test

old/run_whisper_finetuning.py

@@ -0,0 +1,813 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers Whisper model for automatic speech recognition"""
+
+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
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric, AUdio
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+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
+from transformers import WhisperFeatureExtractor
+from transformers import WhisperTokenizer
+from transformers import WhisperProcessor
+from transformers import WhisperForConditionalGeneration
+
+from transformers import Seq2SeqTrainingArguments
+from transformers import Seq2SeqTrainer
+
+import evaluate
+
+logger = logging.getLogger(__name__)
+
+
+# ** Check if this is needed...
+
+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"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer 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_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'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@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=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    # )
+    train_split_name: str = field(
+        default="train",
+        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."
+        },
+    )
+
+
+@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
+
+
+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
+    alphabet = set()
+
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        alphabet.update(all_text)
+
+    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: {"vocab": list(set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]))}, vocabs.values()
+    # )["vocab"][0]
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(alphabet)))}
+
+    # 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
+
+
+def make_dataset(seed=42):
+    # Pre-processing dataset
+    import re
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...Vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text}
+
+    def filter_nst(entry):
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        return True
+
+    def filter_npsc(entry):
+        # False if there are digits in the text
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    # TODO NST_hesitate
+
+    split = len(npsc["train"]) / (len(npsc["train"]) + len(npsc["validation"]))  # Use same train/val ratio as NPSC
+    nst_train = nst["train"].train_test_split(train_size=split, seed=seed)
+    nst["train"] = nst_train["train"]
+    nst["validation"] = nst_train["test"]
+
+    nst = nst.filter(filter_nst).map(map_nst).shuffle(seed=seed)
+    npsc = npsc.filter(filter_npsc).map(map_npsc).shuffle(seed=seed)
+
+    npsc_base = npsc.remove_columns([col for col in npsc["train"].column_names if col not in ["text", "audio"]])
+    nst_base = nst.remove_columns([col for col in nst["train"].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in "train", "validation", "test":
+        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+
+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)
+
+    # 1. First, let's load the dataset
+    raw_datasets = make_dataset(seed=training_args.seed)
+
+    if training_args.do_train:
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    # chars_to_ignore_regex = (
+    #    f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+    # )
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'
+
+    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",
+        )
+
+    # 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
+    )
+
+    # 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,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "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
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        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"],
+        )
+
+    # 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}
+
+    # 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}")
+        return
+
+    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)
+
+    # 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["validation"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # 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:
+        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 "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)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

run_speech_recognition_whisper_pere.py → old/run_whisper_finetuning_old.py

@@ -22,7 +22,7 @@ import re
 import sys
 import warnings
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional,Union
+from typing import Any, Dict, List, Optional, Union
 import evaluate
 
 import numpy as np
@@ -47,9 +47,11 @@ 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
 
+
 def list_field(default=None, metadata=None):
     return field(default_factory=lambda: default, metadata=metadata)
 
+
 @dataclass
 class ModelArguments:
     """
@@ -57,21 +59,25 @@ class ModelArguments:
     """
 
     model_name_or_path: str = field(
-        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+        metadata={
+            "help": "Path to pretrained model or model identifier from huggingface.co/models"}
     )
     language: str = field(
         metadata={"help": "Whisper specific language"}
     )
     task: str = field(
-        metadata={"help": "Whisper specific task, i.e., 'transcribe' or 'translate'"}
+        metadata={
+            "help": "Whisper specific task, i.e., 'transcribe' or 'translate'"}
     )
     tokenizer_name_or_path: Optional[str] = field(
         default=None,
-        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+        metadata={
+            "help": "Path to pretrained tokenizer or tokenizer 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"},
+        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."}
@@ -82,7 +88,8 @@ class ModelArguments:
     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."})
+    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={
@@ -91,7 +98,8 @@ class ModelArguments:
     )
     final_dropout: float = field(
         default=0.0,
-        metadata={"help": "The dropout probability for the final projection layer."},
+        metadata={
+            "help": "The dropout probability for the final projection layer."},
     )
     mask_time_prob: float = field(
         default=0.05,
@@ -116,7 +124,8 @@ class ModelArguments:
         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."})
+    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'."}
     )
@@ -136,7 +145,8 @@ class DataTrainingArguments:
     """
 
