2y71vcu4: saving weights and logs of step 5k

events.out.tfevents.1659700313.t1v-n-eedfb410-w-0.3572327.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c97be7c1da6944777366fdb9b6a215f817d3404702de43e7cb635c65bce877c3
+size 40

events.out.tfevents.1659704423.t1v-n-eedfb410-w-0.2632364.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9acb1bfa25db8551780fabc1f10392872609619a94b58bb4d3da1dc556c4f882
+size 40

events.out.tfevents.1659713857.t1v-n-eedfb410-w-0.1781451.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7be4e8f9f2175b6c57c706283fc3d06d821d847cce440f5ed5561f91c721a68
+size 40

events.out.tfevents.1659741255.t1v-n-eedfb410-w-0.825506.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea9456621b28d25411e36f9a7e20faa6862a27c9b3165c65e01f70413e35ae2f
+size 40

flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c0dfd76c4de572f7ef0d3b62ef8e1512bb9c3814cdffd8f9e8ae54845bd9dc9d
+oid sha256:635cdc48780adc608e330d5ba666a9ea24d6f5425496bf993f1286ead4a35623
 size 3850218852

run.sh

@@ -1,6 +1,6 @@
 WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_ctc.py \
         --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
-        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst" \
+        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst-tpu" \
         --tokenizer_name="./" \
         --output_dir="./" \
         --overwrite_output_dir \
@@ -11,13 +11,13 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
         --precision="full_mixed" \
         --matmul_precision="bfloat16" \
         --multisteps \
-        --learning_rate="2e-4" \
+        --learning_rate="1e-4" \
         --warmup_steps="2000" \
         --length_column_name="input_length" \
         --evaluation_strategy="steps" \
         --text_column_name="text" \
-        --save_steps="4000" \
-        --eval_steps="4000" \
+        --save_steps="5000" \
+        --eval_steps="5000" \
         --logging_steps="100" \
         --layerdrop="0.041" \
         --attention_dropout="0.094" \
@@ -42,7 +42,7 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
         --ctc_zero_infinity \
         --do_lower_case \
         --wandb_project="wav2vec2" \
-        --wandb_name="wav2vec2-1b-npsc-nst" \
+        --wandb_name="wav2vec2-1b-npsc-nst-tpu" \
         --remove_punctuation
 
 

special_tokens_map.json

@@ -343,6 +343,62 @@
       "rstrip": false,
       "single_word": false
     },
+    {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
     {
       "content": "</s>",
       "lstrip": false,

wandb/debug-internal.log

@@ -1 +1 @@
-run-20220803_091109-yit1e59z/logs/debug-internal.log
+run-20220805_230151-2y71vcu4/logs/debug-internal.log

wandb/debug.log

@@ -1 +1 @@
-run-20220803_091109-yit1e59z/logs/debug.log
+run-20220805_230151-2y71vcu4/logs/debug.log

wandb/latest-run

@@ -1 +1 @@
-run-20220803_091109-yit1e59z
+run-20220805_230151-2y71vcu4

wandb/run-20220805_114003-3f1x0hvu/files/code/run_flax_speech_recognition_ctc.py

@@ -0,0 +1,1631 @@
+#!/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,
+    set_seed,
+)
+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."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    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."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.1,
+        metadata={
+            "help": "The spec aug dropout probability for feature encoder representations."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@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": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    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=2,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+            "than this will be filtered."
+        },
+    )
+    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."}
+    )
+    skip_steps: Optional[int] = field(
+        default=0,
+        metadata={
+            "help": "Skip this number of steps. Useful to continue 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_util.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def to_bf16(t):
+    return jax.tree_util.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_util.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, num_log=50, 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))]
+        # we'll log the first 50 predictions for each epoch
+        wandb.log(
+            {
+                f"{prefix}/step_{int(step / 1000)}k": wandb.Table(
+                    columns=["label_str", "pred_str"], data=str_data[:num_log]
+                )
+            },
+            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 make_dataset(data_args, seed=42):
+    # Pre-processing dataset
+    import re
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text}
+
+    def filter_nst(entry):
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        return True
+
+    def filter_npsc(entry):
+        # False if there are digits in the text
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    # TODO NST_hesitate
+
+    split = len(npsc[data_args.train_split_name]) / (len(npsc[data_args.train_split_name]) + len(npsc[data_args.eval_split_name]))  # Use same train/val ratio as NPSC
+    nst_train = nst[data_args.train_split_name].train_test_split(train_size=split, seed=seed)
+    nst[data_args.train_split_name] = nst_train["train"]
+    nst[data_args.eval_split_name] = nst_train["test"]
+
+    nst = nst.filter(filter_nst).map(
+        map_nst,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NST",
+    ).shuffle(seed=seed)
+    npsc = npsc.filter(filter_npsc).map(
+        map_npsc,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NPSC",
+    ).shuffle(seed=seed)
+
+    npsc_base = npsc.remove_columns([col for col in npsc[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+    nst_base = nst.remove_columns([col for col in nst[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in data_args.train_split_name, data_args.eval_split_name, data_args.test_split_name:
+        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+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
+
+    set_seed(training_args.seed)
+    raw_datasets = make_dataset(data_args, seed=training_args.seed)
+   
+    # raw_datasets = DatasetDict()
+
+    # if training_args.do_train:
+    #     raw_datasets[data_args.train_split_name] = 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[data_args.eval_split_name] = 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(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": tokenizer.vocab_size,  # len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    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
+
+    try:
+        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,
+        )
+    except:
+        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,
+            from_pt=True,
+        )
+
+    # 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_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
+    chars_to_ignore = ', ? . ! - ; : " “ % ‘ ” ?'.split(" ")
+    chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
+    # gigaspeech_punctuation = {" <comma>": ",", " <period>": ".", " <questionmark>": "?", " <exclamationpoint>": "!"}
+    # gigaspeech_disfluencies = ["<other>", "<sil>"]
+    # swb_disfluencies = ["[noise]", "[laughter]", "[silence]", "<a_aside>", "<b_aside>", "<e_aside>", "[laughter-",
+    #                 "[vocalized-noise]", "_1"]
+    # swb_punctuations = ["{", "}", "[", "]-", "]"]
+    # earnings_disfluencies = ["<crosstalk>", "<affirmative>", "<inaudible>", "inaudible", "<laugh>", "<unk>"]
+    ignore_segments = ["ignore_time_segment_in_scoring", "<noise>", "<music>", "[noise]", "[laughter]", "[silence]",
+                       "[vocalized-noise]", "<crosstalk>", "<affirmative>", "<inaudible>", "<laugh>", "<other>", "<sil>", ""]
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets[data_args.eval_split_name] = raw_datasets[data_args.eval_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_predict and data_args.max_test_samples is not None:
+        raw_datasets[data_args.test_split_name] = raw_datasets[data_args.test_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_train and data_args.remove_punctuation:
+
+        def remove_punctuation(batch):
+            batch[text_column_name] = (
+                re.sub(chars_to_ignore_regex, "", batch[text_column_name]).replace("'", "").replace('"', "")
+            )
+
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].map(
+            remove_punctuation,
+            num_proc=data_args.preprocessing_num_workers,
+            desc="removing punctuation from train split",
+        )
+
+    # 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):
+        # process audio
+        try:
+            sample = batch[audio_column_name]
+        except ValueError:
+            sample = {"array": np.array([0.]), "sampling_rate": feature_extractor.sampling_rate}
+        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"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+
+        # if dataset_name == "google/xtreme_s":
+        #     # Finally, we tokenize the processed text
+        #     batch["labels"] = tokenizer(input_str).input_ids
+        #     batch["labels_length"] = len(batch["labels"])
+        #     return batch
+
+        # # Common Voice 9
+        # 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]
+        # # normalize quotation marks
+        # input_str = re.sub(r'["“”]', '"', input_str)
+        # # normalize apostrophes
+        # input_str = re.sub(r"[’']", "'", input_str)
+        # # normalize hyphens
+        # input_str = re.sub(r"[—–]", "-", input_str)
+        # # replace double quotation marks with single
+        # input_str = input_str.replace('""', '"')
+        # if dataset_name == "mozilla-foundation/common_voice_9_0" and len(input_str):
+        #     # for CV9, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # TEDLIUM-3
+        # # delete the <unk> token from the text and replace spaced apostrophes with un-spaced
+        # input_str = input_str.replace("<unk>", "").replace(" '", "'")
+
+        # # 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)
+        # if dataset_name == "speechcolab/gigaspeech" and len(input_str):
+        #     # for GS, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # SWB
+        # for disfluency in swb_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # remove parenthesised text (test data only)
+        # input_str = re.sub("[\(].*?[\)]", "", input_str)
+        # for punctuation in swb_punctuations:
+        #     input_str = input_str.replace(punctuation, "")
+        # # replace anomalous words with their correct transcriptions
+        # 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]])
+
+        # # Earnings 22
+        # for disfluency in earnings_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # replace mal-formatted ellipsis
+        # input_str = input_str.replace("…", ".")
+
+        # JIWER compliance
+        # 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 data with inputs shorter than min_input_length or longer than max_input_length
+    def is_audio_in_length_range(length):
+        return length > min_input_length and length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.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 length > min_target_length  # and length < max_target_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+        is_labels_in_length_range,
+        num_proc=num_workers,
+        input_columns=["labels_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[data_args.train_split_name])
+        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_util.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_util.tree_map(jnp.add, accum_grad, grad), loss
+
+            # create an initial state for accumulating losses, num labels and gradients
+            init_grad = jax.tree_util.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_util.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_util.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:
+        p_eval_step = jax.pmap(eval_step, "batch")
+
+    def run_evaluation(step):
+        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[data_args.eval_split_name], 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[data_args.eval_split_name][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_util.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)
+            # 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_util.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)
+
+    skip_epochs = data_args.skip_steps // (num_train_samples // batch_size_per_update)
+    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}")
+    logger.info(f"  Skipping: {data_args.skip_steps} steps ({skip_epochs} epochs)")
+
+    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()
+
+            if epoch < skip_epochs:
+                logger.info(f"Skipping epoch {epoch + 1}")
+                continue
+
+            # 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[data_args.train_split_name], batch_size_per_update, input_rng)
+            train_batch_idx = generate_batch_splits(train_samples_idx, batch_size_per_update)
+
+            if data_args.skip_steps > cur_step:
+                logger.info(f"Skipping {data_args.skip_steps - (epoch * (num_train_samples // batch_size_per_update))} steps...")
+            # 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):
+                cur_step = epoch * (num_train_samples // batch_size_per_update) + step
+                if cur_step <= data_args.skip_steps:
+                    continue
+
+                samples = [vectorized_datasets[data_args.train_split_name][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)
+
+
+                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=data_args.train_split_name)
+                    # 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)
+
+                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)
+                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)
+
+    # TODO: collapse 'do_predict' into the run_evaluation function
+    if training_args.do_predict:
+        for split in [data_args.test_split_name]:
+            # ======================== 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_util.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, 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()

wandb/run-20220805_114003-3f1x0hvu/files/config.yaml

@@ -0,0 +1,33 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_flax_speech_recognition_ctc.py
+    framework: huggingface
+    huggingface_version: 4.21.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1659699604
+    t:
+      1:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      3:
+      - 13
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.21.0
+      8:
+      - 5

wandb/run-20220805_114003-3f1x0hvu/files/diff.patch

@@ -0,0 +1,53 @@
+diff --git a/run.sh b/run.sh
+index 9cc498e..604b808 100755
+--- a/run.sh
++++ b/run.sh
+@@ -5,8 +5,8 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --output_dir="./" \
+         --overwrite_output_dir \
+         --num_train_epochs="40" \
+-        --per_device_train_batch_size="2" \
+-        --per_device_eval_batch_size="2" \
++        --per_device_train_batch_size="4" \
++        --per_device_eval_batch_size="4" \
+         --gradient_accumulation_steps="1" \
+         --precision="full_mixed" \
+         --matmul_precision="bfloat16" \
+@@ -16,8 +16,8 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --length_column_name="input_length" \
+         --evaluation_strategy="steps" \
+         --text_column_name="text" \
+-        --save_steps="4000" \
+-        --eval_steps="4000" \
++        --save_steps="5000" \
++        --eval_steps="5000" \
+         --logging_steps="100" \
+         --layerdrop="0.041" \
+         --attention_dropout="0.094" \
+diff --git a/wandb/debug-internal.log b/wandb/debug-internal.log
+index 50a0b69..4934fdf 120000
+--- a/wandb/debug-internal.log
++++ b/wandb/debug-internal.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug-internal.log
+
++run-20220805_114003-3f1x0hvu/logs/debug-internal.log
+
+diff --git a/wandb/debug.log b/wandb/debug.log
+index 746223d..3328f24 120000
+--- a/wandb/debug.log
++++ b/wandb/debug.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug.log
+
++run-20220805_114003-3f1x0hvu/logs/debug.log
+
+diff --git a/wandb/latest-run b/wandb/latest-run
+index be58b40..d627a18 120000
+--- a/wandb/latest-run
++++ b/wandb/latest-run
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z
+
++run-20220805_114003-3f1x0hvu
+

wandb/run-20220805_114003-3f1x0hvu/files/output.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5174df6a0b92bb2375fae8ef696b8a151681c1be2513b177260019f859192dd2
+size 224929

wandb/run-20220805_114003-3f1x0hvu/files/requirements.txt

@@ -0,0 +1,158 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+appdirs==1.4.4
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+audioread==2.1.9
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+cffi==1.15.1
+charset-normalizer==2.0.10
+chex==0.1.3
+click==8.0.3
+cloud-tpu-client==0.10
+cloud-tpu-profiler==2.4.0
+clu==0.0.6
+colorama==0.4.5
+commonmark==0.9.1
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==2.4.0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+etils==0.6.0
+exceptiongroup==1.0.0rc8
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.5.3
+fonttools==4.28.5
+frozenlist==1.2.0
+fsspec==2021.11.1
+future==0.18.2
+gast==0.4.0
+gitdb==4.0.9
+gitpython==3.1.26
+google-api-core==1.31.5
+google-api-python-client==1.8.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.4.6
+google-auth==2.3.3
+google-pasta==0.2.0
+googleapis-common-protos==1.54.0
+grpcio==1.43.0
+h5py==3.6.0
+httplib2==0.20.2
+huggingface-hub==0.2.1
+hypothesis==6.53.0
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.3.15
+jaxlib==0.3.15
+jedi==0.18.1
+jiwer==2.3.0
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+librosa==0.9.2
+libtpu-nightly==0.1.dev20220722
+llvmlite==0.39.0
+markdown==3.3.6
+matplotlib-inline==0.1.3
+matplotlib==3.5.1
+ml-collections==0.1.0
+msgpack==1.0.3
+multidict==5.2.0
+multiprocess==0.70.12.2
+numba==0.56.0
+numpy==1.22.0
+oauth2client==4.1.3
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.3
+packaging==21.3
+pandas==1.3.5
+parso==0.8.3
+pathtools==0.1.2
+pexpect==4.8.0
+pickleshare==0.7.5
+pillow==9.0.0
+pip==22.2.1
+pkg-resources==0.0.0
+pooch==1.6.0
+promise==2.3
+prompt-toolkit==3.0.24
+protobuf==3.19.1
+psutil==5.9.0
+ptyprocess==0.7.0
+pyarrow==6.0.1
+pyasn1-modules==0.2.8
+pyasn1==0.4.8
+pycparser==2.21
+pyctcdecode==0.4.0
+pygments==2.11.1
+pygtrie==2.5.0
+pyparsing==3.0.6
+python-dateutil==2.8.2
+python-levenshtein==0.12.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+resampy==0.3.1
+responses==0.18.0
+rich==11.2.0
+rsa==4.8
+sacremoses==0.0.46
+scikit-learn==1.1.1
+scipy==1.7.3
+sentry-sdk==1.5.2
+setuptools==44.0.0
+shortuuid==1.0.8
+six==1.16.0
+smmap==5.0.0
+sortedcontainers==2.4.0
+soundfile==0.10.3.post1
+sox==1.4.1
+subprocess32==3.5.4
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+tensorboard==2.7.0
+tensorflow-cpu==2.7.0
+tensorflow-datasets==4.4.0
+tensorflow-estimator==2.7.0
+tensorflow-io-gcs-filesystem==0.23.1
+tensorflow-metadata==1.5.0
+tensorflow==2.7.0
+tensorstore==0.1.21
+termcolor==1.1.0
+threadpoolctl==3.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+torch==1.12.0
+torchaudio==0.12.0+cpu
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.21.0
+typing-extensions==4.3.0
+uritemplate==3.0.1
+urllib3==1.26.7
+wandb==0.12.9
+wcwidth==0.2.5
+werkzeug==2.0.2
+wheel==0.37.1
+wrapt==1.13.3
+xxhash==2.0.2
+yarl==1.7.2
+yaspin==2.1.0
+zipp==3.7.0

wandb/run-20220805_114003-3f1x0hvu/files/wandb-metadata.json

@@ -0,0 +1,69 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-08-05T11:40:07.382258",
+    "startedAt": "2022-08-05T11:40:03.911271",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--model_name_or_path=facebook/wav2vec2-xls-r-1b",
+        "--hub_model_id=NbAiLab/wav2vec2-1b-npsc-nst",
+        "--tokenizer_name=./",
+        "--output_dir=./",
+        "--overwrite_output_dir",
+        "--num_train_epochs=40",
+        "--per_device_train_batch_size=4",
+        "--per_device_eval_batch_size=4",
+        "--gradient_accumulation_steps=1",
+        "--precision=full_mixed",
+        "--matmul_precision=bfloat16",
+        "--multisteps",
+        "--learning_rate=2e-4",
+        "--warmup_steps=2000",
+        "--length_column_name=input_length",
+        "--evaluation_strategy=steps",
+        "--text_column_name=text",
+        "--save_steps=5000",
+        "--eval_steps=5000",
+        "--logging_steps=100",
+        "--layerdrop=0.041",
+        "--attention_dropout=0.094",
+        "--activation_dropout=0.055",
+        "--hidden_dropout=0.047",
+        "--save_total_limit=5",
+        "--freeze_feature_encoder",
+        "--feat_proj_dropout=0.04",
+        "--mask_time_prob=0.082",
+        "--mask_time_length=10",
+        "--mask_feature_prob=0.25",
+        "--mask_feature_length=64",
+        "--gradient_checkpointing",
+        "--min_duration_in_seconds=0.5",
+        "--max_duration_in_seconds=30.0",
+        "--use_auth_token",
+        "--seed=42",
+        "--group_by_length",
+        "--do_train",
+        "--do_eval",
+        "--push_to_hub",
+        "--preprocessing_num_workers=32",
+        "--ctc_zero_infinity",
+        "--do_lower_case",
+        "--wandb_project=wav2vec2",
+        "--wandb_name=wav2vec2-1b-npsc-nst",
+        "--remove_punctuation"
+    ],
+    "state": "running",
+    "program": "run_flax_speech_recognition_ctc.py",
+    "codePath": "run_flax_speech_recognition_ctc.py",
+    "git": {
+        "remote": "https://huggingface.co/NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "commit": "e2b1320cc68c3ce129a1d654965e0d3eb44e0558"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/wav2vec2-1b-npsc-nst",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220805_114003-3f1x0hvu/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb": {"runtime": 716}}

wandb/run-20220805_114003-3f1x0hvu/logs/debug-internal.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eeba4400c2433c15bfaf1d69d20d5af658a679b9f4cf75713f00e9883d3620e0
+size 42862

wandb/run-20220805_114003-3f1x0hvu/logs/debug.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:361ad7147270e83596fd398076999e7c2b9ad69eb4d2afc82355d01434e95eda
+size 5845

wandb/run-20220805_114003-3f1x0hvu/run-3f1x0hvu.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e68f244da66ea282bb58485e3253a0680c39e3f27e0d1c9cd1851021d6a64342
+size 240293

