first

.gitattributes

@@ -25,3 +25,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.txt filter=lfs diff=lfs merge=lfs -text
+*unigram*.* filter=lfs diff=lfs merge=lfs -text

NbAiLab_NPSC_16K_mp3_nynorsk_test_eval_results.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ed28ce85f9486af2150732dc9833fd7c292c124e0100e8f9f20b25fdd1bcfd4
+size 49

NbAiLab_NPSC_16K_mp3_nynorsk_test_eval_results_no_lang_model.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ad6da4ab72c542e88f1b751239f053f2d0a3dd03cc51b5c4600b44b967e977b
+size 49

add_kenlm.py

@@ -0,0 +1,34 @@
+import argparse
+from transformers import AutoProcessor
+from transformers import Wav2Vec2ProcessorWithLM
+from pyctcdecode import build_ctcdecoder
+
+
+def main(args):
+    processor = AutoProcessor.from_pretrained(args.model_name_or_path)
+    vocab_dict = processor.tokenizer.get_vocab()
+    sorted_vocab_dict = {
+        k.lower(): v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])
+    }
+    decoder = build_ctcdecoder(
+        labels=list(sorted_vocab_dict.keys()),
+        kenlm_model_path=args.kenlm_model_path,
+    )
+    processor_with_lm = Wav2Vec2ProcessorWithLM(
+        feature_extractor=processor.feature_extractor,
+        tokenizer=processor.tokenizer,
+        decoder=decoder,
+    )
+    processor_with_lm.save_pretrained(args.model_name_or_path)
+    print(f"Run: ~/bin/build_binary language_model/*.arpa language_model/5gram.bin -T $(pwd) && rm language_model/*.arpa")
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model_name_or_path', default="./", help='Model name or path. Defaults to ./')
+    parser.add_argument('--kenlm_model_path', required=True, help='Path to KenLM arpa file.')
+    args = parser.parse_args()
+    return args
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 32, "</s>": 33}

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 40.0,
+    "eval_loss": 0.35698655247688293,
+    "eval_runtime": 52.1875,
+    "eval_samples": 825,
+    "eval_samples_per_second": 15.808,
+    "eval_steps_per_second": 0.996,
+    "eval_wer": 0.21414389713529355,
+    "train_loss": 0.896209309619406,
+    "train_runtime": 25745.3803,
+    "train_samples": 7354,
+    "train_samples_per_second": 11.426,
+    "train_steps_per_second": 0.357
+}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e5", "\u00e6", "\u00f8", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

config.json

@@ -0,0 +1,115 @@
+{
+  "_name_or_path": "KBLab/wav2vec2-large-voxrex",
+  "activation_dropout": 0.055,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.094,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 768,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": false,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.04,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.047,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.041,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "mask_feature_length": 64,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.25,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.082,
+  "mask_time_selection": "static",
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "num_negatives": 100,
+  "output_hidden_size": 1024,
+  "pad_token_id": 31,
+  "proj_codevector_dim": 768,
+  "tdnn_dilation": [
+    1,
+    2,
+    3,
+    1,
+    1
+  ],
+  "tdnn_dim": [
+    512,
+    512,
+    512,
+    512,
+    1500
+  ],
+  "tdnn_kernel": [
+    5,
+    3,
+    3,
+    1,
+    1
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.17.0.dev0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 34,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower()) + " "
+    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('\s', ' ', text)
+    text = re.sub('<ee>', 'eee', text)
+    text = re.sub('<qq>', 'qqq', text)
+    text = re.sub('<mm>', 'mmm', text)
+    text = re.sub('<inaudible>', 'xxx', text)
+    text = re.sub('[<>]', '', text)
+
+    # # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # # note that order is important here!
+    # token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    # for t in token_sequences_to_ignore:
+    #     text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["text"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

eval_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 40.0,
+    "eval_loss": 0.35698655247688293,
+    "eval_runtime": 52.1875,
+    "eval_samples": 825,
+    "eval_samples_per_second": 15.808,
+    "eval_steps_per_second": 0.996,
+    "eval_wer": 0.21414389713529355
+}

language_model/5gram.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab572020f4871ec611ad95a92071040bafe77c4633de2d26166006749e964129
+size 3981756555

language_model/attrs.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}

language_model/unigrams.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48aeb02d80e42421057e95161188779f1cdace619908c0feeab767d1cadb44cb
+size 14892613

