run_flax_speech_recognition_ctc.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2022 The HuggingFace Team All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the Flax library models for connectionist temporal classification (CTC) speech recognition.
"""
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.

import logging
import math
import os
import re
import sys
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional, Union

import datasets
import numpy as np
from datasets import DatasetDict, load_dataset, load_metric
from tqdm import tqdm

import flax
import jax
import jax.numpy as jnp
import optax
import transformers
import wandb as wandb
from flax import core, jax_utils, struct, traverse_util
from flax.jax_utils import unreplicate, pad_shard_unpad
from flax.training.common_utils import get_metrics, shard, shard_prng_key
from huggingface_hub import Repository
from models import Wav2Vec2Config, FlaxWav2Vec2ForCTC
from optax._src import linear_algebra
from transformers import (
    AutoFeatureExtractor,
    AutoProcessor,
    AutoTokenizer,
    HfArgumentParser,
    TrainingArguments,
    is_tensorboard_available,
)
from transformers.file_utils import get_full_repo_name
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("4.17.0.dev0")

require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")

logger = logging.getLogger(__name__)


@flax.struct.dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    feature_extractor_name: Optional[str] = field(
        default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
            "with private models)."
        },
    )
    freeze_feature_encoder: bool = field(
        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
    )
    activation_dropout: float = field(
        default=0.1,
        metadata={
            "help": "The hidden activation dropout probability in the embeddings, encoder, and pooler."
        },
    )
    hidden_dropout: float = field(
        default=0.1,
        metadata={
            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
        },
    )
    feat_proj_dropout: float = field(
        default=0.0,
        metadata={
            "help": "The feat proj dropout probability for feature encoder representations."
        },
    )
    mask_time_prob: float = field(
        default=0.1,
        metadata={
            "help": "The spec aug dropout probability for feature encoder representations."
        },
    )


@flax.struct.dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

    dataset_name: str = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    text_column: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
    )
    dataset_cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Path to cache directory for saving and loading datasets"}
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    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 evaluation examples to this "
            "value if set."
        },
    )
    max_test_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of test examples to this "
            "value if set."
        },
    )
    audio_column_name: str = field(
        default="audio",
        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
    )
    text_column_name: str = field(
        default="text",
        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
    )
    max_duration_in_seconds: float = field(
        default=20.0,
        metadata={
            "help": "Filter audio files in the training set that are longer than `max_duration_in_seconds` seconds"
        },
    )
    min_duration_in_seconds: float = field(
        default=0.0, metadata={"help": "Filter audio files in the training set that are shorter than `min_duration_in_seconds` seconds"}
    )
    max_label_length: Optional[int] = field(
        default=512,
        metadata={
            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
            "than this will be filtered."
        },
    )
    min_label_length: Optional[int] = field(
        default=0,
        metadata={
            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
            "than this will be filtered."
        },
    )
    max_eval_duration_in_seconds: float = field(
        default=None,
        metadata={
            "help": "Filter audio files in the eval/test set that are longer than `max_duration_in_seconds` seconds"
        },
    )
    pad_input_to_multiple_of: Optional[int] = field(
        default=32000,
        metadata={
            "help": "If set will pad the input sequence to a multiple of the provided value. "
            "This is important to avoid triggering recompilations on TPU."
        },
    )
    pad_target_to_multiple_of: Optional[int] = field(
        default=None,
        metadata={
            "help": "If set will pad the target sequence to a multiple of the provided value. "
            "This is important to avoid triggering recompilations on TPU."
        },
    )
    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"
        },
    )
    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'"
        },
    )
    do_lower_case: bool = field(
        default=True,
        metadata={"help": "Whether the target text should be lower cased."},
    )
    wandb_project: str = field(
        default="flax-speech-recognition-ctc",
        metadata={"help": "The name of the wandb project."},
    )
    wandb_name: str = field(
        default=None,
        metadata={"help": "The name of the wandb run."},
    )
    wandb_job_type: str = field(
        default="CTC",
        metadata={"help": "The name of the wandb job type."},
    )
    test_split_name: str = field(
        default="test",
        metadata={"help": "The name of the test data set split to use (via the datasets library). Defaults to 'test'"},
    )
    remove_punctuation: bool = field(
        default=False, metadata={"help": "Whether or not to remove punctuation during training."}
    )


# @flax.struct.dataclass
@dataclass
class FlaxTrainingArguments(TrainingArguments):
    precision: str = field(
        default="full",
        metadata={
            "help": "Whether to enable mixed-precision training. If true, the optimizer is stored in half-precision (bfloat16) and computations are executed in half-precision"
            "**Note that this only specifies the dtype of the computation and optimizer state. It does not influence the dtype of model parameters.**"
        },
    )
    matmul_precision: str = field(
        default="default",
        metadata={
            "help": "Default floating-point precision of internal computations used in TPU matrix multiplications and convolutions. "
            "This configuration option controls the default precision for JAX operations that take an optional precision argument (e.g. `lax.conv_general_dilated` and `lax.dot`). "
            "This configuration option does not change the behaviours of such calls with explicit precision arguments; "
            "it only changes the behaviors of calls with no such argument provided. "
            "One of `['highest', 'float32', 'high', 'bfloat16_3x', 'default', 'bfloat16', 'fastest', None]`."
        },
    )
    multisteps: bool = field(
        default=False,
        metadata={
            "help": "Whether to use Optax MultiSteps for gradient accumulation. If `False` (default) and `gradient_accumulation_steps > 1`, "
            "a custom gradient accumulation implementation will be employed."
        },
    )


def to_fp32(t):
    return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)


def to_bf16(t):
    return jax.tree_map(lambda x: x.astype(jnp.bfloat16) if x.dtype == jnp.float32 else x, t)


class MixedPrecisionTrainState(struct.PyTreeNode):
    """Train state for use with a single Optax optimizer.
    Adapted from flax train_state https://github.com/google/flax/blob/main/flax/training/train_state.py

    Synopsis::

        state = TrainState.create(
            apply_fn=model.apply,
            params=variables['params'],
            tx=tx)
        grad_fn = jax.grad(make_loss_fn(state.apply_fn))
        for batch in data:
          grads = grad_fn(state.params, batch)
          state = state.apply_gradients(grads=grads)

    Args:
      step: Counter starts at 0 and is incremented by every call to
        `.apply_gradients()`.
      apply_fn: Usually set to `model.apply()`. Kept in this dataclass for
        convenience to have a shorter params list for the `train_step()` function
        in your training loop.
      params: The parameters to be updated by `tx` and used by `apply_fn`.
      tx: An Optax gradient transformation.
      opt_state: The state for `tx`.
      dropout_rng: PRNG key for stochastic operations.
      bf16: Whether to use bf16 16-bit (mixed) precision training instead of 32-bit training.
    """

    step: int
    apply_fn: Callable = struct.field(pytree_node=False)
    get_attention_mask_fn: Callable = struct.field(pytree_node=False)
    params: core.FrozenDict[str, Any]
    tx: optax.GradientTransformation = struct.field(pytree_node=False)
    opt_state: optax.OptState
    dropout_rng: jnp.ndarray
    max_grad_norm: Optional[float] = 1.0

    def apply_gradients(self, *, grads, to_dtype, **kwargs):
        """Updates `step`, `params`, `opt_state` and `**kwargs` in return value.