wandb/run-20220805_124834-3ep5xqhh/files/code/run_flax_speech_recognition_ctc.py

@@ -0,0 +1,1631 @@
+#!/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,
+    set_seed,
+)
+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."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    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."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.1,
+        metadata={
+            "help": "The spec aug dropout probability for feature encoder representations."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@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": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    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=2,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+            "than this will be filtered."
+        },
+    )
+    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."}
+    )
+    skip_steps: Optional[int] = field(
+        default=0,
+        metadata={
+            "help": "Skip this number of steps. Useful to continue 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_util.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def to_bf16(t):
+    return jax.tree_util.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_util.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, num_log=50, 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))]
+        # we'll log the first 50 predictions for each epoch
+        wandb.log(
+            {
+                f"{prefix}/step_{int(step / 1000)}k": wandb.Table(
+                    columns=["label_str", "pred_str"], data=str_data[:num_log]
+                )
+            },
+            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 make_dataset(data_args, seed=42):
+    # Pre-processing dataset
+    import re
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text}
+
+    def filter_nst(entry):
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        return True
+
+    def filter_npsc(entry):
+        # False if there are digits in the text
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    # TODO NST_hesitate
+
+    split = len(npsc[data_args.train_split_name]) / (len(npsc[data_args.train_split_name]) + len(npsc[data_args.eval_split_name]))  # Use same train/val ratio as NPSC
+    nst_train = nst[data_args.train_split_name].train_test_split(train_size=split, seed=seed)
+    nst[data_args.train_split_name] = nst_train["train"]
+    nst[data_args.eval_split_name] = nst_train["test"]
+
+    nst = nst.filter(filter_nst).map(
+        map_nst,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NST",
+    ).shuffle(seed=seed)
+    npsc = npsc.filter(filter_npsc).map(
+        map_npsc,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NPSC",
+    ).shuffle(seed=seed)
+
+    npsc_base = npsc.remove_columns([col for col in npsc[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+    nst_base = nst.remove_columns([col for col in nst[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in data_args.train_split_name, data_args.eval_split_name, data_args.test_split_name:
+        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+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
+
+    set_seed(training_args.seed)
+    raw_datasets = make_dataset(data_args, seed=training_args.seed)
+   
+    # raw_datasets = DatasetDict()
+
+    # if training_args.do_train:
+    #     raw_datasets[data_args.train_split_name] = 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[data_args.eval_split_name] = 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(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": tokenizer.vocab_size,  # len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    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
+
+    try:
+        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,
+        )
+    except:
+        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,
+            from_pt=True,
+        )
+
+    # 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_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
+    chars_to_ignore = ', ? . ! - ; : " “ % ‘ ” ?'.split(" ")
+    chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
+    # gigaspeech_punctuation = {" <comma>": ",", " <period>": ".", " <questionmark>": "?", " <exclamationpoint>": "!"}
+    # gigaspeech_disfluencies = ["<other>", "<sil>"]
+    # swb_disfluencies = ["[noise]", "[laughter]", "[silence]", "<a_aside>", "<b_aside>", "<e_aside>", "[laughter-",
+    #                 "[vocalized-noise]", "_1"]
+    # swb_punctuations = ["{", "}", "[", "]-", "]"]
+    # earnings_disfluencies = ["<crosstalk>", "<affirmative>", "<inaudible>", "inaudible", "<laugh>", "<unk>"]
+    ignore_segments = ["ignore_time_segment_in_scoring", "<noise>", "<music>", "[noise]", "[laughter]", "[silence]",
+                       "[vocalized-noise]", "<crosstalk>", "<affirmative>", "<inaudible>", "<laugh>", "<other>", "<sil>", ""]
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets[data_args.eval_split_name] = raw_datasets[data_args.eval_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_predict and data_args.max_test_samples is not None:
+        raw_datasets[data_args.test_split_name] = raw_datasets[data_args.test_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_train and data_args.remove_punctuation:
+
+        def remove_punctuation(batch):
+            batch[text_column_name] = (
+                re.sub(chars_to_ignore_regex, "", batch[text_column_name]).replace("'", "").replace('"', "")
+            )
+
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].map(
+            remove_punctuation,
+            num_proc=data_args.preprocessing_num_workers,
+            desc="removing punctuation from train split",
+        )
+
+    # 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):
+        # process audio
+        try:
+            sample = batch[audio_column_name]
+        except ValueError:
+            sample = {"array": np.array([0.]), "sampling_rate": feature_extractor.sampling_rate}
+        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"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+
+        # if dataset_name == "google/xtreme_s":
+        #     # Finally, we tokenize the processed text
+        #     batch["labels"] = tokenizer(input_str).input_ids
+        #     batch["labels_length"] = len(batch["labels"])
+        #     return batch
+
+        # # Common Voice 9
+        # 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]
+        # # normalize quotation marks
+        # input_str = re.sub(r'["“”]', '"', input_str)
+        # # normalize apostrophes
+        # input_str = re.sub(r"[’']", "'", input_str)
+        # # normalize hyphens
+        # input_str = re.sub(r"[—–]", "-", input_str)
+        # # replace double quotation marks with single
+        # input_str = input_str.replace('""', '"')
+        # if dataset_name == "mozilla-foundation/common_voice_9_0" and len(input_str):
+        #     # for CV9, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # TEDLIUM-3
+        # # delete the <unk> token from the text and replace spaced apostrophes with un-spaced
+        # input_str = input_str.replace("<unk>", "").replace(" '", "'")
+
+        # # 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)
+        # if dataset_name == "speechcolab/gigaspeech" and len(input_str):
+        #     # for GS, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # SWB
+        # for disfluency in swb_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # remove parenthesised text (test data only)
+        # input_str = re.sub("[\(].*?[\)]", "", input_str)
+        # for punctuation in swb_punctuations:
+        #     input_str = input_str.replace(punctuation, "")
+        # # replace anomalous words with their correct transcriptions
+        # 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]])
+
+        # # Earnings 22
+        # for disfluency in earnings_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # replace mal-formatted ellipsis
+        # input_str = input_str.replace("…", ".")
+
+        # JIWER compliance
+        # 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 data with inputs shorter than min_input_length or longer than max_input_length
+    def is_audio_in_length_range(length):
+        return length > min_input_length and length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.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 length > min_target_length  # and length < max_target_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+        is_labels_in_length_range,
+        num_proc=num_workers,
+        input_columns=["labels_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[data_args.train_split_name])
+        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_util.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_util.tree_map(jnp.add, accum_grad, grad), loss
+
+            # create an initial state for accumulating losses, num labels and gradients
+            init_grad = jax.tree_util.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_util.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_util.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:
+        p_eval_step = jax.pmap(eval_step, "batch")
+
+    def run_evaluation(step):
+        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[data_args.eval_split_name], 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[data_args.eval_split_name][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_util.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)
+            # 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_util.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)
+
+    skip_epochs = data_args.skip_steps // (num_train_samples // batch_size_per_update)
+    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}")
+    logger.info(f"  Skipping: {data_args.skip_steps} steps ({skip_epochs} epochs)")
+
+    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()
+
+            if epoch < skip_epochs:
+                logger.info(f"Skipping epoch {epoch + 1}")
+                continue
+
+            # 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[data_args.train_split_name], batch_size_per_update, input_rng)
+            train_batch_idx = generate_batch_splits(train_samples_idx, batch_size_per_update)
+
+            if data_args.skip_steps > cur_step:
+                logger.info(f"Skipping {data_args.skip_steps - (epoch * (num_train_samples // batch_size_per_update))} steps...")
+            # 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):
+                cur_step = epoch * (num_train_samples // batch_size_per_update) + step
+                if cur_step <= data_args.skip_steps:
+                    continue
+
+                samples = [vectorized_datasets[data_args.train_split_name][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)
+
+
+                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=data_args.train_split_name)
+                    # 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)
+
+                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)
+                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)
+
+    # TODO: collapse 'do_predict' into the run_evaluation function
+    if training_args.do_predict:
+        for split in [data_args.test_split_name]:
+            # ======================== 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_util.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, 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()

wandb/run-20220805_124834-3ep5xqhh/files/config.yaml

@@ -0,0 +1,33 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_flax_speech_recognition_ctc.py
+    framework: huggingface
+    huggingface_version: 4.21.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1659703714
+    t:
+      1:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      3:
+      - 13
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.21.0
+      8:
+      - 5

wandb/run-20220805_124834-3ep5xqhh/files/diff.patch

@@ -0,0 +1,105 @@
+diff --git a/config.json b/config.json
+index 260219f..246b797 100644
+--- a/config.json
++++ b/config.json
+@@ -5,7 +5,7 @@
+   "add_adapter": false,
+   "apply_spec_augment": true,
+   "architectures": [
+-    "Wav2Vec2ForCTC"
++    "Wav2Vec2ForPreTraining"
+   ],
+   "attention_dropout": 0.094,
+   "bos_token_id": 1,
+diff --git a/run.sh b/run.sh
+index 9cc498e..5068965 100755
+--- a/run.sh
++++ b/run.sh
+@@ -1,12 +1,12 @@
+ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_ctc.py \
+         --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+-        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst" \
++        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst-tpu" \
+         --tokenizer_name="./" \
+         --output_dir="./" \
+         --overwrite_output_dir \
+         --num_train_epochs="40" \
+-        --per_device_train_batch_size="2" \
+-        --per_device_eval_batch_size="2" \
++        --per_device_train_batch_size="4" \
++        --per_device_eval_batch_size="4" \
+         --gradient_accumulation_steps="1" \
+         --precision="full_mixed" \
+         --matmul_precision="bfloat16" \
+@@ -16,8 +16,8 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --length_column_name="input_length" \
+         --evaluation_strategy="steps" \
+         --text_column_name="text" \
+-        --save_steps="4000" \
+-        --eval_steps="4000" \
++        --save_steps="5000" \
++        --eval_steps="5000" \
+         --logging_steps="100" \
+         --layerdrop="0.041" \
+         --attention_dropout="0.094" \
+@@ -42,7 +42,7 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --ctc_zero_infinity \
+         --do_lower_case \
+         --wandb_project="wav2vec2" \
+-        --wandb_name="wav2vec2-1b-npsc-nst" \
++        --wandb_name="wav2vec2-1b-npsc-nst-tpu" \
+         --remove_punctuation
+ 
+ 
+diff --git a/special_tokens_map.json b/special_tokens_map.json
+index 89389bf..c3eacb7 100644
+--- a/special_tokens_map.json
++++ b/special_tokens_map.json
+@@ -343,6 +343,20 @@
+       "rstrip": false,
+       "single_word": false
+     },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
+     {
+       "content": "</s>",
+       "lstrip": false,
+diff --git a/wandb/debug-internal.log b/wandb/debug-internal.log
+index 50a0b69..d716bbe 120000
+--- a/wandb/debug-internal.log
++++ b/wandb/debug-internal.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug-internal.log
+
++run-20220805_124834-3ep5xqhh/logs/debug-internal.log
+
+diff --git a/wandb/debug.log b/wandb/debug.log
+index 746223d..92d0ec0 120000
+--- a/wandb/debug.log
++++ b/wandb/debug.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug.log
+
++run-20220805_124834-3ep5xqhh/logs/debug.log
+
+diff --git a/wandb/latest-run b/wandb/latest-run
+index be58b40..075d875 120000
+--- a/wandb/latest-run
++++ b/wandb/latest-run
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z
+
++run-20220805_124834-3ep5xqhh
+

wandb/run-20220805_124834-3ep5xqhh/files/output.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de21dd0aba5898c5acdb62ceee14797ef9d861ef1edb14f2c99b96ab509c9ccf
+size 237607

wandb/run-20220805_124834-3ep5xqhh/files/requirements.txt

@@ -0,0 +1,158 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+appdirs==1.4.4
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+audioread==2.1.9
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+cffi==1.15.1
+charset-normalizer==2.0.10
+chex==0.1.3
+click==8.0.3
+cloud-tpu-client==0.10
+cloud-tpu-profiler==2.4.0
+clu==0.0.6
+colorama==0.4.5
+commonmark==0.9.1
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==2.4.0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+etils==0.6.0
+exceptiongroup==1.0.0rc8
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.5.3
+fonttools==4.28.5
+frozenlist==1.2.0
+fsspec==2021.11.1
+future==0.18.2
+gast==0.4.0
+gitdb==4.0.9
+gitpython==3.1.26
+google-api-core==1.31.5
+google-api-python-client==1.8.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.4.6
+google-auth==2.3.3
+google-pasta==0.2.0
+googleapis-common-protos==1.54.0
+grpcio==1.43.0
+h5py==3.6.0
+httplib2==0.20.2
+huggingface-hub==0.2.1
+hypothesis==6.53.0
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.3.15
+jaxlib==0.3.15
+jedi==0.18.1
+jiwer==2.3.0
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+librosa==0.9.2
+libtpu-nightly==0.1.dev20220722
+llvmlite==0.39.0
+markdown==3.3.6
+matplotlib-inline==0.1.3
+matplotlib==3.5.1
+ml-collections==0.1.0
+msgpack==1.0.3
+multidict==5.2.0
+multiprocess==0.70.12.2
+numba==0.56.0
+numpy==1.22.0
+oauth2client==4.1.3
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.3
+packaging==21.3
+pandas==1.3.5
+parso==0.8.3
+pathtools==0.1.2
+pexpect==4.8.0
+pickleshare==0.7.5
+pillow==9.0.0
+pip==22.2.1
+pkg-resources==0.0.0
+pooch==1.6.0
+promise==2.3
+prompt-toolkit==3.0.24
+protobuf==3.19.1
+psutil==5.9.0
+ptyprocess==0.7.0
+pyarrow==6.0.1
+pyasn1-modules==0.2.8
+pyasn1==0.4.8
+pycparser==2.21
+pyctcdecode==0.4.0
+pygments==2.11.1
+pygtrie==2.5.0
+pyparsing==3.0.6
+python-dateutil==2.8.2
+python-levenshtein==0.12.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+resampy==0.3.1
+responses==0.18.0
+rich==11.2.0
+rsa==4.8
+sacremoses==0.0.46
+scikit-learn==1.1.1
+scipy==1.7.3
+sentry-sdk==1.5.2
+setuptools==44.0.0
+shortuuid==1.0.8
+six==1.16.0
+smmap==5.0.0
+sortedcontainers==2.4.0
+soundfile==0.10.3.post1
+sox==1.4.1
+subprocess32==3.5.4
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+tensorboard==2.7.0
+tensorflow-cpu==2.7.0
+tensorflow-datasets==4.4.0
+tensorflow-estimator==2.7.0
+tensorflow-io-gcs-filesystem==0.23.1
+tensorflow-metadata==1.5.0
+tensorflow==2.7.0
+tensorstore==0.1.21
+termcolor==1.1.0
+threadpoolctl==3.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+torch==1.12.0
+torchaudio==0.12.0+cpu
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.21.0
+typing-extensions==4.3.0
+uritemplate==3.0.1
+urllib3==1.26.7
+wandb==0.12.9
+wcwidth==0.2.5
+werkzeug==2.0.2
+wheel==0.37.1
+wrapt==1.13.3
+xxhash==2.0.2
+yarl==1.7.2
+yaspin==2.1.0
+zipp==3.7.0

wandb/run-20220805_124834-3ep5xqhh/files/wandb-metadata.json

@@ -0,0 +1,69 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-08-05T12:48:37.490835",
+    "startedAt": "2022-08-05T12:48:34.040204",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--model_name_or_path=facebook/wav2vec2-xls-r-1b",
+        "--hub_model_id=NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "--tokenizer_name=./",
+        "--output_dir=./",
+        "--overwrite_output_dir",
+        "--num_train_epochs=40",
+        "--per_device_train_batch_size=4",
+        "--per_device_eval_batch_size=4",
+        "--gradient_accumulation_steps=1",
+        "--precision=full_mixed",
+        "--matmul_precision=bfloat16",
+        "--multisteps",
+        "--learning_rate=2e-4",
+        "--warmup_steps=2000",
+        "--length_column_name=input_length",
+        "--evaluation_strategy=steps",
+        "--text_column_name=text",
+        "--save_steps=5000",
+        "--eval_steps=5000",
+        "--logging_steps=100",
+        "--layerdrop=0.041",
+        "--attention_dropout=0.094",
+        "--activation_dropout=0.055",
+        "--hidden_dropout=0.047",
+        "--save_total_limit=5",
+        "--freeze_feature_encoder",
+        "--feat_proj_dropout=0.04",
+        "--mask_time_prob=0.082",
+        "--mask_time_length=10",
+        "--mask_feature_prob=0.25",
+        "--mask_feature_length=64",
+        "--gradient_checkpointing",
+        "--min_duration_in_seconds=0.5",
+        "--max_duration_in_seconds=30.0",
+        "--use_auth_token",
+        "--seed=42",
+        "--group_by_length",
+        "--do_train",
+        "--do_eval",
+        "--push_to_hub",
+        "--preprocessing_num_workers=32",
+        "--ctc_zero_infinity",
+        "--do_lower_case",
+        "--wandb_project=wav2vec2",
+        "--wandb_name=wav2vec2-1b-npsc-nst-tpu",
+        "--remove_punctuation"
+    ],
+    "state": "running",
+    "program": "run_flax_speech_recognition_ctc.py",
+    "codePath": "run_flax_speech_recognition_ctc.py",
+    "git": {
+        "remote": "https://huggingface.co/NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "commit": "e2b1320cc68c3ce129a1d654965e0d3eb44e0558"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/wav2vec2-1b-npsc-nst",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220805_124834-3ep5xqhh/files/wandb-summary.json

