initial commit

deep_run.sh

@@ -0,0 +1,44 @@
+deepspeed run_speech_recognition_seq2seq_deepspeed.py \
+	--deepspeed="ds_config.json" \
+	--model_name_or_path="openai/whisper-large-v2" \
+	--dataset_name="KBLab/rixvox" \
+	--dataset_config_name="train" \
+    	--evalset_name="mozilla-foundation/common_voice_11_0" \
+    	--evalset_config_name="sv-SE" \
+	--language="swedish" \
+	--train_split_name="train" \
+	--eval_split_name="test" \
+	--model_index_name="Whisper Large Swedish Rixvox" \
+	--text_column_name="text" \
+    	--eval_text_column_name="sentence" \
+	--max_steps="20000" \
+	--output_dir="./" \
+	--per_device_train_batch_size="64" \
+	--gradient_accumulation_steps="2" \
+	--per_device_eval_batch_size="32" \
+	--logging_steps="25" \
+	--learning_rate="1e-5" \
+	--seed="42" \
+	--warmup_steps="500" \
+	--evaluation_strategy="steps" \
+	--eval_steps="1000" \
+	--save_strategy="steps" \
+	--save_steps="1000" \
+	--generation_max_length="225" \
+	--length_column_name="input_length" \
+	--max_duration_in_seconds="30" \
+	--freeze_feature_encoder="False" \
+	--report_to="tensorboard" \
+	--metric_for_best_model="wer" \
+	--greater_is_better="False" \
+	--load_best_model_at_end \
+	--gradient_checkpointing \
+	--fp16 \
+	--overwrite_output_dir \
+	--do_train \
+	--do_eval \
+	--predict_with_generate \
+	--do_normalize_eval \
+	--streaming="True" \
+	--use_auth_token \
+	--push_to_hub

ds_config.json

@@ -0,0 +1,50 @@
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+            "lr": "auto",
+            "betas": "auto",
+            "eps": "auto",
+            "weight_decay": "auto"
+        }
+    },
+
+   "scheduler": {
+         "type": "WarmupDecayLR",
+         "params": {
+             "last_batch_iteration": -1,
+             "total_num_steps": "auto",
+             "warmup_min_lr": "auto",
+             "warmup_max_lr": "auto",
+             "warmup_num_steps": "auto"
+         }
+     },
+
+    "zero_optimization": {
+        "stage": 2,
+        "offload_optimizer": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "allgather_partitions": true,
+        "allgather_bucket_size": 2e8,
+        "overlap_comm": true,
+        "reduce_scatter": true,
+        "reduce_bucket_size": 2e8,
+        "contiguous_gradients": true
+    },
+
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto"
+}

requirements.txt

@@ -0,0 +1,14 @@
+
+torch>=1.7
+git+https://github.com/huggingface/transformers
+git+https://github.com/huggingface/datasets
+librosa
+jiwer
+evaluate>=0.3.0
+more-itertools
+tensorboard
+deepspeed
+accelerate
+pysrt
+prefetch_generator
+audiomentations

run_speech_recognition_seq2seq_deepspeed.py

@@ -0,0 +1,911 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2022 The HuggingFace 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
+# limitations under the License.
+"""
+Fine-tuning the library models for sequence to sequence speech recognition
+with 🤗 Datasets' streaming mode.
+"""
+# This progam was modified by Michael Kamfonas ([email protected]) on Dec 11 2022
+# - added options for drpout, gradient_checkpointing, use_cache, stopping_strategy and streaming
+# - restructured it to enable both streaming and non-streaming modes 
+# - allows concatenation of mutiple datasets (single-string comma-separated) for interleaving
+#     The following params must have the same number of comma-separated (,) elements:
+#       dataset_name, 
+#       dataset_config_name, 
+#       train_split_name and eval_split_name (each element plus-separated (+) for multiple splits), 
+#       text_column_name and audio_column_name 
+ 
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+
+import datasets
+import torch
+from datasets import Audio, DatasetDict, IterableDatasetDict, interleave_datasets, load_dataset
+from torch.utils.data import IterableDataset
+
+import evaluate
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForSpeechSeq2Seq,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Seq2SeqTrainer,
+    Seq2SeqTrainingArguments,
+    TrainerCallback,
+    set_seed,
+)
+from transformers.trainer_pt_utils import IterableDatasetShard
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version, send_example_telemetry
+from transformers.utils.versions import require_version
+from transformers.models.whisper.english_normalizer import BasicTextNormalizer
+from prefetch_generator import BackgroundGenerator
+import torchaudio.transforms as T
+import numpy as np
+from transformers import WhisperTokenizer
+
+TEXT_COL_NAME="text"
+AUDIO_COL_NAME="audio"
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.25.0.dev0")
+
+require_version("datasets>=1.18.2", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
+
+require_version("prefetch-generator>=1.0.3", "To fix: pip install prefetch-generator")
+
+logger = logging.getLogger(__name__)
+
+
+@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"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Will use the token generated when running `huggingface-cli login` (necessary to use this script "
+                "with private models)."