        Note that internally this function calls `.tx.update()` followed by a call
        to `optax.apply_updates()` to update `params` and `opt_state`.

        Args:
          grads: Gradients that have the same pytree structure as `.params`.
          **kwargs: Additional dataclass attributes that should be `.replace()`-ed.

        Returns:
          An updated instance of `self` with `step` incremented by one, `params`
          and `opt_state` updated by applying `grads`, and additional attributes
          replaced as specified by `kwargs`.
        """

        # clip gradients by global l2 norm
        casted_max_grad_norm = to_dtype(self.max_grad_norm)
        g_norm = linear_algebra.global_norm(grads)
        g_norm = jnp.maximum(casted_max_grad_norm, g_norm)
        grads = jax.tree_map(lambda t: (t / g_norm) * casted_max_grad_norm, grads)

        # perform update step in fp32 and subsequently downcast optimizer states if mixed precision training
        # grads and opt_state in bf16 (need to upcast), params in fp32 (leave as is)
        updates, new_opt_state = self.tx.update(to_fp32(grads), to_fp32(self.opt_state), self.params)

        new_params = optax.apply_updates(self.params, updates)
        return self.replace(
            step=self.step + 1,
            params=new_params,
            opt_state=to_dtype(new_opt_state),
            **kwargs,
        )

    @classmethod
    def create(cls, *, apply_fn, params, tx, to_dtype, **kwargs):
        """Creates a new instance with `step=0` and initialized `opt_state`."""
        # downcast optimizer state to bf16 if mixed-precision training
        opt_state = tx.init(to_dtype(params)) if tx is not None else None
        return cls(
            step=0,
            apply_fn=apply_fn,
            params=params,
            tx=tx,
            opt_state=opt_state,
            **kwargs,
        )

    def replicate(self):
        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))


@flax.struct.dataclass
class FlaxDataCollatorSpeechSeq2SeqWithPadding:
    """
    Data collator that will dynamically pad the inputs received.
    Args:
        processor ([`Wav2Vec2Processor`])
            The processor used for proccessing the data.
        decoder_start_token_id (:obj: `int`)
            The begin-of-sentence of the decoder.
        input_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
            Select a strategy to pad the returned input 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).
        target_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
            Select a strategy to pad the returned target sequences (according to the model's padding side and padding index).
            See above for details.
        max_input_length (:obj:`float`, `optional`):
            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
        pad_input_to_multiple_of (:obj:`int`, `optional`):
            If set will pad the input 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).
        pad_target_to_multiple_of (:obj:`int`, `optional`):
            If set will pad the target 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: Any
    input_padding: Union[bool, str] = "longest"
    label_padding: Union[bool, str] = "max_length"
    pad_input_to_multiple_of: Optional[int] = None
    pad_to_multiple_of_label: Optional[int] = None
    max_input_length: Optional[float] = None
    max_label_length: Optional[float] = None

    def __call__(self, features: List[Dict[str, Union[List[int], np.ndarray]]]) -> Dict[str, np.ndarray]:
        # split inputs and labels since they have to be of different lengths 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]

        # reformat list to dict and set to pytorch format
        batch = self.processor.feature_extractor.pad(
            input_features,
            max_length=self.max_input_length,
            padding=self.input_padding,
            pad_to_multiple_of=self.pad_input_to_multiple_of,
            return_tensors="np",
        )

        labels_batch = self.processor.tokenizer.pad(
            label_features,
            max_length=self.max_label_length,
            padding=self.label_padding,
            pad_to_multiple_of=self.pad_to_multiple_of_label,
            return_tensors="np",
        )

        labels = labels_batch["input_ids"]
        labels = np.ma.array(labels, mask=np.not_equal(labels_batch.attention_mask, 1))
        labels = labels.filled(fill_value=-100)

        batch["labels"] = labels

        return batch


def get_grouped_indices(
    dataset, batch_size: int, rng: Optional[List[int]] = None, mega_batch_mult: Optional[int] = None
) -> np.array:
    """
    Adapted from the `get_length_grouped_indices` function in the PyTorch Trainer utils file (https://github.com/huggingface/transformers/blob/main/src/transformers/trainer_pt_utils.py#L486)
    Function that returns a list of indices in which each slice of `batch_size` consecutive indices correspond to elements of similar
    lengths. To do this, the indices are:

    - randomly permuted (if a JAX rng is specified)
    - grouped in mega-batches of size `mega_batch_mult * batch_size`
    - sorted by length in each mega-batch

    The result is the concatenation of all mega-batches, with the batch of `batch_size` containing the element of
    maximum length placed first, so that an OOM happens sooner rather than later.
    """
    lengths = dataset["input_length"]

    # Default for mega_batch_mult: 50 or the number to get 4 megabatches, whichever is smaller.
    if mega_batch_mult is None:
        mega_batch_mult = min(len(lengths) // (batch_size * 4), 50)
        # Just in case, for tiny datasets
        if mega_batch_mult == 0:
            mega_batch_mult = 1

    # We need to use JAX for the random permutation as the PRNG key will be set based on the seed outside of the sampler.
    num_samples = len(lengths)
    indices = jax.random.permutation(rng, np.arange(num_samples)) if rng is not None else np.arange(num_samples)

    megabatch_size = mega_batch_mult * batch_size
    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
    megabatches = [list(sorted(megabatch, key=lambda i: lengths[i], reverse=True)) for megabatch in megabatches]

