init

README.md

@@ -0,0 +1,105 @@
+---
+language:
+- he
+tags:
+- automatic-speech-recognition
+- robust-speech-event
+- he
+- generated_from_trainer
+model-index:
+- name: wav2vec2-xls-r-300m-hebrew
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# wav2vec2-xls-r-300m-hebrew
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the private dataset with stats:
+
+| split  |size | n_samples | duration(hrs)|   |
+|---|---|---|---|---|
+|train|4.19gb| 20306  | 28  |   |
+|dev  |1.05gb|  5076 |  7 |   |
+
+
+It achieves the following results on the evaluation set:
+- Loss: 0.5438
+- Wer: 0.1773
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0003
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 64
+- total_eval_batch_size: 16
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 1000
+- num_epochs: 100.0
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|
+| No log        | 3.15  | 1000  | 0.5203          | 0.4333 |
+| 1.4284        | 6.31  | 2000  | 0.4816          | 0.3951 |
+| 1.4284        | 9.46  | 3000  | 0.4315          | 0.3546 |
+| 1.283         | 12.62 | 4000  | 0.4278          | 0.3404 |
+| 1.283         | 15.77 | 5000  | 0.4090          | 0.3054 |
+| 1.1777        | 18.93 | 6000  | 0.3893          | 0.3006 |
+| 1.1777        | 22.08 | 7000  | 0.3968          | 0.2857 |
+| 1.0994        | 25.24 | 8000  | 0.3892          | 0.2751 |
+| 1.0994        | 28.39 | 9000  | 0.4061          | 0.2690 |
+| 1.0323        | 31.54 | 10000 | 0.4114          | 0.2507 |
+| 1.0323        | 34.7  | 11000 | 0.4021          | 0.2508 |
+| 0.9623        | 37.85 | 12000 | 0.4032          | 0.2378 |
+| 0.9623        | 41.01 | 13000 | 0.4148          | 0.2374 |
+| 0.9077        | 44.16 | 14000 | 0.4350          | 0.2323 |
+| 0.9077        | 47.32 | 15000 | 0.4515          | 0.2246 |
+| 0.8573        | 50.47 | 16000 | 0.4474          | 0.2180 |
+| 0.8573        | 53.63 | 17000 | 0.4649          | 0.2171 |
+| 0.8083        | 56.78 | 18000 | 0.4455          | 0.2102 |
+| 0.8083        | 59.94 | 19000 | 0.4587          | 0.2092 |
+| 0.769         | 63.09 | 20000 | 0.4794          | 0.2012 |
+| 0.769         | 66.25 | 21000 | 0.4845          | 0.2007 |
+| 0.7308        | 69.4  | 22000 | 0.4937          | 0.2008 |
+| 0.7308        | 72.55 | 23000 | 0.4920          | 0.1895 |
+| 0.6927        | 75.71 | 24000 | 0.5179          | 0.1911 |
+| 0.6927        | 78.86 | 25000 | 0.5202          | 0.1877 |
+| 0.6622        | 82.02 | 26000 | 0.5266          | 0.1840 |
+| 0.6622        | 85.17 | 27000 | 0.5351          | 0.1854 |
+| 0.6315        | 88.33 | 28000 | 0.5373          | 0.1811 |
+| 0.6315        | 91.48 | 29000 | 0.5331          | 0.1792 |
+| 0.6075        | 94.64 | 30000 | 0.5390          | 0.1779 |
+| 0.6075        | 97.79 | 31000 | 0.5459          | 0.1773 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.11.0

