Scripts

README.md

@@ -0,0 +1,55 @@
+---
+license: apache-2.0
+tags:
+- whisper-event
+- generated_from_trainer
+datasets:
+- kul-speech-lab/CGN
+metrics:
+- wer
+model-index:
+- name: Whisper Large CGN
+  results:
+  - task:
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: kul-speech-lab/CGN
+      type: kul-speech-lab/CGN
+      config: cgn-dev.py
+      split: test
+    metrics:
+    - name: Wer
+      type: wer
+      value: 9.6159
+---
+
+# Whisper Large CGN
+
+This model is a fine-tuned version of [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) on the kul-speech-lab/CGN dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.23932012915611267
+- Wer: 9.615871912312803
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 1e-05
+- train_batch_size: 32
+- eval_batch_size: 16
+- gradient_accumulation_steps: 2
+- seed: 42
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 500
+- training_steps: 15000
+- mixed_precision_training: Native AMP
+
+### Framework versions
+
+- Transformers 4.26.0.dev0
+- Pytorch 1.13.0
+- Datasets 2.7.1.dev0
+- Tokenizers 0.13.2
+
+Whisper large model finetuned on Flemish part of Corpus Gesproken Nederlands (CGN).

scripts/cgn-dev.py

@@ -0,0 +1,94 @@
+import numpy as np
+import kaldiio
+
+import datasets
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+_DESCRIPTION = "Corpus Gesproken Nederlands"
+
+_FILEPATHS = {
+    "fbank_pitch": "/esat/spchtemp/scratch/jponcele/espnet2/dump/fbank_pitch/dev_s",
+    "raw": "/esat/spchtemp/scratch/jponcele/espnet2/dump/raw/dev_s"
+}
+
+_FEATURES_NAME = {
+    "fbank_pitch": "feats.scp",
+    "raw": "wav.scp"
+}
+
+
+class CGNConfig(datasets.BuilderConfig):
+    def __init__(self, **kwargs):
+        """
+        Args:
+          data_dir: `string`, the path to the folder containing the files in the
+            downloaded .tar
+          citation: `string`, citation for the data set
+          url: `string`, url for information about the data set
+          **kwargs: keyword arguments forwarded to super.
+        """
+        super(CGNConfig, self).__init__(version=datasets.Version("2.6.1", ""), **kwargs)
+
+
+class CGN(datasets.GeneratorBasedBuilder):
+
+    DEFAULT_WRITER_BATCH_SIZE = 256
+    DEFAULT_CONFIG_NAME = "raw"
+    BUILDER_CONFIGS = [
+        CGNConfig(name="raw", description="All Components")
+    ]
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(
+                {
+                    "audio": datasets.Value("string"),
+                    "text": datasets.Value("string"),
+                    "id": datasets.Value("string"),
+                }
+            ),
+            supervised_keys=("text",),
+        )
+
+    def _split_generators(self, _):
+
+        return [
+            datasets.SplitGenerator(
+                name="test",
+                gen_kwargs={}
+            )
+        ]
+
+    def _generate_examples(self):
+
+        data_dirs = [_FILEPATHS[self.config.name]]
+        for data_dir in data_dirs:
+            with open(f"{data_dir}/text", "r") as txtfile:
+                lines = txtfile.readlines()
+            texts = {line.rstrip().split(' ')[0]: ' '.join(line.rstrip().split(' ')[1:]) for line in lines if len(line.rstrip().split(' ')) > 1}
+
+            featfile = f"{data_dir}/{_FEATURES_NAME[self.config.name]}"
+
+            with open(featfile, "r") as txtfile:
+                feats_generator = dict(map(lambda s: s.strip().split(maxsplit=1), txtfile))
+
+            #if featfile.endswith(".scp"):
+            #    feats_generator = kaldiio.load_scp(featfile)
+            #elif featfile.endswith(".npz"):
+            #    feats_generator = np.load(featfile)
+
+            for key, (uttid, transcript) in enumerate(texts.items()):
+                if uttid not in feats_generator:
+                    logger.warning(f"Missing utterance: {uttid}")
+                    continue
+
+                wav = feats_generator[uttid]
+                #if isinstance(feats, tuple):
+                #    sr, feats = feats
+                #feats = np.expand_dims(feats, axis=1)
+
+                yield key, {"audio": wav, "text": transcript, "id": uttid}