log_NbAiLab_NPSC_16K_mp3_nynorsk_test_predictions.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b98d304c7f36c039cd2e791141a6f8822405cf949216ff660d73a9d0f19390d6
+size 67295

log_NbAiLab_NPSC_16K_mp3_nynorsk_test_predictions_no_lang_model.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5dcc454a467aa8b06d24b95425a90188b6426ac8fdb66086bd790b236bac0adc
+size 67397

log_NbAiLab_NPSC_16K_mp3_nynorsk_test_targets.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0755805652d10110cf798b5558d40def38c504be77ea23d85c49b564214e4035
+size 68478

log_NbAiLab_NPSC_16K_mp3_nynorsk_test_targets_no_lang_model.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0755805652d10110cf798b5558d40def38c504be77ea23d85c49b564214e4035
+size 68478

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8d5d272bf5e836fcfa6de6249caafb0e0ba5f080ffcb6054fdb5b983a7b5d21
+size 1262063089

run.sh

@@ -0,0 +1,39 @@
+python run_speech_recognition_ctc.py \
+        --dataset_name="NbAiLab/NPSC" \
+        --model_name_or_path="KBLab/wav2vec2-large-voxrex" \
+        --dataset_config_name="16K_mp3_nynorsk" \
+        --output_dir="./" \
+        --overwrite_output_dir \
+        --num_train_epochs="80" \
+        --per_device_train_batch_size="16" \
+        --per_device_eval_batch_size="16" \
+        --gradient_accumulation_steps="2" \
+        --learning_rate="1e-4" \
+        --warmup_steps="2000" \
+        --length_column_name="input_length" \
+        --evaluation_strategy="steps" \
+        --text_column_name="text" \
+        --save_steps="500" \
+        --eval_steps="500" \
+        --logging_steps="100" \
+        --layerdrop="0.041" \
+        --attention_dropout="0.094" \
+        --activation_dropout="0.055" \
+        --hidden_dropout="0.047" \
+        --save_total_limit="3" \
+        --freeze_feature_encoder \
+        --feat_proj_dropout="0.04" \
+        --mask_time_prob="0.082" \
+        --mask_time_length="10" \
+        --mask_feature_prob="0.25" \
+        --mask_feature_length="64" \
+        --gradient_checkpointing \
+        --min_duration_in_seconds="0.5" \
+        --max_duration_in_seconds="30.0" \
+        --use_auth_token \
+        --seed="42" \
+        --fp16 \
+        --group_by_length \
+        --do_train --do_eval \
+        --push_to_hub \
+        --preprocessing_num_workers="32"

run_speech_recognition_ctc.py

@@ -0,0 +1,793 @@
+#!/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 CTC 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
+
+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
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.16.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    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+validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@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
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+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)
+
+    # Pre-processing dataset
+    import re
+    def filter_numeric(entry):
+        return  (
+            "0" not in entry["text"]
+            and "1" not in entry["text"]
+            and "2" not in entry["text"]
+            and "3" not in entry["text"]
+            and "4" not in entry["text"]
+            and "5" not in entry["text"]
+            and "6" not in entry["text"]
+            and "7" not in entry["text"]
+            and "8" not in entry["text"]
+            and "9" not in entry["text"]
+        )
+
+    #def filter_inaudible(entry):
+    #   return not re.search("\d|<inaudible>", entry["text"], flags=re.IGNORECASE)
+    #
+    #def filter_nynorsk(entry):
+    #    return re.search("nn-no", entry["sentence_language_code"], flags=re.IGNORECASE)
+
+    def filter_tooshort(entry):
+        #print(f"The audio sample ({entry["audio"]["path"]}) is too small, and has been omitted. "
+        return (len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)
+
+    def map_dataset(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
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        ).shuffle()
+        raw_datasets["train"] = raw_datasets["train"].filter(filter_numeric).filter(filter_tooshort)
+        raw_datasets["train"] = raw_datasets["train"].map(map_dataset)
+
+        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:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        ).shuffle()
+        raw_datasets["eval"] = raw_datasets["eval"].filter(filter_numeric).filter(filter_tooshort)
+        raw_datasets["eval"] = raw_datasets["eval"].map(map_dataset)
+
+        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["eval"] 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()

runs/Feb05_15-59-35_dante/1644073207.5513232/events.out.tfevents.1644073207.dante.2625060.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:362da1f3278e08c8d2885e51c17415f55316029cb4db2b1a674844b79d1f8ad0
+size 4756

runs/Feb05_15-59-35_dante/events.out.tfevents.1644073207.dante.2625060.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:609a294ed52ebef007124dbb2231bcbce9e7219085c29d846fac7e37b72d50e4
+size 46177