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 30, "</s>": 31}

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.5438345074653625,
+    "eval_runtime": 140.268,
+    "eval_samples": 5076,
+    "eval_samples_per_second": 36.188,
+    "eval_steps_per_second": 2.267,
+    "eval_wer": 0.177349387392344,
+    "train_loss": 0.8928292760036721,
+    "train_runtime": 80759.6589,
+    "train_samples": 20306,
+    "train_samples_per_second": 25.144,
+    "train_steps_per_second": 0.393
+}

config.json

@@ -0,0 +1,107 @@
+{
+  "_name_or_path": "imvladikon/wav2vec2-xls-r-300m-hebrew",
+  "activation_dropout": 0.1,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.0,
+  "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": 0.0,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.0,
+  "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_prob": 0.75,
+  "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": 29,
+  "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": 32,
+  "xvector_output_dim": 512
+}

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31ddb630e651b5670137ac169c70f3befd37058b7e77c57183510edbf5c313f9
+size 1262054897

run_train.py

@@ -0,0 +1,981 @@
+# !/usr/bin/env python
+# coding=utf-8
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+import torchaudio
+from datasets import DatasetDict, ReadInstruction, load_dataset, load_metric, concatenate_datasets
+
+try:
+    import bitsandbytes as bnb
+
+    BNB_AVAILABLE = True
+except:
+    BNB_AVAILABLE = False
+try:
+    import wandb
+
+    WANDB_AVAILABLE = True
+except:
+    WANDB_AVAILABLE = False
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainerCallback, TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+
+try:
+    from torch_audiomentations import (
+        Compose,
+        AddGaussianNoise,
+        AddGaussianSNR,
+        ClippingDistortion,
+        FrequencyMask,
+        Gain,
+        LoudnessNormalization,
+        Normalize,
+        PitchShift,
+        PolarityInversion,
+        Shift,
+        TimeMask,
+        TimeStretch,
+    )
+
+    AUDIOMENTATIONS_AVAILABLE = True
+except:
+    AUDIOMENTATIONS_AVAILABLE = False
+try:
+    from transformers import AutoProcessor
+except:
+    pass
+from transformers.trainer_pt_utils import get_parameter_names
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.16.0")
+
+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'."
+        },
+    )
+
+
+@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_path: str = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."
+        }
+    )
+    dataset_name: str = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."
+        },
+    )
+    dataset_config_name: str = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."
+        },
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="validation",
+        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'"
+        },
+    )
+    wav_filesize_column_name: str = field(
+        default=None,
+        metadata={
+            "help": "The name of the dataset column containing the wav filesize. Defaults is None"
+        },
+    )
+    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"
+        },
+    )
+    print_samples: bool = field(
+        default=False,
+        metadata={
+            "help": "Print row with validation inference results to stdout after each epoch"
+        },
+    )
+    use_augmentations: bool = field(
+        default=False,
+        metadata={
+            "help": "Use data augmentation during training"
+        },
+    )
+    use_auth_token: str = field(
+        default="",
+        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."
+        },
+    )
+
+
+class Augmentator:
+
+    def __init__(
+            self,
+            apply_gaussian_noise_with_p=0.1,
+            apply_gain_with_p=0.1,
+            apply_pitch_shift_with_p=0.1,
+            apply_time_stretch_with_p=0.1,
+            augment_proba=0.1,
+            sample_rate=16_000
+    ):
+        self.augmentator_fn = None
+        self.sample_rate = sample_rate
+        self.augment_proba = augment_proba
+        all_p = (
+                apply_gaussian_noise_with_p
+                + apply_gain_with_p
+                + apply_pitch_shift_with_p
+                + apply_time_stretch_with_p
+        )
+        if AUDIOMENTATIONS_AVAILABLE and all_p > 0:
+            self.augmentator_fn = Compose([
+                TimeStretch(min_rate=0.8, max_rate=1.2, leave_length_unchanged=False,
+                            p=apply_time_stretch_with_p),
+                PitchShift(min_semitones=-1, max_semitones=1,
+                           p=apply_pitch_shift_with_p),
+                Gain(min_gain_in_db=-1, max_gain_in_db=1, p=apply_gain_with_p),
+                AddGaussianNoise(min_amplitude=0.0001, max_amplitude=0.001,
+                                 p=apply_gaussian_noise_with_p),
+            ])
+
