update experiments scripts

train-experiments.py

@@ -0,0 +1,835 @@
+import pandas as pd
+from tqdm.auto import tqdm
+import random
+from p_tqdm import p_map
+from datasets import load_dataset, load_metric, Audio
+from datasets import load_from_disk, concatenate_datasets
+import torchaudio
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+from datasets import concatenate_datasets, load_dataset
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric, Dataset
+
+import bitsandbytes as bnb
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_pt_utils import get_parameter_names
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+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": ""}, default="hf-test/xls-r-dummy"
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "hf-test/xls-r-dummy"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability 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."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+
+
+# In[4]:
+
+
+@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_name: str = field(
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default="ab", metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_split_name: str = field(
+        default="train+validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    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_eval_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: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+# In[5]:
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            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: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    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,
+            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,
+                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
+
+# download the augmented Dataset from 
+# https://huggingface.co/datasets/bakrianoo/arabic-cv8-augmented
+
+base_path = "/workspace/cv-corpus-8.0-2022-01-19"
+
+# load augmented datasets
+train_ar_df = pd.read_csv(f"{base_path}/train.tsv", sep="\t")
+train_ar_df["audio"] = train_ar_df["path"]
+
+test_ar_df = pd.read_csv(f"{base_path}/test.tsv", sep="\t")
+test_ar_df["audio"] = test_ar_df["path"]
+
+train_ar_df = train_ar_df.sample(frac=1, random_state=101, ignore_index=True)
+
+raw_datasets = DatasetDict()
+
+# select Dataset range
+from_rows = 0
+to_rows = 500_000
+
+saved_vecs_path = f"{base_path}/saved_vec_dataset-{from_rows}-{to_rows}.ds"
+
+raw_datasets["train"] = Dataset.from_pandas(train_ar_df.iloc[from_rows:to_rows])
+raw_datasets["eval"] = Dataset.from_pandas(test_ar_df)
+
+# Audio casting
+raw_datasets["train"] = raw_datasets["train"].cast_column("audio", datasets.features.Audio(sampling_rate=16000))
+raw_datasets["eval"] = raw_datasets["eval"].cast_column("audio", datasets.features.Audio(sampling_rate=16000))
+
+
+parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+
+model_args, data_args, training_args = parser.parse_dict({
+        "dataset_name": "mozilla-foundation/common_voice_8_0",
+        "model_name_or_path": "facebook/wav2vec2-xls-r-300m",
+        "dataset_config_name": "ar",
+        "overwrite_output_dir": False,
+
+        # "preprocessing_only": True,
+    
+        "output_dir": f"{base_path}/output",
+        "text_column_name": "sentence",
+        
+        "freeze_feature_encoder": True,
+        "gradient_checkpointing": True,
+        "group_by_length": False,
+        "push_to_hub": False,
+        "use_auth_token": True,
+        "do_train": True,
+        "do_eval": True,
+        
+        "per_device_train_batch_size":32,
+        "gradient_accumulation_steps":1,
+        "per_device_eval_batch_size":10,
+        
+        "metric_for_best_model":'wer',
+        "evaluation_strategy":"steps",
+        "eval_steps":1000,
+        "logging_strategy":"steps",
+        "logging_steps":500,
+        "save_strategy":"steps",
+        "save_steps":1000,
+        "num_train_epochs":10,
+        "fp16":True,
+        "learning_rate":2e-4,
+        "warmup_steps":1000,
+        "save_total_limit":8,
+        "chars_to_ignore": [':', 'T', '؟', 'ۖ', '…', 'x', 'چ', '?', '.', 'ْ', 'g', '☭', 'w', ';', ',', 'a', 'ۙ', 'e', '`', '“', '!', 'n', 's', '؛', 'ﺃ', 'r', 'ٓ', 'c', '-', 't', 'u', 'l', 'o', '»', 'ٰ', 'ۗ', 'h', 'ڨ', 'ۚ', 'S', '—', 'ٌ', 'm', '”', 'd', 'ۛ', 'H', 'ُ', 'ﻻ', 'y', 'M', 'ھ', 'ک', 'ٍ', 'A', 'ۘ', 'ِ', '–', 'i', 'f', "'", 'ً', '«', 'َ'] + ['\\', '(',')','-','b','c','d','e','g','i','k','p','q','r','u','v','x'],
+
+    })
+
+
+# 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.