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer", "processor_class": "Wav2Vec2ProcessorWithLM"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 40.0,
+    "train_loss": 0.896209309619406,
+    "train_runtime": 25745.3803,
+    "train_samples": 7354,
+    "train_samples_per_second": 11.426,
+    "train_steps_per_second": 0.357
+}

trainer_state.json

@@ -0,0 +1,739 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 40.0,
+  "global_step": 9200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.43,
+      "learning_rate": 3.675e-06,
+      "loss": 10.3748,
+      "step": 100
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 7.35e-06,
+      "loss": 9.8141,
+      "step": 200
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 1.1099999999999999e-05,
+      "loss": 5.6549,
+      "step": 300
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 1.485e-05,
+      "loss": 3.4055,
+      "step": 400
+    },
+    {
+      "epoch": 2.17,
+      "learning_rate": 1.8599999999999998e-05,
+      "loss": 3.086,
+      "step": 500
+    },
+    {
+      "epoch": 2.17,
+      "eval_loss": 3.0773141384124756,
+      "eval_runtime": 54.4742,
+      "eval_samples_per_second": 15.145,
+      "eval_steps_per_second": 0.955,
+      "eval_wer": 1.0,
+      "step": 500
+    },
+    {
+      "epoch": 2.61,
+      "learning_rate": 2.2349999999999998e-05,
+      "loss": 2.9844,
+      "step": 600
+    },
+    {
+      "epoch": 3.04,
+      "learning_rate": 2.6099999999999997e-05,
+      "loss": 2.957,
+      "step": 700
+    },
+    {
+      "epoch": 3.48,
+      "learning_rate": 2.985e-05,
+      "loss": 2.912,
+      "step": 800
+    },
+    {
+      "epoch": 3.91,
+      "learning_rate": 3.36e-05,
+      "loss": 2.8929,
+      "step": 900
+    },
+    {
+      "epoch": 4.35,
+      "learning_rate": 3.735e-05,
+      "loss": 2.8532,
+      "step": 1000
+    },
+    {
+      "epoch": 4.35,
+      "eval_loss": 2.839279890060425,
+      "eval_runtime": 54.2914,
+      "eval_samples_per_second": 15.196,
+      "eval_steps_per_second": 0.958,
+      "eval_wer": 1.0,
+      "step": 1000
+    },
+    {
+      "epoch": 4.78,
+      "learning_rate": 4.11e-05,
+      "loss": 2.595,
+      "step": 1100
+    },
+    {
+      "epoch": 5.22,
+      "learning_rate": 4.484999999999999e-05,
+      "loss": 1.9229,
+      "step": 1200
+    },
+    {
+      "epoch": 5.65,
+      "learning_rate": 4.8599999999999995e-05,
+      "loss": 1.3521,
+      "step": 1300
+    },
+    {
+      "epoch": 6.09,
+      "learning_rate": 5.234999999999999e-05,
+      "loss": 1.0772,
+      "step": 1400
+    },
+    {
+      "epoch": 6.52,
+      "learning_rate": 5.6099999999999995e-05,
+      "loss": 0.9738,
+      "step": 1500
+    },
+    {
+      "epoch": 6.52,
+      "eval_loss": 0.728294849395752,
+      "eval_runtime": 54.3803,
+      "eval_samples_per_second": 15.171,
+      "eval_steps_per_second": 0.956,
+      "eval_wer": 0.4889855346506571,
+      "step": 1500
+    },
+    {
+      "epoch": 6.96,
+      "learning_rate": 5.985e-05,
+      "loss": 0.891,
+      "step": 1600
+    },
+    {
+      "epoch": 7.39,
+      "learning_rate": 6.359999999999999e-05,
+      "loss": 0.7848,
+      "step": 1700
+    },
+    {
+      "epoch": 7.83,
+      "learning_rate": 6.735e-05,
+      "loss": 0.7527,
+      "step": 1800
+    },
+    {
+      "epoch": 8.26,
+      "learning_rate": 7.11e-05,