+    def __call__(self, input_values: List[float], *args, **kwargs):
+        if AUDIOMENTATIONS_AVAILABLE and self.augmentator_fn is not None:
+            return self.augmentator_fn(samples=np.array(input_values),
+                                       sample_rate=self.sample_rate).tolist()
+        else:
+            return input_values
+
+
+@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
+    augmentator_fn: Optional[Callable] = None
+    use_augmentations: bool = False
+
+    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": self.augmentator_fn(feature["input_values"])
+                if self.use_augmentations
+                else 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,
+        text_column_name: str,
+        train_split_name: str,
+        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[text_column_name])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    print("extract chars")
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets[train_split_name].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    print("make vocab_set")
+    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 speech_file_to_array_fn(batch, audio_column_name, dataset_path=""):
+    if dataset_path:
+        dataset_path = os.path.join(dataset_path, batch[audio_column_name])
+    else:
+        dataset_path = batch[audio_column_name] if isinstance(batch[audio_column_name],
+                                                              str) else \
+            batch[audio_column_name]["path"]
+    speech_array, sampling_rate = torchaudio.load(dataset_path)
+    batch[audio_column_name] = {
+        "array": speech_array[0].numpy(),
+        "sampling_rate": sampling_rate,
+    }
+    return batch
+
+
+class PrintSamplesPredictionCallback(TrainerCallback):
+
+    def __init__(self, processor, eval_dataset):
+        super(PrintSamplesPredictionCallback, self).__init__()
+        self.processor = processor
+        self.eval_dataset = eval_dataset
+        self.metric_fn = load_metric("wer")
+
+    def on_log(
+            self,
+            args: Any,
+            state: Any,
+            control: Any,
+            model: Any,
+            logs: Optional[Any] = None,
+            **kwargs
+    ):
+        """
+        :param args:
+        :param state:
+        :param control:
+        :param model:
+        :param logs:
+        :param kwargs: 'tokenizer', 'optimizer', 'lr_scheduler', 'train_dataloader', 'eval_dataloader'
+        :return:
+        """
+        if state.is_local_process_zero:
+            columns = ["id", "prediction", "reference", "audio", "wer"]
+            data = []
+            for idx, row in enumerate(self.eval_dataset):
+                input_dict = self.processor(row["input_values"],
+                                            return_tensors="pt", padding=True)
+                logits = model(input_dict.input_values.to(model.device)).logits
+                pred_ids = torch.argmax(logits, dim=-1)[0]
+                prediction = self.processor.decode(pred_ids)
+                print(f"Prediction: {prediction}")
+                reference = row['references'].lower()
+                print(f"\nReference: {reference}")
+
+                if WANDB_AVAILABLE:
+                    
+                    audio, sample_rate = tuple(row["audio"].values())
+                    audio = wandb.Audio(np.squeeze(audio),
+                                        sample_rate=sample_rate)
+                    wer = self.metric_fn.compute(
+                        predictions=[prediction],
+                        references=[reference],
+                    )
+
+                    data.append([idx, prediction, reference, audio, wer])
+            if WANDB_AVAILABLE:
+                table = wandb.Table(data=data, columns=columns)
+                wandb.run.log({"audio_predictions": table})
+
+
+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)
+
+    train_split_name = data_args.train_split_name
+    eval_split_name = data_args.eval_split_name
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict({
+        train_split_name: None,
+        eval_split_name: None,
+    })
+
+    if data_args.dataset_path:
+        raw_datasets = load_dataset(
+            "csv",
+            data_files={
+                train_split_name: os.path.join(data_args.dataset_path, "train-all.csv"),
+                eval_split_name: os.path.join(data_args.dataset_path, "eval-all.csv"),
+            },
+        )
+
+    if training_args.do_train:
+        if raw_datasets[train_split_name] is None:
+            raw_datasets[train_split_name] = 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,
+            )
+   
+        if data_args.audio_column_name not in raw_datasets[train_split_name].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset. "
+                "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_split_name].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset. "
