="0.047" \ + --layerdrop="0.041" \ + --mask_time_prob="0.082" \ + --gradient_checkpointing \ + --learning_rate="3e-4" \ + --do_train --do_eval + + + #--model_name_or_path="facebook/wav2vec2-large-xlsr-53" \ + #--overwrite_output_dir \ + #--model_name_or_path="/home/othrif/projects/wav2vec2/finetune-xlsr/models/ar/msa/wav2vec2-large-xlsr-arabic" \ + + + + diff --git a/preprocessor_config.json b/preprocessor_config.json new file mode 100644 index 0000000..0886a48 --- /dev/null +++ b/preprocessor_config.json @@ -0,0 +1,8 @@ +{ + "do_normalize": true, + "feature_size": 1, + "padding_side": "right", + "padding_value": 0.0, + "return_attention_mask": true, + "sampling_rate": 16000 +} diff --git a/pytorch_model.bin b/pytorch_model.bin new file mode 100644 index 0000000..9032fac --- /dev/null +++ b/pytorch_model.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3a226cc02178fad03b0e8311c290e48dd677d144e66b148c221e43b4e104617 +size 1262138839 diff --git a/run_common_voice.py b/run_common_voice.py new file mode 100644 index 0000000..8572443 --- /dev/null +++ b/run_common_voice.py @@ -0,0 +1,518 @@ +#!/usr/bin/env python3 +import json +import logging +import os +import re +import sys +from dataclasses import dataclass, field +from typing import Any, Dict, List, Optional, Union + +import datasets +import numpy as np +import torch +import torchaudio +from packaging import version +from torch import nn + +import transformers +from transformers import ( + HfArgumentParser, + Trainer, + TrainingArguments, + Wav2Vec2CTCTokenizer, + Wav2Vec2FeatureExtractor, + Wav2Vec2ForCTC, + Wav2Vec2Processor, + is_apex_available, + set_seed, +) +from transformers.trainer_utils import get_last_checkpoint, is_main_process + + +if is_apex_available(): + from apex import amp + +import tnkeeh as tn + +if version.parse(torch.__version__) >= version.parse("1.6"): + _is_native_amp_available = True + from torch.cuda.amp import autocast + +logger = logging.getLogger(__name__) + + +def list_field(default=None, metadata=None): + return field(default_factory=lambda: default, metadata=metadata) + + +@dataclass +class ModelArguments: + """ + Arguments pertaining to which model/config/tokenizer we are going to fine-tune from. + """ + + model_name_or_path: str = field( + metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"} + ) + cache_dir: Optional[str] = field( + default=None, + metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"}, + ) + freeze_feature_extractor: Optional[bool] = field( + default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."} + ) + attention_dropout: Optional[float] = field( + default=0.1, metadata={"help": "The dropout ratio for the attention probabilities."} + ) + activation_dropout: Optional[float] = field( + default=0.1, metadata={"help": "The dropout ratio for activations inside the fully connected layer."} + ) + hidden_dropout: Optional[float] = field( + default=0.1, + metadata={ + "help": "The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler." + }, + ) + feat_proj_dropout: Optional[float] = field( + default=0.1, + metadata={"help": "The dropout probabilitiy for all 1D convolutional layers in feature extractor."}, + ) + mask_time_prob: Optional[float] = field( + default=0.05, + metadata={ + "help": "Propability of each feature vector along the time axis to be chosen as the start of the vector" + "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature" + "vectors will be masked along the time axis. This is only relevant if ``apply_spec_augment is True``." + }, + ) + gradient_checkpointing: Optional[bool] = field( + default=True, + metadata={ + "help": "If True, use gradient checkpointing to save memory at the expense of slower backward pass." + }, + ) + layerdrop: Optional[float] = field(default=0.0, metadata={"help": "The LayerDrop probability."}) + + +@dataclass +class DataTrainingArguments: + """ + Arguments pertaining to what data we are going to input our model for training and eval. + + Using `HfArgumentParser` we can turn this class + into argparse arguments to be able to specify them on + the command line. + """ + + dataset_config_name: Optional[str] = field( + default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."} + ) + train_split_name: Optional[str] = field( + default="train+validation", + metadata={ + "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'" + }, + ) + overwrite_cache: bool = field( + default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."} + ) + preprocessing_num_workers: Optional[int] = field( + default=None, + metadata={"help": "The number of processes to use for the preprocessing."