+      "loss": 0.7279,
+      "step": 1900
+    },
+    {
+      "epoch": 8.7,
+      "learning_rate": 7.484999999999999e-05,
+      "loss": 0.6763,
+      "step": 2000
+    },
+    {
+      "epoch": 8.7,
+      "eval_loss": 0.5340386033058167,
+      "eval_runtime": 53.7943,
+      "eval_samples_per_second": 15.336,
+      "eval_steps_per_second": 0.967,
+      "eval_wer": 0.36617188238630993,
+      "step": 2000
+    },
+    {
+      "epoch": 9.13,
+      "learning_rate": 7.4e-05,
+      "loss": 0.6165,
+      "step": 2100
+    },
+    {
+      "epoch": 9.57,
+      "learning_rate": 7.295833333333332e-05,
+      "loss": 0.6021,
+      "step": 2200
+    },
+    {
+      "epoch": 10.0,
+      "learning_rate": 7.191666666666666e-05,
+      "loss": 0.5904,
+      "step": 2300
+    },
+    {
+      "epoch": 10.43,
+      "learning_rate": 7.087499999999999e-05,
+      "loss": 0.5469,
+      "step": 2400
+    },
+    {
+      "epoch": 10.87,
+      "learning_rate": 6.983333333333333e-05,
+      "loss": 0.5303,
+      "step": 2500
+    },
+    {
+      "epoch": 10.87,
+      "eval_loss": 0.45213818550109863,
+      "eval_runtime": 53.7095,
+      "eval_samples_per_second": 15.36,
+      "eval_steps_per_second": 0.968,
+      "eval_wer": 0.31398317103148343,
+      "step": 2500
+    },
+    {
+      "epoch": 11.3,
+      "learning_rate": 6.879166666666667e-05,
+      "loss": 0.5061,
+      "step": 2600
+    },
+    {
+      "epoch": 11.74,
+      "learning_rate": 6.775e-05,
+      "loss": 0.4928,
+      "step": 2700
+    },
+    {
+      "epoch": 12.17,
+      "learning_rate": 6.670833333333333e-05,
+      "loss": 0.4693,
+      "step": 2800
+    },
+    {
+      "epoch": 12.61,
+      "learning_rate": 6.566666666666666e-05,
+      "loss": 0.4733,
+      "step": 2900
+    },
+    {
+      "epoch": 13.04,
+      "learning_rate": 6.4625e-05,
+      "loss": 0.4765,
+      "step": 3000
+    },
+    {
+      "epoch": 13.04,
+      "eval_loss": 0.4181167483329773,
+      "eval_runtime": 53.0777,
+      "eval_samples_per_second": 15.543,
+      "eval_steps_per_second": 0.98,
+      "eval_wer": 0.2853361066464971,
+      "step": 3000
+    },
+    {
+      "epoch": 13.48,
+      "learning_rate": 6.358333333333332e-05,
+      "loss": 0.4396,
+      "step": 3100
+    },
+    {
+      "epoch": 13.91,
+      "learning_rate": 6.254166666666666e-05,
+      "loss": 0.4359,
+      "step": 3200
+    },
+    {
+      "epoch": 14.35,
+      "learning_rate": 6.149999999999999e-05,
+      "loss": 0.4173,
+      "step": 3300
+    },
+    {
+      "epoch": 14.78,
+      "learning_rate": 6.045833333333333e-05,
+      "loss": 0.408,
+      "step": 3400
+    },
+    {
+      "epoch": 15.22,
+      "learning_rate": 5.941666666666666e-05,
+      "loss": 0.4219,
+      "step": 3500
+    },
+    {
+      "epoch": 15.22,
+      "eval_loss": 0.41555172204971313,
+      "eval_runtime": 52.8273,
+      "eval_samples_per_second": 15.617,
+      "eval_steps_per_second": 0.984,
+      "eval_wer": 0.2933724118370048,
+      "step": 3500
+    },
+    {
+      "epoch": 15.65,
+      "learning_rate": 5.8374999999999996e-05,
+      "loss": 0.4022,
+      "step": 3600
+    },
+    {
+      "epoch": 16.09,
+      "learning_rate": 5.733333333333333e-05,
+      "loss": 0.3932,
+      "step": 3700
+    },
+    {
+      "epoch": 16.52,
+      "learning_rate": 5.629166666666666e-05,
+      "loss": 0.374,
+      "step": 3800