scripts/jobfile_eval.job

@@ -0,0 +1,34 @@
+Universe = vanilla
+
+RequestCpus = 8
+RequestMemory = 30G
++RequestWallTime = 30000
+request_GPUs = 0
+
+Requirements = (Has_avx)\
+               &&(machine!="audioslave.esat.kuleuven.be")\
+               &&(machine!="z4-demo.esat.kuleuven.be")\
+               &&(machine!="xyzzy.esat.kuleuven.be")\
+               &&(machine!="z8-demo.esat.kuleuven.be")\
+               &&(machine!="stadius-nc-4.esat.kuleuven.be")\
+               &&(machine!="stadius-nc-5.esat.kuleuven.be")\
+               &&(machine!="aristoteles.esat.kuleuven.be")\
+               &&(machine!="andromeda.esat.kuleuven.be")\
+               &&(machine!="cauchy.esat.kuleuven.be")\
+               &&(machine!="euler.esat.kuleuven.be")\
+               &&(machine!="gauss.esat.kuleuven.be")\
+               &&(machine!="fig.esat.kuleuven.be")\
+               &&(machine!="nasu.esat.kuleuven.be")\
+               &&(machine!="fehu.esat.kuleuven.be")
+
+NiceUser = true
+initialdir = /users/spraak/jponcele/whisper/finetune_event/whisper_runs 
+
+Executable = /esat/spchtemp/scratch/jponcele/anaconda3/envs/whisper/bin/python 
+Arguments = "run_eval_whisper_streaming_local.py --model_size=$(size) --dataset=$(dataset)"
+
+Log          = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).log
+Output       = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).out
+Error        = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).err
+
+Queue

scripts/jobfile_eval_gpu.job

@@ -0,0 +1,21 @@
+Universe = vanilla
+
+RequestCpus = 4
+RequestMemory = 30G
++RequestWallTime = 30000
+request_GPUs = 1
+
+Requirements = (Has_avx)\
+               &&((GPUs_GlobalMemoryMb >= 15000)||(CUDAGlobalMemoryMb >= 15000))
+
+NiceUser = true
+initialdir = /users/spraak/jponcele/whisper/finetune_event/whisper_runs 
+
+Executable = /esat/spchtemp/scratch/jponcele/anaconda3/envs/whisper/bin/python 
+Arguments = "run_eval_whisper_streaming_local.py --model_size=$(size) --dataset=$(dataset) --device=gpu --batch_size=$(bs)"
+
+Log          = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).log
+Output       = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).out
+Error        = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-eval-$(dataset)-$(size).err
+
+Queue

scripts/jobfile_large.job

@@ -0,0 +1,20 @@
+Universe = vanilla
+
+RequestCpus = 8
+RequestMemory = 30G
++RequestWallTime = 43000
+request_GPUs = 1
+
+Requirements = (machine=="aristoteles.esat.kuleuven.be")
+
+NiceUser = false
+initialdir = /users/spraak/jponcele/whisper/finetune_event/whisper_runs 
+
+Executable = run-cgn-$(size).sh
+Arguments = ""
+
+Log          = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-$(size).log
+Output       = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-$(size).out
+Error        = /esat/audioslave/jponcele/whisper/finetuning_event/CGN/condor/condor-$(size).err
+
+Queue

scripts/run-cgn-large.sh

@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+export PYTHONPATH=""
+source /esat/spchtemp/scratch/jponcele/anaconda3/bin/activate whisper
+python --version
+
+#--dataset_config_name="nl" \
+#--language="dutch" \
+
+python run_speech_recognition_seq2seq_streaming.py \
+--model_name_or_path="openai/whisper-large-v2" \
+--dataset_name="kul-speech-lab/CGN" \
+--train_split_name="train+validation" \
+--eval_split_name="test" \
+--language="dutch" \
+--task="transcribe" \
+--model_index_name="Whisper Large (v2) CGN" \
+--max_steps="15000" \
+--output_dir="/esat/audioslave/jponcele/whisper/finetuning_event/CGN/large" \
+--per_device_train_batch_size="32" \
+--per_device_eval_batch_size="16" \
+--gradient_accumulation_steps="2" \
+--logging_steps="100" \
+--learning_rate="1e-5" \
+--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" \
+--text_column_name="transcription" \
+--freeze_feature_encoder="False" \
+--report_to="tensorboard" \
+--metric_for_best_model="wer" \
+--greater_is_better="False" \
+--load_best_model_at_end \
+--gradient_checkpointing \
+--fp16 \
+--do_train \
+--do_eval \
+--predict_with_generate \
+--do_normalize_eval \
+--streaming \
+--use_auth_token
+# --overwrite_output_dir
+#  --push_to_hub