+                "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_split_name] = raw_datasets[train_split_name].select(
+                range(data_args.max_train_samples)
+            )
+
+        if data_args.wav_filesize_column_name is not None:
+            raw_datasets[train_split_name] = raw_datasets[train_split_name].sort(
+                data_args.wav_filesize_column_name, reverse=True)
+
+    if training_args.do_eval:
+        if raw_datasets[eval_split_name] is None:
+            raw_datasets[eval_split_name] = 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,
+            )
+   
+        if data_args.max_eval_samples is not None:
+            raw_datasets[eval_split_name] = raw_datasets[eval_split_name].select(
+                range(data_args.max_eval_samples)
+            )
+        if data_args.wav_filesize_column_name is not None:
+            raw_datasets[eval_split_name] = raw_datasets[eval_split_name].sort(
+                data_args.wav_filesize_column_name, reverse=True)
+
+    # 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 = {}
+
+    # 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.
+    with open(os.path.join(tokenizer_name_or_path, "vocab.json"), "r") as fin:
+        print("loading tokenizer")
+        print(fin.read())
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+
+    # derive max & min input length for sample rate & max duration
+    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
+
+    raw_datasets[train_split_name] = raw_datasets[train_split_name].map(
+        speech_file_to_array_fn,
+        num_proc=num_workers,
+        fn_kwargs={"dataset_path": data_args.dataset_path,
+                   "audio_column_name": audio_column_name},
+    )
+    raw_datasets[eval_split_name] = raw_datasets[eval_split_name].map(
+        speech_file_to_array_fn,
+        num_proc=num_workers,
+        fn_kwargs={"dataset_path": data_args.dataset_path,
+                   "audio_column_name": audio_column_name},
+    )
+
+    # 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[data_args.text_column_name],
+                                    **additional_kwargs).input_ids
+        return batch
+
+    print(f"Vectorizing")
+
+    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",
+        )
+
+    # 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,
+        augmentator_fn=Augmentator(),
+        use_augmentations=data_args.use_augmentations
+    )
+
+    decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
+    decay_parameters = [name for name in decay_parameters if "bias" not in name]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if n in decay_parameters],
+            "weight_decay": training_args.weight_decay,
+        },
+        {
+            "params": [
+                p for n, p in model.named_parameters() if n not in decay_parameters
+            ],
+            "weight_decay": 0.0,
+        },
+    ]
+    trainer_kwargs = {}
+    if BNB_AVAILABLE:
+        optimizer = bnb.optim.Adam8bit(
+            params=optimizer_grouped_parameters,
+            betas=(training_args.adam_beta1, training_args.adam_beta2),
+            eps=training_args.adam_epsilon,
+        )
+        trainer_kwargs["optimizers"] = (optimizer, None)
+
+    samples_to_log = [
+        {
+            **vectorized_datasets[eval_split_name][i],
+            "references": raw_datasets[eval_split_name][i][data_args.text_column_name],
+            "audio": raw_datasets[eval_split_name][i][data_args.audio_column_name],
+        } for i in range(5)
+    ]
+
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets[
+            train_split_name] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets[
+            eval_split_name] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        **trainer_kwargs,
+        callbacks=[PrintSamplesPredictionCallback(
+            processor=processor,
+            eval_dataset=samples_to_log)] if data_args.print_samples and training_args.do_eval else None,
+    )
+
+    # 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_split_name])
+        )
+        metrics["train_samples"] = min(
+            max_train_samples, len(vectorized_datasets[train_split_name])
+        )
+
+        trainer.log_metrics(train_split_name, metrics)
+        trainer.save_metrics(train_split_name, 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_split_name])
+        )
+        metrics["eval_samples"] = min(max_eval_samples,
+                                      len(vectorized_datasets[eval_split_name]))
+
+        trainer.log_metrics(eval_split_name, metrics)
+        trainer.save_metrics(eval_split_name, 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 = {
+        "language": "he",
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", "robust-speech-event", "he"],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+    }
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