    # The rest is to get the biggest batch first.
    # Since each megabatch is sorted by descending length, the longest element is the first
    megabatch_maximums = [lengths[megabatch[0]] for megabatch in megabatches]
    max_idx = np.argmax(megabatch_maximums).item()
    # Switch to put the longest batch in first position
    # (note that this is different to the PT grouped sampler in which we only put the longest element in the first position, and not its batch)
    megabatches[0], megabatches[max_idx] = megabatches[max_idx], megabatches[0]

    megabatches = np.array([i for megabatch in megabatches for i in megabatch])

    return megabatches


def generate_batch_splits(samples_idx: np.ndarray, batch_size: int, drop_last=True) -> np.ndarray:
    """Generate batches of data for a specified batch size from sample indices. If the dataset size is not divisible by
    the batch size and `drop_last` is `True`, the last incomplete batch is dropped. Else, it is returned."""
    num_samples = len(samples_idx)
    if drop_last:
        samples_to_remove = num_samples % batch_size
        if samples_to_remove != 0:
            samples_idx = samples_idx[:-samples_to_remove]
        sections_split = num_samples // batch_size
        samples_idx = samples_idx.reshape((sections_split, batch_size))
    else:
        sections_split = math.ceil(num_samples / batch_size)
        samples_idx = np.array_split(samples_idx, sections_split)
    return samples_idx


def write_train_metric(summary_writer, train_metrics, train_time, step):
    summary_writer.scalar("train_time", train_time, step)

    train_metrics = get_metrics(train_metrics)
    for key, vals in train_metrics.items():
        tag = f"train_{key}"
        for i, val in enumerate(vals):
            summary_writer.scalar(tag, val, step - len(vals) + i + 1)


def write_eval_metric(summary_writer, eval_metrics, step, pred_str=None):
    for metric_name, value in eval_metrics.items():
        summary_writer.scalar(f"eval_{metric_name}", value, step)

    if pred_str is not None:
        # write output actual predictions for debugging
        summary_writer.text("eval_predictions", "\n".join(pred_str), step)


def write_wandb_log(metrics, step, prefix=None):
    if jax.process_index() == 0:
        log_metrics = {}
        for k, v in metrics.items():
            if "layer" in k:
                log_metrics[f"{k}/"] = v
            elif prefix is not None:
                log_metrics[f"{prefix}/{k}"] = v
            else:
                log_metrics[k] = v
        wandb.log(log_metrics, step)


def write_wandb_pred(pred_str, label_str, step, final_step=False, prefix="eval"):
    if jax.process_index() == 0:
        # convert str data to a wandb compatible format
        str_data = [[label_str[i], pred_str[i]] for i in range(len(pred_str))]
        if not final_step:
            # we'll log the first 50 predictions for each intermediate epoch
            wandb.log(
                {
                    f"{prefix}/step_{int(step / 1000)}k": wandb.Table(
                        columns=["label_str", "pred_str"], data=str_data[:50]
                    )
                },
                step,
            )
        else:
            # we'll log all predictions for the last epoch
            wandb.log(
                {
                    f"{prefix}/step_{int(step / 1000)}k_all": wandb.Table(
                        columns=["label_str", "pred_str"], data=str_data
                    )
                },
                step,
            )


def create_learning_rate_fn(
    num_train_steps: int, num_warmup_steps: int, learning_rate: float
) -> Callable[[int], jnp.array]:
    """Returns a linear warmup, linear_decay learning rate function."""
    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
    decay_fn = optax.linear_schedule(
        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
    )
    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
    return schedule_fn


def ctc_loss(
    logits,
    logits_attention_mask,
    labels,
    blank_id,
    loss_reduction="mean",
    output_emission_dict=False,
    log_epsilon=-100000.0,
):
    """Computes CTC loss.
    This function performs forward computation over an FSA with `N * 2` states
    where `N` is the max number of labels. The states are split into two groups:
    Phi states and emission states. a phi-state accepts repetition of
    phi (blank)-symbols and transits to emission state when the correct label is
    observed. An emission state accepts repetition of the label and transits to
    the next phi states at any time (so called epsilon-transition).
    Below, `B` denotes the batch size, `T` denotes the time steps in `logits`,
    and `N` denotes the time steps in `labels`.
    Args:
      logits: (B, T, K)-array containing log-probabilities of each class.
      logitpaddings: (B, T)-array. Padding indicators for `logits`.
      labels: (B, N)-array containing reference integer labels.
      labelpaddings: (B, N)-array. Padding indicators for `labels`. Currently,
        `labels` must be right-padded, i.e. each row of `labelpaddings` must be
        repetition of zeroes, followed by repetition of ones.
      blank_id: Id for blank token.
      loss_reduction: one of "mean", "sum", "default"
        - "none": no reduction is applied.
        - "mean": output loss will be divided by target lengths and then the
          mean over the batch is taken.
        - "sum": output loss are summed over batch
      output_emission_dict: whether to output additional information about the emission probs
    Returns:
      A pair of `(per_seq_loss, aux)`.
      per_seq_loss:
        (B,)-array containing loss values for each sequence in the batch.
      aux: Dictionary containing interim variables used for computing losses.
        aux['logalpha_phi']: (T, B, N+1)-array. Log-forward-probabilities of each
          phi-state corresponding to the n-th label.
        aux['logalpha_emit']: (T, B, N)-array. Log-forward-probabilities of each
          emission-state corresponding to the n-th label.
        aux['logprobs_phi']: (T, B, 1)-array. Probability of the phi-symbol
          corresponding to each time frame.
        aux['logprobs_emit']: (T, B, N)-array. Probability of the n-th label
          corresponding to each time frame.
    """
    # label paddings are indicated by -100
    labelpaddings = labels < 0
    # logit paddings are the inverse of attention_mask
    logitpaddings = ~logits_attention_mask

    # Copied from https://github.com/tensorflow/lingvo/blob/master/lingvo/jax/layers/ctc_objectives.py
    batchsize, unused_maxinputlen, num_classes = logits.shape
    batchsize_, maxlabellen = labels.shape

    logprobs = jax.nn.log_softmax(logits)
    labellens = maxlabellen - jnp.sum(labelpaddings, axis=1).astype(jnp.int32)