+
+# 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)
+
+
+### Load Dataset
+
+
+chars_to_ignore_regex = (
+    f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+)
+text_column_name = data_args.text_column_name
+
+
+def remove_special_characters(batch):
+    if chars_to_ignore_regex is not None:
+        batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+    else:
+        batch["target_text"] = batch[text_column_name].lower() + " "
+    return batch
+
+with training_args.main_process_first(desc="dataset map special characters removal"):
+    
+    raw_datasets = raw_datasets.map(
+        remove_special_characters,
+        remove_columns=[text_column_name],
+        desc="remove special characters from datasets",
+    )
+
+
+data_args.word_delimiter_token
+
+
+# save special tokens for tokenizer
+word_delimiter_token = data_args.word_delimiter_token
+unk_token = data_args.unk_token
+pad_token = data_args.pad_token
+
+# 3. Next, let's load the config as we might need it to create
+# the tokenizer
+# load config
+config = AutoConfig.from_pretrained(
+    model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+)
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+    
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+raw_datasets["train"] = raw_datasets["train"].remove_columns("file_id")
+
+
+# 4. Next, if no tokenizer file is defined,
+# we create the vocabulary of the model by extracting all unique characters from
+# the training and evaluation datasets
+# We need to make sure that only first rank saves vocabulary
+# make sure all processes wait until vocab is created
+tokenizer_name_or_path = model_args.tokenizer_name_or_path
+tokenizer_kwargs = {}
+if tokenizer_name_or_path is None:
+    # save vocab in training output dir
+    tokenizer_name_or_path = training_args.output_dir
+
+    vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+    with training_args.main_process_first():
+        if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+            os.remove(vocab_file)
+
+    with training_args.main_process_first(desc="dataset map vocabulary creation"):
+        if not os.path.isfile(vocab_file):
+            os.makedirs(tokenizer_name_or_path, exist_ok=True)
+            vocab_dict = create_vocabulary_from_data(
+                raw_datasets,
+                word_delimiter_token=word_delimiter_token,
+                unk_token=unk_token,
+                pad_token=pad_token,
+            )
+
+            # save vocab dict to be loaded into tokenizer
+            with open(vocab_file, "w") as file:
+                json.dump(vocab_dict, file)
+
+    # if tokenizer has just been created
+    # it is defined by `tokenizer_class` if present in config else by `model_type`
+    tokenizer_kwargs = {
+        "config": config if config.tokenizer_class is not None else None,
+        "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+        "unk_token": unk_token,
+        "pad_token": pad_token,
+        "word_delimiter_token": word_delimiter_token,
+    }
+
+
+# 5. Now we can instantiate the feature extractor, tokenizer and model
+# Note for distributed training, the .from_pretrained methods guarantee that only
+# one local process can concurrently download model & vocab.
+
+# load feature_extractor and tokenizer
+tokenizer = AutoTokenizer.from_pretrained(
+    tokenizer_name_or_path,
+    use_auth_token=data_args.use_auth_token,
+    **tokenizer_kwargs,
+)
+feature_extractor = AutoFeatureExtractor.from_pretrained(
+    model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+)
+
+
+# adapt config
+config.update(
+    {
+        "feat_proj_dropout": model_args.feat_proj_dropout,
+        "attention_dropout": model_args.attention_dropout,
+        "hidden_dropout": model_args.hidden_dropout,
+        "final_dropout": model_args.final_dropout,
+        "mask_time_prob": model_args.mask_time_prob,
+        "mask_time_length": model_args.mask_time_length,
+        "mask_feature_prob": model_args.mask_feature_prob,
+        "mask_feature_length": model_args.mask_feature_length,
+        "gradient_checkpointing": training_args.gradient_checkpointing,
+        "layerdrop": model_args.layerdrop,
+        "ctc_loss_reduction": model_args.ctc_loss_reduction,
+        "pad_token_id": tokenizer.pad_token_id,
+        "vocab_size": len(tokenizer),
+        "activation_dropout": model_args.activation_dropout,
+    }
+)
+
+
+# create model
+model = AutoModelForCTC.from_pretrained(
+    model_args.model_name_or_path,
+    cache_dir=model_args.cache_dir,
+    config=config,
+    use_auth_token=data_args.use_auth_token,
+)
+
+# freeze encoder
+if model_args.freeze_feature_encoder:
+    model.freeze_feature_encoder()
+
+
+# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+# so that we just need to set the correct target sampling rate and normalize the input
+# via the `feature_extractor`
+
+# make sure that dataset decodes audio with correct sampling rate
+dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+if dataset_sampling_rate != feature_extractor.sampling_rate:
+    raw_datasets = raw_datasets.cast_column(
+        data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+    )
+
+# derive max & min input length for sample rate & max duration
+max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+
+audio_column_name = data_args.audio_column_name
+num_workers = data_args.preprocessing_num_workers
+
+# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+phoneme_language = data_args.phoneme_language
+
+
+# Preprocessing the datasets.
+# We need to read the audio files as arrays and tokenize the targets.