run_train.sh

@@ -0,0 +1,39 @@
+export CUDA_VISIBLE_DEVICES="0,1"
+
+python -m torch.distributed.launch --nproc_per_node=2 run_train.py \
+	--dataset_name="imvladikon/hebrew_speech_???" \
+	--use_auth_token="???" \
+    --audio_column_name="audio" \
+    --text_column_name="sentence" \
+	--model_name_or_path="imvladikon/wav2vec2-xls-r-300m-hebrew" \
+    --tokenizer_name_or_path="./wav2vec2-xls-r-300m-hebrew" \
+	--output_dir="./wav2vec2-xls-r-300m-hebrew" \
+	--overwrite_output_dir \
+	--evaluation_strategy="steps" \
+	--length_column_name="input_length" \
+	--gradient_checkpointing \
+	--fp16 \
+	--group_by_length \
+	--num_train_epochs="100" \
+	--per_device_train_batch_size="8" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="4" \
+	--learning_rate="3e-4" \
+	--warmup_steps="1000" \
+	--save_steps="1000" \
+	--eval_steps="1000" \
+	--preprocessing_num_workers="$(nproc)" \
+	--logging_steps="2000" \
+	--layerdrop="0.0" \
+	--activation_dropout="0.1" \
+	--save_total_limit="3" \
+	--freeze_feature_encoder \
+	--feat_proj_dropout="0.0" \
+	--mask_time_prob="0.75" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.25" \
+	--mask_feature_length="64" \
+	--do_train --do_eval \
+	--print_samples \
+	--use_augmentations \
+    --push_to_hub

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": "./wav2vec2-xls-r-300m-hebrew", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 100.0,
+    "train_loss": 0.8928292760036721,
+    "train_runtime": 80759.6589,
+    "train_samples": 20306,
+    "train_samples_per_second": 25.144,
+    "train_steps_per_second": 0.393
+}