@@ -0,0 +1 @@
+{"train/grad_norm": 7.09375, "layer_grad_norm/": {"lm_head": {"bias": 0.1953125, "kernel": 2.984375}, "wav2vec2": {"encoder": {"layer_norm": {"bias": 0.1513671875, "scale": 0.1123046875}, "layers": {"0": {"attention": {"k_proj": {"bias": 0.0001430511474609375, "kernel": 0.1162109375}, "out_proj": {"bias": 0.119140625, "kernel": 0.74609375}, "q_proj": {"bias": 0.015380859375, "kernel": 0.1611328125}, "v_proj": {"bias": 0.11328125, "kernel": 0.859375}}, "feed_forward": {"intermediate_dense": {"bias": 0.08544921875, "kernel": 1.296875}, "output_dense": {"bias": 0.026123046875, "kernel": 1.140625}}, "final_layer_norm": {"bias": 0.267578125, "scale": 0.4375}, "layer_norm": {"bias": 0.208984375, "scale": 0.330078125}}, "1": {"attention": {"k_proj": {"bias": 0.00010538101196289062, "kernel": 0.1025390625}, "out_proj": {"bias": 0.0284423828125, "kernel": 0.431640625}, "q_proj": {"bias": 0.008056640625, "kernel": 0.10400390625}, "v_proj": {"bias": 0.04248046875, "kernel": 0.31640625}}, "feed_forward": {"intermediate_dense": {"bias": 0.040283203125, "kernel": 0.6328125}, "output_dense": {"bias": 0.0263671875, "kernel": 0.55859375}}, "final_layer_norm": {"bias": 0.07666015625, "scale": 0.078125}, "layer_norm": {"bias": 0.07958984375, "scale": 0.111328125}}, "10": {"attention": {"k_proj": {"bias": 7.200241088867188e-05, "kernel": 0.185546875}, "out_proj": {"bias": 0.017333984375, "kernel": 0.30859375}, "q_proj": {"bias": 0.01123046875, "kernel": 0.203125}, "v_proj": {"bias": 0.0263671875, "kernel": 0.296875}}, "feed_forward": {"intermediate_dense": {"bias": 0.0281982421875, "kernel": 0.4921875}, "output_dense": {"bias": 0.01611328125, "kernel": 0.36328125}}, "final_layer_norm": {"bias": 0.0498046875, "scale": 0.047119140625}, "layer_norm": {"bias": 0.055419921875, "scale": 0.05712890625}}, "11": {"attention": {"k_proj": {"bias": 0.00011920928955078125, "kernel": 0.25}, "out_proj": {"bias": 0.0166015625, "kernel": 0.3984375}, "q_proj": {"bias": 0.01312255859375, "kernel": 0.232421875}, "v_proj": {"bias": 0.027099609375, "kernel": 0.390625}}, "feed_forward": {"intermediate_dense": {"bias": 0.025390625, "kernel": 0.451171875}, "output_dense": {"bias": 0.0159912109375, "kernel": 0.31640625}}, "final_layer_norm": {"bias": 0.04296875, "scale": 0.041015625}, "layer_norm": {"bias": 0.05322265625, "scale": 0.072265625}}, "12": {"attention": {"k_proj": {"bias": 6.866455078125e-05, "kernel": 0.181640625}, "out_proj": {"bias": 0.0162353515625, "kernel": 0.3046875}, "q_proj": {"bias": 0.00927734375, "kernel": 0.17578125}, "v_proj": {"bias": 0.02490234375, "kernel": 0.3046875}}, "feed_forward": {"intermediate_dense": {"bias": 0.026123046875, "kernel": 0.44921875}, "output_dense": {"bias": 0.015380859375, "kernel": 0.30859375}}, "final_layer_norm": {"bias": 0.043701171875, "scale": 0.04296875}, "layer_norm": {"bias": 0.044921875, "scale": 0.056396484375}}, "13": {"attention": {"k_proj": {"bias": 0.0001392364501953125, "kernel": 0.22265625}, "out_proj": {"bias": 0.0162353515625, "kernel": 0.392578125}, "q_proj": {"bias": 0.01123046875, "kernel": 0.216796875}, "v_proj": {"bias": 0.027099609375, "kernel": 0.4140625}}, "feed_forward": {"intermediate_dense": {"bias": 0.024658203125, "kernel": 0.421875}, "output_dense": {"bias": 0.0155029296875, "kernel": 0.30078125}}, "final_layer_norm": {"bias": 0.043212890625, "scale": 0.045166015625}, "layer_norm": {"bias": 0.04443359375, "scale": 0.07470703125}}, "14": {"attention": {"k_proj": {"bias": 0.0002117156982421875, "kernel": 0.146484375}, "out_proj": {"bias": 0.01556396484375, "kernel": 0.3359375}, "q_proj": {"bias": 0.0072021484375, "kernel": 0.142578125}, "v_proj": {"bias": 0.0238037109375, "kernel": 0.34765625}}, "feed_forward": {"intermediate_dense": {"bias": 0.024658203125, "kernel": 0.419921875}, "output_dense": {"bias": 0.01513671875, "kernel": 0.3046875}}, "final_layer_norm": {"bias": 0.04345703125, "scale": 0.04638671875}, "layer_norm": {"bias": 0.038330078125, "scale": 0.04052734375}}, "15": {"attention": {"k_proj": {"bias": 0.0002269744873046875, "kernel": 0.2421875}, "out_proj": {"bias": 0.014892578125, "kernel": 0.4140625}, "q_proj": {"bias": 0.01318359375, "kernel": 0.228515625}, "v_proj": {"bias": 0.024658203125, "kernel": 0.376953125}}, "feed_forward": {"intermediate_dense": {"bias": 0.021240234375, "kernel": 0.359375}, "output_dense": {"bias": 0.0146484375, "kernel": 0.283203125}}, "final_layer_norm": {"bias": 0.036376953125, "scale": 0.03466796875}, "layer_norm": {"bias": 0.0439453125, "scale": 0.0634765625}}, "16": {"attention": {"k_proj": {"bias": 9.34600830078125e-05, "kernel": 0.15234375}, "out_proj": {"bias": 0.01470947265625, "kernel": 0.283203125}, "q_proj": {"bias": 0.00775146484375, "kernel": 0.138671875}, "v_proj": {"bias": 0.021728515625, "kernel": 0.28125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0205078125, "kernel": 0.35546875}, "output_dense": {"bias": 0.0142822265625, "kernel": 0.2734375}}, "final_layer_norm": {"bias": 0.0341796875, "scale": 0.02978515625}, "layer_norm": {"bias": 0.037353515625, "scale": 0.03955078125}}, "17": {"attention": {"k_proj": {"bias": 5.435943603515625e-05, "kernel": 0.1240234375}, "out_proj": {"bias": 0.0152587890625, "kernel": 0.2265625}, "q_proj": {"bias": 0.00628662109375, "kernel": 0.1181640625}, "v_proj": {"bias": 0.021484375, "kernel": 0.23828125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0206298828125, "kernel": 0.353515625}, "output_dense": {"bias": 0.0146484375, "kernel": 0.271484375}}, "final_layer_norm": {"bias": 0.035400390625, "scale": 0.0308837890625}, "layer_norm": {"bias": 0.036865234375, "scale": 0.0400390625}}, "18": {"attention": {"k_proj": {"bias": 0.0001049041748046875, "kernel": 0.1767578125}, "out_proj": {"bias": 0.01470947265625, "kernel": 0.30078125}, "q_proj": {"bias": 0.00860595703125, "kernel": 0.158203125}, "v_proj": {"bias": 0.02197265625, "kernel": 0.28515625}}, "feed_forward": {"intermediate_dense": {"bias": 0.019775390625, "kernel": 0.33984375}, "output_dense": {"bias": 0.01409912109375, "kernel": 0.265625}}, "final_layer_norm": {"bias": 0.0341796875, "scale": 0.030517578125}, "layer_norm": {"bias": 0.0380859375, "scale": 0.0439453125}}, "19": {"attention": {"k_proj": {"bias": 6.580352783203125e-05, "kernel": 0.126953125}, "out_proj": {"bias": 0.01446533203125, "kernel": 0.2255859375}, "q_proj": {"bias": 0.006866455078125, "kernel": 0.1171875}, "v_proj": {"bias": 0.0201416015625, "kernel": 0.234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.018798828125, "kernel": 0.33203125}, "output_dense": {"bias": 0.01416015625, "kernel": 0.263671875}}, "final_layer_norm": {"bias": 0.0306396484375, "scale": 0.02734375}, "layer_norm": {"bias": 0.03173828125, "scale": 0.02734375}}, "2": {"attention": {"k_proj": {"bias": 0.00013065338134765625, "kernel": 0.15625}, "out_proj": {"bias": 0.0289306640625, "kernel": 0.48828125}, "q_proj": {"bias": 0.010986328125, "kernel": 0.154296875}, "v_proj": {"bias": 0.04638671875, "kernel": 0.421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.04296875, "kernel": 0.75390625}, "output_dense": {"bias": 0.02587890625, "kernel": 0.56640625}}, "final_layer_norm": {"bias": 0.0791015625, "scale": 0.076171875}, "layer_norm": {"bias": 0.07763671875, "scale": 0.154296875}}, "20": {"attention": {"k_proj": {"bias": 1.9788742065429688e-05, "kernel": 0.0703125}, "out_proj": {"bias": 0.014892578125, "kernel": 0.146484375}, "q_proj": {"bias": 0.003662109375, "kernel": 0.0673828125}, "v_proj": {"bias": 0.019775390625, "kernel": 0.162109375}}, "feed_forward": {"intermediate_dense": {"bias": 0.01953125, "kernel": 0.34765625}, "output_dense": {"bias": 0.01470947265625, "kernel": 0.26171875}}, "final_layer_norm": {"bias": 0.032470703125, "scale": 0.0296630859375}, "layer_norm": {"bias": 0.0308837890625, "scale": 0.0230712890625}}, "21": {"attention": {"k_proj": {"bias": 6.151199340820312e-05, "kernel": 0.10546875}, "out_proj": {"bias": 0.014892578125, "kernel": 0.224609375}, "q_proj": {"bias": 0.00518798828125, "kernel": 0.09521484375}, "v_proj": {"bias": 0.021240234375, "kernel": 0.24609375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0196533203125, "kernel": 0.353515625}, "output_dense": {"bias": 0.0147705078125, "kernel": 0.263671875}}, "final_layer_norm": {"bias": 0.03271484375, "scale": 0.03369140625}, "layer_norm": {"bias": 0.03076171875, "scale": 0.0322265625}}, "22": {"attention": {"k_proj": {"bias": 3.409385681152344e-05, "kernel": 0.1044921875}, "out_proj": {"bias": 0.015625, "kernel": 0.19921875}, "q_proj": {"bias": 0.00518798828125, "kernel": 0.10205078125}, "v_proj": {"bias": 0.021484375, "kernel": 0.20703125}}, "feed_forward": {"intermediate_dense": {"bias": 0.02001953125, "kernel": 0.353515625}, "output_dense": {"bias": 0.01544189453125, "kernel": 0.26171875}}, "final_layer_norm": {"bias": 0.033447265625, "scale": 0.03857421875}, "layer_norm": {"bias": 0.03515625, "scale": 0.04052734375}}, "23": {"attention": {"k_proj": {"bias": 0.0001430511474609375, "kernel": 0.158203125}, "out_proj": {"bias": 0.015869140625, "kernel": 0.34765625}, "q_proj": {"bias": 0.00762939453125, "kernel": 0.14453125}, "v_proj": {"bias": 0.02392578125, "kernel": 0.32421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.01953125, "kernel": 0.34375}, "output_dense": {"bias": 0.015625, "kernel": 0.25390625}}, "final_layer_norm": {"bias": 0.0322265625, "scale": 0.034912109375}, "layer_norm": {"bias": 0.03759765625, "scale": 0.033203125}}, "24": {"attention": {"k_proj": {"bias": 8.58306884765625e-05, "kernel": 0.150390625}, "out_proj": {"bias": 0.0147705078125, "kernel": 0.26953125}, "q_proj": {"bias": 0.0072021484375, "kernel": 0.1337890625}, "v_proj": {"bias": 0.0234375, "kernel": 0.27734375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0196533203125, "kernel": 0.353515625}, "output_dense": {"bias": 0.014404296875, "kernel": 0.240234375}}, "final_layer_norm": {"bias": 0.03466796875, "scale": 0.03466796875}, "layer_norm": {"bias": 0.041015625, "scale": 0.03369140625}}, "25": {"attention": {"k_proj": {"bias": 9.34600830078125e-05, "kernel": 0.109375}, "out_proj": {"bias": 0.0147705078125, "kernel": 0.251953125}, "q_proj": {"bias": 0.0057373046875, "kernel": 0.1044921875}, "v_proj": {"bias": 0.0224609375, "kernel": 0.26171875}}, "feed_forward": {"intermediate_dense": {"bias": 0.019287109375, "kernel": 0.35546875}, "output_dense": {"bias": 0.014404296875, "kernel": 0.236328125}}, "final_layer_norm": {"bias": 0.033203125, "scale": 0.034912109375}, "layer_norm": {"bias": 0.03466796875, "scale": 0.029296875}}, "26": {"attention": {"k_proj": {"bias": 6.29425048828125e-05, "kernel": 0.12451171875}, "out_proj": {"bias": 0.0145263671875, "kernel": 0.248046875}, "q_proj": {"bias": 0.0064697265625, "kernel": 0.12451171875}, "v_proj": {"bias": 0.022705078125, "kernel": 0.2578125}}, "feed_forward": {"intermediate_dense": {"bias": 0.018798828125, "kernel": 0.32421875}, "output_dense": {"bias": 0.01422119140625, "kernel": 0.23046875}}, "final_layer_norm": {"bias": 0.031982421875, "scale": 0.03173828125}, "layer_norm": {"bias": 0.03564453125, "scale": 0.0296630859375}}, "27": {"attention": {"k_proj": {"bias": 0.000133514404296875, "kernel": 0.14453125}, "out_proj": {"bias": 0.01348876953125, "kernel": 0.30859375}, "q_proj": {"bias": 0.0072021484375, "kernel": 0.1328125}, "v_proj": {"bias": 0.0224609375, "kernel": 0.306640625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0184326171875, "kernel": 0.30859375}, "output_dense": {"bias": 0.0135498046875, "kernel": 0.228515625}}, "final_layer_norm": {"bias": 0.0322265625, "scale": 0.0361328125}, "layer_norm": {"bias": 0.039306640625, "scale": 0.03369140625}}, "28": {"attention": {"k_proj": {"bias": 0.00012063980102539062, "kernel": 0.130859375}, "out_proj": {"bias": 0.0128173828125, "kernel": 0.291015625}, "q_proj": {"bias": 0.00640869140625, "kernel": 0.12890625}, "v_proj": {"bias": 0.0213623046875, "kernel": 0.298828125}}, "feed_forward": {"intermediate_dense": {"bias": 0.01708984375, "kernel": 0.29296875}, "output_dense": {"bias": 0.012939453125, "kernel": 0.21875}}, "final_layer_norm": {"bias": 0.0299072265625, "scale": 0.03125}, "layer_norm": {"bias": 0.0361328125, "scale": 0.044921875}}, "29": {"attention": {"k_proj": {"bias": 9.822845458984375e-05, "kernel": 0.134765625}, "out_proj": {"bias": 0.0123291015625, "kernel": 0.24609375}, "q_proj": {"bias": 0.006439208984375, "kernel": 0.1279296875}, "v_proj": {"bias": 0.019775390625, "kernel": 0.251953125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0174560546875, "kernel": 0.32421875}, "output_dense": {"bias": 0.0120849609375, "kernel": 0.2197265625}}, "final_layer_norm": {"bias": 0.028076171875, "scale": 0.0302734375}, "layer_norm": {"bias": 0.037841796875, "scale": 0.031982421875}}, "3": {"attention": {"k_proj": {"bias": 0.00019168853759765625, "kernel": 0.2353515625}, "out_proj": {"bias": 0.027099609375, "kernel": 0.62890625}, "q_proj": {"bias": 0.01446533203125, "kernel": 0.228515625}, "v_proj": {"bias": 0.04345703125, "kernel": 0.5546875}}, "feed_forward": {"intermediate_dense": {"bias": 0.04248046875, "kernel": 0.7421875}, "output_dense": {"bias": 0.02490234375, "kernel": 0.5390625}}, "final_layer_norm": {"bias": 0.078125, "scale": 0.0732421875}, "layer_norm": {"bias": 0.07177734375, "scale": 0.076171875}}, "30": {"attention": {"k_proj": {"bias": 0.0001049041748046875, "kernel": 0.15625}, "out_proj": {"bias": 0.0118408203125, "kernel": 0.275390625}, "q_proj": {"bias": 0.007781982421875, "kernel": 0.15625}, "v_proj": {"bias": 0.01953125, "kernel": 0.2890625}}, "feed_forward": {"intermediate_dense": {"bias": 0.016845703125, "kernel": 0.3203125}, "output_dense": {"bias": 0.0115966796875, "kernel": 0.203125}}, "final_layer_norm": {"bias": 0.028076171875, "scale": 0.02880859375}, "layer_norm": {"bias": 0.0303955078125, "scale": 0.042724609375}}, "31": {"attention": {"k_proj": {"bias": 0.00010251998901367188, "kernel": 0.14453125}, "out_proj": {"bias": 0.01104736328125, "kernel": 0.236328125}, "q_proj": {"bias": 0.00714111328125, "kernel": 0.1416015625}, "v_proj": {"bias": 0.0181884765625, "kernel": 0.25390625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0159912109375, "kernel": 0.3046875}, "output_dense": {"bias": 0.0107421875, "kernel": 0.193359375}}, "final_layer_norm": {"bias": 0.0263671875, "scale": 0.025146484375}, "layer_norm": {"bias": 0.030517578125, "scale": 0.04638671875}}, "32": {"attention": {"k_proj": {"bias": 0.0001010894775390625, "kernel": 0.13671875}, "out_proj": {"bias": 0.0103759765625, "kernel": 0.21484375}, "q_proj": {"bias": 0.006561279296875, "kernel": 0.13671875}, "v_proj": {"bias": 0.0166015625, "kernel": 0.244140625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0155029296875, "kernel": 0.30078125}, "output_dense": {"bias": 0.0096435546875, "kernel": 0.1806640625}}, "final_layer_norm": {"bias": 0.0283203125, "scale": 0.02734375}, "layer_norm": {"bias": 0.0281982421875, "scale": 0.0283203125}}, "33": {"attention": {"k_proj": {"bias": 0.00011444091796875, "kernel": 0.16015625}, "out_proj": {"bias": 0.0096435546875, "kernel": 0.2109375}, "q_proj": {"bias": 0.00775146484375, "kernel": 0.158203125}, "v_proj": {"bias": 0.0159912109375, "kernel": 0.232421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.014404296875, "kernel": 0.27734375}, "output_dense": {"bias": 0.0091552734375, "kernel": 0.1796875}}, "final_layer_norm": {"bias": 0.0274658203125, "scale": 0.028564453125}, "layer_norm": {"bias": 0.0263671875, "scale": 0.033447265625}}, "34": {"attention": {"k_proj": {"bias": 0.0001583099365234375, "kernel": 0.15234375}, "out_proj": {"bias": 0.0086669921875, "kernel": 0.21484375}, "q_proj": {"bias": 0.00689697265625, "kernel": 0.14453125}, "v_proj": {"bias": 0.014892578125, "kernel": 0.234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0133056640625, "kernel": 0.251953125}, "output_dense": {"bias": 0.0081787109375, "kernel": 0.171875}}, "final_layer_norm": {"bias": 0.025146484375, "scale": 0.031494140625}, "layer_norm": {"bias": 0.026123046875, "scale": 0.030517578125}}, "35": {"attention": {"k_proj": {"bias": 0.00021076202392578125, "kernel": 0.134765625}, "out_proj": {"bias": 0.007598876953125, "kernel": 0.21484375}, "q_proj": {"bias": 0.006011962890625, "kernel": 0.1318359375}, "v_proj": {"bias": 0.0113525390625, "kernel": 0.1982421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.01068115234375, "kernel": 0.201171875}, "output_dense": {"bias": 0.0074462890625, "kernel": 0.1494140625}}, "final_layer_norm": {"bias": 0.01953125, "scale": 0.0235595703125}, "layer_norm": {"bias": 0.0245361328125, "scale": 0.03076171875}}, "36": {"attention": {"k_proj": {"bias": 0.000156402587890625, "kernel": 0.126953125}, "out_proj": {"bias": 0.00714111328125, "kernel": 0.1630859375}, "q_proj": {"bias": 0.006072998046875, "kernel": 0.1279296875}, "v_proj": {"bias": 0.01019287109375, "kernel": 0.15234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.009765625, "kernel": 0.181640625}, "output_dense": {"bias": 0.006927490234375, "kernel": 0.1337890625}}, "final_layer_norm": {"bias": 0.017333984375, "scale": 0.0185546875}, "layer_norm": {"bias": 0.021240234375, "scale": 0.02001953125}}, "37": {"attention": {"k_proj": {"bias": 0.00011348724365234375, "kernel": 0.126953125}, "out_proj": {"bias": 0.006622314453125, "kernel": 0.1650390625}, "q_proj": {"bias": 0.0074462890625, "kernel": 0.1298828125}, "v_proj": {"bias": 0.010498046875, "kernel": 0.16015625}}, "feed_forward": {"intermediate_dense": {"bias": 0.009521484375, "kernel": 0.177734375}, "output_dense": {"bias": 0.0064697265625, "kernel": 0.126953125}}, "final_layer_norm": {"bias": 0.017333984375, "scale": 0.01611328125}, "layer_norm": {"bias": 0.02587890625, "scale": 0.03857421875}}, "38": {"attention": {"k_proj": {"bias": 0.00012683868408203125, "kernel": 0.1201171875}, "out_proj": {"bias": 0.0062255859375, "kernel": 0.14453125}, "q_proj": {"bias": 0.005462646484375, "kernel": 0.11474609375}, "v_proj": {"bias": 0.00982666015625, "kernel": 0.15234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.00885009765625, "kernel": 0.1611328125}, "output_dense": {"bias": 0.00604248046875, "kernel": 0.12060546875}}, "final_layer_norm": {"bias": 0.0166015625, "scale": 0.016845703125}, "layer_norm": {"bias": 0.0224609375, "scale": 0.021484375}}, "39": {"attention": {"k_proj": {"bias": 9.1552734375e-05, "kernel": 0.103515625}, "out_proj": {"bias": 0.00567626953125, "kernel": 0.150390625}, "q_proj": {"bias": 0.00439453125, "kernel": 0.1005859375}, "v_proj": {"bias": 0.0091552734375, "kernel": 0.1494140625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0078125, "kernel": 0.146484375}, "output_dense": {"bias": 0.0054931640625, "kernel": 0.119140625}}, "final_layer_norm": {"bias": 0.01531982421875, "scale": 0.02001953125}, "layer_norm": {"bias": 0.021484375, "scale": 0.02099609375}}, "4": {"attention": {"k_proj": {"bias": 0.0002593994140625, "kernel": 0.263671875}, "out_proj": {"bias": 0.0255126953125, "kernel": 0.65625}, "q_proj": {"bias": 0.01495361328125, "kernel": 0.255859375}, "v_proj": {"bias": 0.039306640625, "kernel": 0.58984375}}, "feed_forward": {"intermediate_dense": {"bias": 0.041015625, "kernel": 0.6875}, "output_dense": {"bias": 0.02392578125, "kernel": 0.51171875}}, "final_layer_norm": {"bias": 0.0712890625, "scale": 0.07177734375}, "layer_norm": {"bias": 0.06640625, "scale": 0.08837890625}}, "40": {"attention": {"k_proj": {"bias": 5.245208740234375e-05, "kernel": 0.0908203125}, "out_proj": {"bias": 0.005584716796875, "kernel": 0.12353515625}, "q_proj": {"bias": 0.00445556640625, "kernel": 0.0947265625}, "v_proj": {"bias": 0.00848388671875, "kernel": 0.1201171875}}, "feed_forward": {"intermediate_dense": {"bias": 