+            )
+        },
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Deprecated - Whether to freeze the feature encoder layers of the model."}
+    )
+    freeze_encoder: bool = field(
+        default=False, metadata={"help": "Whether to freeze the entire encoder of the seq2seq model."}
+    )
+    forced_decoder_ids: List[List[int]] = field(
+        default=None,
+        metadata={
+            "help": (
+                "A list of pairs of integers which indicates a mapping from generation indices to token indices "
+                "that will be forced before sampling. For example, [[0, 123]] means the first generated token "
+                "will always be a token of index 123."
+            )
+        },
+    )
+    suppress_tokens: List[int] = field(
+        default=None, metadata={"help": "A list of tokens that will be suppressed at generation."}
+    )
+    model_index_name: str = field(default=None, metadata={"help": "Pretty name for the model card."})
+
+    ## added by Michael Kamfonas
+    use_cache: bool = field(
+        default=False, metadata={"help": "Whether to use cache."}
+    )
+
+    dropout: float = field(
+        default = 0.0, metadata = {"help": "dropout probability."}
+    )
+
+    attention_dropout: float = field(
+        default = 0.0, metadata = {"help": "attention_dropout probability."}
+    )
+
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: str = field(
+        default=None, 
+        metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, 
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    evalset_name: str = field(
+        default=None, 
+        metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    evalset_config_name: Optional[str] = field(
+        default=None, 
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the transcription."},
+    )
+    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 evaluation examples to this "
+                "value if set."
+            )
+        },
+    )
+    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'"},
+    )
+    eval_text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the evalset column containing the text data. Defaults to 'text'"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": (
+                "Truncate 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"}
+    )
+    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'"
+        },
+    )
+    do_lower_case: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+    do_remove_punctuation: bool = field(
+        default=False,
+        metadata={"help": "Whether the target text should be striped of punctuation."},
+    )
+    do_normalize_eval: bool = field(
+        default=True,
+        metadata={"help": "Whether to normalise the references and predictions in the eval WER calculation."},
+    )
+    language: str = field(
+        default=None,
+        metadata={
+            "help": (
+                "Language for multilingual fine-tuning. This argument should be set for multilingual fine-tuning "
+                "only. For English speech recognition, it should be set to `None`."
+            )
+        },
+    )
+    task: str = field(
+        default="transcribe",
+        metadata={"help": "Task, either `transcribe` for speech recognition or `translate` for speech translation."},
+    )
+    shuffle_buffer_size: Optional[int] = field(
+        default=500,
+        metadata={
+            "help": (
+                "The number of streamed examples to download before shuffling them. The large the buffer, "
+                "the closer it is to real offline shuffling."
+            )
+        },
+    )
+    stopping_strategy: Optional[str] = field(
+        default="all_exhausted",
+        metadata={
+             "help": "Strategy used to consume interleaved data. Default = 'all_exhausted'"
+        }
+    )
+    streaming: bool = field(
+        default=True,
+        metadata={"help": "Whether to use streaming mode to load and pre-process the data."},
+    )
+
+@dataclass
+class DataCollatorSpeechSeq2SeqWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor ([`WhisperProcessor`])
+            The processor used for processing the data.
+        decoder_start_token_id (`int`)
+            The begin-of-sentence of the decoder.
+    """
+
+    processor: Any
+    decoder_start_token_id: int
+
+    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
+        model_input_name = self.processor.model_input_names[0]
+        input_features = [{model_input_name: feature[model_input_name]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.feature_extractor.pad(input_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)
+
+        # if bos token is appended in previous tokenization step,
+        # cut bos token here as it's append later anyways
+        if (labels[:, 0] == self.decoder_start_token_id).all().cpu().item():
+            labels = labels[:, 1:]
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def load_streaming_dataset(dataset_name, dataset_config_name, text_column_name, split="train", **kwargs):
+    """
+    Utility function to load a dataset in streaming mode. For datasets with multiple splits,
+    each split is loaded individually and then splits combined by taking alternating examples from
+    each (interleaving).