trainer_state.json

@@ -0,0 +1,394 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 99.99842519685039,
+  "global_step": 31700,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 3.15,
+      "eval_loss": 0.5203462243080139,
+      "eval_runtime": 141.5768,
+      "eval_samples_per_second": 35.853,
+      "eval_steps_per_second": 2.246,
+      "eval_wer": 0.4333326279704594,
+      "step": 1000
+    },
+    {
+      "epoch": 6.31,
+      "learning_rate": 0.0009674592833876221,
+      "loss": 1.4284,
+      "step": 2000
+    },
+    {
+      "epoch": 6.31,
+      "eval_loss": 0.48156219720840454,
+      "eval_runtime": 142.5636,
+      "eval_samples_per_second": 35.605,
+      "eval_steps_per_second": 2.231,
+      "eval_wer": 0.3950949065746873,
+      "step": 2000
+    },
+    {
+      "epoch": 9.46,
+      "eval_loss": 0.4314565062522888,
+      "eval_runtime": 139.8322,
+      "eval_samples_per_second": 36.301,
+      "eval_steps_per_second": 2.274,
+      "eval_wer": 0.3545506485811626,
+      "step": 3000
+    },
+    {
+      "epoch": 12.62,
+      "learning_rate": 0.0009023452768729642,
+      "loss": 1.283,
+      "step": 4000
+    },
+    {
+      "epoch": 12.62,
+      "eval_loss": 0.42783403396606445,
+      "eval_runtime": 141.0912,
+      "eval_samples_per_second": 35.977,
+      "eval_steps_per_second": 2.254,
+      "eval_wer": 0.34039401570137756,
+      "step": 4000
+    },
+    {
+      "epoch": 15.77,
+      "eval_loss": 0.40902915596961975,
+      "eval_runtime": 140.9561,
+      "eval_samples_per_second": 36.011,
+      "eval_steps_per_second": 2.256,
+      "eval_wer": 0.3053939098969465,
+      "step": 5000
+    },
+    {
+      "epoch": 18.93,
+      "learning_rate": 0.0008372312703583062,
+      "loss": 1.1777,
+      "step": 6000
+    },
+    {
+      "epoch": 18.93,
+      "eval_loss": 0.3892641067504883,
+      "eval_runtime": 139.9163,
+      "eval_samples_per_second": 36.279,
+      "eval_steps_per_second": 2.273,
+      "eval_wer": 0.30056922783926193,
+      "step": 6000
+    },
+    {
+      "epoch": 22.08,
+      "eval_loss": 0.3967570960521698,
+      "eval_runtime": 139.2725,
+      "eval_samples_per_second": 36.447,
+      "eval_steps_per_second": 2.283,
+      "eval_wer": 0.28565080305563195,
+      "step": 7000
+    },
+    {
+      "epoch": 25.24,
+      "learning_rate": 0.0007720846905537459,
+      "loss": 1.0994,
+      "step": 8000
+    },
+    {
+      "epoch": 25.24,
+      "eval_loss": 0.3892391324043274,
+      "eval_runtime": 138.7844,
+      "eval_samples_per_second": 36.575,
+      "eval_steps_per_second": 2.291,
+      "eval_wer": 0.27509152083289246,
+      "step": 8000
+    },
+    {
+      "epoch": 28.39,
+      "eval_loss": 0.4061281681060791,
+      "eval_runtime": 139.2312,
+      "eval_samples_per_second": 36.457,
+      "eval_steps_per_second": 2.284,
+      "eval_wer": 0.2689760247159151,
+      "step": 9000
+    },
+    {
+      "epoch": 31.54,
+      "learning_rate": 0.000706970684039088,
+      "loss": 1.0323,
+      "step": 10000
+    },
+    {
+      "epoch": 31.54,
+      "eval_loss": 0.41136494278907776,
+      "eval_runtime": 139.4432,
+      "eval_samples_per_second": 36.402,
+      "eval_steps_per_second": 2.28,
+      "eval_wer": 0.25065069725120087,
+      "step": 10000
+    },
+    {
+      "epoch": 34.7,
+      "eval_loss": 0.40214526653289795,
+      "eval_runtime": 139.4093,
+      "eval_samples_per_second": 36.411,
+      "eval_steps_per_second": 2.281,
+      "eval_wer": 0.2508411452271621,
+      "step": 11000
+    },
+    {
+      "epoch": 37.85,
+      "learning_rate": 0.00064185667752443,