scripts/run_eval_whisper_streaming_local.py

@@ -0,0 +1,151 @@
+import os
+import numpy as np
+import re
+import argparse
+
+os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["CUDA_VISIBLE_DEVICES"].replace("CUDA", "")
+
+from transformers import pipeline
+from transformers.models.whisper.english_normalizer import BasicTextNormalizer
+from datasets import load_dataset, Audio
+
+whisper_norm = BasicTextNormalizer()
+
+def simple_norm(utt):
+    norm_utt = re.sub(r'[^\w\s]', '', utt)  # remove punctualisation
+    norm_utt = " ".join(norm_utt.split())  # remove whitespaces
+    norm_utt = norm_utt.lower()
+    return norm_utt
+
+def data(dataset):
+    for i, item in enumerate(dataset):
+        yield {**item["audio"], "reference": item["text"], "utt_id": item["id"]}
+
+def get_ckpt(path, ckpt_id):
+    if ckpt_id != 0:
+        model = os.path.join(path, "checkpoint-%i" % ckpt)
+    else:
+        dirs = [d for d in os.listdir(path) if d.startswith("checkpoint-")]
+        ckpts = [int(d.split('-')[-1]) for d in dirs]
+        last_ckpt = sorted(ckpts)[-1]
+        model = os.path.join(path, "checkpoint-%s" % last_ckpt)
+    return model
+
+def main(args):
+    batch_size = args.batch_size
+    
+    if args.device == "cpu":
+        device_id = -1
+    elif args.device == "gpu":
+        device_id = 0
+    else:
+        raise NotImplementedError("unknown device %s, should be cpu/gpu" % args.device)
+
+    model_dir = os.path.join(args.expdir, args.model_size)
+    #model = os.path.join(get_ckpt(model_dir, args.checkpoint), 'pytorch_model.bin')
+    #model = get_ckpt(model_dir, args.checkpoint)
+    
+    model = model_dir
+    #model = "openai/whisper-tiny"
+    
+    whisper_asr = pipeline(
+        "automatic-speech-recognition", model=model, device=device_id
+    )
+
+    whisper_asr.model.config.forced_decoder_ids = (
+        whisper_asr.tokenizer.get_decoder_prompt_ids(
+            language=args.language, task="transcribe"
+        )
+    )
+
+    if args.dataset == 'cgn-dev':
+        dataset_path = "./cgn-dev/cgn-dev.py"
+    elif args.dataset == 'subs-annot':
+        dataset_path = "./subs-annot/subs-annot.py"
+    else:
+        raise NotImplementedError('unknown dataset %s' % args.dataset)
+
+    cache_dir = "/esat/audioslave/jponcele/hf_cache"
+    dataset = load_dataset(dataset_path, name="raw", split="test", cache_dir=cache_dir, streaming=True)
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
+
+    utterances = []
+    predictions = []
+    references = []
+
+    # run streamed inference
+    for out in whisper_asr(data(dataset), batch_size=batch_size):
+        predictions.append(out["text"])
+        utterances.append(out["utt_id"][0])
+        references.append(out["reference"][0])
+        #break
+
+    result_dir = os.path.join(args.expdir, "results", args.dataset)
+    os.makedirs(result_dir, exist_ok=True)
+
+    with open(os.path.join(result_dir, "whisper_%s.txt" % args.model_size), "w") as pd:
+        for i, utt in enumerate(utterances):
+            pred = predictions[i]
+            pd.write(utt + ' ' + pred + '\n')
+
+    with open(os.path.join(result_dir, "whisper_%s_normW.txt" % args.model_size), "w") as pd:
+        for i, utt in enumerate(utterances):
+            pred = whisper_norm(predictions[i])
+            pd.write(utt + ' ' + pred + '\n')
+
+    with open(os.path.join(result_dir, "whisper_%s_normS.txt" % args.model_size), "w") as pd:
+        for i, utt in enumerate(utterances):
+            pred = simple_norm(predictions[i])
+            pd.write(utt + ' ' + pred + '\n')
+
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--expdir",
+        type=str,
+        default="/esat/audioslave/jponcele/whisper/finetuning_event/CGN",
+        help="Directory with finetuned models",
+    )
+    parser.add_argument(
+        "--model_size",
+        type=str,
+        default="tiny",
+        help="Model size",
+    )
+    parser.add_argument(
+        "--checkpoint",
+        type=int,
+        default=0,
+        help="Load specific checkpoint. 0 means latest",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="cgn-dev",
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cpu",
+        help="cpu/gpu",
+    )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=16,
+        help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--language",
+        type=str,
+        default="dutch",
+        help="Two letter language code for the transcription language, e.g. use 'en' for English.",
+    )
+
+    args = parser.parse_args()
+    main(args)
+