    # repeat[b, n] == 1.0 when label[b, n] == label[b, n+1].
    repeat = (labels[:, :-1] == labels[:, 1:]).astype(jnp.float32)
    repeat = jnp.pad(repeat, ((0, 0), (0, 1)))

    logprobs_phi = logprobs[:, :, blank_id : blank_id + 1]  # [B, T, 1]
    logprobs_phi = jnp.transpose(logprobs_phi, (1, 0, 2))  # [T, B, 1]

    one_hot = jax.nn.one_hot(labels, num_classes=num_classes)  # [B, N, K]
    logprobs_emit = jnp.einsum("btk,bnk->btn", logprobs, one_hot)
    logprobs_emit = jnp.transpose(logprobs_emit, (1, 0, 2))  # [T, B, N]

    logalpha_phi_init = jnp.ones((batchsize, maxlabellen + 1)) * log_epsilon  # [B, N]
    logalpha_phi_init = logalpha_phi_init.at[:, 0].set(0.0)
    logalpha_emit_init = jnp.ones((batchsize, maxlabellen)) * log_epsilon  # [B, N]

    def loop_body(prev, x):
        prev_phi, prev_emit = prev
        # emit-to-phi epsilon transition, except if the next label is repetition
        prev_phi_orig = prev_phi
        prev_phi = prev_phi.at[:, 1:].set(jnp.logaddexp(prev_phi[:, 1:], prev_emit + log_epsilon * repeat))

        logprob_emit, logprob_phi, pad = x

        # phi-to-emit transition
        next_emit = jnp.logaddexp(prev_phi[:, :-1] + logprob_emit, prev_emit + logprob_emit)
        # self-loop transition
        next_phi = prev_phi + logprob_phi
        # emit-to-phi blank transition only when the next label is repetition
        next_phi = next_phi.at[:, 1:].set(
            jnp.logaddexp(next_phi[:, 1:], prev_emit + logprob_phi + log_epsilon * (1.0 - repeat))
        )

        pad = pad.reshape((batchsize, 1))
        next_emit = pad * prev_emit + (1.0 - pad) * next_emit
        next_phi = pad * prev_phi_orig + (1.0 - pad) * next_phi

        return (next_phi, next_emit), (next_phi, next_emit)

    xs = (logprobs_emit, logprobs_phi, logitpaddings.transpose((1, 0)))
    _, (logalpha_phi, logalpha_emit) = jax.lax.scan(loop_body, (logalpha_phi_init, logalpha_emit_init), xs)

    # last row needs to be updated with the last epsilon transition
    logalpha_phi_last = logalpha_phi[-1].at[:, 1:].set(jnp.logaddexp(logalpha_phi[-1, :, 1:], logalpha_emit[-1]))
    logalpha_phi = logalpha_phi.at[-1].set(logalpha_phi_last)

    # extract per_seq_loss
    one_hot = jax.nn.one_hot(labellens, num_classes=maxlabellen + 1)  # [B, N+1]
    per_seq_loss = -jnp.einsum("bn,bn->b", logalpha_phi_last, one_hot)

    if loss_reduction == "mean":
        target_lengths = labelpaddings.shape[-1] - labelpaddings.sum(axis=-1)
        loss = (per_seq_loss / target_lengths).mean()
    elif loss_reduction == "sum":
        loss = per_seq_loss.sum()
    else:
        loss = per_seq_loss

    if not output_emission_dict:
        return loss

    return loss, {
        "logalpha_phi": logalpha_phi,
        "logalpha_emit": logalpha_emit,
        "logprobs_phi": logprobs_phi,
        "logprobs_emit": logprobs_emit,
    }


def main():
    # 1. Parse input arguments
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.
    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, FlaxTrainingArguments))

    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()

    # 2. Setup logging
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )
    # Set the verbosity to info of the Transformers logger.
    # We only want one process per machine to log things on the screen.
    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
    if jax.process_index() == 0:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # Set up wandb run
    if jax.process_index() == 0:
        wandb.init(project=data_args.wandb_project, name=data_args.wandb_name, job_type=data_args.wandb_job_type)

    logger.info("Training/evaluation parameters %s", training_args)

    # Set the default TPU matmul precision and display the number of devices
    jax.config.update("jax_default_matmul_precision", training_args.matmul_precision)
    logger.info(f"JAX devices: {jax.device_count()}, matmul precision: {training_args.matmul_precision}")

    # 4. Load 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,
            cache_dir=data_args.dataset_cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )

    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,
            cache_dir=data_args.dataset_cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )

    if training_args.do_predict:
        test_split = data_args.test_split_name.split("+")
        for split in test_split:
            raw_datasets[split] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=split,
                cache_dir=data_args.dataset_cache_dir,
                use_auth_token=True if model_args.use_auth_token else None,
            )

    if not training_args.do_train and not training_args.do_eval and not training_args.do_predict:
        raise ValueError(
            "Cannot not train, not do evaluation and not do prediction. At least one of "
            "training, evaluation or prediction has to be done."
        )

    # if not training, there is no need to run multiple epochs
    if not training_args.do_train:
        training_args.num_train_epochs = 1

    if data_args.audio_column_name not in next(iter(raw_datasets.values())).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(next(iter(raw_datasets.values())).column_names)}."
        )

    if data_args.text_column_name not in next(iter(raw_datasets.values())).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(next(iter(raw_datasets.values())).column_names)}."
        )

    # 5. Load pretrained model, tokenizer, and feature extractor
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    config = Wav2Vec2Config.from_pretrained(
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    feature_extractor = AutoFeatureExtractor.from_pretrained(
        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    # update config according to training args, model args, and tokenizer attributes
    config.update(
        {
            "gradient_checkpointing": training_args.gradient_checkpointing,
            "activation_dropout": model_args.activation_dropout,
            "hidden_dropout": model_args.hidden_dropout,
            "feat_proj_dropout": model_args.feat_proj_dropout,
            "mask_time_prob": model_args.mask_time_prob,
            "vocab_size": tokenizer.vocab_size,
        }
    )

    if tokenizer.do_lower_case and data_args.dataset_name != "librispeech_asr":
        raise ValueError(
            "Setting the tokenizer attribute `do_lower_case` to `True` converts all input strings to "
            "uppercase prior to tokenization. This should only be done when the tokenizer is built on an uppercased corpus,"
            "i.e. for the dataset `librispeech_asr` only. If your dataset is not `librispeech_asr`, the tokenizer is mostly likely "
            "built on an lowercased corpus. In this case, set `tokenizer.do_lower_case` to ``False`."
        )

    if training_args.precision == "full_mixed":
        dtype = jnp.bfloat16
        training_args.mixed_precision = True
    elif training_args.precision == "half_mixed":
        dtype = jnp.bfloat16
        training_args.mixed_precision = False
    else:
        dtype = jnp.float32
        training_args.mixed_precision = False

    model = FlaxWav2Vec2ForCTC.from_pretrained(
        model_args.model_name_or_path,
        config=config,
        dtype=dtype,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )

    # 6. Resample speech dataset ALWAYS
    raw_datasets = raw_datasets.cast_column(
        data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
    )