+def prepare_dataset(batch):
+    # load audio
+    sample = batch[audio_column_name]
+
+    inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+    batch["input_values"] = inputs.input_values[0]
+    batch["input_length"] = len(batch["input_values"])
+
+    # encode targets
+    additional_kwargs = {}
+    if phoneme_language is not None:
+        additional_kwargs["phonemizer_lang"] = phoneme_language
+
+    batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+    return batch
+
+def vectorizing_record(audio_path, target_text):
+    batch = {}
+    
+    array, sampling_rate = torchaudio.load(audio_path, format="mp3")
+    
+    batch["input_values"] = array.mean(axis=0)
+    batch["input_length"] = len(array)
+
+    # encode targets
+    additional_kwargs = {}
+    if phoneme_language is not None:
+        additional_kwargs["phonemizer_lang"] = phoneme_language
+
+    batch["labels"] = tokenizer(target_text, **additional_kwargs).input_ids
+    return batch
+
+
+# In[ ]:
+
+print(f"========\n\n{num_workers}\n\n========")
+with training_args.main_process_first(desc="dataset map preprocessing"):
+    saved_vecs_path = f"{base_path}/saved_vec_dataset-{from_rows}-{to_rows}.ds"
+    if not os.path.exists(saved_vecs_path):
+
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+        
+        # save to local disk
+        vectorized_datasets.save_to_disk(saved_vecs_path)
+    else:    
+        # read from disk
+        vectorized_datasets = load_from_disk(saved_vecs_path)
+
+print(vectorized_datasets)
+
+# 7. Next, we can prepare the training.
+# Let's use word error rate (WER) as our evaluation metric,
+# instantiate a data collator and the trainer
+
+# Define evaluation metrics during training, *i.e.* word error rate, character error rate
+eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+vectorized_datasets["train"] = vectorized_datasets["train"].remove_columns("input_length")
+vectorized_datasets["eval"] = vectorized_datasets["eval"].remove_columns("input_length")
+
+# for large datasets it is advised to run the preprocessing on a
+# single machine first with ``args.preprocessing_only`` since there will mostly likely
+# be a timeout when running the script in distributed mode.
+# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+# cached dataset
+if data_args.preprocessing_only:
+    logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+
+
+
+def compute_metrics(pred):
+    pred_logits = pred.predictions
+    pred_ids = np.argmax(pred_logits, axis=-1)
+
+    pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+    pred_str = tokenizer.batch_decode(pred_ids)
+
+    # we do not want to group tokens when computing the metrics
+    label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+    metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+    return metrics
+
+# Now save everything to be able to create a single processor later
+if is_main_process(training_args.local_rank):
+    # save feature extractor, tokenizer and config
+    feature_extractor.save_pretrained(training_args.output_dir)
+    tokenizer.save_pretrained(training_args.output_dir)
+    config.save_pretrained(training_args.output_dir)
+
+try:
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+except (OSError, KeyError):
+    warnings.warn(
+        "Loading a processor from a feature extractor config that does not"
+        " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+        " attribute to your `preprocessor_config.json` file to suppress this warning: "
+        " `'processor_class': 'Wav2Vec2Processor'`",
+        FutureWarning,
+    )
+    processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+# Instantiate custom data collator
+data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+
+decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
+decay_parameters = [name for name in decay_parameters if "bias" not in name]
+
+optimizer_grouped_parameters = [
+    {
+        "params": [p for n, p in model.named_parameters() if n in decay_parameters],
+        "weight_decay": training_args.weight_decay,
+    },
+    {
+        "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
+        "weight_decay": 0.0,
+    },
+]
+
+optimizer = bnb.optim.Adam8bit(
+    params=optimizer_grouped_parameters,
+    lr=training_args.learning_rate,
+    betas=(training_args.adam_beta1, training_args.adam_beta2),
+    eps=training_args.adam_epsilon,
+)
+
+optimizers = (optimizer, None)
+
+
+# Initialize Trainer
+trainer = Trainer(
+    model=model,
+    data_collator=data_collator,
+    args=training_args,
+    compute_metrics=compute_metrics,
+    train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+    eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+    tokenizer=feature_extractor,
+    optimizers=optimizers,
+)
+
+
+
+# 8. Finally, we can start training
+
+# Training
+if training_args.do_train and not data_args.preprocessing_only:
+
+    # use last checkpoint if exist
+    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()
+
+    metrics = train_result.metrics
+    max_train_samples = (
+        data_args.max_train_samples
+        if data_args.max_train_samples is not None
+        else len(vectorized_datasets["train"])
+    )
+    metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+    trainer.log_metrics("train", metrics)
+    trainer.save_metrics("train", metrics)
+    trainer.save_state()
+
+
+# Evaluation
+results = {}
+if training_args.do_eval and not data_args.preprocessing_only:
+    logger.info("*** Evaluate ***")
+    metrics = trainer.evaluate()
+    max_eval_samples = (
+        data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+    )
+    metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+    trainer.log_metrics("eval", metrics)
+    trainer.save_metrics("eval", metrics)
+
+# Write model card and (optionally) push to hub
+config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+kwargs = {
+    "finetuned_from": model_args.model_name_or_path,
+    "tasks": "speech-recognition",
+    "tags": ["automatic-speech-recognition", data_args.dataset_name],
+    "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+    "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+}
+
+if not data_args.preprocessing_only:
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    print(results)
+