scripts/run_speech_recognition_seq2seq_streaming.py

@@ -0,0 +1,634 @@
+#!/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.
+"""
+# You can also adapt this script for your own sequence to sequence speech
+# recognition task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+
+os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["CUDA_VISIBLE_DEVICES"].replace("CUDA", "")
+
+import datasets
+import torch
+from datasets import 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.models.whisper.english_normalizer import BasicTextNormalizer
+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
+
+
+# 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")
+
+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": "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."})
+
+
+@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)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    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'"},
+    )
+    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."
+            )
+        },
+    )
+    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_maybe_streaming_dataset(dataset_name, dataset_config_name, split="train", streaming=True, **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=streaming, **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=streaming, **kwargs)
+        return dataset
+
+
+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)
+        logger.info(f"Last CHECKPOINT {last_checkpoint}")
+
+        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)
+
+    # 4. Load dataset
+    raw_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_maybe_streaming_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=True if model_args.use_auth_token else None,
+            streaming=data_args.streaming,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_maybe_streaming_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=True if model_args.use_auth_token else None,
+            streaming=data_args.streaming,
+        )
+
+    raw_datasets_features = list(next(iter(raw_datasets.values())).features.keys())
+
+    if data_args.audio_column_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 data_args.text_column_name not in raw_datasets_features:
+        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_features)}."
+        )
+
+    # 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})
+
+    if training_args.gradient_checkpointing:
+        config.update({"use_cache": False})
+
+    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,
+    )
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    if model_args.freeze_encoder:
+        model.freeze_encoder()
+
+    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)
+
+    # 6. Resample speech dataset if necessary
+    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)
+        )
+
+    # 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 = data_args.audio_column_name
+    text_column_name = data_args.text_column_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
+        return batch
+
+    with training_args.main_process_first(desc="dataset map pre-processing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=raw_datasets_features,
+        ).with_format("torch")
+
+        if training_args.do_train and data_args.streaming:
+            # manually shuffle if streaming (done by the trainer for non-streaming)
+            vectorized_datasets["train"] = vectorized_datasets["train"].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
+
+    if training_args.do_train:
+        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
+            is_audio_in_length_range,
+            input_columns=["input_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)
+
+    # Initialize Trainer
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        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
+
+
+if __name__ == "__main__":
+    main()

scripts/subs-annot.py

@@ -0,0 +1,94 @@
+import numpy as np
+import kaldiio
+
+import datasets
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+_DESCRIPTION = "Annotated Subtitles"
+
+_FILEPATHS = {
+    "fbank_pitch": "/esat/spchtemp/scratch/jponcele/espnet2/dump/fbank_pitch/subs_annot",
+    "raw": "/esat/spchtemp/scratch/jponcele/espnet2/dump/raw/subs_annot"
+}
+
+_FEATURES_NAME = {
+    "fbank_pitch": "feats.scp",
+    "raw": "wav.scp"
+}
+
+
+class CGNConfig(datasets.BuilderConfig):
+    def __init__(self, **kwargs):
+        """
+        Args:
+          data_dir: `string`, the path to the folder containing the files in the
+            downloaded .tar
+          citation: `string`, citation for the data set
+          url: `string`, url for information about the data set
+          **kwargs: keyword arguments forwarded to super.
+        """
+        super(CGNConfig, self).__init__(version=datasets.Version("2.6.1", ""), **kwargs)
+
+
+class CGN(datasets.GeneratorBasedBuilder):
+
+    DEFAULT_WRITER_BATCH_SIZE = 256
+    DEFAULT_CONFIG_NAME = "raw"
+    BUILDER_CONFIGS = [
+        CGNConfig(name="raw", description="All Components")
+    ]
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(
+                {
+                    "audio": datasets.Value("string"),
+                    "text": datasets.Value("string"),
+                    "id": datasets.Value("string"),
+                }
+            ),
+            supervised_keys=("text",),
+        )
+
+    def _split_generators(self, _):
+
+        return [
+            datasets.SplitGenerator(
+                name="test",
+                gen_kwargs={}
+            )
+        ]
+
+    def _generate_examples(self):
+
+        data_dirs = [_FILEPATHS[self.config.name]]
+        for data_dir in data_dirs:
+            with open(f"{data_dir}/text", "r") as txtfile:
+                lines = txtfile.readlines()
+            texts = {line.rstrip().split(' ')[0]: ' '.join(line.rstrip().split(' ')[1:]) for line in lines if len(line.rstrip().split(' ')) > 1}
+
+            featfile = f"{data_dir}/{_FEATURES_NAME[self.config.name]}"
+
+            with open(featfile, "r") as txtfile:
+                feats_generator = dict(map(lambda s: s.strip().split(maxsplit=1), txtfile))
+
+            #if featfile.endswith(".scp"):
+            #    feats_generator = kaldiio.load_scp(featfile)
+            #elif featfile.endswith(".npz"):
+            #    feats_generator = np.load(featfile)
+
+            for key, (uttid, transcript) in enumerate(texts.items()):
+                if uttid not in feats_generator:
+                    logger.warning(f"Missing utterance: {uttid}")
+                    continue
+
+                wav = feats_generator[uttid]
+                #if isinstance(feats, tuple):
+                #    sr, feats = feats
+                #feats = np.expand_dims(feats, axis=1)
+
+                yield key, {"audio": wav, "text": transcript, "id": uttid}