}, + ) + max_train_samples: Optional[int] = field( + default=None, + metadata={ + "help": "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + }, + ) + max_val_samples: Optional[int] = field( + default=None, + metadata={ + "help": "For debugging purposes or quicker training, truncate the number of validation examples to this " + "value if set." + }, + ) + chars_to_ignore: List[str] = list_field( + default=[",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�"], + metadata={"help": "A list of characters to remove from the transcripts."}, + ) + + +@dataclass +class DataCollatorCTCWithPadding: + """ + Data collator that will dynamically pad the inputs received. + Args: + processor (:class:`~transformers.Wav2Vec2Processor`) + The processor used for proccessing the data. + padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`): + Select a strategy to pad the returned sequences (according to the model's padding side and padding index) + among: + * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single + sequence if provided). + * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the + maximum acceptable input length for the model if that argument is not provided. + * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of + different lengths). + max_length (:obj:`int`, `optional`): + Maximum length of the ``input_values`` of the returned list and optionally padding length (see above). + max_length_labels (:obj:`int`, `optional`): + Maximum length of the ``labels`` returned list and optionally padding length (see above). + pad_to_multiple_of (:obj:`int`, `optional`): + If set will pad the sequence to a multiple of the provided value. + This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= + 7.5 (Volta). + """ + + processor: Wav2Vec2Processor + padding: Union[bool, str] = True + max_length: Optional[int] = None + max_length_labels: Optional[int] = None + pad_to_multiple_of: Optional[int] = None + pad_to_multiple_of_labels: Optional[int] = None + + 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 lenghts and need + # different padding methods + input_features = [{"input_values": feature["input_values"]} for feature in features] + label_features = [{"input_ids": feature["labels"]} for feature in features] + + batch = self.processor.pad( + input_features, + padding=self.padding, + max_length=self.max_length, + pad_to_multiple_of=self.pad_to_multiple_of, + return_tensors="pt", + ) + with self.processor.as_target_processor(): + labels_batch = self.processor.pad( + label_features, + padding=self.padding, + max_length=self.max_length_labels, + pad_to_multiple_of=self.pad_to_multiple_of_labels, + 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) + + batch["labels"] = labels + + return batch + + +class CTCTrainer(Trainer): + def training_step(self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor, Any]]) -> torch.Tensor: + """ + Perform a training step on a batch of inputs. + + Subclass and override to inject custom behavior. + + Args: + model (:obj:`nn.Module`): + The model to train. + inputs (:obj:`Dict[str, Union[torch.Tensor, Any]]`): + The inputs and targets of the model. + + The dictionary will be unpacked before being fed to the model. Most models expect the targets under the + argument :obj:`labels`. Check your model's documentation for all accepted arguments. + + Return: + :obj:`torch.Tensor`: The tensor with training loss on this batch. + """ + + model.train() + inputs = self._prepare_inputs(inputs) + + if self.use_amp: + with autocast(): + loss = self.compute_loss(model, inputs) + else: + loss = self.compute_loss(model, inputs) + + if self.args.n_gpu > 1: + if model.module.config.ctc_loss_reduction == "mean": + loss = loss.mean() + elif model.module.config.ctc_loss_reduction == "sum": + loss = loss.sum() / (inputs["labels"] >= 0).sum() + else: + raise ValueError(f"{model.config.ctc_loss_reduction} is not valid. Choose one of ['mean', 'sum']") + + if self.args.gradient_accumulation_steps > 1: + loss = loss / self.args.gradient_accumulation_steps + + if self.use_amp: + self.scaler.scale(loss).backward() + elif self.use_apex: + with amp.scale_loss(loss, self.optimizer) as scaled_loss: + scaled_loss.backward() + elif self.deepspeed: + self.deepspeed.backward(loss) + else: + loss.backward() + + return loss.detach() + + +def main(): + # See all possible arguments in src/transformers/training_args.py + # or by passing the --help flag to this script. + # We now keep distinct sets of args, for a cleaner separation of concerns. + + parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments)) + if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): + # If we pass only one argument to the script and it's the path to a json file, + # let's parse it to get our arguments. + model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1])) + else: + model_args, data_args, training_args = parser.parse_args_into_dataclasses() + + # Detecting last checkpoint. + last_checkpoint = None + if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir: + last_checkpoint = get_last_checkpoint(training_args.output_dir) + if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0: + raise ValueError( + f"Output directory ({training_args.output_dir}) already exists and is not empty. " + "Use --overwrite_output_dir to overcome." + ) + elif last_checkpoint is not None: + logger.info( + f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " + "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." + ) + + # Setup logging + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + handlers=[logging.StreamHandler(sys.stdout)], + ) + logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN) + + # Log on each process the small summary: + logger.warning( + f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}" + + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}" + ) + # Set the verbosity to info of the Transformers logger (on main process only): + if is_main_process(training_args.local_rank): + transformers.utils.logging.set_verbosity_info() + logger.info("Training/evaluation parameters %s", training_args) + + # Set seed before initializing model. + set_seed(training_args.seed) + + # Get the datasets: + train_dataset = datasets.load_dataset( + "common_voice", data_args.dataset_config_name, split=data_args.train_split_name + ) + eval_dataset = datasets.load_dataset("common_voice", data_args.dataset_config_name, split="test") + + # Create and save tokenizer + #chars_to_ignore_regex = f'[{"".join(data_args.chars_to_ignore)}]' + chars_to_ignore_regex = '[\؛\—\_get\«\»\ـ\ـ\,\?\.\!\-\;\:\"\“\%\‘\”\�\#\،\☭,\؟\'ۚ\چ\ڨ\ﺃ\ھ\ﻻ\'ۖ]' + + def remove_special_characters(batch): + batch["text"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).lower() + " " + return batch + + # For arabic diacritics + cleander = tn.Tnkeeh(remove_diacritics=True) + train_dataset = cleander.clean_hf_dataset(train_dataset, 'sentence') + eval_dataset = cleander.clean_hf_dataset(eval_dataset, 'sentence') + + train_dataset = train_dataset.map(remove_special_characters, remove_columns=["sentence"]) + eval_dataset = eval_dataset.map(remove_special_characters, remove_columns=["sentence"]) + + def extract_all_chars(batch): + all_text = " ".join(batch["text"]) + vocab = list(set(all_text)) + return {"vocab": [vocab], "all_text": [all_text]} + + vocab_train = train_dataset.map( + extract_all_chars, + batched=True, + batch_size=-1, + keep_in_memory=True, + remove_columns=train_dataset.column_names, + ) + vocab_test = train_dataset.map( + extract_all_chars, + batched=True, + batch_size=-1, + keep_in_memory=True, + remove_columns=eval_dataset.column_names, + ) + + vocab_list = list(set(vocab_train["vocab"][0]) | set(vocab_test["vocab"][0])) + vocab_dict = {v: k for k, v in enumerate(vocab_list)} + vocab_dict["|"] = vocab_dict[" "] + del vocab_dict[" "] + vocab_dict["[UNK]"] = len(vocab_dict) + vocab_dict["[PAD]"] = len(vocab_dict) + + with open("vocab.json", "w") as vocab_file: + json.dump(vocab_dict, vocab_file) + + # Load pretrained model and tokenizer + # + # Distributed training: + # The .from_pretrained methods guarantee that only one local process can concurrently + # download model & vocab. + tokenizer = Wav2Vec2CTCTokenizer( + "vocab.json", + unk_token="[UNK]", + pad_token="[PAD]", + word_delimiter_token="|", + ) + feature_extractor = Wav2Vec2FeatureExtractor( + feature_size=1, sampling_rate=16_000, padding_value=0.0, do_normalize=True, return_attention_mask=True + ) + processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer) + model = Wav2Vec2ForCTC.from_pretrained( + model_args.model_name_or_path, + cache_dir=model_args.cache_dir, + activation_dropout=model_args.activation_dropout, + attention_dropout=model_args.attention_dropout, + hidden_dropout=model_args.hidden_dropout, + feat_proj_dropout=model_args.feat_proj_dropout, + mask_time_prob=model_args.mask_time_prob, + gradient_checkpointing=model_args.gradient_checkpointing, + layerdrop=model_args.layerdrop, + ctc_loss_reduction="mean", + pad_token_id=processor.tokenizer.pad_token_id, + vocab_size=len(processor.tokenizer), + ) + + if data_args.max_train_samples is not None: + train_dataset = train_dataset.select(range(data_args.max_train_samples)) + + if data_args.max_val_samples is not None: + eval_dataset = eval_dataset.select(range(data_args.max_val_samples)) + + resampler = torchaudio.transforms.Resample(48_000, 16_000) + + # Preprocessing the datasets. + # We need to read the aduio files as arrays and tokenize the targets. + def speech_file_to_array_fn(batch): + speech_array, sampling_rate = torchaudio.load(batch["path"]) + batch["speech"] = resampler(speech_array).squeeze().numpy() + batch["sampling_rate"] = 16_000 + batch["target_text"] = batch["text"] + return batch + + train_dataset = train_dataset.map( + speech_file_to_array_fn, + remove_columns=train_dataset.column_names, + num_proc=data_args.preprocessing_num_workers, + ) + eval_dataset = eval_dataset.map( + speech_file_to_array_fn, + remove_columns=eval_dataset.column_names, + num_proc=data_args.preprocessing_num_workers, + ) + + def prepare_dataset(batch): + # check that all files have the correct sampling rate + assert ( + len(set(batch["sampling_rate"])) == 1 + ), f"Make sure all inputs have the same sampling rate of {processor.feature_extractor.sampling_rate}." + batch["input_values"] = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0]).input_values + # Setup the processor for targets + with processor.as_target_processor(): + batch["labels"] = processor(batch["target_text"]).input_ids + return batch + + train_dataset = train_dataset.map( + prepare_dataset, + remove_columns=train_dataset.column_names, + batch_size=training_args.per_device_train_batch_size, + batched=True, + num_proc=data_args.preprocessing_num_workers, + ) + eval_dataset = eval_dataset.map( + prepare_dataset, + remove_columns=eval_dataset.column_names, + batch_size=training_args.per_device_train_batch_size, + batched=True, + num_proc=data_args.preprocessing_num_workers, + ) + + # Metric + wer_metric = datasets.load_metric("wer") + + def compute_metrics(pred): + pred_logits = pred.predictions + pred_ids = np.argmax(pred_logits, axis=-1) + + pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id + + pred_str = processor.batch_decode(pred_ids) + # we do not want to group tokens when computing the metrics + label_str = processor.batch_decode(pred.label_ids, group_tokens=False) + + wer = wer_metric.compute(predictions=pred_str, references=label_str) + + return {"wer": wer} + + if model_args.freeze_feature_extractor: + model.freeze_feature_extractor() + + # Data collator + data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True) + + # Initialize our Trainer + trainer = CTCTrainer( + model=model, + data_collator=data_collator, + args=training_args, + compute_metrics=compute_metrics, + train_dataset=train_dataset if training_args.do_train else None, + eval_dataset=eval_dataset if training_args.do_eval else None, + tokenizer=processor.feature_extractor, + ) + + # Training + if training_args.do_train: + if last_checkpoint is not None: + checkpoint = last_checkpoint + elif os.path.isdir(model_args.model_name_or_path): + checkpoint = model_args.model_name_or_path + else: + checkpoint = None + train_result = trainer.train(resume_from_checkpoint=checkpoint) + trainer.save_model() + + # save the feature_extractor and the tokenizer + if is_main_process(training_args.local_rank): + processor.save_pretrained(training_args.output_dir) + + metrics = train_result.metrics + max_train_samples = ( + data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset) + ) + metrics["train_samples"] = min(max_train_samples, len(train_dataset)) + + trainer.log_metrics("train", metrics) + trainer.save_metrics("train", metrics) + trainer.save_state() + + # Evaluation + results = {} + if training_args.do_eval: + logger.info("*** Evaluate ***") + metrics = trainer.evaluate() + max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset) + metrics["eval_samples"] = min(max_val_samples, len(eval_dataset)) + + trainer.log_metrics("eval", metrics) + trainer.save_metrics("eval", metrics) + + return results + + +if __name__ == "__main__": + main() diff --git a/special_tokens_map.json b/special_tokens_map.json new file mode 100644 index 0000000..9abf719 --- /dev/null +++ b/special_tokens_map.json @@ -0,0 +1 @@ +{"bos_token": "", "eos_token": "", "unk_token": "[UNK]", "pad_token": "[PAD]"} \ No newline at end of file diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000..a2a8340 --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1 @@ +{"unk_token": "[UNK]", "bos_token": "", "eos_token": "", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|"} \ No newline at end of file diff --git a/vocab.json b/vocab.json new file mode 100644 index 0000000..0fc6857 --- /dev/null +++ b/vocab.json @@ -0,0 +1 @@ +{"ف": 0, "ی": 1, "إ": 2, "ا": 3, "ز": 4, "ء": 5, "ک": 6, "م": 7, "ص": 8, "ع": 10, "ؤ": 11, "ذ": 12, "ج": 13, "ٰ": 14, "ك": 15, "ش": 16, "ت": 17, "ه": 18, "غ": 19, "آ": 20, "ض": 21, "ظ": 22, "ل": 23, "ث": 24, "ٌ": 25, "ي": 26, "أ": 27, "ِ": 28, "ر": 29, "ْ": 30, "ة": 31, "ى": 32, "د": 33, "ُ": 34, "ً": 35, "و": 36, "ق": 37, "خ": 38, "ن": 39, "ط": 40, "ح": 41, "ئ": 42, "ّ": 43, "س": 44, "ب": 45, "ٍ": 46, "َ": 47, "|": 9, "[UNK]": 48, "[PAD]": 49} \ No newline at end of file