+    },
+    {
+      "epoch": 16.96,
+      "learning_rate": 5.5249999999999994e-05,
+      "loss": 0.3697,
+      "step": 3900
+    },
+    {
+      "epoch": 17.39,
+      "learning_rate": 5.420833333333333e-05,
+      "loss": 0.3564,
+      "step": 4000
+    },
+    {
+      "epoch": 17.39,
+      "eval_loss": 0.3925444483757019,
+      "eval_runtime": 54.1337,
+      "eval_samples_per_second": 15.24,
+      "eval_steps_per_second": 0.961,
+      "eval_wer": 0.2509218114777347,
+      "step": 4000
+    },
+    {
+      "epoch": 17.83,
+      "learning_rate": 5.316666666666666e-05,
+      "loss": 0.3784,
+      "step": 4100
+    },
+    {
+      "epoch": 18.26,
+      "learning_rate": 5.212499999999999e-05,
+      "loss": 0.3805,
+      "step": 4200
+    },
+    {
+      "epoch": 18.7,
+      "learning_rate": 5.1083333333333326e-05,
+      "loss": 0.3419,
+      "step": 4300
+    },
+    {
+      "epoch": 19.13,
+      "learning_rate": 5.0041666666666666e-05,
+      "loss": 0.3344,
+      "step": 4400
+    },
+    {
+      "epoch": 19.57,
+      "learning_rate": 4.899999999999999e-05,
+      "loss": 0.3282,
+      "step": 4500
+    },
+    {
+      "epoch": 19.57,
+      "eval_loss": 0.3824474811553955,
+      "eval_runtime": 54.4571,
+      "eval_samples_per_second": 15.15,
+      "eval_steps_per_second": 0.955,
+      "eval_wer": 0.24203460338470265,
+      "step": 4500
+    },
+    {
+      "epoch": 20.0,
+      "learning_rate": 4.795833333333333e-05,
+      "loss": 0.3447,
+      "step": 4600
+    },
+    {
+      "epoch": 20.43,
+      "learning_rate": 4.691666666666666e-05,
+      "loss": 0.3308,
+      "step": 4700
+    },
+    {
+      "epoch": 20.87,
+      "learning_rate": 4.5875e-05,
+      "loss": 0.3056,
+      "step": 4800
+    },
+    {
+      "epoch": 21.3,
+      "learning_rate": 4.483333333333333e-05,
+      "loss": 0.3069,
+      "step": 4900
+    },
+    {
+      "epoch": 21.74,
+      "learning_rate": 4.379166666666666e-05,
+      "loss": 0.3118,
+      "step": 5000
+    },
+    {
+      "epoch": 21.74,
+      "eval_loss": 0.36364665627479553,
+      "eval_runtime": 52.4627,
+      "eval_samples_per_second": 15.725,
+      "eval_steps_per_second": 0.991,
+      "eval_wer": 0.2354164696984022,
+      "step": 5000
+    },
+    {
+      "epoch": 22.17,
+      "learning_rate": 4.2749999999999996e-05,
+      "loss": 0.3108,
+      "step": 5100
+    },
+    {
+      "epoch": 22.61,
+      "learning_rate": 4.170833333333333e-05,
+      "loss": 0.3034,
+      "step": 5200
+    },
+    {
+      "epoch": 23.04,
+      "learning_rate": 4.066666666666667e-05,
+      "loss": 0.3132,
+      "step": 5300
+    },
+    {
+      "epoch": 23.48,
+      "learning_rate": 3.9624999999999994e-05,
+      "loss": 0.3024,
+      "step": 5400
+    },
+    {
+      "epoch": 23.91,
+      "learning_rate": 3.8583333333333334e-05,
+      "loss": 0.2919,
+      "step": 5500
+    },
+    {
+      "epoch": 23.91,
+      "eval_loss": 0.3614845275878906,
+      "eval_runtime": 52.502,
+      "eval_samples_per_second": 15.714,
+      "eval_steps_per_second": 0.99,
+      "eval_wer": 0.22813652264347167,
+      "step": 5500
+    },
+    {
+      "epoch": 24.35,
+      "learning_rate": 3.754166666666666e-05,
+      "loss": 0.2868,
+      "step": 5600
+    },
+    {
+      "epoch": 24.78,
+      "learning_rate": 3.65e-05,
+      "loss": 0.2929,
+      "step": 5700