0.0074462890625, "kernel": 0.1279296875}, "output_dense": {"bias": 0.005401611328125, "kernel": 0.1064453125}}, "final_layer_norm": {"bias": 0.0164794921875, "scale": 0.033203125}, "layer_norm": {"bias": 0.0196533203125, "scale": 0.01953125}}, "41": {"attention": {"k_proj": {"bias": 6.151199340820312e-05, "kernel": 0.076171875}, "out_proj": {"bias": 0.00537109375, "kernel": 0.154296875}, "q_proj": {"bias": 0.0035247802734375, "kernel": 0.0791015625}, "v_proj": {"bias": 0.0091552734375, "kernel": 0.162109375}}, "feed_forward": {"intermediate_dense": {"bias": 0.006561279296875, "kernel": 0.119140625}, "output_dense": {"bias": 0.00518798828125, "kernel": 0.10009765625}}, "final_layer_norm": {"bias": 0.01611328125, "scale": 0.0238037109375}, "layer_norm": {"bias": 0.0205078125, "scale": 0.027587890625}}, "42": {"attention": {"k_proj": {"bias": 1.3828277587890625e-05, "kernel": 0.030029296875}, "out_proj": {"bias": 0.00537109375, "kernel": 0.09033203125}, "q_proj": {"bias": 0.001678466796875, "kernel": 0.029541015625}, "v_proj": {"bias": 0.00665283203125, "kernel": 0.08349609375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0059814453125, "kernel": 0.109375}, "output_dense": {"bias": 0.005279541015625, "kernel": 0.0869140625}}, "final_layer_norm": {"bias": 0.0145263671875, "scale": 0.02001953125}, "layer_norm": {"bias": 0.013427734375, "scale": 0.0194091796875}}, "43": {"attention": {"k_proj": {"bias": 4.798173904418945e-06, "kernel": 0.017578125}, "out_proj": {"bias": 0.005401611328125, "kernel": 0.06103515625}, "q_proj": {"bias": 0.0010223388671875, "kernel": 0.017822265625}, "v_proj": {"bias": 0.006103515625, "kernel": 0.0615234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.005767822265625, "kernel": 0.1123046875}, "output_dense": {"bias": 0.005462646484375, "kernel": 0.0859375}}, "final_layer_norm": {"bias": 0.01220703125, "scale": 0.016845703125}, "layer_norm": {"bias": 0.01129150390625, "scale": 0.01507568359375}}, "44": {"attention": {"k_proj": {"bias": 5.8710575103759766e-06, "kernel": 0.0185546875}, "out_proj": {"bias": 0.005584716796875, "kernel": 0.0654296875}, "q_proj": {"bias": 0.00104522705078125, "kernel": 0.01904296875}, "v_proj": {"bias": 0.0062255859375, "kernel": 0.06494140625}}, "feed_forward": {"intermediate_dense": {"bias": 0.005279541015625, "kernel": 0.1142578125}, "output_dense": {"bias": 0.00567626953125, "kernel": 0.080078125}}, "final_layer_norm": {"bias": 0.0106201171875, "scale": 0.01275634765625}, "layer_norm": {"bias": 0.01318359375, "scale": 0.011962890625}}, "45": {"attention": {"k_proj": {"bias": 5.543231964111328e-06, "kernel": 0.017333984375}, "out_proj": {"bias": 0.005767822265625, "kernel": 0.06591796875}, "q_proj": {"bias": 0.001068115234375, "kernel": 0.0177001953125}, "v_proj": {"bias": 0.007110595703125, "kernel": 0.06982421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.00518798828125, "kernel": 0.111328125}, "output_dense": {"bias": 0.005889892578125, "kernel": 0.080078125}}, "final_layer_norm": {"bias": 0.01031494140625, "scale": 0.0118408203125}, "layer_norm": {"bias": 0.0174560546875, "scale": 0.01611328125}}, "46": {"attention": {"k_proj": {"bias": 8.702278137207031e-06, "kernel": 0.01904296875}, "out_proj": {"bias": 0.00604248046875, "kernel": 0.0693359375}, "q_proj": {"bias": 0.001190185546875, "kernel": 0.018310546875}, "v_proj": {"bias": 0.00830078125, "kernel": 0.0810546875}}, "feed_forward": {"intermediate_dense": {"bias": 0.00494384765625, "kernel": 0.095703125}, "output_dense": {"bias": 0.00616455078125, "kernel": 0.076171875}}, "final_layer_norm": {"bias": 0.01171875, "scale": 0.0203857421875}, "layer_norm": {"bias": 0.02294921875, "scale": 0.0159912109375}}, "47": {"attention": {"k_proj": {"bias": 1.2278556823730469e-05, "kernel": 0.020263671875}, "out_proj": {"bias": 0.00634765625, "kernel": 0.06689453125}, "q_proj": {"bias": 0.00138092041015625, "kernel": 0.018798828125}, "v_proj": {"bias": 0.0098876953125, "kernel": 0.08935546875}}, "feed_forward": {"intermediate_dense": {"bias": 0.00445556640625, "kernel": 0.0703125}, "output_dense": {"bias": 0.006561279296875, "kernel": 0.072265625}}, "final_layer_norm": {"bias": 0.01416015625, "scale": 0.0301513671875}, "layer_norm": {"bias": 0.0303955078125, "scale": 0.03271484375}}, "5": {"attention": {"k_proj": {"bias": 0.00010967254638671875, "kernel": 0.228515625}, "out_proj": {"bias": 0.025390625, "kernel": 0.42578125}, "q_proj": {"bias": 0.01287841796875, "kernel": 0.224609375}, "v_proj": {"bias": 0.0400390625, "kernel": 0.41796875}}, "feed_forward": {"intermediate_dense": {"bias": 0.0400390625, "kernel": 0.6484375}, "output_dense": {"bias": 0.02392578125, "kernel": 0.48046875}}, "final_layer_norm": {"bias": 0.0712890625, "scale": 0.0703125}, "layer_norm": {"bias": 0.068359375, "scale": 0.08984375}}, "6": {"attention": {"k_proj": {"bias": 0.00018310546875, "kernel": 0.3203125}, "out_proj": {"bias": 0.023193359375, "kernel": 0.5703125}, "q_proj": {"bias": 0.01806640625, "kernel": 0.296875}, "v_proj": {"bias": 0.0419921875, "kernel": 0.578125}}, "feed_forward": {"intermediate_dense": {"bias": 0.037109375, "kernel": 0.62890625}, "output_dense": {"bias": 0.0216064453125, "kernel": 0.4375}}, "final_layer_norm": {"bias": 0.06396484375, "scale": 0.06103515625}, "layer_norm": {"bias": 0.072265625, "scale": 0.095703125}}, "7": {"attention": {"k_proj": {"bias": 0.0001773834228515625, "kernel": 0.25}, "out_proj": {"bias": 0.0218505859375, "kernel": 0.515625}, "q_proj": {"bias": 0.013916015625, "kernel": 0.240234375}, "v_proj": {"bias": 0.03662109375, "kernel": 0.5078125}}, "feed_forward": {"intermediate_dense": {"bias": 0.035400390625, "kernel": 0.625}, "output_dense": {"bias": 0.02001953125, "kernel": 0.4296875}}, "final_layer_norm": {"bias": 0.061767578125, "scale": 0.06103515625}, "layer_norm": {"bias": 0.06689453125, "scale": 0.083984375}}, "8": {"attention": {"k_proj": {"bias": 0.000164031982421875, "kernel": 0.2353515625}, "out_proj": {"bias": 0.0201416015625, "kernel": 0.45703125}, "q_proj": {"bias": 0.01220703125, "kernel": 0.22265625}, "v_proj": {"bias": 0.0322265625, "kernel": 0.43359375}}, "feed_forward": {"intermediate_dense": {"bias": 0.032470703125, "kernel": 0.57421875}, "output_dense": {"bias": 0.0185546875, "kernel": 0.41015625}}, "final_layer_norm": {"bias": 0.0576171875, "scale": 0.06298828125}, "layer_norm": {"bias": 0.05712890625, "scale": 0.078125}}, "9": {"attention": {"k_proj": {"bias": 0.000255584716796875, "kernel": 0.29296875}, "out_proj": {"bias": 0.017822265625, "kernel": 0.5859375}, "q_proj": {"bias": 0.013916015625, "kernel": 0.27734375}, "v_proj": {"bias": 0.030029296875, "kernel": 0.58203125}}, "feed_forward": {"intermediate_dense": {"bias": 0.02685546875, "kernel": 0.482421875}, "output_dense": {"bias": 0.0172119140625, "kernel": 0.3671875}}, "final_layer_norm": {"bias": 0.043701171875, "scale": 0.04638671875}, "layer_norm": {"bias": 0.0556640625, "scale": 0.06689453125}}}, "pos_conv_embed": {"conv": {"bias": 0.1005859375, "weight_g": 0.09716796875, "weight_v": 0.75}}}, "feature_extractor": {"conv_layers": {"0": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "1": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "2": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "3": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "4": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "5": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "6": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}}}, "feature_projection": {"layer_norm": {"bias": 0.349609375, "scale": 0.431640625}, "projection": {"bias": 0.162109375, "kernel": 2.96875}}, "masked_spec_embed": 0.0}}, "layer_param_norm/": {"lm_head": {"bias": 0.068152517080307, "kernel": 5.452093124389648}, "wav2vec2": {"encoder": {"layer_norm": {"bias": 0.8908292055130005, "scale": 22.55011749267578}, "layers": {"0": {"attention": {"k_proj": {"bias": 0.2033224105834961, "kernel": 26.631282806396484}, "out_proj": {"bias": 1.5543559789657593, "kernel": 25.581758499145508}, "q_proj": {"bias": 1.3539581298828125, "kernel": 26.875282287597656}, "v_proj": {"bias": 0.3526538014411926, "kernel": 26.176233291625977}}, "feed_forward": {"intermediate_dense": {"bias": 1.820350170135498, "kernel": 96.90775299072266}, "output_dense": {"bias": 1.0441815853118896, "kernel": 92.76568603515625}}, "final_layer_norm": {"bias": 1.3226207494735718, "scale": 19.917068481445312}, "layer_norm": {"bias": 3.2080445289611816, "scale": 15.740410804748535}}, "1": {"attention": {"k_proj": {"bias": 0.20793935656547546, "kernel": 40.073883056640625}, "out_proj": {"bias": 1.325477957725525, "kernel": 42.14421081542969}, "q_proj": {"bias": 2.8881866931915283, "kernel": 39.97734069824219}, "v_proj": {"bias": 0.29305994510650635, "kernel": 40.59188461303711}}, "feed_forward": {"intermediate_dense": {"bias": 1.7098056077957153, "kernel": 94.91741943359375}, "output_dense": {"bias": 0.8041924238204956, "kernel": 85.20159912109375}}, "final_layer_norm": {"bias": 1.2720082998275757, "scale": 18.5848331451416}, "layer_norm": {"bias": 1.8610038757324219, "scale": 18.784591674804688}}, "10": {"attention": {"k_proj": {"bias": 0.23071043193340302, "kernel": 47.817657470703125}, "out_proj": {"bias": 1.260962724685669, "kernel": 50.55366516113281}, "q_proj": {"bias": 2.428488254547119, "kernel": 47.79916000366211}, "v_proj": {"bias": 0.31947579979896545, "kernel": 50.702056884765625}}, "feed_forward": {"intermediate_dense": {"bias": 1.7117397785186768, "kernel": 98.74652862548828}, "output_dense": {"bias": 0.5619156360626221, "kernel": 92.54859161376953}}, "final_layer_norm": {"bias": 2.2949581146240234, "scale": 20.386058807373047}, "layer_norm": {"bias": 1.7546437978744507, "scale": 22.054208755493164}}, "11": {"attention": {"k_proj": {"bias": 0.2500627934932709, "kernel": 47.698875427246094}, "out_proj": {"bias": 1.1115316152572632, "kernel": 49.86783981323242}, "q_proj": {"bias": 2.4702019691467285, "kernel": 47.44954299926758}, "v_proj": {"bias": 0.3839545249938965, "kernel": 50.3409423828125}}, "feed_forward": {"intermediate_dense": {"bias": 1.7618391513824463, "kernel": 99.50746154785156}, "output_dense": {"bias": 0.5432395935058594, "kernel": 94.19831085205078}}, "final_layer_norm": {"bias": 2.2942888736724854, "scale": 20.414934158325195}, "layer_norm": {"bias": 1.7346091270446777, "scale": 22.51268768310547}}, "12": {"attention": {"k_proj": {"bias": 0.23851683735847473, "kernel": 48.23259735107422}, "out_proj": {"bias": 1.0888965129852295, "kernel": 50.09956741333008}, "q_proj": {"bias": 2.3461594581604004, "kernel": 48.003665924072266}, "v_proj": {"bias": 0.3631010949611664, "kernel": 50.46704864501953}}, "feed_forward": {"intermediate_dense": {"bias": 1.803997278213501, "kernel": 100.34449768066406}, "output_dense": {"bias": 0.5307854413986206, "kernel": 95.76693725585938}}, "final_layer_norm": {"bias": 2.242570400238037, "scale": 20.37445640563965}, "layer_norm": {"bias": 1.8127485513687134, "scale": 22.940025329589844}}, "13": {"attention": {"k_proj": {"bias": 0.23810593783855438, "kernel": 50.11332702636719}, "out_proj": {"bias": 1.0852420330047607, "kernel": 49.833316802978516}, "q_proj": {"bias": 2.3189704418182373, "kernel": 49.995277404785156}, "v_proj": {"bias": 0.37377816438674927, "kernel": 49.975830078125}}, "feed_forward": {"intermediate_dense": {"bias": 1.856095552444458, "kernel": 100.97918701171875}, "output_dense": {"bias": 0.5486018657684326, "kernel": 96.21040344238281}}, "final_layer_norm": {"bias": 2.139707088470459, "scale": 20.528472900390625}, "layer_norm": {"bias": 1.9281730651855469, "scale": 23.274127960205078}}, "14": {"attention": {"k_proj": {"bias": 0.2723248600959778, "kernel": 50.209617614746094}, "out_proj": {"bias": 1.2504926919937134, "kernel": 48.13068771362305}, "q_proj": {"bias": 2.3632712364196777, "kernel": 50.29181671142578}, "v_proj": {"bias": 0.35957953333854675, "kernel": 47.67962646484375}}, "feed_forward": {"intermediate_dense": {"bias": 1.8947381973266602, "kernel": 101.58688354492188}, "output_dense": {"bias": 0.5726581811904907, "kernel": 97.57473754882812}}, "final_layer_norm": {"bias": 2.2654528617858887, "scale": 20.67388153076172}, "layer_norm": {"bias": 2.0739710330963135, "scale": 23.353639602661133}}, "15": {"attention": {"k_proj": {"bias": 0.2381843477487564, "kernel": 50.35422897338867}, "out_proj": {"bias": 1.3084888458251953, "kernel": 48.870052337646484}, "q_proj": {"bias": 2.5118086338043213, "kernel": 50.4517822265625}, "v_proj": {"bias": 0.4177994132041931, "kernel": 48.54165267944336}}, "feed_forward": {"intermediate_dense": {"bias": 1.90169095993042, "kernel": 101.39517211914062}, "output_dense": {"bias": 0.728958010673523, "kernel": 98.26675415039062}}, "final_layer_norm": {"bias": 2.2084929943084717, "scale": 20.776508331298828}, "layer_norm": {"bias": 2.2970352172851562, "scale": 23.787940979003906}}, "16": {"attention": {"k_proj": {"bias": 0.20325782895088196, "kernel": 50.171905517578125}, "out_proj": {"bias": 1.232241153717041, "kernel": 48.22490692138672}, "q_proj": {"bias": 2.600128173828125, "kernel": 50.07044982910156}, "v_proj": {"bias": 0.36339572072029114, "kernel": 47.90862274169922}}, "feed_forward": {"intermediate_dense": {"bias": 1.8926692008972168, "kernel": 102.01324462890625}, "output_dense": {"bias": 0.7524616718292236, "kernel": 99.08518981933594}}, "final_layer_norm": {"bias": 2.253352642059326, "scale": 21.251562118530273}, "layer_norm": {"bias": 2.222353458404541, "scale": 22.561655044555664}}, "17": {"attention": {"k_proj": {"bias": 0.19596010446548462, "kernel": 50.29098892211914}, "out_proj": {"bias": 1.1670058965682983, "kernel": 47.52079772949219}, "q_proj": {"bias": 2.6935129165649414, "kernel": 50.364601135253906}, "v_proj": {"bias": 0.4015589952468872, "kernel": 47.18798828125}}, "feed_forward": {"intermediate_dense": {"bias": 1.9025218486785889, "kernel": 103.06841278076172}, "output_dense": {"bias": 0.7684204578399658, "kernel": 99.4727783203125}}, "final_layer_norm": {"bias": 2.332897663116455, "scale": 21.796119689941406}, "layer_norm": {"bias": 2.1179938316345215, "scale": 21.90562629699707}}, "18": {"attention": {"k_proj": {"bias": 0.21306441724300385, "kernel": 50.75135803222656}, "out_proj": {"bias": 1.2695965766906738, "kernel": 48.55261993408203}, "q_proj": {"bias": 2.566439390182495, "kernel": 51.118408203125}, "v_proj": {"bias": 0.42802894115448, "kernel": 48.09583282470703}}, "feed_forward": {"intermediate_dense": {"bias": 1.9341740608215332, "kernel": 103.33280944824219}, "output_dense": {"bias": 0.885681688785553, "kernel": 101.10433959960938}}, "final_layer_norm": {"bias": 2.4191389083862305, "scale": 21.78841781616211}, "layer_norm": {"bias": 2.3001561164855957, "scale": 23.88214874267578}}, "19": {"attention": {"k_proj": {"bias": 0.19877497851848602, "kernel": 49.8792839050293}, "out_proj": {"bias": 1.2443280220031738, "kernel": 48.387168884277344}, "q_proj": {"bias": 2.83821439743042, "kernel": 50.26993179321289}, "v_proj": {"bias": 0.39237794280052185, "kernel": 47.64271545410156}}, "feed_forward": {"intermediate_dense": {"bias": 1.984041452407837, "kernel": 103.87921142578125}, "output_dense": {"bias": 0.9528074860572815, "kernel": 101.99952697753906}}, "final_layer_norm": {"bias": 2.3636670112609863, "scale": 22.150318145751953}, "layer_norm": {"bias": 2.209664821624756, "scale": 22.944355010986328}}, "2": {"attention": {"k_proj": {"bias": 0.22293740510940552, "kernel": 46.508689880371094}, "out_proj": {"bias": 1.2349822521209717, "kernel": 44.43158721923828}, "q_proj": {"bias": 3.102576732635498, "kernel": 46.24449920654297}, "v_proj": {"bias": 0.32414352893829346, "kernel": 44.35133361816406}}, "feed_forward": {"intermediate_dense": {"bias": 1.729475498199463, "kernel": 99.74760437011719}, "output_dense": {"bias": 0.6975946426391602, "kernel": 88.37123107910156}}, "final_layer_norm": {"bias": 1.5292998552322388, "scale": 21.104074478149414}, "layer_norm": {"bias": 1.7049822807312012, "scale": 21.620132446289062}}, "20": {"attention": {"k_proj": {"bias": 0.17897281050682068, "kernel": 49.828556060791016}, "out_proj": {"bias": 1.2720823287963867, "kernel": 47.74470520019531}, "q_proj": {"bias": 2.7510714530944824, "kernel": 50.551265716552734}, "v_proj": {"bias": 0.36270517110824585, "kernel": 46.669921875}}, "feed_forward": {"intermediate_dense": {"bias": 1.978695273399353, "kernel": 105.0912857055664}, "output_dense": {"bias": 1.0657472610473633, "kernel": 102.6424331665039}}, "final_layer_norm": {"bias": 2.3712844848632812, "scale": 23.086246490478516}, "layer_norm": {"bias": 2.184242010116577, "scale": 23.00891876220703}}, "21": {"attention": {"k_proj": {"bias": 0.1918519139289856, "kernel": 50.296329498291016}, "out_proj": {"bias": 1.3105319738388062, "kernel": 47.7496337890625}, "q_proj": {"bias": 2.719822883605957, "kernel": 51.077423095703125}, "v_proj": {"bias": 0.4093472957611084, "kernel": 46.88066101074219}}, "feed_forward": {"intermediate_dense": {"bias": 2.0147135257720947, "kernel": 105.262451171875}, "output_dense": {"bias": 1.1490479707717896, "kernel": 102.94585418701172}}, "final_layer_norm": {"bias": 2.3835391998291016, "scale": 22.749361038208008}, "layer_norm": {"bias": 2.236501455307007, "scale": 23.257896423339844}}, "22": {"attention": {"k_proj": {"bias": 0.20655860006809235, "kernel": 50.70651626586914}, "out_proj": {"bias": 1.236067533493042, "kernel": 47.31515884399414}, "q_proj": {"bias": 2.776611566543579, "kernel": 51.042152404785156}, "v_proj": {"bias": 0.359075129032135, "kernel": 47.19292449951172}}, "feed_forward": {"intermediate_dense": {"bias": 1.9531865119934082, "kernel": 105.61796569824219}, "output_dense": {"bias": 1.1664657592773438, "kernel": 102.2547607421875}}, "final_layer_norm": {"bias": 2.2776260375976562, "scale": 22.2520694732666}, "layer_norm": {"bias": 2.2378594875335693, "scale": 22.298477172851562}}, "23": {"attention": {"k_proj": {"bias": 0.2479013055562973, "kernel": 51.734161376953125}, "out_proj": {"bias": 1.3774523735046387, "kernel": 48.390968322753906}, "q_proj": {"bias": 2.6543586254119873, "kernel": 51.83258056640625}, "v_proj": {"bias": 0.5280159115791321, "kernel": 49.03071594238281}}, "feed_forward": {"intermediate_dense": {"bias": 1.9238386154174805, "kernel": 105.43772888183594}, "output_dense": {"bias": 1.142749547958374, "kernel": 103.04647064208984}}, "final_layer_norm": {"bias": 2.552396297454834, "scale": 22.24652862548828}, "layer_norm": {"bias": 2.7072105407714844, "scale": 23.703632354736328}}, "24": {"attention": {"k_proj": {"bias": 0.2229425311088562, "kernel": 50.472808837890625}, "out_proj": {"bias": 1.435978651046753, "kernel": 50.262542724609375}, "q_proj": {"bias": 2.7695631980895996, "kernel": 50.45745849609375}, "v_proj": {"bias": 0.48706698417663574, "kernel": 50.34023666381836}}, "feed_forward": {"intermediate_dense": {"bias": 2.035344123840332, "kernel": 104.8966293334961}, "output_dense": {"bias": 1.183034896850586, "kernel": 105.89578247070312}}, "final_layer_norm": {"bias": 2.6251282691955566, "scale": 22.29348373413086}, "layer_norm": {"bias": 2.4621849060058594, "scale": 23.277725219726562}}, "25": {"attention": {"k_proj": {"bias": 0.21065406501293182, "kernel": 50.781959533691406}, "out_proj": {"bias": 1.2442545890808105, "kernel": 48.13123321533203}, "q_proj": {"bias": 2.870298147201538, "kernel": 50.558074951171875}, "v_proj": {"bias": 0.5664586424827576, "kernel": 48.67639923095703}}, "feed_forward": {"intermediate_dense": {"bias": 1.9374840259552002, "kernel": 105.15123748779297}, "output_dense": {"bias": 1.0432946681976318, "kernel": 105.82373809814453}}, "final_layer_norm": {"bias": 2.360422134399414, "scale": 22.809398651123047}, "layer_norm": {"bias": 2.59881591796875, "scale": 