    # 7. Preprocessing the datasets.
    # We need to read the audio files as arrays and tokenize the targets.
    max_input_length = int(data_args.max_duration_in_seconds * feature_extractor.sampling_rate)
    min_input_length = int(data_args.min_duration_in_seconds * feature_extractor.sampling_rate)
    max_eval_input_length = int(data_args.max_eval_duration_in_seconds * feature_extractor.sampling_rate) if data_args.max_eval_duration_in_seconds else None
    max_target_length = data_args.max_label_length
    min_target_length = data_args.min_label_length
    pad_input_to_multiple_of = data_args.pad_input_to_multiple_of
    audio_column_name = data_args.audio_column_name
    num_workers = data_args.preprocessing_num_workers
    text_column_name = data_args.text_column_name
    model_input_name = feature_extractor.model_input_names[0]
    do_lower_case = data_args.do_lower_case
    dataset_name = data_args.dataset_name
    tedlium_contractions = [" 's", " 't", " 're", " 've", " 'm", " 'll", " 'd", " 'clock", " 'all"]
    gigaspeech_punctuation = {" <comma>": ",", " <period>": ".", " <questionmark>": "?", " <exclamationpoint>": "!"}
    gigaspeech_disfluencies = ["<other>", "<sil>"]
    swb_disfluencies = ["[noise]", "[laughter]", "[silence]", "[vocalized-noise]", "<a_aside>", "<b_aside>", "<e_aside>",
                        "[laughter-", "_1", "[laugh]", "[sigh]", "[cough]", "[mn]", "[breath]", "[lipsmack]",
                        "[sneeze]", "[skip]", "[pause]", "(%hesitation)", "(%HESITATION)"]
    swb_punctuations = ["{", "}", "[", "]-", "]", "((", "))", "(", ")"]
    earnings_disfluencies = ["<noise>", "<crosstalk>", "<affirmative>", "<inaudible>", "inaudible", "<laugh>"]
    ignore_segments = ["ignore_time_segment_in_scoring", "<noise>", "<music>", "[noise]", "[laughter]", "[silence]",
                       "[vocalized-noise]", "<crosstalk>", "<affirmative>", "<inaudible>", "<laugh>", "<other>", "<sil>", ""]
    ignore_segments += swb_disfluencies

    if training_args.do_train and 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 and data_args.max_eval_samples is not None:
        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))

    if training_args.do_predict and data_args.max_test_samples is not None:
        for split in test_split:
            raw_datasets[split] = raw_datasets[split].select(range(data_args.max_eval_samples))

    # filter data where the targets are ignored in scoring
    def is_target_labels(input_str):
        return input_str.lower() not in ignore_segments

    raw_datasets = raw_datasets.filter(
            is_target_labels,
            num_proc=num_workers,
            input_columns=[text_column_name],
            desc="filtering data where the targets are ignored in scoring",
        )

    def prepare_dataset(batch):
        # Pre-process audio
        try:
            sample = batch[audio_column_name]
        except ValueError:
            # E22: some samples are empty (no audio). Reading the empty audio array will trigger
            # a soundfile ValueError. For now, we'll manually set these arrays to a zero array.
            # They will be filtered in the subsequent filtering stage and so are
            # explicitly ignored during training.
            sample = {"array": np.array([0.]), "sampling_rate": feature_extractor.sampling_rate}

        # normalise audio (mean, std) to (0, 1)
        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
        # process audio length
        batch[model_input_name] = inputs.input_values[0]
        batch["input_length"] = len(batch["input_values"])

        # 'Error correction' of targets
        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]

        # LibriSpeech ASR
        if dataset_name == "librispeech_asr":
            pass  # no error correction necessary

        # VoxPopuli
        if dataset_name == "google/xtreme_s":
            pass  # no error correction necessary

        # Common Voice 9
        if dataset_name == "mozilla-foundation/common_voice_9_0":
            if input_str.startswith('"') and input_str.endswith('"'):
                # we can remove trailing quotation marks as they do not affect the transcription
                input_str = input_str[1:-1]
            # replace double quotation marks with single
            input_str = input_str.replace('""', '"')

        # TED-LIUM (Release 3)
        if dataset_name == "LIUM/tedlium":
            # delete the <unk> token from the text
            input_str = input_str.replace("<unk>", "")
            # replace spaced apostrophes with un-spaced (it 's -> it's)
            for contraction in tedlium_contractions:
                input_str = input_str.replace(contraction, contraction[1:])

        # GigaSpeech
        if dataset_name == "speechcolab/gigaspeech":
            for disfluency in gigaspeech_disfluencies:
                input_str = input_str.replace(disfluency, "")
            # convert spelled out punctuation to symbolic form
            for punctuation, replacement in gigaspeech_punctuation.items():
                input_str = input_str.replace(punctuation, replacement)

        # SWB: hide the path to the private HF dataset
        if "switchboard" in dataset_name:
            # In one conversation people speak some German phrases that are tagged as
            # <german (( ja wohl )) > -- we remove these
            input_str = re.sub("<[^>]*>", "", input_str)

            # Remove junk tokens
            for disfluency in swb_disfluencies:
                input_str = input_str.replace(disfluency, "")

            # Replace partially pronounced words (square brackets + hyphen): westmin[ster]- to westmin- or -[go]ing to -ing
            # Replace anomalous words (square brackets + backslack): [lemguini/linguini] to linguini
            # Replace the combo of the two: [lem[guini]-/linguini] to lem-
            # Example: we [ah/are] -[go]ing to westmin[ster]- for [lem[guini]-/linguini]
            # Target: we ah -ing to westmin- for lem-
            # Treat anomalous words first then destroy the content of all square brackets (partially pronounced words)

            # First treat partially pronounced anomalous words by removing correct word: [lem[guini]-/linguini] to [lem[guini]-
            input_str = re.sub(r"\-\/.*?\]", "-", input_str)

            # Now replace anomalous words with their correct transcriptions: [lemguini/linguini] to linguini
            split_str = input_str.split("/")
            if len(split_str) > 1:
                input_str = " ".join(
                    [" ".join([" ".join(i.split(" ")[:-1]) for i in split_str])] + [split_str[-1].split(" ")[-1]])

            # Remove the trailing brackets on the start/end of words
            processed_str = []
            for word in input_str.split():
                if word[0] == "[":
                    processed_str.append(word[1:])
                elif word[-1] == "]":
                    processed_str.append(word[:-1])
                else:
                    processed_str.append(word)

            # Stick the processed words back together
            input_str = " ".join(processed_str)

            # Now we can remove all words in square brackets: -[go]ing to -ing
            input_str = re.sub(r"\-\[(.*?)\]", "-", input_str)