+    },
+    {
+      "epoch": 25.22,
+      "learning_rate": 3.545833333333333e-05,
+      "loss": 0.2956,
+      "step": 5800
+    },
+    {
+      "epoch": 25.65,
+      "learning_rate": 3.4416666666666665e-05,
+      "loss": 0.2818,
+      "step": 5900
+    },
+    {
+      "epoch": 26.09,
+      "learning_rate": 3.3375e-05,
+      "loss": 0.2961,
+      "step": 6000
+    },
+    {
+      "epoch": 26.09,
+      "eval_loss": 0.35476088523864746,
+      "eval_runtime": 52.7192,
+      "eval_samples_per_second": 15.649,
+      "eval_steps_per_second": 0.986,
+      "eval_wer": 0.2254892691689515,
+      "step": 6000
+    },
+    {
+      "epoch": 26.52,
+      "learning_rate": 3.233333333333333e-05,
+      "loss": 0.2816,
+      "step": 6100
+    },
+    {
+      "epoch": 26.96,
+      "learning_rate": 3.1291666666666664e-05,
+      "loss": 0.2788,
+      "step": 6200
+    },
+    {
+      "epoch": 27.39,
+      "learning_rate": 3.0249999999999997e-05,
+      "loss": 0.272,
+      "step": 6300
+    },
+    {
+      "epoch": 27.83,
+      "learning_rate": 2.920833333333333e-05,
+      "loss": 0.2756,
+      "step": 6400
+    },
+    {
+      "epoch": 28.26,
+      "learning_rate": 2.8166666666666662e-05,
+      "loss": 0.284,
+      "step": 6500
+    },
+    {
+      "epoch": 28.26,
+      "eval_loss": 0.3526020646095276,
+      "eval_runtime": 52.8188,
+      "eval_samples_per_second": 15.619,
+      "eval_steps_per_second": 0.984,
+      "eval_wer": 0.2208565755885412,
+      "step": 6500
+    },
+    {
+      "epoch": 28.7,
+      "learning_rate": 2.7125e-05,
+      "loss": 0.2721,
+      "step": 6600
+    },
+    {
+      "epoch": 29.13,
+      "learning_rate": 2.608333333333333e-05,
+      "loss": 0.2601,
+      "step": 6700
+    },
+    {
+      "epoch": 29.57,
+      "learning_rate": 2.5041666666666664e-05,
+      "loss": 0.2462,
+      "step": 6800
+    },
+    {
+      "epoch": 30.0,
+      "learning_rate": 2.3999999999999997e-05,
+      "loss": 0.271,
+      "step": 6900
+    },
+    {
+      "epoch": 30.43,
+      "learning_rate": 2.295833333333333e-05,
+      "loss": 0.2566,
+      "step": 7000
+    },
+    {
+      "epoch": 30.43,
+      "eval_loss": 0.35256850719451904,
+      "eval_runtime": 52.6434,
+      "eval_samples_per_second": 15.671,
+      "eval_steps_per_second": 0.988,
+      "eval_wer": 0.22047839652075257,
+      "step": 7000
+    },
+    {
+      "epoch": 30.87,
+      "learning_rate": 2.1916666666666663e-05,
+      "loss": 0.2612,
+      "step": 7100
+    },
+    {
+      "epoch": 31.3,
+      "learning_rate": 2.0874999999999996e-05,
+      "loss": 0.2524,
+      "step": 7200
+    },
+    {
+      "epoch": 31.74,
+      "learning_rate": 1.983333333333333e-05,
+      "loss": 0.254,
+      "step": 7300
+    },
+    {
+      "epoch": 32.17,
+      "learning_rate": 1.879166666666666e-05,
+      "loss": 0.2446,
+      "step": 7400
+    },
+    {
+      "epoch": 32.61,
+      "learning_rate": 1.7749999999999998e-05,
+      "loss": 0.2422,
+      "step": 7500
+    },
+    {
+      "epoch": 32.61,
+      "eval_loss": 0.35690072178840637,
+      "eval_runtime": 53.8708,
+      "eval_samples_per_second": 15.314,
+      "eval_steps_per_second": 0.965,
+      "eval_wer": 0.2172638744445495,
+      "step": 7500
+    },
+    {
+      "epoch": 33.04,
+      "learning_rate": 1.670833333333333e-05,
+      "loss": 0.2549,
+      "step": 7600