22.279644012451172}}, "26": {"attention": {"k_proj": {"bias": 0.22183993458747864, "kernel": 51.04751205444336}, "out_proj": {"bias": 1.1648166179656982, "kernel": 48.92171096801758}, "q_proj": {"bias": 2.8587160110473633, "kernel": 50.80378723144531}, "v_proj": {"bias": 0.475290447473526, "kernel": 49.529388427734375}}, "feed_forward": {"intermediate_dense": {"bias": 2.041374683380127, "kernel": 104.55793762207031}, "output_dense": {"bias": 1.0075979232788086, "kernel": 102.93330383300781}}, "final_layer_norm": {"bias": 2.006486415863037, "scale": 21.653133392333984}, "layer_norm": {"bias": 2.478868246078491, "scale": 22.6640625}}, "27": {"attention": {"k_proj": {"bias": 0.43229615688323975, "kernel": 51.916725158691406}, "out_proj": {"bias": 1.4090557098388672, "kernel": 50.325904846191406}, "q_proj": {"bias": 2.6127448081970215, "kernel": 51.7978515625}, "v_proj": {"bias": 0.5855052471160889, "kernel": 50.790611267089844}}, "feed_forward": {"intermediate_dense": {"bias": 2.196718692779541, "kernel": 102.87921142578125}, "output_dense": {"bias": 0.8832440972328186, "kernel": 102.62045288085938}}, "final_layer_norm": {"bias": 2.280376434326172, "scale": 20.95566177368164}, "layer_norm": {"bias": 2.5605921745300293, "scale": 23.58169937133789}}, "28": {"attention": {"k_proj": {"bias": 0.45622092485427856, "kernel": 52.793487548828125}, "out_proj": {"bias": 1.4313125610351562, "kernel": 50.987464904785156}, "q_proj": {"bias": 2.7933201789855957, "kernel": 52.45018005371094}, "v_proj": {"bias": 0.465924471616745, "kernel": 51.342491149902344}}, "feed_forward": {"intermediate_dense": {"bias": 2.139850616455078, "kernel": 102.88671875}, "output_dense": {"bias": 0.7735686302185059, "kernel": 104.78712463378906}}, "final_layer_norm": {"bias": 2.202572822570801, "scale": 21.290332794189453}, "layer_norm": {"bias": 2.0484907627105713, "scale": 24.346450805664062}}, "29": {"attention": {"k_proj": {"bias": 0.21946273744106293, "kernel": 49.203609466552734}, "out_proj": {"bias": 1.4073214530944824, "kernel": 53.4610595703125}, "q_proj": {"bias": 2.729091167449951, "kernel": 49.02506637573242}, "v_proj": {"bias": 0.42470136284828186, "kernel": 53.37601089477539}}, "feed_forward": {"intermediate_dense": {"bias": 2.1310319900512695, "kernel": 103.58599090576172}, "output_dense": {"bias": 0.9014177322387695, "kernel": 109.07768249511719}}, "final_layer_norm": {"bias": 2.3953723907470703, "scale": 22.446285247802734}, "layer_norm": {"bias": 2.1818008422851562, "scale": 25.359933853149414}}, "3": {"attention": {"k_proj": {"bias": 0.26286524534225464, "kernel": 50.66347885131836}, "out_proj": {"bias": 1.402914047241211, "kernel": 47.055397033691406}, "q_proj": {"bias": 2.7446746826171875, "kernel": 50.88633728027344}, "v_proj": {"bias": 0.3097341060638428, "kernel": 47.41233825683594}}, "feed_forward": {"intermediate_dense": {"bias": 1.7387281656265259, "kernel": 101.31109619140625}, "output_dense": {"bias": 0.6584216356277466, "kernel": 91.24249267578125}}, "final_layer_norm": {"bias": 1.818251132965088, "scale": 21.199321746826172}, "layer_norm": {"bias": 1.888195514678955, "scale": 23.423542022705078}}, "30": {"attention": {"k_proj": {"bias": 0.3396183252334595, "kernel": 51.11002731323242}, "out_proj": {"bias": 1.2126681804656982, "kernel": 49.820674896240234}, "q_proj": {"bias": 2.8068060874938965, "kernel": 51.180885314941406}, "v_proj": {"bias": 0.4783928692340851, "kernel": 50.18727493286133}}, "feed_forward": {"intermediate_dense": {"bias": 2.06760835647583, "kernel": 104.07792663574219}, "output_dense": {"bias": 0.8458258509635925, "kernel": 108.11724090576172}}, "final_layer_norm": {"bias": 2.230673313140869, "scale": 23.58123779296875}, "layer_norm": {"bias": 2.303117036819458, "scale": 25.078384399414062}}, "31": {"attention": {"k_proj": {"bias": 0.4084402322769165, "kernel": 49.54199981689453}, "out_proj": {"bias": 1.1298390626907349, "kernel": 50.54130554199219}, "q_proj": {"bias": 2.605281352996826, "kernel": 49.6292724609375}, "v_proj": {"bias": 0.5218612551689148, "kernel": 50.690101623535156}}, "feed_forward": {"intermediate_dense": {"bias": 2.1498327255249023, "kernel": 102.74031829833984}, "output_dense": {"bias": 1.0199881792068481, "kernel": 105.45809936523438}}, "final_layer_norm": {"bias": 2.1171905994415283, "scale": 23.458181381225586}, "layer_norm": {"bias": 2.2760095596313477, "scale": 24.785846710205078}}, "32": {"attention": {"k_proj": {"bias": 0.29190176725387573, "kernel": 48.523231506347656}, "out_proj": {"bias": 1.1341676712036133, "kernel": 49.7735481262207}, "q_proj": {"bias": 2.8203587532043457, "kernel": 48.477691650390625}, "v_proj": {"bias": 0.39520299434661865, "kernel": 50.0748405456543}}, "feed_forward": {"intermediate_dense": {"bias": 2.0911455154418945, "kernel": 101.61611938476562}, "output_dense": {"bias": 1.0767910480499268, "kernel": 104.828369140625}}, "final_layer_norm": {"bias": 2.1094062328338623, "scale": 23.85711669921875}, "layer_norm": {"bias": 2.2725131511688232, "scale": 25.117982864379883}}, "33": {"attention": {"k_proj": {"bias": 0.30211469531059265, "kernel": 48.42399597167969}, "out_proj": {"bias": 1.1708109378814697, "kernel": 49.60016632080078}, "q_proj": {"bias": 2.959620952606201, "kernel": 48.389739990234375}, "v_proj": {"bias": 0.41753631830215454, "kernel": 49.881370544433594}}, "feed_forward": {"intermediate_dense": {"bias": 2.096595048904419, "kernel": 99.9796142578125}, "output_dense": {"bias": 1.0682170391082764, "kernel": 103.541259765625}}, "final_layer_norm": {"bias": 2.007560968399048, "scale": 23.61741828918457}, "layer_norm": {"bias": 2.442840814590454, "scale": 25.323089599609375}}, "34": {"attention": {"k_proj": {"bias": 0.3150138854980469, "kernel": 47.643341064453125}, "out_proj": {"bias": 1.423769235610962, "kernel": 51.03999328613281}, "q_proj": {"bias": 2.888617515563965, "kernel": 47.66726303100586}, "v_proj": {"bias": 0.38617265224456787, "kernel": 50.97981262207031}}, "feed_forward": {"intermediate_dense": {"bias": 2.187166213989258, "kernel": 98.79045104980469}, "output_dense": {"bias": 1.0054786205291748, "kernel": 102.78819274902344}}, "final_layer_norm": {"bias": 1.9427646398544312, "scale": 23.23785972595215}, "layer_norm": {"bias": 2.5134119987487793, "scale": 25.711688995361328}}, "35": {"attention": {"k_proj": {"bias": 0.40803784132003784, "kernel": 49.40678405761719}, "out_proj": {"bias": 1.349858045578003, "kernel": 49.86473846435547}, "q_proj": {"bias": 2.581186294555664, "kernel": 49.70142364501953}, "v_proj": {"bias": 0.4777841567993164, "kernel": 49.73640823364258}}, "feed_forward": {"intermediate_dense": {"bias": 2.2715296745300293, "kernel": 97.324462890625}, "output_dense": {"bias": 0.8976683616638184, "kernel": 101.4354248046875}}, "final_layer_norm": {"bias": 2.017836093902588, "scale": 23.347476959228516}, "layer_norm": {"bias": 2.329078435897827, "scale": 26.2806396484375}}, "36": {"attention": {"k_proj": {"bias": 0.2764836549758911, "kernel": 46.63423156738281}, "out_proj": {"bias": 1.38099205493927, "kernel": 51.14894104003906}, "q_proj": {"bias": 2.6797680854797363, "kernel": 46.574954986572266}, "v_proj": {"bias": 0.36416226625442505, "kernel": 51.36273956298828}}, "feed_forward": {"intermediate_dense": {"bias": 2.155928134918213, "kernel": 96.31394958496094}, "output_dense": {"bias": 0.9205336570739746, "kernel": 101.03855895996094}}, "final_layer_norm": {"bias": 1.7051949501037598, "scale": 23.84616470336914}, "layer_norm": {"bias": 2.04105806350708, "scale": 25.763103485107422}}, "37": {"attention": {"k_proj": {"bias": 0.5623797178268433, "kernel": 45.71692657470703}, "out_proj": {"bias": 1.6310793161392212, "kernel": 51.138118743896484}, "q_proj": {"bias": 2.395171880722046, "kernel": 45.76018524169922}, "v_proj": {"bias": 0.3581632673740387, "kernel": 51.03189468383789}}, "feed_forward": {"intermediate_dense": {"bias": 2.0483310222625732, "kernel": 95.48307800292969}, "output_dense": {"bias": 0.9227561354637146, "kernel": 100.72814178466797}}, "final_layer_norm": {"bias": 1.4903086423873901, "scale": 24.2398738861084}, "layer_norm": {"bias": 2.0101869106292725, "scale": 25.782228469848633}}, "38": {"attention": {"k_proj": {"bias": 0.6424748301506042, "kernel": 43.90583038330078}, "out_proj": {"bias": 1.3247406482696533, "kernel": 50.61767578125}, "q_proj": {"bias": 2.314286231994629, "kernel": 43.89386749267578}, "v_proj": {"bias": 0.40888553857803345, "kernel": 50.517494201660156}}, "feed_forward": {"intermediate_dense": {"bias": 1.9870116710662842, "kernel": 93.46450805664062}, "output_dense": {"bias": 0.906753420829773, "kernel": 98.9403076171875}}, "final_layer_norm": {"bias": 1.5285555124282837, "scale": 24.95897674560547}, "layer_norm": {"bias": 2.1943142414093018, "scale": 26.587631225585938}}, "39": {"attention": {"k_proj": {"bias": 0.6710178256034851, "kernel": 43.74102020263672}, "out_proj": {"bias": 1.6112127304077148, "kernel": 50.4604377746582}, "q_proj": {"bias": 2.1047635078430176, "kernel": 44.102073669433594}, "v_proj": {"bias": 0.39144566655158997, "kernel": 50.153480529785156}}, "feed_forward": {"intermediate_dense": {"bias": 1.9708259105682373, "kernel": 91.68063354492188}, "output_dense": {"bias": 0.9833765029907227, "kernel": 99.22885131835938}}, "final_layer_norm": {"bias": 1.6258282661437988, "scale": 25.595191955566406}, "layer_norm": {"bias": 2.1547136306762695, "scale": 27.168075561523438}}, "4": {"attention": {"k_proj": {"bias": 0.2694867253303528, "kernel": 53.26494598388672}, "out_proj": {"bias": 1.5952560901641846, "kernel": 48.508522033691406}, "q_proj": {"bias": 2.533445358276367, "kernel": 53.46202850341797}, "v_proj": {"bias": 0.35670536756515503, "kernel": 48.84002685546875}}, "feed_forward": {"intermediate_dense": {"bias": 1.7226250171661377, "kernel": 100.87029266357422}, "output_dense": {"bias": 0.8173469305038452, "kernel": 92.466064453125}}, "final_layer_norm": {"bias": 1.9236550331115723, "scale": 20.729034423828125}, "layer_norm": {"bias": 1.9882155656814575, "scale": 23.948692321777344}}, "40": {"attention": {"k_proj": {"bias": 0.6092542409896851, "kernel": 42.94449996948242}, "out_proj": {"bias": 1.5522677898406982, "kernel": 49.07390594482422}, "q_proj": {"bias": 2.02258038520813, "kernel": 43.671382904052734}, "v_proj": {"bias": 0.4459694027900696, "kernel": 48.660919189453125}}, "feed_forward": {"intermediate_dense": {"bias": 1.822402000427246, "kernel": 89.84906005859375}, "output_dense": {"bias": 1.0344607830047607, "kernel": 96.44818878173828}}, "final_layer_norm": {"bias": 1.7864675521850586, "scale": 24.866188049316406}, "layer_norm": {"bias": 2.107056140899658, "scale": 26.691997528076172}}, "41": {"attention": {"k_proj": {"bias": 1.6814589500427246, "kernel": 40.39122772216797}, "out_proj": {"bias": 1.3104321956634521, "kernel": 50.64059066772461}, "q_proj": {"bias": 1.7168387174606323, "kernel": 41.10041427612305}, "v_proj": {"bias": 0.39700430631637573, "kernel": 49.57951354980469}}, "feed_forward": {"intermediate_dense": {"bias": 1.9485447406768799, "kernel": 86.61335754394531}, "output_dense": {"bias": 1.0688600540161133, "kernel": 95.44161224365234}}, "final_layer_norm": {"bias": 2.2872514724731445, "scale": 28.325698852539062}, "layer_norm": {"bias": 2.095804452896118, "scale": 28.469776153564453}}, "42": {"attention": {"k_proj": {"bias": 0.8075435757637024, "kernel": 36.95627212524414}, "out_proj": {"bias": 1.3640947341918945, "kernel": 44.83252716064453}, "q_proj": {"bias": 1.5312429666519165, "kernel": 38.272701263427734}, "v_proj": {"bias": 0.6266118288040161, "kernel": 43.17291259765625}}, "feed_forward": {"intermediate_dense": {"bias": 1.7192597389221191, "kernel": 85.57884979248047}, "output_dense": {"bias": 1.1186156272888184, "kernel": 93.6015853881836}}, "final_layer_norm": {"bias": 1.9876140356063843, "scale": 29.61162757873535}, "layer_norm": {"bias": 1.5676913261413574, "scale": 27.305191040039062}}, "43": {"attention": {"k_proj": {"bias": 1.219111442565918, "kernel": 33.429054260253906}, "out_proj": {"bias": 1.350442886352539, "kernel": 41.2076416015625}, "q_proj": {"bias": 1.3615591526031494, "kernel": 34.21763610839844}, "v_proj": {"bias": 0.5368777513504028, "kernel": 39.10566711425781}}, "feed_forward": {"intermediate_dense": {"bias": 1.737384557723999, "kernel": 84.75312805175781}, "output_dense": {"bias": 0.8779321312904358, "kernel": 91.51695251464844}}, "final_layer_norm": {"bias": 1.9532383680343628, "scale": 31.82501220703125}, "layer_norm": {"bias": 1.704296588897705, "scale": 25.475486755371094}}, "44": {"attention": {"k_proj": {"bias": 2.4949541091918945, "kernel": 34.01667022705078}, "out_proj": {"bias": 1.1109570264816284, "kernel": 44.92375946044922}, "q_proj": {"bias": 1.3070473670959473, "kernel": 34.36231231689453}, "v_proj": {"bias": 0.38874804973602295, "kernel": 44.021461486816406}}, "feed_forward": {"intermediate_dense": {"bias": 1.812990665435791, "kernel": 83.68305969238281}, "output_dense": {"bias": 0.8185780644416809, "kernel": 89.15675354003906}}, "final_layer_norm": {"bias": 1.9702775478363037, "scale": 33.993751525878906}, "layer_norm": {"bias": 1.5985355377197266, "scale": 25.498332977294922}}, "45": {"attention": {"k_proj": {"bias": 2.0529022216796875, "kernel": 33.83161163330078}, "out_proj": {"bias": 0.9920564889907837, "kernel": 48.51708984375}, "q_proj": {"bias": 1.3887860774993896, "kernel": 33.97296905517578}, "v_proj": {"bias": 0.4346076250076294, "kernel": 48.68134307861328}}, "feed_forward": {"intermediate_dense": {"bias": 1.9257898330688477, "kernel": 80.37091827392578}, "output_dense": {"bias": 0.9466295838356018, "kernel": 84.56382751464844}}, "final_layer_norm": {"bias": 1.6998722553253174, "scale": 32.70838165283203}, "layer_norm": {"bias": 1.5237869024276733, "scale": 24.026718139648438}}, "46": {"attention": {"k_proj": {"bias": 1.5453729629516602, "kernel": 34.99570846557617}, "out_proj": {"bias": 0.7569196224212646, "kernel": 50.94432067871094}, "q_proj": {"bias": 1.5501949787139893, "kernel": 35.083892822265625}, "v_proj": {"bias": 0.37976646423339844, "kernel": 51.70716857910156}}, "feed_forward": {"intermediate_dense": {"bias": 1.9667569398880005, "kernel": 74.7667236328125}, "output_dense": {"bias": 1.100508451461792, "kernel": 74.91085815429688}}, "final_layer_norm": {"bias": 1.6546576023101807, "scale": 28.219646453857422}, "layer_norm": {"bias": 1.342303991317749, "scale": 22.941308975219727}}, "47": {"attention": {"k_proj": {"bias": 0.31063246726989746, "kernel": 37.36616897583008}, "out_proj": {"bias": 0.6397010087966919, "kernel": 45.241363525390625}, "q_proj": {"bias": 1.6792919635772705, "kernel": 37.94398880004883}, "v_proj": {"bias": 0.3538339138031006, "kernel": 46.22050476074219}}, "feed_forward": {"intermediate_dense": {"bias": 2.031810760498047, "kernel": 72.15599060058594}, "output_dense": {"bias": 0.6105413436889648, "kernel": 68.52323150634766}}, "final_layer_norm": {"bias": 1.5555658340454102, "scale": 23.038026809692383}, "layer_norm": {"bias": 1.0648114681243896, "scale": 20.196226119995117}}, "5": {"attention": {"k_proj": {"bias": 0.23068922758102417, "kernel": 48.48977279663086}, "out_proj": {"bias": 1.5695205926895142, "kernel": 49.58379364013672}, "q_proj": {"bias": 2.636868953704834, "kernel": 48.62889862060547}, "v_proj": {"bias": 0.3259868621826172, "kernel": 50.311851501464844}}, "feed_forward": {"intermediate_dense": {"bias": 1.649585485458374, "kernel": 100.93144226074219}, "output_dense": {"bias": 0.8427544832229614, "kernel": 91.76881408691406}}, "final_layer_norm": {"bias": 2.1725611686706543, "scale": 20.89813232421875}, "layer_norm": {"bias": 2.0255513191223145, "scale": 23.038808822631836}}, "6": {"attention": {"k_proj": {"bias": 0.3303305506706238, "kernel": 50.40907287597656}, "out_proj": {"bias": 1.554028868675232, "kernel": 49.0322265625}, "q_proj": {"bias": 2.703887701034546, "kernel": 50.87598419189453}, "v_proj": {"bias": 0.32478368282318115, "kernel": 49.51765441894531}}, "feed_forward": {"intermediate_dense": {"bias": 1.620659351348877, "kernel": 100.0206069946289}, "output_dense": {"bias": 0.7007467150688171, "kernel": 91.39401245117188}}, "final_layer_norm": {"bias": 2.4702341556549072, "scale": 20.376611709594727}, "layer_norm": {"bias": 1.9957983493804932, "scale": 23.686267852783203}}, "7": {"attention": {"k_proj": {"bias": 0.3042152523994446, "kernel": 49.961692810058594}, "out_proj": {"bias": 1.3766974210739136, "kernel": 49.27983856201172}, "q_proj": {"bias": 2.458479881286621, "kernel": 50.367801666259766}, "v_proj": {"bias": 0.4104377031326294, "kernel": 49.22698211669922}}, "feed_forward": {"intermediate_dense": {"bias": 1.627145767211914, "kernel": 99.72712707519531}, "output_dense": {"bias": 0.5415225625038147, "kernel": 91.11566162109375}}, "final_layer_norm": {"bias": 2.324662685394287, "scale": 20.59018325805664}, "layer_norm": {"bias": 1.9219211339950562, "scale": 22.409427642822266}}, "8": {"attention": {"k_proj": {"bias": 0.292106032371521, "kernel": 49.483314514160156}, "out_proj": {"bias": 1.1988368034362793, "kernel": 49.76782989501953}, "q_proj": {"bias": 2.43380069732666, "kernel": 49.27770233154297}, "v_proj": {"bias": 0.336185097694397, "kernel": 49.90403747558594}}, "feed_forward": {"intermediate_dense": {"bias": 1.6834964752197266, "kernel": 99.28549194335938}, "output_dense": {"bias": 0.5039160251617432, "kernel": 90.5440902709961}}, "final_layer_norm": {"bias": 2.276369571685791, "scale": 20.358753204345703}, "layer_norm": {"bias": 1.8604187965393066, "scale": 22.826251983642578}}, "9": {"attention": {"k_proj": {"bias": 0.3227398991584778, "kernel": 50.32160186767578}, "out_proj": {"bias": 1.4219694137573242, "kernel": 50.671417236328125}, "q_proj": {"bias": 2.3705806732177734, "kernel": 50.517311096191406}, "v_proj": {"bias": 0.3579637110233307, "kernel": 51.04570770263672}}, "feed_forward": {"intermediate_dense": {"bias": 1.76076340675354, "kernel": 97.91405487060547}, "output_dense": {"bias": 0.6380627155303955, "kernel": 90.99038696289062}}, "final_layer_norm": {"bias": 2.2011656761169434, "scale": 19.671579360961914}, "layer_norm": {"bias": 1.946717381477356, "scale": 24.423818588256836}}}, "pos_conv_embed": {"conv": {"bias": 5.6507768630981445, "weight_g": 8.974504470825195, "weight_v": 85.95291137695312}}}, "feature_extractor": {"conv_layers": {"0": {"conv": {"bias": 2.0290679931640625, "kernel": 20.55536460876465}, "layer_norm": {"bias": 4.550922393798828, "scale": 16.167570114135742}}, "1": {"conv": {"bias": 1.7790228128433228, "kernel": 51.24136734008789}, "layer_norm": {"bias": 5.962646961212158, "scale": 23.268157958984375}}, "2": {"conv": {"bias": 1.140576720237732, "kernel": 46.50312042236328}, "layer_norm": {"bias": 4.176670551300049, "scale": 20.370853424072266}}, "3": {"conv": {"bias": 0.6725863218307495, "kernel": 44.397525787353516}, "layer_norm": {"bias": 3.888174533843994, "scale": 17.53795051574707}}, "4": {"conv": {"bias": 0.6373162269592285, "kernel": 41.314056396484375}, "layer_norm": {"bias": 2.385471820831299, "scale": 16.34571647644043}}, "5": {"conv": {"bias": 0.5147221684455872, "kernel": 37.479759216308594}, "layer_norm": {"bias": 2.020900011062622, "scale": 17.064470291137695}}, "6": {"conv": {"bias": 0.4947893023490906, "kernel": 40.64780044555664}, "layer_norm": {"bias": 0.5876954793930054, "scale": 19.058603286743164}}}}, "feature_projection": {"layer_norm": {"bias": 6.348620891571045, "scale": 16.567195892333984}, "projection": {"bias": 1.7358717918395996, "kernel": 35.4318733215332}}, "masked_spec_embed": 11.914372444152832}}, "train/learning_rate": 0.0001997361978283152, "train/loss": 1.570224642753601, "train/param_norm": 1196.265869140625, "_runtime": 7703, "_timestamp": 1659711417, "_step": 2500, "_wandb": {"runtime": 7705}}