     dataset_name: str = field(
-        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."}
     )
     dataset_config_name: str = field(
         default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
@@ -155,11 +165,13 @@ class DataTrainingArguments:
     )
     audio_column_name: str = field(
         default="audio",
-        metadata={"help": "The name of the dataset column containing the audio data. Defaults to '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'"},
+        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."}
@@ -188,7 +200,8 @@ class DataTrainingArguments:
     )
     eval_metrics: List[str] = list_field(
         default=["wer"],
-        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+        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,
@@ -229,7 +242,8 @@ class DataTrainingArguments:
     )
     predict_with_generate: bool = field(
         default=True,
-        metadata={"help": "Output tokens in addition to loss and digits for calculating metrics"},
+        metadata={
+            "help": "Output tokens in addition to loss and digits for calculating metrics"},
     )
     generation_max_length: int = field(
         default=225,
@@ -259,16 +273,21 @@ class DataCollatorSpeechSeq2SeqWithPadding:
     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")
+        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]
+        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")
+        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)
+        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
@@ -279,15 +298,16 @@ class DataCollatorSpeechSeq2SeqWithPadding:
         return batch
 
 
-def main():    
+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))
+    parser = HfArgumentParser(
+        (ModelArguments, DataTrainingArguments, TrainingArguments))
     model_args, data_args, training_args = parser.parse_args_into_dataclasses()
- 
 
     # Metrics
+
     def compute_metrics(pred):
         pred_ids = pred.predictions
         label_ids = pred.label_ids
@@ -315,16 +335,16 @@ def main():
         # encode target text to label ids
         batch["labels"] = tokenizer(batch["sentence"]).input_ids
         return batch
-    
+
     def print_training_arguments(model_args, data_args, training_args):
         print("Starting with the following parameters:")
         print("\n* Model arguments:")
-        pprint(vars(model_args),indent=2)
+        pprint(vars(model_args), indent=2)
         print("\n* Data arguments")
-        pprint(vars(data_args),indent=2)
+        pprint(vars(data_args), indent=2)
         print("\n* Training arguments")
-        pprint(vars(training_args),indent=2)
-    
+        pprint(vars(training_args), indent=2)
+
     # TODO - Might use this function later
     # def make_dataset(training_args, data_args):
     #     seed = training_args.seed or 42
@@ -338,28 +358,26 @@ def main():
     # Load dataset
     speech_data = DatasetDict()
     speech_data["train"] = load_dataset(
-        data_args.dataset_name, data_args.dataset_config_name, split="train", use_auth_token=True)
+        data_args.dataset_name, data_args.dataset_config_name, split="train", streaming=True, use_auth_token=True)
     speech_data["test"] = load_dataset(
-        data_args.dataset_name, data_args.dataset_config_name, split="test", use_auth_token=True)
+        data_args.dataset_name, data_args.dataset_config_name, split="test", streaming=True, use_auth_token=True)
 
     # TODO - Implement streaming and include this
     # speech_data = make_dataset(training_args, data_args)
 
-
-    # Adapt dataset - Change column names and delete extra data
-    # Map columns
+    breakpoint()
+    
+    # Rename columns
     if "audio" not in speech_data.column_names["train"]:
-        speech_data = speech_data.rename_column(source, "audio")
+        speech_data = speech_data.rename_column(
+            data_args.audio_column_name, "audio")
 
     if "sentence" not in speech_data.column_names["train"]:
-        speech_data = speech_data.rename_column(target, "sentence")
-
-    # Remove not needed columns 
-    remove_list = [i for i in speech_data.column_names["train"]
-                if i not in ["audio", "sentence"]]
-
-    speech_data = speech_data.remove_columns(remove_list)
+        speech_data = speech_data.rename_column(data_args.text_column_name, "sentence")
 
+    # Remove not needed columns
+    speech_data = speech_data.remove_columns(
+        [i for i in speech_data.column_names["train"] if i not in ["audio", "sentence"]])
 
     # Initialise
     feature_extractor = WhisperFeatureExtractor.from_pretrained(
@@ -375,12 +393,10 @@ def main():
     speech_data = speech_data.map(
         prepare_dataset, remove_columns=speech_data.column_names["train"], num_proc=1)
 
-
-
     # Metrics
     metric = evaluate.load("wer")
 
-    #Detecting last checkpoint.
+    # 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)
@@ -393,8 +409,8 @@ def main():
             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."
-        
-    ) 
+
+            )
 