+    """
+    if "+" in split:
+        # load multiple splits separated by the `+` symbol with streaming mode
+        dataset_splits = [
+            load_dataset(dataset_name, dataset_config_name, split=split_name, streaming=True, **kwargs)
+            for split_name in split.split("+")
+        ]
+        # interleave multiple splits to form one dataset
+        interleaved_dataset = interleave_datasets(dataset_splits)
+        return interleaved_dataset
+    else:
+        # load a single split *with* streaming mode
+        dataset = load_dataset(dataset_name, dataset_config_name, split=split, streaming=True, **kwargs)
+        dataset = dataset.rename_column(text_column_name, TEXT_COL_NAME)
+        return dataset
+
+def load_multiple_streaming_datasets(
+    dataset_names: List,
+    dataset_config_names: List,
+    splits: Optional[List] = None,
+    text_column_names: Optional[List] = None,
+    audio_column_names: Optional[List] = None,
+    sampling_rate: Optional[int] = 16000,
+    stopping_strategy: Optional[str] = "all_exhausted",
+    streaming = True,
+    **kwargs
+    ):
+
+    if len(dataset_names) != len(dataset_config_names):
+        raise ValueError(
+            f"Ensure one config is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(dataset_config_names)} configs."
+        )
+
+    if splits is not None and len(splits) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one train_split is passed for each dataset, got {len(dataset_names)} datasets and {len(splits)} splits."
+        )
+
+    if text_column_names is not None and len(text_column_names) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one text column name is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(text_column_names)} text column names."
+        )
+
+    if audio_column_names is not None and len(audio_column_names) != len(dataset_names):
+        raise ValueError(
+            f"Ensure one text column name is passed for each dataset, got {len(dataset_names)} datasets and"
+            f" {len(audio_column_names)} text column names."
+        )
+
+    splits = splits if splits is not None \
+        else ["train" for i in range(len(dataset_names))]
+
+    text_column_names = (
+        text_column_names if text_column_names is not None \
+            else [TEXT_COL_NAME for i in range(len(dataset_names))]
+    )
+
+    audio_column_names = (
+        audio_column_names if audio_column_names is not None \
+            else [AUDIO_COL_NAME for i in range(len(dataset_names))]
+    )
+
+    all_data_splits = []
+    # iterate over the datasets we want to interleave
+    for dset, cfgNm, splt,  txtColNm, audColNm in zip(dataset_names,dataset_config_names,\
+                                                splits,text_column_names, audio_column_names):
+
+        dset_splits = [load_dataset(dset, cfgNm, split=c, streaming=streaming, **kwargs) \
+            for c in splt.split('+') if c != '-']
+
+        if streaming:
+            dset_splits = [ds if TEXT_COL_NAME  in ds.features else ds.rename_column(txtColNm, TEXT_COL_NAME) \
+                for ds in dset_splits ]
+            dset_splits = [ds if AUDIO_COL_NAME in ds.features else ds.rename_column(audColNm, AUDIO_COL_NAME) \
+                for ds in dset_splits]
+
+            if len(dset_splits)>0 and sampling_rate != next(iter(dset_splits[0]))[AUDIO_COL_NAME]['sampling_rate']:
+                dset_splits = [ds.cast_column(AUDIO_COL_NAME, Audio(sampling_rate)) for ds in dset_splits]
+        else:
+
+            dset_splits = [ds if TEXT_COL_NAME  in ds.column_names else ds.rename_column(txtColNm, TEXT_COL_NAME) \
+                for ds in dset_splits ]
+            dset_splits = [ds if AUDIO_COL_NAME in ds.column_names else ds.rename_column(audColNm, AUDIO_COL_NAME) \
+                for ds in dset_splits]
+        
+        if len(dset_splits)>0 and sampling_rate != next(iter(dset_splits[0]))[AUDIO_COL_NAME]['sampling_rate']:
+            dset_splits = [ds.cast_column(AUDIO_COL_NAME, Audio(sampling_rate)) for ds in dset_splits]
+
+        cols2keep = set([AUDIO_COL_NAME, TEXT_COL_NAME])
+
+        dset_splits = [ds.remove_columns(set(ds.features.keys()) - cols2keep) for ds in dset_splits]
+
+        all_data_splits +=   dset_splits
+
+    return interleave_datasets(all_data_splits, stopping_strategy=stopping_strategy)
+
+def main():
+    # 1. Parse input arguments
+    # 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, Seq2SeqTrainingArguments))
+
+    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()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your Python/PyTorch versions.