wandb/run-20220805_124834-3ep5xqhh/logs/debug-internal.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3da4ff9acc37b96e42cab317a91c67a7f027a3b091f75199a9bb88d1da66494
+size 488530

wandb/run-20220805_124834-3ep5xqhh/logs/debug.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48fa6c02e11489989284e294f2bdd85e2a91addb6c39bc72e5836b6a56820370
+size 6061

wandb/run-20220805_124834-3ep5xqhh/run-3ep5xqhh.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d8246b0d8a9370c41447785409866c16f57fb082173b041d6c86503a87ba094
+size 2319945

wandb/run-20220805_152536-2fzkf8n5/files/code/run_flax_speech_recognition_ctc.py

@@ -0,0 +1,1631 @@
+#!/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,
+    set_seed,
+)
+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."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    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."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.1,
+        metadata={
+            "help": "The spec aug dropout probability for feature encoder representations."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@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": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    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=2,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+            "than this will be filtered."
+        },
+    )
+    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."}
+    )
+    skip_steps: Optional[int] = field(
+        default=0,
+        metadata={
+            "help": "Skip this number of steps. Useful to continue 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_util.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def to_bf16(t):
+    return jax.tree_util.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_util.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, num_log=50, 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))]
+        # we'll log the first 50 predictions for each epoch
+        wandb.log(
+            {
+                f"{prefix}/step_{int(step / 1000)}k": wandb.Table(
+                    columns=["label_str", "pred_str"], data=str_data[:num_log]
+                )
+            },
+            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 make_dataset(data_args, seed=42):
+    # Pre-processing dataset
+    import re
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text}
+
+    def filter_nst(entry):
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        return True
+
+    def filter_npsc(entry):
+        # False if there are digits in the text
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    # TODO NST_hesitate
+
+    split = len(npsc[data_args.train_split_name]) / (len(npsc[data_args.train_split_name]) + len(npsc[data_args.eval_split_name]))  # Use same train/val ratio as NPSC
+    nst_train = nst[data_args.train_split_name].train_test_split(train_size=split, seed=seed)
+    nst[data_args.train_split_name] = nst_train["train"]
+    nst[data_args.eval_split_name] = nst_train["test"]
+
+    nst = nst.filter(filter_nst).map(
+        map_nst,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NST",
+    ).shuffle(seed=seed)
+    npsc = npsc.filter(filter_npsc).map(
+        map_npsc,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NPSC",
+    ).shuffle(seed=seed)
+
+    npsc_base = npsc.remove_columns([col for col in npsc[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+    nst_base = nst.remove_columns([col for col in nst[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in data_args.train_split_name, data_args.eval_split_name, data_args.test_split_name:
+        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+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
+
+    set_seed(training_args.seed)
+    raw_datasets = make_dataset(data_args, seed=training_args.seed)
+   
+    # raw_datasets = DatasetDict()
+
+    # if training_args.do_train:
+    #     raw_datasets[data_args.train_split_name] = 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[data_args.eval_split_name] = 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(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": tokenizer.vocab_size,  # len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    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
+
+    try:
+        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,
+        )
+    except:
+        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,
+            from_pt=True,
+        )
+
+    # 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_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
+    chars_to_ignore = ', ? . ! - ; : " “ % ‘ ” ?'.split(" ")
+    chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
+    # gigaspeech_punctuation = {" <comma>": ",", " <period>": ".", " <questionmark>": "?", " <exclamationpoint>": "!"}
+    # gigaspeech_disfluencies = ["<other>", "<sil>"]
+    # swb_disfluencies = ["[noise]", "[laughter]", "[silence]", "<a_aside>", "<b_aside>", "<e_aside>", "[laughter-",
+    #                 "[vocalized-noise]", "_1"]
+    # swb_punctuations = ["{", "}", "[", "]-", "]"]
+    # earnings_disfluencies = ["<crosstalk>", "<affirmative>", "<inaudible>", "inaudible", "<laugh>", "<unk>"]
+    ignore_segments = ["ignore_time_segment_in_scoring", "<noise>", "<music>", "[noise]", "[laughter]", "[silence]",
+                       "[vocalized-noise]", "<crosstalk>", "<affirmative>", "<inaudible>", "<laugh>", "<other>", "<sil>", ""]
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets[data_args.eval_split_name] = raw_datasets[data_args.eval_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_predict and data_args.max_test_samples is not None:
+        raw_datasets[data_args.test_split_name] = raw_datasets[data_args.test_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_train and data_args.remove_punctuation:
+
+        def remove_punctuation(batch):
+            batch[text_column_name] = (
+                re.sub(chars_to_ignore_regex, "", batch[text_column_name]).replace("'", "").replace('"', "")
+            )
+
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].map(
+            remove_punctuation,
+            num_proc=data_args.preprocessing_num_workers,
+            desc="removing punctuation from train split",
+        )
+
+    # 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):
+        # process audio
+        try:
+            sample = batch[audio_column_name]
+        except ValueError:
+            sample = {"array": np.array([0.]), "sampling_rate": feature_extractor.sampling_rate}
+        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"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+
+        # if dataset_name == "google/xtreme_s":
+        #     # Finally, we tokenize the processed text
+        #     batch["labels"] = tokenizer(input_str).input_ids
+        #     batch["labels_length"] = len(batch["labels"])
+        #     return batch
+
+        # # Common Voice 9
+        # 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]
+        # # normalize quotation marks
+        # input_str = re.sub(r'["“”]', '"', input_str)
+        # # normalize apostrophes
+        # input_str = re.sub(r"[’']", "'", input_str)
+        # # normalize hyphens
+        # input_str = re.sub(r"[—–]", "-", input_str)
+        # # replace double quotation marks with single
+        # input_str = input_str.replace('""', '"')
+        # if dataset_name == "mozilla-foundation/common_voice_9_0" and len(input_str):
+        #     # for CV9, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # TEDLIUM-3
+        # # delete the <unk> token from the text and replace spaced apostrophes with un-spaced
+        # input_str = input_str.replace("<unk>", "").replace(" '", "'")
+
+        # # 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)
+        # if dataset_name == "speechcolab/gigaspeech" and len(input_str):
+        #     # for GS, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # SWB
+        # for disfluency in swb_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # remove parenthesised text (test data only)
+        # input_str = re.sub("[\(].*?[\)]", "", input_str)
+        # for punctuation in swb_punctuations:
+        #     input_str = input_str.replace(punctuation, "")
+        # # replace anomalous words with their correct transcriptions
+        # 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]])
+
+        # # Earnings 22
+        # for disfluency in earnings_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # replace mal-formatted ellipsis
+        # input_str = input_str.replace("…", ".")
+
+        # JIWER compliance
+        # 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 data with inputs shorter than min_input_length or longer than max_input_length
+    def is_audio_in_length_range(length):
+        return length > min_input_length and length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.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 length > min_target_length  # and length < max_target_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+        is_labels_in_length_range,
+        num_proc=num_workers,
+        input_columns=["labels_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[data_args.train_split_name])
+        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_util.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_util.tree_map(jnp.add, accum_grad, grad), loss
+
+            # create an initial state for accumulating losses, num labels and gradients
+            init_grad = jax.tree_util.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_util.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_util.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:
+        p_eval_step = jax.pmap(eval_step, "batch")
+
+    def run_evaluation(step):
+        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[data_args.eval_split_name], 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[data_args.eval_split_name][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_util.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)
+            # 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_util.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)
+
+    skip_epochs = data_args.skip_steps // (num_train_samples // batch_size_per_update)
+    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}")
+    logger.info(f"  Skipping: {data_args.skip_steps} steps ({skip_epochs} epochs)")
+
+    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()
+
+            if epoch < skip_epochs:
+                logger.info(f"Skipping epoch {epoch + 1}")
+                continue
+
+            # 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[data_args.train_split_name], batch_size_per_update, input_rng)
+            train_batch_idx = generate_batch_splits(train_samples_idx, batch_size_per_update)
+
+            if data_args.skip_steps > cur_step:
+                logger.info(f"Skipping {data_args.skip_steps - (epoch * (num_train_samples // batch_size_per_update))} steps...")
+            # 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):
+                cur_step = epoch * (num_train_samples // batch_size_per_update) + step
+                if cur_step <= data_args.skip_steps:
+                    continue
+
+                samples = [vectorized_datasets[data_args.train_split_name][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)
+
+
+                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=data_args.train_split_name)
+                    # 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)
+
+                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)
+                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)
+
+    # TODO: collapse 'do_predict' into the run_evaluation function
+    if training_args.do_predict:
+        for split in [data_args.test_split_name]:
+            # ======================== 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_util.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, 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()

wandb/run-20220805_152536-2fzkf8n5/files/config.yaml

@@ -0,0 +1,33 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_flax_speech_recognition_ctc.py
+    framework: huggingface
+    huggingface_version: 4.21.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1659713136
+    t:
+      1:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      3:
+      - 13
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.21.0
+      8:
+      - 5

wandb/run-20220805_152536-2fzkf8n5/files/diff.patch

@@ -0,0 +1,119 @@
+diff --git a/config.json b/config.json
+index 260219f..246b797 100644
+--- a/config.json
++++ b/config.json
+@@ -5,7 +5,7 @@
+   "add_adapter": false,
+   "apply_spec_augment": true,
+   "architectures": [
+-    "Wav2Vec2ForCTC"
++    "Wav2Vec2ForPreTraining"
+   ],
+   "attention_dropout": 0.094,
+   "bos_token_id": 1,
+diff --git a/run.sh b/run.sh
+index 9cc498e..d5f5166 100755
+--- a/run.sh
++++ b/run.sh
+@@ -1,13 +1,13 @@
+ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_ctc.py \
+         --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+-        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst" \
++        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst-tpu" \
+         --tokenizer_name="./" \
+         --output_dir="./" \
+         --overwrite_output_dir \
+         --num_train_epochs="40" \
+         --per_device_train_batch_size="2" \
+         --per_device_eval_batch_size="2" \
+-        --gradient_accumulation_steps="1" \
++        --gradient_accumulation_steps="2" \
+         --precision="full_mixed" \
+         --matmul_precision="bfloat16" \
+         --multisteps \
+@@ -16,8 +16,8 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --length_column_name="input_length" \
+         --evaluation_strategy="steps" \
+         --text_column_name="text" \
+-        --save_steps="4000" \
+-        --eval_steps="4000" \
++        --save_steps="5000" \
++        --eval_steps="5000" \
+         --logging_steps="100" \
+         --layerdrop="0.041" \
+         --attention_dropout="0.094" \
+@@ -42,7 +42,7 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --ctc_zero_infinity \
+         --do_lower_case \
+         --wandb_project="wav2vec2" \
+-        --wandb_name="wav2vec2-1b-npsc-nst" \
++        --wandb_name="wav2vec2-1b-npsc-nst-tpu" \
+         --remove_punctuation
+ 
+ 
+diff --git a/special_tokens_map.json b/special_tokens_map.json
+index 89389bf..81394dd 100644
+--- a/special_tokens_map.json
++++ b/special_tokens_map.json
+@@ -343,6 +343,34 @@
+       "rstrip": false,
+       "single_word": false
+     },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
+     {
+       "content": "</s>",
+       "lstrip": false,
+diff --git a/wandb/debug-internal.log b/wandb/debug-internal.log
+index 50a0b69..6eb3065 120000
+--- a/wandb/debug-internal.log
++++ b/wandb/debug-internal.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug-internal.log
+
++run-20220805_152536-2fzkf8n5/logs/debug-internal.log
+
+diff --git a/wandb/debug.log b/wandb/debug.log
+index 746223d..31f5db8 120000
+--- a/wandb/debug.log
++++ b/wandb/debug.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug.log
+
++run-20220805_152536-2fzkf8n5/logs/debug.log
+
+diff --git a/wandb/latest-run b/wandb/latest-run
+index be58b40..d408175 120000
+--- a/wandb/latest-run
++++ b/wandb/latest-run
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z
+
++run-20220805_152536-2fzkf8n5
+

wandb/run-20220805_152536-2fzkf8n5/files/output.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40ff34b9c765c40e17b2e182bf4b3c8d4c775695283737622845e6848a90e685
+size 231583

wandb/run-20220805_152536-2fzkf8n5/files/requirements.txt

@@ -0,0 +1,158 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+appdirs==1.4.4
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+audioread==2.1.9
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+cffi==1.15.1
+charset-normalizer==2.0.10
+chex==0.1.3
+click==8.0.3
+cloud-tpu-client==0.10
+cloud-tpu-profiler==2.4.0
+clu==0.0.6
+colorama==0.4.5
+commonmark==0.9.1
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==2.4.0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+etils==0.6.0
+exceptiongroup==1.0.0rc8
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.5.3
+fonttools==4.28.5
+frozenlist==1.2.0
+fsspec==2021.11.1
+future==0.18.2
+gast==0.4.0
+gitdb==4.0.9
+gitpython==3.1.26
+google-api-core==1.31.5
+google-api-python-client==1.8.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.4.6
+google-auth==2.3.3
+google-pasta==0.2.0
+googleapis-common-protos==1.54.0
+grpcio==1.43.0
+h5py==3.6.0
+httplib2==0.20.2
+huggingface-hub==0.2.1
+hypothesis==6.53.0
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.3.15
+jaxlib==0.3.15
+jedi==0.18.1
+jiwer==2.3.0
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+librosa==0.9.2
+libtpu-nightly==0.1.dev20220722
+llvmlite==0.39.0
+markdown==3.3.6
+matplotlib-inline==0.1.3
+matplotlib==3.5.1
+ml-collections==0.1.0
+msgpack==1.0.3
+multidict==5.2.0
+multiprocess==0.70.12.2
+numba==0.56.0
+numpy==1.22.0
+oauth2client==4.1.3
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.3
+packaging==21.3
+pandas==1.3.5
+parso==0.8.3
+pathtools==0.1.2
+pexpect==4.8.0
+pickleshare==0.7.5
+pillow==9.0.0
+pip==22.2.1
+pkg-resources==0.0.0
+pooch==1.6.0
+promise==2.3
+prompt-toolkit==3.0.24
+protobuf==3.19.1
+psutil==5.9.0
+ptyprocess==0.7.0
+pyarrow==6.0.1
+pyasn1-modules==0.2.8
+pyasn1==0.4.8
+pycparser==2.21
+pyctcdecode==0.4.0
+pygments==2.11.1
+pygtrie==2.5.0
+pyparsing==3.0.6
+python-dateutil==2.8.2
+python-levenshtein==0.12.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+resampy==0.3.1
+responses==0.18.0
+rich==11.2.0
+rsa==4.8
+sacremoses==0.0.46
+scikit-learn==1.1.1
+scipy==1.7.3
+sentry-sdk==1.5.2
+setuptools==44.0.0
+shortuuid==1.0.8
+six==1.16.0
+smmap==5.0.0
+sortedcontainers==2.4.0
+soundfile==0.10.3.post1
+sox==1.4.1
+subprocess32==3.5.4
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+tensorboard==2.7.0
+tensorflow-cpu==2.7.0
+tensorflow-datasets==4.4.0
+tensorflow-estimator==2.7.0
+tensorflow-io-gcs-filesystem==0.23.1
+tensorflow-metadata==1.5.0
+tensorflow==2.7.0
+tensorstore==0.1.21
+termcolor==1.1.0
+threadpoolctl==3.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+torch==1.12.0
+torchaudio==0.12.0+cpu
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.21.0
+typing-extensions==4.3.0
+uritemplate==3.0.1
+urllib3==1.26.7
+wandb==0.12.9
+wcwidth==0.2.5
+werkzeug==2.0.2
+wheel==0.37.1
+wrapt==1.13.3
+xxhash==2.0.2
+yarl==1.7.2
+yaspin==2.1.0
+zipp==3.7.0