+      "loss": 0.9623,
+      "step": 12000
+    },
+    {
+      "epoch": 37.85,
+      "eval_loss": 0.40321338176727295,
+      "eval_runtime": 139.9917,
+      "eval_samples_per_second": 36.259,
+      "eval_steps_per_second": 2.272,
+      "eval_wer": 0.2378060393169266,
+      "step": 12000
+    },
+    {
+      "epoch": 41.01,
+      "eval_loss": 0.4147748053073883,
+      "eval_runtime": 139.3612,
+      "eval_samples_per_second": 36.423,
+      "eval_steps_per_second": 2.282,
+      "eval_wer": 0.23744630425122204,
+      "step": 13000
+    },
+    {
+      "epoch": 44.16,
+      "learning_rate": 0.0005767100977198697,
+      "loss": 0.9077,
+      "step": 14000
+    },
+    {
+      "epoch": 44.16,
+      "eval_loss": 0.4350396394729614,
+      "eval_runtime": 138.8108,
+      "eval_samples_per_second": 36.568,
+      "eval_steps_per_second": 2.291,
+      "eval_wer": 0.23230420890026873,
+      "step": 14000
+    },
+    {
+      "epoch": 47.32,
+      "eval_loss": 0.4514589309692383,
+      "eval_runtime": 138.719,
+      "eval_samples_per_second": 36.592,
+      "eval_steps_per_second": 2.292,
+      "eval_wer": 0.22464396808938358,
+      "step": 15000
+    },
+    {
+      "epoch": 50.47,
+      "learning_rate": 0.0005115960912052118,
+      "loss": 0.8573,
+      "step": 16000
+    },
+    {
+      "epoch": 50.47,
+      "eval_loss": 0.4473990499973297,
+      "eval_runtime": 140.3605,
+      "eval_samples_per_second": 36.164,
+      "eval_steps_per_second": 2.266,
+      "eval_wer": 0.21797828893074042,
+      "step": 16000
+    },
+    {
+      "epoch": 53.63,
+      "eval_loss": 0.4649062752723694,
+      "eval_runtime": 137.3039,
+      "eval_samples_per_second": 36.969,
+      "eval_steps_per_second": 2.316,
+      "eval_wer": 0.21713185348202382,
+      "step": 17000
+    },
+    {
+      "epoch": 56.78,
+      "learning_rate": 0.00044651465798045605,
+      "loss": 0.8083,
+      "step": 18000
+    },
+    {
+      "epoch": 56.78,
+      "eval_loss": 0.44551119208335876,
+      "eval_runtime": 139.4699,
+      "eval_samples_per_second": 36.395,
+      "eval_steps_per_second": 2.28,
+      "eval_wer": 0.2102334045749836,
+      "step": 18000
+    },
+    {
+      "epoch": 59.94,
+      "eval_loss": 0.4586869478225708,
+      "eval_runtime": 139.1403,
+      "eval_samples_per_second": 36.481,
+      "eval_steps_per_second": 2.285,
+      "eval_wer": 0.20917536026408787,
+      "step": 19000
+    },
+    {
+      "epoch": 63.09,
+      "learning_rate": 0.00038140065146579803,
+      "loss": 0.769,
+      "step": 20000
+    },
+    {
+      "epoch": 63.09,
+      "eval_loss": 0.4793929159641266,
+      "eval_runtime": 139.562,
+      "eval_samples_per_second": 36.371,
+      "eval_steps_per_second": 2.279,
+      "eval_wer": 0.20117654527371606,
+      "step": 20000
+    },
+    {
+      "epoch": 66.25,
+      "eval_loss": 0.4844733476638794,
+      "eval_runtime": 138.9678,
+      "eval_samples_per_second": 36.526,
+      "eval_steps_per_second": 2.288,
+      "eval_wer": 0.20073216666313987,
+      "step": 21000
+    },
+    {
+      "epoch": 69.4,
+      "learning_rate": 0.00031628664495114006,
+      "loss": 0.7308,
+      "step": 22000
+    },
+    {
+      "epoch": 69.4,
+      "eval_loss": 0.49372631311416626,
+      "eval_runtime": 139.7927,
+      "eval_samples_per_second": 36.311,
+      "eval_steps_per_second": 2.275,
+      "eval_wer": 0.20075332754935776,
+      "step": 22000
+    },
+    {
+      "epoch": 72.55,
+      "eval_loss": 0.4920376241207123,
+      "eval_runtime": 138.7644,
+      "eval_samples_per_second": 36.58,
+      "eval_steps_per_second": 2.292,