+    send_example_telemetry("run_speech_recognition_seq2seq_streaming", model_args, data_args)
+
+    # 2. 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)],
+    )
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    datasets.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    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}"
+    )
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # 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)
+
+    # 3. Detecting last checkpoint and eventually continue from 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 and training_args.resume_from_checkpoint is 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."
+            )
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 5. Load pretrained model, tokenizer, and feature extractor
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    config.update({ "forced_decoder_ids": model_args.forced_decoder_ids, 
+                    "suppress_tokens": model_args.suppress_tokens})
+
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    model = AutoModelForSpeechSeq2Seq.from_pretrained(
+        model_args.model_name_or_path,
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    model.config.use_cache = model_args.use_cache
+    model.config.dropout = model_args.dropout
+    model.config.attention_dropout = model_args.attention_dropout
+    if training_args.gradient_checkpointing:
+        model.gradient_checkpointing_enable()
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    # deprecated
+    #if model_args.freeze_feature_encoder:
+    #    model.freeze_feature_encoder()
+
+    if model_args.freeze_encoder:
+        model.freeze_encoder()
+        model.model.encoder.gradient_checkpointing = False
+
+    if data_args.language is not None:
+        # We only need to set the task id when the language is specified (i.e. in a multilingual setting)
+        tokenizer.set_prefix_tokens(language=data_args.language, task=data_args.task)
+
+
+    # 4. Load dataset
+    raw_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_multiple_streaming_datasets(
+            dataset_names=data_args.dataset_name.split(","), 
+            dataset_config_names=data_args.dataset_config_name.split(","),
+            splits = data_args.train_split_name.split(","),
+            text_column_names = data_args.text_column_name.split(","),
+            sampling_rate  = feature_extractor.sampling_rate, 
+            streaming=data_args.streaming,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_streaming_dataset(
+            data_args.evalset_name,
+            data_args.evalset_config_name,
+            text_column_name = data_args.eval_text_column_name,
+            split=data_args.eval_split_name,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    raw_datasets_features = list(next(iter(raw_datasets.values())).features.keys())
+
+    if AUDIO_COL_NAME not in raw_datasets_features:
+        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_features)}."
+        )
+
+    if TEXT_COL_NAME not in raw_datasets_features:
+        raise ValueError(
+            f"--text_column_name {TEXT_COL_NAME} not found in dataset. "
+            "Make sure to set `--text_column_name` to the the respective correct text columns."
+        )
+
+    # 6. Resample eval common voice dataset as it has sampling_rate 48000
+    raw_datasets['eval'] = raw_datasets['eval'].cast_column("audio", datasets.features.Audio(sampling_rate=16000))
+
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    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 = AUDIO_COL_NAME
+    text_column_name = TEXT_COL_NAME
+    model_input_name = feature_extractor.model_input_names[0]
+    do_lower_case = data_args.do_lower_case
+    do_remove_punctuation = data_args.do_remove_punctuation
+    normalizer = BasicTextNormalizer()  # 'official' text normalizer from OpenAI
+
+    if data_args.max_train_samples is not None:
+        raw_datasets["train"] = (
+            raw_datasets["train"].take(data_args.max_train_samples)
+            if data_args.streaming
+            else raw_datasets["train"].select(range(data_args.max_train_samples))
+        )
+
+    if data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = (
+            raw_datasets["eval"].take(data_args.max_eval_samples)
+            if data_args.streaming
+            else raw_datasets["eval"].select(range(data_args.max_eval_samples))
+        )
+
+    def prepare_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.get(model_input_name)[0]
+        batch["input_length"] = len(sample["array"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+        if do_remove_punctuation:
+            input_str = normalizer(input_str).strip()
+        batch["labels"] = tokenizer(input_str).input_ids
+
+        # compute labels length **with** special tokens! -> total label length
+        batch["labels_length"] = len(batch["labels"])
+
+        return batch
+
+
+    ts_tokenizer = WhisperTSTokenizer("openai/whisper-medium", language="sv")
+