wandb/run-20220805_152536-2fzkf8n5/files/wandb-metadata.json

@@ -0,0 +1,69 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-08-05T15:25:40.370466",
+    "startedAt": "2022-08-05T15:25:36.782402",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--model_name_or_path=facebook/wav2vec2-xls-r-1b",
+        "--hub_model_id=NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "--tokenizer_name=./",
+        "--output_dir=./",
+        "--overwrite_output_dir",
+        "--num_train_epochs=40",
+        "--per_device_train_batch_size=2",
+        "--per_device_eval_batch_size=2",
+        "--gradient_accumulation_steps=2",
+        "--precision=full_mixed",
+        "--matmul_precision=bfloat16",
+        "--multisteps",
+        "--learning_rate=2e-4",
+        "--warmup_steps=2000",
+        "--length_column_name=input_length",
+        "--evaluation_strategy=steps",
+        "--text_column_name=text",
+        "--save_steps=5000",
+        "--eval_steps=5000",
+        "--logging_steps=100",
+        "--layerdrop=0.041",
+        "--attention_dropout=0.094",
+        "--activation_dropout=0.055",
+        "--hidden_dropout=0.047",
+        "--save_total_limit=5",
+        "--freeze_feature_encoder",
+        "--feat_proj_dropout=0.04",
+        "--mask_time_prob=0.082",
+        "--mask_time_length=10",
+        "--mask_feature_prob=0.25",
+        "--mask_feature_length=64",
+        "--gradient_checkpointing",
+        "--min_duration_in_seconds=0.5",
+        "--max_duration_in_seconds=30.0",
+        "--use_auth_token",
+        "--seed=42",
+        "--group_by_length",
+        "--do_train",
+        "--do_eval",
+        "--push_to_hub",
+        "--preprocessing_num_workers=32",
+        "--ctc_zero_infinity",
+        "--do_lower_case",
+        "--wandb_project=wav2vec2",
+        "--wandb_name=wav2vec2-1b-npsc-nst-tpu",
+        "--remove_punctuation"
+    ],
+    "state": "running",
+    "program": "run_flax_speech_recognition_ctc.py",
+    "codePath": "run_flax_speech_recognition_ctc.py",
+    "git": {
+        "remote": "https://huggingface.co/NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "commit": "e2b1320cc68c3ce129a1d654965e0d3eb44e0558"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/wav2vec2-1b-npsc-nst",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220805_152536-2fzkf8n5/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb": {"runtime": 1183}}

wandb/run-20220805_152536-2fzkf8n5/logs/debug-internal.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c137bfb0e98ec07c29d675736207e73a2a80e52db6b05737195bcc03723ca10a
+size 51309

wandb/run-20220805_152536-2fzkf8n5/logs/debug.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38b16fcef53f7c67610a4fe7d00663568b5a7c2f40e719c37aa96391ea55eb3e
+size 6063

wandb/run-20220805_152536-2fzkf8n5/run-2fzkf8n5.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a29f323346e50c6ded2df9a6993e83112cafa162f35900a0bbd05f05dddc5e29
+size 250453

wandb/run-20220805_230151-2y71vcu4/files/code/run_flax_speech_recognition_ctc.py

@@ -0,0 +1,1631 @@
+#!/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,
+    set_seed,
+)
+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."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    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."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.1,
+        metadata={
+            "help": "The spec aug dropout probability for feature encoder representations."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@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": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    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=2,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+            "than this will be filtered."
+        },
+    )
+    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."}
+    )
+    skip_steps: Optional[int] = field(
+        default=0,
+        metadata={
+            "help": "Skip this number of steps. Useful to continue 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_util.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def to_bf16(t):
+    return jax.tree_util.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_util.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, num_log=50, 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))]
+        # we'll log the first 50 predictions for each epoch
+        wandb.log(
+            {
+                f"{prefix}/step_{int(step / 1000)}k": wandb.Table(
+                    columns=["label_str", "pred_str"], data=str_data[:num_log]
+                )
+            },
+            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 make_dataset(data_args, seed=42):
+    # Pre-processing dataset
+    import re
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text}
+
+    def filter_nst(entry):
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        return True
+
+    def filter_npsc(entry):
+        # False if there are digits in the text
+        if not ((len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        batch["text"] = re.sub('<inaudible>', 'xxx', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    # TODO NST_hesitate
+
+    split = len(npsc[data_args.train_split_name]) / (len(npsc[data_args.train_split_name]) + len(npsc[data_args.eval_split_name]))  # Use same train/val ratio as NPSC
+    nst_train = nst[data_args.train_split_name].train_test_split(train_size=split, seed=seed)
+    nst[data_args.train_split_name] = nst_train["train"]
+    nst[data_args.eval_split_name] = nst_train["test"]
+
+    nst = nst.filter(filter_nst).map(
+        map_nst,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NST",
+    ).shuffle(seed=seed)
+    npsc = npsc.filter(filter_npsc).map(
+        map_npsc,
+        num_proc=data_args.preprocessing_num_workers,
+        desc="filtering NPSC",
+    ).shuffle(seed=seed)
+
+    npsc_base = npsc.remove_columns([col for col in npsc[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+    nst_base = nst.remove_columns([col for col in nst[data_args.train_split_name].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in data_args.train_split_name, data_args.eval_split_name, data_args.test_split_name:
+        probs = np.array([len(nst_base[split]), len(npsc_base[split])])  # Weight by number of examples
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split]], probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+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
+
+    set_seed(training_args.seed)
+    raw_datasets = make_dataset(data_args, seed=training_args.seed)
+   
+    # raw_datasets = DatasetDict()
+
+    # if training_args.do_train:
+    #     raw_datasets[data_args.train_split_name] = 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[data_args.eval_split_name] = 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(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": tokenizer.vocab_size,  # len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    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
+
+    try:
+        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,
+        )
+    except:
+        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,
+            from_pt=True,
+        )
+
+    # 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_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
+    chars_to_ignore = ', ? . ! - ; : " “ % ‘ ” ?'.split(" ")
+    chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
+    # gigaspeech_punctuation = {" <comma>": ",", " <period>": ".", " <questionmark>": "?", " <exclamationpoint>": "!"}
+    # gigaspeech_disfluencies = ["<other>", "<sil>"]
+    # swb_disfluencies = ["[noise]", "[laughter]", "[silence]", "<a_aside>", "<b_aside>", "<e_aside>", "[laughter-",
+    #                 "[vocalized-noise]", "_1"]
+    # swb_punctuations = ["{", "}", "[", "]-", "]"]
+    # earnings_disfluencies = ["<crosstalk>", "<affirmative>", "<inaudible>", "inaudible", "<laugh>", "<unk>"]
+    ignore_segments = ["ignore_time_segment_in_scoring", "<noise>", "<music>", "[noise]", "[laughter]", "[silence]",
+                       "[vocalized-noise]", "<crosstalk>", "<affirmative>", "<inaudible>", "<laugh>", "<other>", "<sil>", ""]
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets[data_args.eval_split_name] = raw_datasets[data_args.eval_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_predict and data_args.max_test_samples is not None:
+        raw_datasets[data_args.test_split_name] = raw_datasets[data_args.test_split_name].select(range(data_args.max_eval_samples))
+
+    if training_args.do_train and data_args.remove_punctuation:
+
+        def remove_punctuation(batch):
+            batch[text_column_name] = (
+                re.sub(chars_to_ignore_regex, "", batch[text_column_name]).replace("'", "").replace('"', "")
+            )
+
+        raw_datasets[data_args.train_split_name] = raw_datasets[data_args.train_split_name].map(
+            remove_punctuation,
+            num_proc=data_args.preprocessing_num_workers,
+            desc="removing punctuation from train split",
+        )
+
+    # 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):
+        # process audio
+        try:
+            sample = batch[audio_column_name]
+        except ValueError:
+            sample = {"array": np.array([0.]), "sampling_rate": feature_extractor.sampling_rate}
+        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"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+
+        # if dataset_name == "google/xtreme_s":
+        #     # Finally, we tokenize the processed text
+        #     batch["labels"] = tokenizer(input_str).input_ids
+        #     batch["labels_length"] = len(batch["labels"])
+        #     return batch
+
+        # # Common Voice 9
+        # 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]
+        # # normalize quotation marks
+        # input_str = re.sub(r'["“”]', '"', input_str)
+        # # normalize apostrophes
+        # input_str = re.sub(r"[’']", "'", input_str)
+        # # normalize hyphens
+        # input_str = re.sub(r"[—–]", "-", input_str)
+        # # replace double quotation marks with single
+        # input_str = input_str.replace('""', '"')
+        # if dataset_name == "mozilla-foundation/common_voice_9_0" and len(input_str):
+        #     # for CV9, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # TEDLIUM-3
+        # # delete the <unk> token from the text and replace spaced apostrophes with un-spaced
+        # input_str = input_str.replace("<unk>", "").replace(" '", "'")
+
+        # # 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)
+        # if dataset_name == "speechcolab/gigaspeech" and len(input_str):
+        #     # for GS, we'll normalize the text to always finish with punctuation
+        #     if input_str[-1] not in [".", "?", "!"]:
+        #         input_str = input_str + "."
+
+        # # SWB
+        # for disfluency in swb_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # remove parenthesised text (test data only)
+        # input_str = re.sub("[\(].*?[\)]", "", input_str)
+        # for punctuation in swb_punctuations:
+        #     input_str = input_str.replace(punctuation, "")
+        # # replace anomalous words with their correct transcriptions
+        # 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]])
+
+        # # Earnings 22
+        # for disfluency in earnings_disfluencies:
+        #     input_str = input_str.replace(disfluency, "")
+        # # replace mal-formatted ellipsis
+        # input_str = input_str.replace("…", ".")
+
+        # JIWER compliance
+        # 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 data with inputs shorter than min_input_length or longer than max_input_length
+    def is_audio_in_length_range(length):
+        return length > min_input_length and length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.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 length > min_target_length  # and length < max_target_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+        is_labels_in_length_range,
+        num_proc=num_workers,
+        input_columns=["labels_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[data_args.train_split_name])
+        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_util.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_util.tree_map(jnp.add, accum_grad, grad), loss
+
+            # create an initial state for accumulating losses, num labels and gradients
+            init_grad = jax.tree_util.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_util.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_util.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:
+        p_eval_step = jax.pmap(eval_step, "batch")
+
+    def run_evaluation(step):
+        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[data_args.eval_split_name], 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[data_args.eval_split_name][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_util.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)
+            # 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_util.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)
+
+    skip_epochs = data_args.skip_steps // (num_train_samples // batch_size_per_update)
+    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}")
+    logger.info(f"  Skipping: {data_args.skip_steps} steps ({skip_epochs} epochs)")
+
+    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()
+
+            if epoch < skip_epochs:
+                logger.info(f"Skipping epoch {epoch + 1}")
+                continue
+
+            # 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[data_args.train_split_name], batch_size_per_update, input_rng)
+            train_batch_idx = generate_batch_splits(train_samples_idx, batch_size_per_update)
+
+            if data_args.skip_steps > cur_step:
+                logger.info(f"Skipping {data_args.skip_steps - (epoch * (num_train_samples // batch_size_per_update))} steps...")
+            # 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):
+                cur_step = epoch * (num_train_samples // batch_size_per_update) + step
+                if cur_step <= data_args.skip_steps:
+                    continue
+
+                samples = [vectorized_datasets[data_args.train_split_name][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)
+
+
+                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=data_args.train_split_name)
+                    # 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)
+
+                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)
+                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)
+
+    # TODO: collapse 'do_predict' into the run_evaluation function
+    if training_args.do_predict:
+        for split in [data_args.test_split_name]:
+            # ======================== 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_util.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, 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()

wandb/run-20220805_230151-2y71vcu4/files/config.yaml

@@ -0,0 +1,27 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_flax_speech_recognition_ctc.py
+    framework: huggingface
+    huggingface_version: 4.21.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1659740511
+    t:
+      1:
+      - 1
+      - 2
+      - 3
+      - 11
+      - 12
+      3:
+      - 13
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.21.0
+      8:
+      - 5

wandb/run-20220805_230151-2y71vcu4/files/diff.patch

@@ -0,0 +1,131 @@
+diff --git a/config.json b/config.json
+index 260219f..246b797 100644
+--- a/config.json
++++ b/config.json
+@@ -5,7 +5,7 @@
+   "add_adapter": false,
+   "apply_spec_augment": true,
+   "architectures": [
+-    "Wav2Vec2ForCTC"
++    "Wav2Vec2ForPreTraining"
+   ],
+   "attention_dropout": 0.094,
+   "bos_token_id": 1,
+diff --git a/run.sh b/run.sh
+index 9cc498e..8758978 100755
+--- a/run.sh
++++ b/run.sh
+@@ -1,6 +1,6 @@
+ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_ctc.py \
+         --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+-        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst" \
++        --hub_model_id="NbAiLab/wav2vec2-1b-npsc-nst-tpu" \
+         --tokenizer_name="./" \
+         --output_dir="./" \
+         --overwrite_output_dir \
+@@ -11,13 +11,13 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --precision="full_mixed" \
+         --matmul_precision="bfloat16" \
+         --multisteps \
+-        --learning_rate="2e-4" \
++        --learning_rate="1e-4" \
+         --warmup_steps="2000" \
+         --length_column_name="input_length" \
+         --evaluation_strategy="steps" \
+         --text_column_name="text" \
+-        --save_steps="4000" \
+-        --eval_steps="4000" \
++        --save_steps="5000" \
++        --eval_steps="5000" \
+         --logging_steps="100" \
+         --layerdrop="0.041" \
+         --attention_dropout="0.094" \
+@@ -42,7 +42,7 @@ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_flax_speech_recognition_c
+         --ctc_zero_infinity \
+         --do_lower_case \
+         --wandb_project="wav2vec2" \
+-        --wandb_name="wav2vec2-1b-npsc-nst" \
++        --wandb_name="wav2vec2-1b-npsc-nst-tpu" \
+         --remove_punctuation
+ 
+ 
+diff --git a/special_tokens_map.json b/special_tokens_map.json
+index 89389bf..308786b 100644
+--- a/special_tokens_map.json
++++ b/special_tokens_map.json
+@@ -343,6 +343,48 @@
+       "rstrip": false,
+       "single_word": false
+     },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "</s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
++    {
++      "content": "<s>",
++      "lstrip": false,
++      "normalized": true,
++      "rstrip": false,
++      "single_word": false
++    },
+     {
+       "content": "</s>",
+       "lstrip": false,
+diff --git a/wandb/debug-internal.log b/wandb/debug-internal.log
+index 50a0b69..23926ef 120000
+--- a/wandb/debug-internal.log
++++ b/wandb/debug-internal.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug-internal.log
+
++run-20220805_230151-2y71vcu4/logs/debug-internal.log
+
+diff --git a/wandb/debug.log b/wandb/debug.log
+index 746223d..279853d 120000
+--- a/wandb/debug.log
++++ b/wandb/debug.log
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z/logs/debug.log
+
++run-20220805_230151-2y71vcu4/logs/debug.log
+
+diff --git a/wandb/latest-run b/wandb/latest-run
+index be58b40..f069a7a 120000
+--- a/wandb/latest-run
++++ b/wandb/latest-run
+@@ -1 +1 @@
+-run-20220803_091109-yit1e59z
+
++run-20220805_230151-2y71vcu4
+

wandb/run-20220805_230151-2y71vcu4/files/output.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:864734adde2a9711d54e4c8f0ab743d16890f9d9702b4457b261c095c32ea5e0
+size 239534

wandb/run-20220805_230151-2y71vcu4/files/requirements.txt

@@ -0,0 +1,158 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+appdirs==1.4.4
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+audioread==2.1.9
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+cffi==1.15.1
+charset-normalizer==2.0.10
+chex==0.1.3
+click==8.0.3
+cloud-tpu-client==0.10
+cloud-tpu-profiler==2.4.0
+clu==0.0.6
+colorama==0.4.5
+commonmark==0.9.1
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==2.4.0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+etils==0.6.0
+exceptiongroup==1.0.0rc8
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.5.3
+fonttools==4.28.5
+frozenlist==1.2.0
+fsspec==2021.11.1
+future==0.18.2
+gast==0.4.0
+gitdb==4.0.9
+gitpython==3.1.26
+google-api-core==1.31.5
+google-api-python-client==1.8.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.4.6
+google-auth==2.3.3
+google-pasta==0.2.0
+googleapis-common-protos==1.54.0
+grpcio==1.43.0
+h5py==3.6.0
+httplib2==0.20.2
+huggingface-hub==0.2.1
+hypothesis==6.53.0
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.3.15
+jaxlib==0.3.15
+jedi==0.18.1
+jiwer==2.3.0
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+librosa==0.9.2
+libtpu-nightly==0.1.dev20220722
+llvmlite==0.39.0
+markdown==3.3.6
+matplotlib-inline==0.1.3
+matplotlib==3.5.1
+ml-collections==0.1.0
+msgpack==1.0.3
+multidict==5.2.0
+multiprocess==0.70.12.2
+numba==0.56.0
+numpy==1.22.0
+oauth2client==4.1.3
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.3
+packaging==21.3
+pandas==1.3.5
+parso==0.8.3
+pathtools==0.1.2
+pexpect==4.8.0
+pickleshare==0.7.5
+pillow==9.0.0
+pip==22.2.1
+pkg-resources==0.0.0
+pooch==1.6.0
+promise==2.3
+prompt-toolkit==3.0.24
+protobuf==3.19.1
+psutil==5.9.0
+ptyprocess==0.7.0
+pyarrow==6.0.1
+pyasn1-modules==0.2.8
+pyasn1==0.4.8
+pycparser==2.21
+pyctcdecode==0.4.0
+pygments==2.11.1
+pygtrie==2.5.0
+pyparsing==3.0.6
+python-dateutil==2.8.2
+python-levenshtein==0.12.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+resampy==0.3.1
+responses==0.18.0
+rich==11.2.0
+rsa==4.8
+sacremoses==0.0.46
+scikit-learn==1.1.1
+scipy==1.7.3
+sentry-sdk==1.5.2
+setuptools==44.0.0
+shortuuid==1.0.8
+six==1.16.0
+smmap==5.0.0
+sortedcontainers==2.4.0
+soundfile==0.10.3.post1
+sox==1.4.1
+subprocess32==3.5.4
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+tensorboard==2.7.0
+tensorflow-cpu==2.7.0
+tensorflow-datasets==4.4.0
+tensorflow-estimator==2.7.0
+tensorflow-io-gcs-filesystem==0.23.1
+tensorflow-metadata==1.5.0
+tensorflow==2.7.0
+tensorstore==0.1.21
+termcolor==1.1.0
+threadpoolctl==3.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+torch==1.12.0
+torchaudio==0.12.0+cpu
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.21.0
+typing-extensions==4.3.0
+uritemplate==3.0.1
+urllib3==1.26.7
+wandb==0.12.9
+wcwidth==0.2.5
+werkzeug==2.0.2
+wheel==0.37.1
+wrapt==1.13.3
+xxhash==2.0.2
+yarl==1.7.2
+yaspin==2.1.0
+zipp==3.7.0

wandb/run-20220805_230151-2y71vcu4/files/wandb-metadata.json

@@ -0,0 +1,69 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-08-05T23:01:55.087413",
+    "startedAt": "2022-08-05T23:01:51.745697",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--model_name_or_path=facebook/wav2vec2-xls-r-1b",
+        "--hub_model_id=NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "--tokenizer_name=./",
+        "--output_dir=./",
+        "--overwrite_output_dir",
+        "--num_train_epochs=40",
+        "--per_device_train_batch_size=2",
+        "--per_device_eval_batch_size=2",
+        "--gradient_accumulation_steps=1",
+        "--precision=full_mixed",
+        "--matmul_precision=bfloat16",
+        "--multisteps",
+        "--learning_rate=1e-4",
+        "--warmup_steps=2000",
+        "--length_column_name=input_length",
+        "--evaluation_strategy=steps",
+        "--text_column_name=text",
+        "--save_steps=5000",
+        "--eval_steps=5000",
+        "--logging_steps=100",
+        "--layerdrop=0.041",
+        "--attention_dropout=0.094",
+        "--activation_dropout=0.055",
+        "--hidden_dropout=0.047",
+        "--save_total_limit=5",
+        "--freeze_feature_encoder",
+        "--feat_proj_dropout=0.04",
+        "--mask_time_prob=0.082",
+        "--mask_time_length=10",
+        "--mask_feature_prob=0.25",
+        "--mask_feature_length=64",
+        "--gradient_checkpointing",
+        "--min_duration_in_seconds=0.5",
+        "--max_duration_in_seconds=30.0",
+        "--use_auth_token",
+        "--seed=42",
+        "--group_by_length",
+        "--do_train",
+        "--do_eval",
+        "--push_to_hub",
+        "--preprocessing_num_workers=32",
+        "--ctc_zero_infinity",
+        "--do_lower_case",
+        "--wandb_project=wav2vec2",
+        "--wandb_name=wav2vec2-1b-npsc-nst-tpu",
+        "--remove_punctuation"
+    ],
+    "state": "running",
+    "program": "run_flax_speech_recognition_ctc.py",
+    "codePath": "run_flax_speech_recognition_ctc.py",
+    "git": {
+        "remote": "https://huggingface.co/NbAiLab/wav2vec2-1b-npsc-nst-tpu",
+        "commit": "e2b1320cc68c3ce129a1d654965e0d3eb44e0558"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/wav2vec2-1b-npsc-nst",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220805_230151-2y71vcu4/files/wandb-summary.json