            # westmin[ster]- to westmin-
            input_str = re.sub(r"\[(.*?)\]\-", "-", input_str)

            # tech[n]ology to tech-ology
            input_str = re.sub(r"\[(.*?)\]", "-", input_str)

            # partially pronounced words are now done!
            # remove erroneous punctuations (curly braces, trailing square brackets, etc.)
            for punctuation in swb_punctuations:
                input_str = input_str.replace(punctuation, "")

        # Earnings 22: still figuring out best segmenting method. Thus, dataset name subject to change
        if "earnings22" in dataset_name:
            for disfluency in earnings_disfluencies:
                input_str = input_str.replace(disfluency, "")

        # SPGISpeech
        if dataset_name == "kensho/spgispeech":
            pass  # no error correction necessary

        # JIWER compliance (for WER/CER calc.)
        # remove multiple spaces
        input_str = re.sub(r"\s\s+", " ", input_str)
        # strip trailing spaces
        input_str = input_str.strip()

        # Finally, we tokenize the processed text
        batch["labels"] = tokenizer(input_str).input_ids
        batch["labels_length"] = len(batch["labels"])
        return batch

    vectorized_datasets = raw_datasets.map(
        prepare_dataset,
        remove_columns=next(iter(raw_datasets.values())).column_names,
        num_proc=num_workers,
        desc="preprocess dataset",
    )

    # filter training data with inputs longer than max_input_length
    def is_audio_in_length_range(length):
        return min_input_length < length < max_input_length

    if training_args.do_train:
        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
            is_audio_in_length_range,
            num_proc=num_workers,
            input_columns=["input_length"],
        )

    # filter data with targets shorter than min_target_length or longer than max_target_length
    def is_labels_in_length_range(length):
        return min_target_length < length < max_target_length

    if training_args.do_train:
        vectorized_datasets["train"] = vectorized_datasets["train"].filter(
            is_labels_in_length_range,
            num_proc=num_workers,
            input_columns=["labels_length"],
        )

    # filter data with targets shorter than 2 tokens (empty sentences)
    def is_labels_greater_than_min(length):
        return length > 2

    vectorized_datasets = vectorized_datasets.filter(
        is_labels_greater_than_min,
        num_proc=num_workers,
        input_columns=["labels_length"],
    )

    if max_eval_input_length is not None:
        # filter training data with inputs longer than max_input_length
        def is_eval_audio_in_length_range(length):
            return min_input_length < length < max_eval_input_length

        if training_args.do_eval:
            vectorized_datasets["eval"] = vectorized_datasets["eval"].filter(
                is_eval_audio_in_length_range,
                num_proc=num_workers,
                input_columns=["input_length"],
            )

        if training_args.do_predict:
            for split in test_split:
                vectorized_datasets[split] = vectorized_datasets[split].filter(
                    is_eval_audio_in_length_range,
                    num_proc=num_workers,
                    input_columns=["input_length"],
                )

    # 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:
        cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
        logger.info(f"Data preprocessing finished. Files cached at {cache}.")
        return

    # 8. Load Metrics
    wer_metric = load_metric("wer")
    cer_metric = load_metric("cer")

    def compute_metrics(pred_ids: List[List[int]], label_ids: List[List[int]]):
        padded_ids = np.where(np.asarray(label_ids) == -100, tokenizer.pad_token_id, np.asarray(label_ids))

        pred_str = tokenizer.batch_decode(pred_ids)
        # we do not want to group tokens when computing the metrics
        label_str = tokenizer.batch_decode(padded_ids, group_tokens=False)

        wer = wer_metric.compute(predictions=pred_str, references=label_str)
        cer = cer_metric.compute(predictions=pred_str, references=label_str)

        return {"wer": wer, "cer": cer}, pred_str, label_str

    # 9. save feature extractor, tokenizer and config
    feature_extractor.save_pretrained(training_args.output_dir)
    tokenizer.save_pretrained(training_args.output_dir)
    config.save_pretrained(training_args.output_dir)

    processor = AutoProcessor.from_pretrained(training_args.output_dir)

    data_collator = FlaxDataCollatorSpeechSeq2SeqWithPadding(
        processor=processor,
        input_padding="longest",
        pad_input_to_multiple_of=pad_input_to_multiple_of,
        max_label_length=data_args.max_label_length,
    )

    # Enable tensorboard only on the master node
    has_tensorboard = is_tensorboard_available()
    if has_tensorboard and jax.process_index() == 0:
        try:
            from flax.metrics.tensorboard import SummaryWriter

            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
        except ImportError as ie:
            has_tensorboard = False
            logger.warning(
                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
            )
    else:
        logger.warning(
            "Unable to display metrics through TensorBoard because the package is not installed: "
            "Please run `pip install tensorboard` to enable."
        )

    # 10. Handle the repository creation
    if training_args.push_to_hub:
        with open(os.path.join(training_args.output_dir, ".gitattributes"), "r+") as f:
            git_lfs_extensions = f.read()
            if "*.wandb" not in git_lfs_extensions:
                f.write("*.wandb filter=lfs diff=lfs merge=lfs -text")
        if training_args.hub_model_id is None:
            repo_name = get_full_repo_name(
                Path(training_args.output_dir).absolute().name, token=training_args.hub_token
            )
        else:
            repo_name = training_args.hub_model_id
        repo = Repository(training_args.output_dir, clone_from=repo_name)

    # 11. Initialize our training
    rng = jax.random.PRNGKey(training_args.seed)
    rng, dropout_rng = jax.random.split(rng)

    # Store some constants
    max_steps = int(training_args.max_steps)
    gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
    batch_size_per_update = train_batch_size * gradient_accumulation_steps
    per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
    to_dtype = to_bf16 if training_args.mixed_precision else to_fp32

    if training_args.do_train:
        num_train_samples = len(vectorized_datasets["train"])
        steps_per_epoch = num_train_samples // batch_size_per_update
        if max_steps > 0:
            num_epochs = -(training_args.max_steps // -steps_per_epoch)
            total_train_steps = max_steps
        else:
            num_epochs = int(training_args.num_train_epochs)
            total_train_steps = steps_per_epoch * num_epochs

        # Create learning rate schedule
        # Create learning rate schedule
        linear_decay_lr_schedule_fn = create_learning_rate_fn(
            total_train_steps,
            training_args.warmup_steps,
            training_args.learning_rate,
        )