+    def prepare_train_dataset(batch):
+        # load and (possibly) resample audio data to 16kHz
+        audio = batch["audio"]
+
+        # compute log-Mel input features from input audio array 
+        batch["input_features"] = processor.feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
+        # compute input length of audio sample in seconds
+        batch["input_length"] = len(audio["array"]) / audio["sampling_rate"]
+
+        batch["labels"] = ts_tokenizer(batch["text"]).input_ids
+
+        # compute labels length **with** special tokens! -> total label length
+        batch["labels_length"] = len(batch["labels"])
+
+        return batch
+
+
+    def prepare_default_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.get(model_input_name)[0]
+        batch["input_length"] = len(sample["array"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+        if do_remove_punctuation:
+            input_str = normalizer(input_str).strip()
+        batch["labels"] = tokenizer(input_str).input_ids
+
+        # compute labels length **with** special tokens! -> total label length
+        batch["labels_length"] = len(batch["labels"])
+
+        return batch
+
+
+    with training_args.main_process_first(desc="dataset map pre-processing"):
+        eval_dataset = raw_datasets['eval'].map(
+            prepare_dataset,
+            remove_columns=list(raw_datasets['eval'].features),
+        ).with_format("torch")
+
+        if training_args.do_train and data_args.streaming:
+            train_dataset = raw_datasets['train'].map(
+                prepare_default_dataset,
+                remove_columns=list(raw_datasets['train'].features),
+            ).with_format("torch")
+
+
+            train_dataset = train_dataset.shuffle(
+                buffer_size=data_args.shuffle_buffer_size,
+                seed=training_args.seed,
+            )
+
+    # filter training data that is shorter than min_input_length or longer than
+    # max_input_length
+    def is_audio_in_length_range(length):
+        return min_input_length < length < max_input_length
+
+    max_label_length = model.config.max_length
+
+    def filter_labels(labels_length):
+        """Filter label sequences longer than max length (448)"""
+        return labels_length < max_label_length
+
+    if training_args.do_train:
+        train_dataset = train_dataset.filter(
+            is_audio_in_length_range,
+            input_columns=["input_length"],
+        )
+        train_dataset = train_dataset.filter(
+            filter_labels,
+            input_columns=["labels_length"],
+        )
+
+    # 8. Load Metric
+    metric = evaluate.load("wer")
+    do_normalize_eval = data_args.do_normalize_eval
+
+    def compute_metrics(pred):
+        pred_ids = pred.predictions
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
+
+        if do_normalize_eval:
+            pred_str = [normalizer(pred) for pred in pred_str]
+            label_str = [normalizer(label) for label in label_str]
+            # filtering step to only evaluate the samples that correspond to non-zero references:
+            pred_str = [pred_str[i] for i in range(len(pred_str)) if len(label_str[i]) > 0]
+            label_str = [label_str[i] for i in range(len(label_str)) if len(label_str[i]) > 0]
+
+        wer = 100 * metric.compute(predictions=pred_str, references=label_str)
+
+        return {"wer": wer}
+
+    # 9. Create a single speech processor
+    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)
+
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+
+    # 10. Define data collator
+    data_collator = DataCollatorSpeechSeq2SeqWithPadding(
+        processor=processor,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+    )
+
+    # 11. Configure Trainer
+    # Trainer callback to reinitialise and reshuffle the streamable datasets at the beginning of each epoch
+    # Only required for streaming: Trainer automatically shuffles non-streaming datasets
+    class ShuffleCallback(TrainerCallback):
+        def on_epoch_begin(self, args, state, control, train_dataloader, **kwargs):
+            if isinstance(train_dataloader.dataset, IterableDatasetShard):
+                pass  # set_epoch() is handled by the Trainer
+            elif isinstance(train_dataloader.dataset, IterableDataset):
+                train_dataloader.dataset.set_epoch(train_dataloader.dataset._epoch + 1)
+
+    # Apply Spec Augment
+    def apply_spec_augment(batch):
+        freq_masking = T.FrequencyMasking(freq_mask_param=27)
+        time_masking = T.TimeMasking(time_mask_param=100)
+
+        spec_input = torch.from_numpy(batch[model_input_name])
+        spec_input_augmented = freq_masking(spec_input)
+        spec_input_augmented = time_masking(spec_input_augmented.reshape(1,80,-1))
+        batch[model_input_name] = spec_input_augmented.reshape(80,-1).numpy()
+        return batch
+
+    #train_dataset = train_dataset.map(apply_spec_augment)
+
+
+    # Create prefetched dataset
+    def prefetch_iterator_train():
+        for example in BackgroundGenerator(train_dataset, max_prefetch=1000):