    # Training
     if training_args.do_train:
@@ -403,22 +419,21 @@ def main():
         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         
+            checkpoint = model_args.model_name_or_path
         else:
             checkpoint = None
-            
+
          # We need to set use_cache=False here if we want to use gradient accumulation
         model = WhisperForConditionalGeneration.from_pretrained(
-                "openai/whisper-small", use_cache=False)
+            "openai/whisper-small", use_cache=False)
 
         # Overriding 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)):
         model.config.forced_decoder_ids = None
         model.config.suppress_tokens = []
-        
+
         # Set seed before initializing model.
         set_seed(training_args.seed)
 
-
         trainer = Seq2SeqTrainer(
             args=training_args,
             model=model,
@@ -428,7 +443,7 @@ def main():
             compute_metrics=compute_metrics,
             tokenizer=processor.feature_extractor,
         )
- 
+
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()
 
@@ -436,23 +451,25 @@ def main():
         trainer.log_metrics("train", metrics)
         trainer.save_metrics("train", metrics)
         trainer.save_state()
-        
+
         if training_args.push_to_hub:
             trainer.push_to_hub(**kwargs)
         else:
             trainer.create_model_card(**kwargs)
- 
+
     # TODO - Look closer into the evaluation and the model card writing.
-    
+
     # Evaluation
     results = {}
     if training_args.do_eval:
         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"])
+            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"]))
+        metrics["eval_samples"] = min(
+            max_eval_samples, len(vectorized_datasets["eval"]))
 
         trainer.log_metrics("eval", metrics)
         trainer.save_metrics("eval", metrics)
@@ -467,15 +484,16 @@ def main():
         "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
         "language": model_args.language,
     }
-    
+
     return results
 
 
-#XLA hook
+# XLA hook
 def _mp_fn(index):
     # For xla_spawn (TPUs)
     print("The XLA is initiated")
     main()
 
+
 if __name__ == "__main__":
     main()

run.sh

@@ -1,31 +1,30 @@
 
-python run_speech_recognition_whisper_pere.py \
+python run_whisper_finetuning.py \
 	--model_name_or_path="openai/whisper-small" \
 	--output_dir="../whisper-testrun1" \
+	--repo_id="NbAiLab/whisper-testrun1" \
 	--overwrite_output_dir=True \
 	--language="Norwegian" \
 	--task="transcribe" \
-	--dataset_name="mozilla-foundation/common_voice_11_0" \
-	--dataset_config="nn-NO" \
-	--output_dir="./whisper-small-hi" \
+	--dataset_name="NbAiLab/NPSC" \
+	--dataset_config="16K_mp3" \
 	--do_train=True \
 	--do_eval=True \
 	--audio_column_name="audio" \
-	--text_column_name="sentence" \
+	--text_column_name="normsentence_text" \
 	--per_device_train_batch_size=16 \
        	--per_device_train_batch_size=16 \
 	--learning_rate=2e-5 \
 	--warmup_steps=500 \
-	--max_steps=5000 \
+	--max_steps=10000 \
 	--gradient_checkpointing=True \
 	--gradient_accumulation_steps=1 \
-	--group_by_length=True \
+	--group_by_length=False \
 	--evaluation_strategy="steps" \
 	--save_steps=1000 \
 	--eval_steps=1000 \
-	--logging_steps=25 \
+	--logging_steps=250 \
 	--fp16=True \
-	--save_steps=1000 \
 	--load_best_model_at_end=True \
 	--metric_for_best_model="wer" \
 	--greater_is_better=False \

run_whisper_finetuning.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python
 # coding=utf-8
-# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -23,43 +22,31 @@ import re
 import sys
 import warnings
 from dataclasses import dataclass, field
-from typing import Dict, List, Optional, Union
+from typing import Any, Dict, List, Optional, Union
+import evaluate
 
-import datasets
 import numpy as np
 import torch
-from datasets import DatasetDict, load_dataset, load_metric, AUdio
+from pprint import pprint
+import evaluate
+from datasets import DatasetDict, load_dataset
+from datasets import Audio
 