+      "eval_wer": 0.1894745751952092,
+      "step": 23000
+    },
+    {
+      "epoch": 75.71,
+      "learning_rate": 0.0002511400651465798,
+      "loss": 0.6927,
+      "step": 24000
+    },
+    {
+      "epoch": 75.71,
+      "eval_loss": 0.5178954005241394,
+      "eval_runtime": 139.799,
+      "eval_samples_per_second": 36.309,
+      "eval_steps_per_second": 2.275,
+      "eval_wer": 0.19114628520642443,
+      "step": 24000
+    },
+    {
+      "epoch": 78.86,
+      "eval_loss": 0.520152747631073,
+      "eval_runtime": 140.3812,
+      "eval_samples_per_second": 36.159,
+      "eval_steps_per_second": 2.265,
+      "eval_wer": 0.18767589986668642,
+      "step": 25000
+    },
+    {
+      "epoch": 82.02,
+      "learning_rate": 0.00018602605863192182,
+      "loss": 0.6622,
+      "step": 26000
+    },
+    {
+      "epoch": 82.02,
+      "eval_loss": 0.5265706181526184,
+      "eval_runtime": 138.1289,
+      "eval_samples_per_second": 36.748,
+      "eval_steps_per_second": 2.302,
+      "eval_wer": 0.18401506655098715,
+      "step": 26000
+    },
+    {
+      "epoch": 85.17,
+      "eval_loss": 0.5350863933563232,
+      "eval_runtime": 140.1605,
+      "eval_samples_per_second": 36.216,
+      "eval_steps_per_second": 2.269,
+      "eval_wer": 0.18541168504136954,
+      "step": 27000
+    },
+    {
+      "epoch": 88.33,
+      "learning_rate": 0.00012091205211726384,
+      "loss": 0.6315,
+      "step": 28000
+    },
+    {
+      "epoch": 88.33,
+      "eval_loss": 0.5373002290725708,
+      "eval_runtime": 138.125,
+      "eval_samples_per_second": 36.749,
+      "eval_steps_per_second": 2.302,
+      "eval_wer": 0.18113718602535075,
+      "step": 28000
+    },
+    {
+      "epoch": 91.48,
+      "eval_loss": 0.5330832600593567,
+      "eval_runtime": 139.0156,
+      "eval_samples_per_second": 36.514,
+      "eval_steps_per_second": 2.288,
+      "eval_wer": 0.17923270626573842,
+      "step": 29000
+    },
+    {
+      "epoch": 94.64,
+      "learning_rate": 5.576547231270358e-05,
+      "loss": 0.6075,
+      "step": 30000
+    },
+    {
+      "epoch": 94.64,
+      "eval_loss": 0.538992166519165,
+      "eval_runtime": 138.185,
+      "eval_samples_per_second": 36.733,
+      "eval_steps_per_second": 2.301,
+      "eval_wer": 0.17787840954779185,
+      "step": 30000
+    },
+    {
+      "epoch": 97.79,
+      "eval_loss": 0.5459240078926086,
+      "eval_runtime": 137.8608,
+      "eval_samples_per_second": 36.82,
+      "eval_steps_per_second": 2.307,
+      "eval_wer": 0.17730706561990817,
+      "step": 31000
+    },
+    {
+      "epoch": 100.0,
+      "step": 31700,
+      "total_flos": 3.173184730349909e+20,
+      "train_loss": 0.8928292760036721,
+      "train_runtime": 80759.6589,
+      "train_samples_per_second": 25.144,
+      "train_steps_per_second": 0.393
+    }
+  ],
+  "max_steps": 31700,
+  "num_train_epochs": 100,
+  "total_flos": 3.173184730349909e+20,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e69a04248e0ba54277463cd3b7be574e9546358e8382b38381a54e5df9b996a
+size 3055

validation_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.5438345074653625,
+    "eval_runtime": 140.268,
+    "eval_samples": 5076,
+    "eval_samples_per_second": 36.188,
+    "eval_steps_per_second": 2.267,
+    "eval_wer": 0.177349387392344
+}

vocab.json

@@ -0,0 +1 @@
+{"א": 1, "ב": 2, "ג": 3, "ד": 4, "ה": 5, "ו": 6, "ז": 7, "ח": 8, "ט": 9, "י": 10, "ך": 11, "כ": 12, "ל": 13, "ם": 14, "מ": 15, "ן": 16, "נ": 17, "ס": 18, "ע": 19, "ף": 20, "פ": 21, "ץ": 22, "צ": 23, "ק": 24, "ר": 25, "ש": 26, "ת": 27, "|": 0, "[UNK]": 28, "[PAD]": 29}