+            yield example
+
+    def prefetch_iterator_eval():
+        for example in BackgroundGenerator(eval_dataset, max_prefetch=1000):
+            yield example
+
+    prefetched_ds_train = datasets.iterable_dataset.IterableDataset.from_generator(prefetch_iterator_train)
+    prefetched_ds_eval = datasets.iterable_dataset.IterableDataset.from_generator(prefetch_iterator_eval)
+
+    # Initialize Trainer
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=prefetched_ds_train if training_args.do_train else None,
+        eval_dataset=prefetched_ds_eval if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics if training_args.predict_with_generate else None,
+        callbacks=[ShuffleCallback()] if data_args.streaming else None,
+    )
+
+    # 12. Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()  # Saves the feature extractor too for easy upload
+
+        metrics = train_result.metrics
+        if data_args.max_train_samples:
+            metrics["train_samples"] = data_args.max_train_samples
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # 13. Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate(
+            metric_key_prefix="eval",
+            max_length=training_args.generation_max_length,
+            num_beams=training_args.generation_num_beams,
+        )
+        if data_args.max_eval_samples:
+            metrics["eval_samples"] = data_args.max_eval_samples
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # 14. Write Training Stats
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "automatic-speech-recognition",
+        "tags": "whisper-event",
+    }
+    if data_args.dataset_name is not None:
+        kwargs["dataset_tags"] = data_args.dataset_name
+        if data_args.dataset_config_name is not None:
+            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
+        else:
+            kwargs["dataset"] = data_args.dataset_name
+        if "common_voice" in data_args.dataset_name:
+            kwargs["language"] = data_args.dataset_config_name[:2]
+        if model_args.model_index_name is not None:
+            kwargs["model_name"] = model_args.model_index_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+class WhisperTSTokenizer():
+
+    def __init__(self, pretrained_model_name_or_path, language="sv", task="transcribe", *init_inputs, **kwargs):
+        self.tokenizer = WhisperTokenizer.from_pretrained(pretrained_model_name_or_path,
+                                                          predict_timestamps=False,
+                                                          language=language,
+                                                          task=task,
+                                                          *init_inputs,
+                                                          **kwargs)
+
+        self.ts_tokenizer = WhisperTokenizer.from_pretrained(pretrained_model_name_or_path,
+                                                             predict_timestamps=True,
+                                                            language=language,
+                                                             task=task,
+                                                             *init_inputs,
+                                                             **kwargs)
+        
+        timestamp_tokens = ["<|{}|>".format(round(number, 2)) for number in np.arange (0.0, 30.02, 0.02)]
+        self.ts_tokenizer.add_tokens(timestamp_tokens)
+
+    def decode(self, input_tokens, output_offsets=True, skip_special_tokens=False):
+        return self.tokenizer.decode(input_tokens,
+                                     output_offsets=output_offsets,
+                                     skip_special_tokens=skip_special_tokens)
+
+    def __call__(self, example_text, return_timestamps = True):
+        
+        if return_timestamps is False:
+            return self.tokenizer(example_text['text'])
+        
+        segments_str = self._segments_to_str(example_text['offsets'])
+        results = self.ts_tokenizer(segments_str)
+        input_ids = []
+        attention_mask = []
+        for input_id, attention in zip(results.input_ids, results.attention_mask):
+            if input_id != 62:
+                input_ids.append(input_id)
+                attention_mask.append(attention)
+
+        results['input_ids'] = input_ids
+        results['attention_mask'] = attention_mask
+        return results
+
+    def _segments_to_str(self, segments):
+        result = ""
+        for segment in segments:
+            if len(result) > 0:  result += " _ "
+            result += self._segment_to_str(segment)
+        result += " "
+        return result
+
+    def _segment_to_str(self, segment):
+        text = segment['text']
+        ts_start = segment['timestamp'][0]
+        ts_end = segment['timestamp'][1]
+        result = "<|{}|>_ {} <|{}|>".format(ts_start, text, ts_end)
+        result = result.replace("  ", " ")
+        return result
+
+
+
+if __name__ == "__main__":
+    main()