-import transformers
 from transformers import (
-    AutoConfig,
-    AutoFeatureExtractor,
-    AutoModelForCTC,
-    AutoProcessor,
-    AutoTokenizer,
     HfArgumentParser,
-    Trainer,
     TrainingArguments,
-    Wav2Vec2Processor,
     set_seed,
+    WhisperFeatureExtractor,
+    WhisperTokenizer,
+    WhisperForConditionalGeneration,
+    WhisperProcessor,
+    Seq2SeqTrainer,
+    Seq2SeqTrainingArguments,
 )
 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
-from transformers import WhisperFeatureExtractor
-from transformers import WhisperTokenizer
-from transformers import WhisperProcessor
-from transformers import WhisperForConditionalGeneration
-
-from transformers import Seq2SeqTrainingArguments
-from transformers import Seq2SeqTrainer
-
-import evaluate
-
-logger = logging.getLogger(__name__)
-
 
-# ** Check if this is needed...
 
 def list_field(default=None, metadata=None):
     return field(default_factory=lambda: default, metadata=metadata)
@@ -72,15 +59,25 @@ class ModelArguments:
     """
 
     model_name_or_path: str = field(
-        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+        metadata={
+            "help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    language: str = field(
+        metadata={"help": "Whisper specific language"}
+    )
+    task: str = field(
+        metadata={
+            "help": "Whisper specific task, i.e., 'transcribe' or 'translate'"}
     )
     tokenizer_name_or_path: Optional[str] = field(
         default=None,
-        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+        metadata={
+            "help": "Path to pretrained tokenizer or tokenizer 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"},
+        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."}
@@ -91,7 +88,8 @@ class ModelArguments:
     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."})
+    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={
@@ -100,7 +98,8 @@ class ModelArguments:
     )
     final_dropout: float = field(
         default=0.0,
-        metadata={"help": "The dropout probability for the final projection layer."},
+        metadata={
+            "help": "The dropout probability for the final projection layer."},
     )
     mask_time_prob: float = field(
         default=0.05,
@@ -125,7 +124,8 @@ class ModelArguments:
         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."})
+    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'."}
     )
@@ -144,12 +144,13 @@ class DataTrainingArguments:
     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=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
-    # )
+    dataset_name: str = field(
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
     train_split_name: str = field(
         default="train",
         metadata={
@@ -164,11 +165,13 @@ class DataTrainingArguments:
     )
     audio_column_name: str = field(
         default="audio",
-        metadata={"help": "The name of the dataset column containing the audio data. Defaults to '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'"},
+        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."}
@@ -197,7 +200,8 @@ class DataTrainingArguments:
     )
     eval_metrics: List[str] = list_field(
         default=["wer"],
-        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+        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,
@@ -236,6 +240,15 @@ class DataTrainingArguments:
         default="|",
         metadata={"help": "The word delimiter token for the tokenizer"},
     )
+    predict_with_generate: bool = field(
+        default=True,
+        metadata={
+            "help": "Output tokens in addition to loss and digits for calculating metrics"},
+    )
+    generation_max_length: int = field(
+        default=225,
+        metadata={"help": "Maximum number of tokens generated"},
+    )
     phoneme_language: Optional[str] = field(
         default=None,
         metadata={
@@ -245,202 +258,133 @@ class DataTrainingArguments:
                     " input audio to a sequence of phoneme sequences."
         },
     )
+    print_training_arguments: bool = field(
+        default=True,
+        metadata={
+            "help": "Prints the training arguments. For debugging"
+        },
+    )
 
 
 @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
+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 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",
-            )
+        # 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)
+        labels = labels_batch["input_ids"].masked_fill(
+            labels_batch.attention_mask.ne(1), -100)
 
-        batch["labels"] = labels
+        # 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
 
 
-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
-    alphabet = set()
-
-    def extract_all_chars(batch):
-        all_text = " ".join(batch["target_text"])
-        alphabet.update(all_text)
-
-    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: {"vocab": list(set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]))}, vocabs.values()
-    # )["vocab"][0]
-
-    vocab_dict = {v: k for k, v in enumerate(sorted(list(alphabet)))}
-
-    # 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
-
-
-def make_dataset(seed=42):
-    # Pre-processing dataset
-    import re
-
-    def map_nst(entry):
-        text = entry["text"].lower()
-        text = text.replace("(...Vær stille under dette opptaket...)", "")
-        text = re.sub('[áàâ]', 'a', text)
-        text = re.sub('[ä]', 'æ', text)
-        text = re.sub('[éèëê]', 'e', text)
-        text = re.sub('[íìïî]', 'i', text)
-        text = re.sub('[óòöô]', 'o', text)
-        text = re.sub('[ö]', 'ø', text)
-        text = re.sub('[ç]', 'c', text)
-        text = re.sub('[úùüû]', 'u', text)
-        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
-        text = re.sub('\s+', ' ', text)
-        return {"text": text}
-
-    def filter_nst(entry):
-        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
-            return False  # Too short
-        if re.match(entry["type"], "pIW|CA"):
-            return False  # Spelling out words
-        return True
-
-    def filter_npsc(entry):
-        # False if there are digits in the text
-        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
-            return False  # Too short
-        if re.search("\d", entry["text"]):
-            return False
-        return True
-
-    def map_npsc(entry):
-        batch = {"text": entry["text"].lower()}
-        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
-        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
-        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
-        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
-        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
-        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
-        batch["text"] = re.sub('[ç]', 'c', batch["text"])
-        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
-        batch["text"] = re.sub('\s', ' ', batch["text"])
-        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
-        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
-        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
-        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
-        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
-        if "<" in batch["text"]:
-            raise ValueError(batch["text"])
-        return batch
+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))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Metrics
 