        # We use Optax's "masking" functionality to not apply weight decay
        # to bias and LayerNorm scale parameters. decay_mask_fn returns a
        # mask boolean with the same structure as the parameters.
        # The mask is True for parameters that should be decayed.
        # Note that this mask is specifically adapted for FlaxWav2Vec2 and FlaxBart.
        # For FlaxT5, one should correct the layer norm parameter naming
        # accordingly - see `run_t5_mlm_flax.py` e.g.
        def decay_mask_fn(params):
            flat_params = traverse_util.flatten_dict(params)
            layer_norm_params = [
                (name, "scale")
                for name in ["layer_norm", "self_attn_layer_norm", "layernorm_embedding", "final_layer_norm"]
            ]
            flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_params) for path in flat_params}
            return traverse_util.unflatten_dict(flat_mask)

        if training_args.adafactor:
            # Create Adafactor optimizer
            optim = optax.adafactor(
                learning_rate=linear_decay_lr_schedule_fn,
                dtype_momentum=jnp.bfloat16 if training_args.mixed_precision else jnp.float32,
                weight_decay_rate=training_args.weight_decay,
                weight_decay_mask=decay_mask_fn,
            )
        else:
            # Create AdamW optimizer
            optim = optax.adamw(
                learning_rate=linear_decay_lr_schedule_fn,
                b1=training_args.adam_beta1,
                b2=training_args.adam_beta2,
                eps=training_args.adam_epsilon,
                weight_decay=training_args.weight_decay,
                mask=decay_mask_fn,
            )

        # Optax MultiSteps for gradient accumulation. We'll only call this optimizer transformation if gradient accumulation is required (i.e. gradient accumulation steps > 1)
        if training_args.multisteps and gradient_accumulation_steps > 1:
            optim = optax.MultiSteps(optim, gradient_accumulation_steps, use_grad_mean=False)
    else:
        num_epochs = 0
        total_train_steps = 0
        num_train_samples = 0
        optim = None

    # Setup train state
    state = MixedPrecisionTrainState.create(
        apply_fn=model.__call__,
        get_attention_mask_fn=model._get_feature_vector_attention_mask,
        params=model.params,
        tx=optim,
        to_dtype=to_dtype,
        dropout_rng=dropout_rng,
        max_grad_norm=training_args.max_grad_norm,
    )

    # Replicate the train state on each device
    state = state.replicate()
    blank_id = model.config.pad_token_id

    # Define gradient update step fn
    def train_step(state, batch):
        # only one single rng per grad step, with or without accumulation, as the graph should be identical over one effective training batch
        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)

        def compute_loss(params, minibatch):
            labels = minibatch.pop("labels")
            logits = state.apply_fn(
                **minibatch,
                params=params,
                dropout_rng=dropout_rng,
                freeze_feature_encoder=model_args.freeze_feature_encoder,
                train=True,
            )[0]
            logits_mask = state.get_attention_mask_fn(logits.shape[1], batch["attention_mask"])
            loss = ctc_loss(logits, logits_mask, labels, blank_id, loss_reduction="mean")

            return loss

        grad_fn = jax.value_and_grad(compute_loss)

        if gradient_accumulation_steps == 1 or training_args.multisteps:
            loss, grad = grad_fn(to_dtype(state.params), batch)

        # Custom gradient accumulation
        else:
            # add a first dimension over gradient_accumulation_steps for minibatch slices
            batch = jax.tree_map(
                lambda x: x.reshape(
                    gradient_accumulation_steps, training_args.per_device_train_batch_size, *x.shape[1::]
                ),
                batch,
            )

            def accum_minibatch_step(accum_grad, minibatch):
                # compute loss, num labels and grad over minibatch and accumulate
                loss, grad = grad_fn(to_dtype(state.params), minibatch)
                return jax.tree_map(jnp.add, accum_grad, grad), loss

            # create an initial state for accumulating losses, num labels and gradients
            init_grad = jax.tree_map(jnp.zeros_like, to_dtype(state.params))
            # loop accum minibatch step over the number of gradient accumulation steps
            grad, loss = jax.lax.scan(accum_minibatch_step, init_grad, batch)

        # update state
        new_state = state.apply_gradients(
            grads=grad,
            dropout_rng=new_dropout_rng,
            to_dtype=to_dtype,
        )

        # compute gradient norms over all layers and globally for detailed monitoring
        layer_grad_norm = jax.tree_map(jnp.linalg.norm, grad)
        logs = {
            "layer_grad_norm": layer_grad_norm,
            "grad_norm": jnp.linalg.norm(jax.tree_util.tree_leaves(layer_grad_norm)),
        }

        # compute parameter norms over all layers and globally for detailed monitoring
        layer_param_norm = jax.tree_map(jnp.linalg.norm, new_state.params)
        logs["layer_param_norm"] = layer_param_norm
        logs["param_norm"] = jnp.linalg.norm(jax.tree_util.tree_leaves(layer_param_norm))

        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
        metrics.update(logs)

        metrics = jax.lax.pmean(metrics, axis_name="batch")
        # metrics = to_fp32(metrics)

        return new_state, metrics

    # Define eval fn
    def eval_step(params, batch):
        labels = batch.pop("labels")
        logits = model(**batch, params=params, train=False)[0]

        logits_mask = model._get_feature_vector_attention_mask(logits.shape[1], batch["attention_mask"])
        loss = ctc_loss(logits, logits_mask, labels, blank_id, loss_reduction="mean")

        pred_ids = jnp.argmax(logits, axis=-1)

        # summarize metrics
        metrics = {"loss": loss}
        metrics = jax.lax.pmean(metrics, axis_name="batch")
        # metrics = to_fp32(metrics)
        return metrics, pred_ids

    # Create parallel version of the train and eval step
    if training_args.do_train:
        p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))

    if training_args.do_eval or training_args.do_predict:
        p_eval_step = jax.pmap(eval_step, "batch")

    def run_evaluation(step, final_step=False):
        if training_args.do_eval:
            # ======================== Evaluating ==============================
            eval_metrics = []
            eval_preds = []
            eval_labels = []

            # Generate eval set by sequentially sampling indices from the eval dataset and grouping by length
            eval_samples_idx = get_grouped_indices(vectorized_datasets["eval"], eval_batch_size)
            eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size, drop_last=False)

            for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
                samples = [vectorized_datasets["eval"][int(idx)] for idx in batch_idx]
                batch = data_collator(samples)
                labels = batch["labels"]

                try:
                    metrics, pred_ids = pad_shard_unpad(p_eval_step)(state.params, batch.data, min_device_batch=per_device_eval_batch_size)
                except TypeError:
                    continue
                eval_preds.extend(jax.device_get(pred_ids.reshape(-1, pred_ids.shape[-1])))
                eval_metrics.append(metrics)

                eval_labels.extend(labels)