+    },
+    {
+      "epoch": 33.48,
+      "learning_rate": 1.5666666666666667e-05,
+      "loss": 0.2547,
+      "step": 7700
+    },
+    {
+      "epoch": 33.91,
+      "learning_rate": 1.4625e-05,
+      "loss": 0.2553,
+      "step": 7800
+    },
+    {
+      "epoch": 34.35,
+      "learning_rate": 1.3583333333333333e-05,
+      "loss": 0.2445,
+      "step": 7900
+    },
+    {
+      "epoch": 34.78,
+      "learning_rate": 1.2541666666666665e-05,
+      "loss": 0.2472,
+      "step": 8000
+    },
+    {
+      "epoch": 34.78,
+      "eval_loss": 0.35918259620666504,
+      "eval_runtime": 52.3221,
+      "eval_samples_per_second": 15.768,
+      "eval_steps_per_second": 0.994,
+      "eval_wer": 0.21660206107591945,
+      "step": 8000
+    },
+    {
+      "epoch": 35.22,
+      "learning_rate": 1.1499999999999998e-05,
+      "loss": 0.2608,
+      "step": 8100
+    },
+    {
+      "epoch": 35.65,
+      "learning_rate": 1.0458333333333333e-05,
+      "loss": 0.2482,
+      "step": 8200
+    },
+    {
+      "epoch": 36.09,
+      "learning_rate": 9.416666666666666e-06,
+      "loss": 0.2464,
+      "step": 8300
+    },
+    {
+      "epoch": 36.52,
+      "learning_rate": 8.374999999999999e-06,
+      "loss": 0.2347,
+      "step": 8400
+    },
+    {
+      "epoch": 36.96,
+      "learning_rate": 7.333333333333333e-06,
+      "loss": 0.2337,
+      "step": 8500
+    },
+    {
+      "epoch": 36.96,
+      "eval_loss": 0.36245495080947876,
+      "eval_runtime": 54.1316,
+      "eval_samples_per_second": 15.241,
+      "eval_steps_per_second": 0.961,
+      "eval_wer": 0.21716932967760236,
+      "step": 8500
+    },
+    {
+      "epoch": 37.39,
+      "learning_rate": 6.291666666666666e-06,
+      "loss": 0.2392,
+      "step": 8600
+    },
+    {
+      "epoch": 37.83,
+      "learning_rate": 5.25e-06,
+      "loss": 0.2347,
+      "step": 8700
+    },
+    {
+      "epoch": 38.26,
+      "learning_rate": 4.208333333333333e-06,
+      "loss": 0.2455,
+      "step": 8800
+    },
+    {
+      "epoch": 38.7,
+      "learning_rate": 3.1666666666666663e-06,
+      "loss": 0.243,
+      "step": 8900
+    },
+    {
+      "epoch": 39.13,
+      "learning_rate": 2.1249999999999996e-06,
+      "loss": 0.2315,
+      "step": 9000
+    },
+    {
+      "epoch": 39.13,
+      "eval_loss": 0.3579612374305725,
+      "eval_runtime": 52.3246,
+      "eval_samples_per_second": 15.767,
+      "eval_steps_per_second": 0.994,
+      "eval_wer": 0.21546752387255366,
+      "step": 9000
+    },
+    {
+      "epoch": 39.57,
+      "learning_rate": 1.0833333333333333e-06,
+      "loss": 0.2489,
+      "step": 9100
+    },
+    {
+      "epoch": 40.0,
+      "learning_rate": 4.166666666666666e-08,
+      "loss": 0.227,
+      "step": 9200
+    },
+    {
+      "epoch": 40.0,
+      "step": 9200,
+      "total_flos": 6.4078116974997864e+19,
+      "train_loss": 0.896209309619406,
+      "train_runtime": 25745.3803,
+      "train_samples_per_second": 11.426,
+      "train_steps_per_second": 0.357
+    }
+  ],
+  "max_steps": 9200,
+  "num_train_epochs": 40,
+  "total_flos": 6.4078116974997864e+19,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c45605b42b58fbf08f0a50a498b37f42abf5163ce81ddbbf3d684bb8da4e8a6
+size 3055

vocab.json

@@ -0,0 +1 @@
+{"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "å": 27, "æ": 28, "ø": 29, "|": 0, "[UNK]": 30, "[PAD]": 31}