-    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
-    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
-    # TODO NST_hesitate
+    def compute_metrics(pred):
+        pred_ids = pred.predictions
+        label_ids = pred.label_ids
 
-    split = len(npsc["train"]) / (len(npsc["train"]) + len(npsc["validation"]))  # Use same train/val ratio as NPSC
-    nst_train = nst["train"].train_test_split(train_size=split, seed=seed)
-    nst["train"] = nst_train["train"]
-    nst["validation"] = nst_train["test"]
+        # replace -100 with the pad_token_id
+        label_ids[label_ids == -100] = tokenizer.pad_token_id
 
-    nst = nst.filter(filter_nst).map(map_nst).shuffle(seed=seed)
-    npsc = npsc.filter(filter_npsc).map(map_npsc).shuffle(seed=seed)
+        # we do not want to group tokens when computing the metrics
+        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
+        label_str = tokenizer.batch_decode(label_ids, skip_special_tokens=True)
 
-    npsc_base = npsc.remove_columns([col for col in npsc["train"].column_names if col not in ["text", "audio"]])
-    nst_base = nst.remove_columns([col for col in nst["train"].column_names if col not in ["text", "audio"]])
+        wer = 100 * metric.compute(predictions=pred_str, references=label_str)
 
-    combined = {}
-    for split in "train", "validation", "test":
-        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
-        probs = (probs / probs.sum()).tolist()
-        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
-        combined[split] = comb
+        return {"wer": wer}
 
-    return datasets.DatasetDict(**combined)
+    # Prepare dataset
+    def prepare_dataset(batch):
+        # load and resample audio data from 48 to 16kHz
+        audio = batch["audio"]
 
+        # compute log-Mel input features from input audio array
+        batch["input_features"] = feature_extractor(
+            audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
 
-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.
+        # encode target text to label ids
+        batch["labels"] = tokenizer(batch["sentence"]).input_ids
+        return batch
 
-    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()
+    def print_training_arguments(model_args, data_args, training_args):
+        print("Starting with the following parameters:")
+        print("\n* Model arguments:")
+        pprint(vars(model_args), indent=2)
+        print("\n* Data arguments")
+        pprint(vars(data_args), indent=2)
+        print("\n* Training arguments")
+        pprint(vars(training_args), indent=2)
+
+    # Print training arguments
+    if data_args.print_training_arguments:
+        print_training_arguments(model_args, data_args, training_args)
+
+    # Load dataset
+    train_dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name, split="train", streaming=True, use_auth_token=True)
+    eval_dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name, split="validation", streaming=True, use_auth_token=True)
+
+    
+    # Rename columns
+    if data_args.audio_column_name != "audio":
+        train_dataset = train_dataset.rename_column(data_args.audio_column_name, "audio")
+        eval_dataset = eval_dataset.rename_column(data_args.audio_column_name, "audio")
+
+    if data_args.text_column_name != "sentence":
+        train_dataset = train_dataset.rename_column(data_args.text_column_name, "sentence")
+        eval_dataset = eval_dataset.rename_column(data_args.text_column_name, "sentence")
+
+    
+    # Initialise
+    feature_extractor = WhisperFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path)
+    tokenizer = WhisperTokenizer.from_pretrained(
+        model_args.model_name_or_path, language=model_args.language, task=model_args.task)
+    processor = WhisperProcessor.from_pretrained(
+        model_args.model_name_or_path, language=model_args.language, task=model_args.task)
+    data_collator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)
+
+    
+    # Prepare data
+    train_dataset = train_dataset.cast_column("audio", Audio(sampling_rate=16000))
+    eval_dataset = eval_dataset.cast_column("audio", Audio(sampling_rate=16000))
+    
+    # TODO Not able to implement in Streaming mode. Can not find a way to list columns. But is is necessary?
+    # train_data = train_data.map(prepare_dataset, remove_columns=train_data.column_names, num_proc=1)
+    
+    train_dataset = train_dataset.map(prepare_dataset)
+    
+    # Metrics
+    metric = evaluate.load("wer")
 