            # normalize eval metrics
            eval_metrics = get_metrics(eval_metrics)
            eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
            eval_metrics = to_fp32(eval_metrics)

            # always run compute metrics
            error_rate_metric, pred_str, label_str = compute_metrics(eval_preds, eval_labels)
            eval_metrics.update(error_rate_metric)
            error_rate_desc = " ".join([f"Eval {key}: {value} |" for key, value in error_rate_metric.items()])

            # Print metrics and update progress bar
            desc = f"Step... ({step}/{total_train_steps} | Eval Loss: {eval_metrics['loss']} | {error_rate_desc})"
            epochs.write(desc)
            epochs.desc = desc

            # Save metrics
            write_wandb_log(eval_metrics, step, prefix="eval")
            write_wandb_pred(pred_str, label_str, step, final_step=final_step)
            # if has_tensorboard and jax.process_index() == 0:
            # write_eval_metric(summary_writer, eval_metrics, step, pred_str=pred_str)

    def save_checkpoint(step):
        # save and push checkpoint to the hub
        if jax.process_index() == 0:
            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
            model.save_pretrained(training_args.output_dir, params=params)
            tokenizer.save_pretrained(training_args.output_dir)
            if training_args.push_to_hub:
                repo.push_to_hub(commit_message=f"{wandb.run.id}: saving weights and logs of step {int(step / 1000)}k", blocking=False)

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {num_train_samples}")
    logger.info(f"  Num Epochs = {num_epochs}")
    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
    logger.info(f"  Num gradient accumulation steps = {gradient_accumulation_steps}")
    logger.info(f"  Total train batch size (w. parallel & distributed) = {batch_size_per_update}")
    logger.info(f"  Total optimization steps = {total_train_steps}")
    logger.info(f"  Gradient checkpointing: {config.gradient_checkpointing}")
    logger.info(f"  Use scan: {config.use_scan}")
    logger.info(f"  Fuse matmuls: {config.fuse_matmuls}")

    train_time = cur_step = 0
    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
    for epoch in epochs:
        if training_args.do_train:
            # ======================== Training ================================
            train_start = time.time()

            # Create sampling rng
            rng, input_rng = jax.random.split(rng)

            # Generate an epoch by randomly shuffling sampling indices from the train dataset and grouping by length
            train_samples_idx = get_grouped_indices(vectorized_datasets["train"], batch_size_per_update, input_rng)
            train_batch_idx = generate_batch_splits(train_samples_idx, batch_size_per_update)

            # Gather the indices for creating the batch and do a training step
            for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1), 1):
                samples = [vectorized_datasets["train"][int(idx)] for idx in batch_idx]
                batch = data_collator(samples)
                batch = shard(batch.data)
                try:
                    state, train_metric = p_train_step(state, batch)
                except TypeError as e:
                    logger.warning("Encountered following error: \n", e)

                cur_step = epoch * (num_train_samples // batch_size_per_update) + step

                if cur_step % training_args.logging_steps == 0:
                    # Save metrics
                    train_metric = unreplicate(train_metric)
                    train_time += time.time() - train_start
                    # need to upcast all device arrays to fp32 for wandb logging (jnp.bfloat16 not supported) -> do this here OR in train_step
                    write_wandb_log(to_fp32(train_metric), cur_step, prefix="train")
                    # we won't log to tensorboard for now (it is fiddly logging param and grad norms on a layer-by-layer basis)
                    # if has_tensorboard and jax.process_index() == 0:
                    # write_train_metric(summary_writer, train_metrics, train_time, cur_step)

                    epochs.write(
                        f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']}, Gradient Norm: {train_metric['grad_norm']})"
                    )

                if cur_step % total_train_steps == 0:
                    break

                if training_args.eval_steps and cur_step % training_args.eval_steps == 0:
                    run_evaluation(cur_step, final_step=False)

                if cur_step % training_args.save_steps == 0:
                    save_checkpoint(cur_step)

            if training_args.eval_steps == 0 and (epoch + 1) != num_epochs:
                # run evaluation at the end of the epoch if eval steps are not specified
                run_evaluation(cur_step, final_step=False)
                save_checkpoint(cur_step)

    if training_args.do_train:
        save_checkpoint(cur_step)

    cur_step = max_steps if max_steps > 0 else cur_step  # set step to max steps so that eval happens in alignment with training

    if training_args.do_eval:
        run_evaluation(cur_step, final_step=True)

    # TODO: collapse 'do_predict' into the run_evaluation function
    if training_args.do_predict:
        for split in test_split:
            # ======================== Evaluating ==============================
            eval_metrics = []
            eval_preds = []
            eval_labels = []

            # Generate eval set by sequentially sampling indices from the test dataset and grouping by length
            eval_samples_idx = get_grouped_indices(vectorized_datasets[split], eval_batch_size)
            eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size, drop_last=False)

            for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc=f"Predicting {split}...", position=2)):
                samples = [vectorized_datasets[split][int(idx)] for idx in batch_idx]
                batch = data_collator(samples)
                labels = batch["labels"]

                metrics, pred_ids = pad_shard_unpad(p_eval_step)(state.params, batch.data, min_device_batch=per_device_eval_batch_size)
                eval_preds.extend(jax.device_get(pred_ids.reshape(-1, pred_ids.shape[-1])))
                eval_metrics.append(metrics)

                eval_labels.extend(labels)

            # normalize eval metrics
            eval_metrics = get_metrics(eval_metrics)
            eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
            eval_metrics = to_fp32(eval_metrics)

            # always run compute metrics
            error_rate_metric, pred_str, label_str = compute_metrics(eval_preds, eval_labels)
            eval_metrics.update(error_rate_metric)
            error_rate_desc = " ".join([f"Eval {key}: {value} |" for key, value in error_rate_metric.items()])

            # Print metrics and update progress bar
            desc = f"Step... ({cur_step}/{total_train_steps} | Eval Loss: {eval_metrics['loss']} | {error_rate_desc})"
            epochs.write(desc)
            epochs.desc = desc

            # Save metrics
            write_wandb_log(eval_metrics, cur_step, prefix=split)
            write_wandb_pred(pred_str, label_str, cur_step, final_step=True, prefix=split)
            # if has_tensorboard and jax.process_index() == 0:
            # write_eval_metric(summary_writer, eval_metrics, cur_step, pred_str=pred_str)


if __name__ == "__main__":
    main()