     # Detecting last checkpoint.
     last_checkpoint = None
@@ -455,302 +399,10 @@ def main():
             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)
-
-    # 1. First, let's load the dataset
-    raw_datasets = make_dataset(seed=training_args.seed)
-
-    if training_args.do_train:
-        if data_args.audio_column_name not in raw_datasets["train"].column_names:
-            raise ValueError(
-                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
-                "Make sure to set `--audio_column_name` to the correct audio column - one of "
-                f"{', '.join(raw_datasets['train'].column_names)}."
-            )
-
-        if data_args.text_column_name not in raw_datasets["train"].column_names:
-            raise ValueError(
-                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
-                "Make sure to set `--text_column_name` to the correct text column - one of "
-                f"{', '.join(raw_datasets['train'].column_names)}."
             )
 
-        if data_args.max_train_samples is not None:
-            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
-
-    if training_args.do_eval:
-        if data_args.max_eval_samples is not None:
-            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
-
-    # 2. We remove some special characters from the datasets
-    # that make training complicated and do not help in transcribing the speech
-    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
-    # that could be easily picked up by the model
-    # chars_to_ignore_regex = (
-    #    f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
-    # )
-    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'
-
-    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",
-        )
-
-    # 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
-    )
-
-    # 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,
-            "ctc_zero_infinity": model_args.ctc_zero_infinity,
-            "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
-
-    with training_args.main_process_first(desc="dataset map preprocessing"):
-        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"],
-        )
-
-    # 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}
-
-    # 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}")
-        return
-
-    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)
-
-    # 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["validation"] if training_args.do_eval else None,
-        tokenizer=feature_extractor,
-    )
-
-    # 8. Finally, we can start training
-
-    # Training
+   # Training
     if training_args.do_train:
 
         # use last checkpoint if exist
@@ -761,30 +413,53 @@ def main():
         else:
             checkpoint = None
 
+         # We need to set use_cache=False here if we want to use gradient accumulation
+        model = WhisperForConditionalGeneration.from_pretrained(
+            "openai/whisper-small", use_cache=False)
+
+        # Overriding 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)):
+        model.config.forced_decoder_ids = None
+        model.config.suppress_tokens = []
+
+        # Set seed before initializing model.
+        set_seed(training_args.seed)
+
+        trainer = Seq2SeqTrainer(
+            args=training_args,
+            model=model,
+            train_dataset=train_dataset.with_format("torch"),
+            eval_dataset=eval_dataset.with_format("torch"),
+            data_collator=data_collator,
+            compute_metrics=compute_metrics,
+            tokenizer=processor.feature_extractor,
+        )
+
         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()
 
+        if training_args.push_to_hub:
+            trainer.push_to_hub(**kwargs)
+        else:
+            trainer.create_model_card(**kwargs)
+
+    # TODO - Look closer into the evaluation and the model card writing.
+
     # Evaluation
     results = {}
     if training_args.do_eval:
         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"])
+            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"]))
+        metrics["eval_samples"] = min(
+            max_eval_samples, len(vectorized_datasets["eval"]))
 
         trainer.log_metrics("eval", metrics)
         trainer.save_metrics("eval", metrics)
@@ -793,21 +468,22 @@ def main():
     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],
+        "tasks": "automatic-speech-recognition",
+        "tags": ["hf-asr-leaderboard", "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()}",
+        "language": model_args.language,
     }
-    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)
 
     return results
 
 
+# XLA hook
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    print("The XLA is initiated")
+    main()
+
+
 if __name__ == "__main__":
-    main()
+    main()

run_xla.sh

@@ -1,4 +1,4 @@
 
-python xla_spawn.py --num_cores=4 run_whisper.py 
+python xla_spawn.py --num_cores=4 run_whisper_finetuining.py