@@ -0,0 +1 @@
+{"train/grad_norm": 8.0625, "layer_grad_norm/": {"lm_head": {"bias": 0.140625, "kernel": 2.40625}, "wav2vec2": {"encoder": {"layer_norm": {"bias": 0.10791015625, "scale": 0.091796875}, "layers": {"0": {"attention": {"k_proj": {"bias": 0.00018215179443359375, "kernel": 0.138671875}, "out_proj": {"bias": 0.119140625, "kernel": 0.86328125}, "q_proj": {"bias": 0.0157470703125, "kernel": 0.1904296875}, "v_proj": {"bias": 0.1064453125, "kernel": 0.84765625}}, "feed_forward": {"intermediate_dense": {"bias": 0.10546875, "kernel": 1.515625}, "output_dense": {"bias": 0.03759765625, "kernel": 1.296875}}, "final_layer_norm": {"bias": 0.30078125, "scale": 0.48828125}, "layer_norm": {"bias": 0.181640625, "scale": 0.314453125}}, "1": {"attention": {"k_proj": {"bias": 0.00012874603271484375, "kernel": 0.123046875}, "out_proj": {"bias": 0.0419921875, "kernel": 0.5390625}, "q_proj": {"bias": 0.01007080078125, "kernel": 0.119140625}, "v_proj": {"bias": 0.0625, "kernel": 0.419921875}}, "feed_forward": {"intermediate_dense": {"bias": 0.051025390625, "kernel": 0.765625}, "output_dense": {"bias": 0.037841796875, "kernel": 0.65625}}, "final_layer_norm": {"bias": 0.095703125, "scale": 0.09326171875}, "layer_norm": {"bias": 0.10546875, "scale": 0.146484375}}, "10": {"attention": {"k_proj": {"bias": 8.726119995117188e-05, "kernel": 0.244140625}, "out_proj": {"bias": 0.02587890625, "kernel": 0.427734375}, "q_proj": {"bias": 0.0137939453125, "kernel": 0.2490234375}, "v_proj": {"bias": 0.040771484375, "kernel": 0.439453125}}, "feed_forward": {"intermediate_dense": {"bias": 0.039306640625, "kernel": 0.66015625}, "output_dense": {"bias": 0.024169921875, "kernel": 0.53125}}, "final_layer_norm": {"bias": 0.06640625, "scale": 0.0654296875}, "layer_norm": {"bias": 0.08251953125, "scale": 0.1181640625}}, "11": {"attention": {"k_proj": {"bias": 0.0001239776611328125, "kernel": 0.26953125}, "out_proj": {"bias": 0.0240478515625, "kernel": 0.50390625}, "q_proj": {"bias": 0.0166015625, "kernel": 0.26171875}, "v_proj": {"bias": 0.03955078125, "kernel": 0.515625}}, "feed_forward": {"intermediate_dense": {"bias": 0.033203125, "kernel": 0.58203125}, "output_dense": {"bias": 0.023193359375, "kernel": 0.4453125}}, "final_layer_norm": {"bias": 0.05322265625, "scale": 0.05712890625}, "layer_norm": {"bias": 0.0703125, "scale": 0.1064453125}}, "12": {"attention": {"k_proj": {"bias": 8.58306884765625e-05, "kernel": 0.197265625}, "out_proj": {"bias": 0.0234375, "kernel": 0.400390625}, "q_proj": {"bias": 0.00970458984375, "kernel": 0.19140625}, "v_proj": {"bias": 0.03662109375, "kernel": 0.431640625}}, "feed_forward": {"intermediate_dense": {"bias": 0.03369140625, "kernel": 0.56640625}, "output_dense": {"bias": 0.0225830078125, "kernel": 0.41796875}}, "final_layer_norm": {"bias": 0.057373046875, "scale": 0.05712890625}, "layer_norm": {"bias": 0.060546875, "scale": 0.059326171875}}, "13": {"attention": {"k_proj": {"bias": 0.00017261505126953125, "kernel": 0.259765625}, "out_proj": {"bias": 0.0235595703125, "kernel": 0.484375}, "q_proj": {"bias": 0.01336669921875, "kernel": 0.240234375}, "v_proj": {"bias": 0.039306640625, "kernel": 0.54296875}}, "feed_forward": {"intermediate_dense": {"bias": 0.032958984375, "kernel": 0.55078125}, "output_dense": {"bias": 0.0230712890625, "kernel": 0.42578125}}, "final_layer_norm": {"bias": 0.05517578125, "scale": 0.06201171875}, "layer_norm": {"bias": 0.06103515625, "scale": 0.06787109375}}, "14": {"attention": {"k_proj": {"bias": 0.000148773193359375, "kernel": 0.19140625}, "out_proj": {"bias": 0.0228271484375, "kernel": 0.4453125}, "q_proj": {"bias": 0.0096435546875, "kernel": 0.1787109375}, "v_proj": {"bias": 0.03466796875, "kernel": 0.45703125}}, "feed_forward": {"intermediate_dense": {"bias": 0.03369140625, "kernel": 0.5625}, "output_dense": {"bias": 0.022705078125, "kernel": 0.43359375}}, "final_layer_norm": {"bias": 0.0556640625, "scale": 0.061279296875}, "layer_norm": {"bias": 0.0537109375, "scale": 0.05419921875}}, "15": {"attention": {"k_proj": {"bias": 0.0001983642578125, "kernel": 0.2490234375}, "out_proj": {"bias": 0.0233154296875, "kernel": 0.5859375}, "q_proj": {"bias": 0.012939453125, "kernel": 0.224609375}, "v_proj": {"bias": 0.0400390625, "kernel": 0.5859375}}, "feed_forward": {"intermediate_dense": {"bias": 0.03271484375, "kernel": 0.53125}, "output_dense": {"bias": 0.0223388671875, "kernel": 0.4296875}}, "final_layer_norm": {"bias": 0.05419921875, "scale": 0.06494140625}, "layer_norm": {"bias": 0.060546875, "scale": 0.0625}}, "16": {"attention": {"k_proj": {"bias": 0.00013446807861328125, "kernel": 0.22265625}, "out_proj": {"bias": 0.0224609375, "kernel": 0.384765625}, "q_proj": {"bias": 0.011962890625, "kernel": 0.2109375}, "v_proj": {"bias": 0.03466796875, "kernel": 0.40234375}}, "feed_forward": {"intermediate_dense": {"bias": 0.03173828125, "kernel": 0.5234375}, "output_dense": {"bias": 0.021484375, "kernel": 0.421875}}, "final_layer_norm": {"bias": 0.054931640625, "scale": 0.0498046875}, "layer_norm": {"bias": 0.057861328125, "scale": 0.05517578125}}, "17": {"attention": {"k_proj": {"bias": 0.0001239776611328125, "kernel": 0.177734375}, "out_proj": {"bias": 0.023193359375, "kernel": 0.330078125}, "q_proj": {"bias": 0.00933837890625, "kernel": 0.154296875}, "v_proj": {"bias": 0.034912109375, "kernel": 0.37109375}}, "feed_forward": {"intermediate_dense": {"bias": 0.03125, "kernel": 0.51171875}, "output_dense": {"bias": 0.0220947265625, "kernel": 0.400390625}}, "final_layer_norm": {"bias": 0.0517578125, "scale": 0.0478515625}, "layer_norm": {"bias": 0.05615234375, "scale": 0.0546875}}, "18": {"attention": {"k_proj": {"bias": 0.000110626220703125, "kernel": 0.201171875}, "out_proj": {"bias": 0.02197265625, "kernel": 0.404296875}, "q_proj": {"bias": 0.0098876953125, "kernel": 0.1806640625}, "v_proj": {"bias": 0.03369140625, "kernel": 0.40625}}, "feed_forward": {"intermediate_dense": {"bias": 0.02978515625, "kernel": 0.4921875}, "output_dense": {"bias": 0.0208740234375, "kernel": 0.408203125}}, "final_layer_norm": {"bias": 0.049560546875, "scale": 0.04296875}, "layer_norm": {"bias": 0.052978515625, "scale": 0.047119140625}}, "19": {"attention": {"k_proj": {"bias": 6.341934204101562e-05, "kernel": 0.13671875}, "out_proj": {"bias": 0.021484375, "kernel": 0.291015625}, "q_proj": {"bias": 0.0076904296875, "kernel": 0.1298828125}, "v_proj": {"bias": 0.03125, "kernel": 0.32421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.028076171875, "kernel": 0.478515625}, "output_dense": {"bias": 0.020751953125, "kernel": 0.41015625}}, "final_layer_norm": {"bias": 0.048583984375, "scale": 0.04296875}, "layer_norm": {"bias": 0.047119140625, "scale": 0.036376953125}}, "2": {"attention": {"k_proj": {"bias": 0.0001392364501953125, "kernel": 0.150390625}, "out_proj": {"bias": 0.04296875, "kernel": 0.56640625}, "q_proj": {"bias": 0.01214599609375, "kernel": 0.1484375}, "v_proj": {"bias": 0.06787109375, "kernel": 0.51953125}}, "feed_forward": {"intermediate_dense": {"bias": 0.057373046875, "kernel": 0.953125}, "output_dense": {"bias": 0.03759765625, "kernel": 0.6953125}}, "final_layer_norm": {"bias": 0.1103515625, "scale": 0.0927734375}, "layer_norm": {"bias": 0.099609375, "scale": 0.130859375}}, "20": {"attention": {"k_proj": {"bias": 3.62396240234375e-05, "kernel": 0.0966796875}, "out_proj": {"bias": 0.02197265625, "kernel": 0.2080078125}, "q_proj": {"bias": 0.0048828125, "kernel": 0.08984375}, "v_proj": {"bias": 0.030517578125, "kernel": 0.24609375}}, "feed_forward": {"intermediate_dense": {"bias": 0.02734375, "kernel": 0.486328125}, "output_dense": {"bias": 0.021484375, "kernel": 0.396484375}}, "final_layer_norm": {"bias": 0.044921875, "scale": 0.03955078125}, "layer_norm": {"bias": 0.04541015625, "scale": 0.0322265625}}, "21": {"attention": {"k_proj": {"bias": 9.584426879882812e-05, "kernel": 0.12158203125}, "out_proj": {"bias": 0.0218505859375, "kernel": 0.283203125}, "q_proj": {"bias": 0.006561279296875, "kernel": 0.11962890625}, "v_proj": {"bias": 0.031005859375, "kernel": 0.3125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0277099609375, "kernel": 0.49609375}, "output_dense": {"bias": 0.0218505859375, "kernel": 0.40234375}}, "final_layer_norm": {"bias": 0.045166015625, "scale": 0.0439453125}, "layer_norm": {"bias": 0.042724609375, "scale": 0.037353515625}}, "22": {"attention": {"k_proj": {"bias": 5.507469177246094e-05, "kernel": 0.138671875}, "out_proj": {"bias": 0.0234375, "kernel": 0.26953125}, "q_proj": {"bias": 0.007659912109375, "kernel": 0.140625}, "v_proj": {"bias": 0.0322265625, "kernel": 0.2890625}}, "feed_forward": {"intermediate_dense": {"bias": 0.029296875, "kernel": 0.51171875}, "output_dense": {"bias": 0.0233154296875, "kernel": 0.400390625}}, "final_layer_norm": {"bias": 0.04736328125, "scale": 0.05126953125}, "layer_norm": {"bias": 0.05029296875, "scale": 0.06640625}}, "23": {"attention": {"k_proj": {"bias": 0.0001430511474609375, "kernel": 0.2080078125}, "out_proj": {"bias": 0.0242919921875, "kernel": 0.462890625}, "q_proj": {"bias": 0.010986328125, "kernel": 0.2109375}, "v_proj": {"bias": 0.03662109375, "kernel": 0.453125}}, "feed_forward": {"intermediate_dense": {"bias": 0.02978515625, "kernel": 0.5234375}, "output_dense": {"bias": 0.0238037109375, "kernel": 0.4140625}}, "final_layer_norm": {"bias": 0.048828125, "scale": 0.04638671875}, "layer_norm": {"bias": 0.05810546875, "scale": 0.06884765625}}, "24": {"attention": {"k_proj": {"bias": 9.441375732421875e-05, "kernel": 0.19140625}, "out_proj": {"bias": 0.0224609375, "kernel": 0.37109375}, "q_proj": {"bias": 0.01025390625, "kernel": 0.189453125}, "v_proj": {"bias": 0.035888671875, "kernel": 0.37890625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0289306640625, "kernel": 0.53125}, "output_dense": {"bias": 0.021728515625, "kernel": 0.390625}}, "final_layer_norm": {"bias": 0.0478515625, "scale": 0.045654296875}, "layer_norm": {"bias": 0.06005859375, "scale": 0.0517578125}}, "25": {"attention": {"k_proj": {"bias": 0.0001239776611328125, "kernel": 0.189453125}, "out_proj": {"bias": 0.0224609375, "kernel": 0.35546875}, "q_proj": {"bias": 0.0106201171875, "kernel": 0.1875}, "v_proj": {"bias": 0.033935546875, "kernel": 0.373046875}}, "feed_forward": {"intermediate_dense": {"bias": 0.02880859375, "kernel": 0.51953125}, "output_dense": {"bias": 0.02197265625, "kernel": 0.3671875}}, "final_layer_norm": {"bias": 0.049072265625, "scale": 0.04443359375}, "layer_norm": {"bias": 0.051025390625, "scale": 0.04296875}}, "26": {"attention": {"k_proj": {"bias": 7.62939453125e-05, "kernel": 0.181640625}, "out_proj": {"bias": 0.022216796875, "kernel": 0.345703125}, "q_proj": {"bias": 0.0108642578125, "kernel": 0.193359375}, "v_proj": {"bias": 0.034912109375, "kernel": 0.359375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0279541015625, "kernel": 0.484375}, "output_dense": {"bias": 0.0220947265625, "kernel": 0.37890625}}, "final_layer_norm": {"bias": 0.0458984375, "scale": 0.04931640625}, "layer_norm": {"bias": 0.052978515625, "scale": 0.0693359375}}, "27": {"attention": {"k_proj": {"bias": 0.00016021728515625, "kernel": 0.24609375}, "out_proj": {"bias": 0.0205078125, "kernel": 0.4453125}, "q_proj": {"bias": 0.013916015625, "kernel": 0.2470703125}, "v_proj": {"bias": 0.034423828125, "kernel": 0.44140625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0286865234375, "kernel": 0.48828125}, "output_dense": {"bias": 0.02099609375, "kernel": 0.3828125}}, "final_layer_norm": {"bias": 0.0478515625, "scale": 0.05126953125}, "layer_norm": {"bias": 0.05859375, "scale": 0.099609375}}, "28": {"attention": {"k_proj": {"bias": 0.00012063980102539062, "kernel": 0.2119140625}, "out_proj": {"bias": 0.019287109375, "kernel": 0.43359375}, "q_proj": {"bias": 0.0130615234375, "kernel": 0.228515625}, "v_proj": {"bias": 0.03076171875, "kernel": 0.412109375}}, "feed_forward": {"intermediate_dense": {"bias": 0.02734375, "kernel": 0.4765625}, "output_dense": {"bias": 0.019775390625, "kernel": 0.380859375}}, "final_layer_norm": {"bias": 0.0458984375, "scale": 0.05615234375}, "layer_norm": {"bias": 0.0546875, "scale": 0.07275390625}}, "29": {"attention": {"k_proj": {"bias": 9.632110595703125e-05, "kernel": 0.2353515625}, "out_proj": {"bias": 0.0184326171875, "kernel": 0.373046875}, "q_proj": {"bias": 0.01165771484375, "kernel": 0.234375}, "v_proj": {"bias": 0.0302734375, "kernel": 0.376953125}}, "feed_forward": {"intermediate_dense": {"bias": 0.027099609375, "kernel": 0.53125}, "output_dense": {"bias": 0.01806640625, "kernel": 0.37109375}}, "final_layer_norm": {"bias": 0.0439453125, "scale": 0.041259765625}, "layer_norm": {"bias": 0.0556640625, "scale": 0.05029296875}}, "3": {"attention": {"k_proj": {"bias": 0.0002079010009765625, "kernel": 0.2421875}, "out_proj": {"bias": 0.03955078125, "kernel": 0.75}, "q_proj": {"bias": 0.016357421875, "kernel": 0.23046875}, "v_proj": {"bias": 0.0634765625, "kernel": 0.671875}}, "feed_forward": {"intermediate_dense": {"bias": 0.05419921875, "kernel": 0.90625}, "output_dense": {"bias": 0.0361328125, "kernel": 0.671875}}, "final_layer_norm": {"bias": 0.09228515625, "scale": 0.0947265625}, "layer_norm": {"bias": 0.0966796875, "scale": 0.171875}}, "30": {"attention": {"k_proj": {"bias": 0.00015354156494140625, "kernel": 0.244140625}, "out_proj": {"bias": 0.0177001953125, "kernel": 0.375}, "q_proj": {"bias": 0.0123291015625, "kernel": 0.2470703125}, "v_proj": {"bias": 0.02880859375, "kernel": 0.39453125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0257568359375, "kernel": 0.5078125}, "output_dense": {"bias": 0.01708984375, "kernel": 0.326171875}}, "final_layer_norm": {"bias": 0.042236328125, "scale": 0.0517578125}, "layer_norm": {"bias": 0.044921875, "scale": 0.04443359375}}, "31": {"attention": {"k_proj": {"bias": 0.00011348724365234375, "kernel": 0.2138671875}, "out_proj": {"bias": 0.01611328125, "kernel": 0.30859375}, "q_proj": {"bias": 0.01141357421875, "kernel": 0.228515625}, "v_proj": {"bias": 0.025146484375, "kernel": 0.3359375}}, "feed_forward": {"intermediate_dense": {"bias": 0.02294921875, "kernel": 0.44921875}, "output_dense": {"bias": 0.0155029296875, "kernel": 0.28515625}}, "final_layer_norm": {"bias": 0.03662109375, "scale": 0.035400390625}, "layer_norm": {"bias": 0.04296875, "scale": 0.054443359375}}, "32": {"attention": {"k_proj": {"bias": 0.0001583099365234375, "kernel": 0.177734375}, "out_proj": {"bias": 0.01495361328125, "kernel": 0.302734375}, "q_proj": {"bias": 0.00897216796875, "kernel": 0.1748046875}, "v_proj": {"bias": 0.0234375, "kernel": 0.326171875}}, "feed_forward": {"intermediate_dense": {"bias": 0.0216064453125, "kernel": 0.4296875}, "output_dense": {"bias": 0.013916015625, "kernel": 0.267578125}}, "final_layer_norm": {"bias": 0.036376953125, "scale": 0.03125}, "layer_norm": {"bias": 0.038330078125, "scale": 0.064453125}}, "33": {"attention": {"k_proj": {"bias": 0.00018310546875, "kernel": 0.1875}, "out_proj": {"bias": 0.0133056640625, "kernel": 0.294921875}, "q_proj": {"bias": 0.0087890625, "kernel": 0.18359375}, "v_proj": {"bias": 0.021728515625, "kernel": 0.318359375}}, "feed_forward": {"intermediate_dense": {"bias": 0.020263671875, "kernel": 0.392578125}, "output_dense": {"bias": 0.0126953125, "kernel": 0.2578125}}, "final_layer_norm": {"bias": 0.036865234375, "scale": 0.041015625}, "layer_norm": {"bias": 0.033447265625, "scale": 0.0361328125}}, "34": {"attention": {"k_proj": {"bias": 0.00014400482177734375, "kernel": 0.203125}, "out_proj": {"bias": 0.01165771484375, "kernel": 0.275390625}, "q_proj": {"bias": 0.009033203125, "kernel": 0.1943359375}, "v_proj": {"bias": 0.01953125, "kernel": 0.2890625}}, "feed_forward": {"intermediate_dense": {"bias": 0.0184326171875, "kernel": 0.357421875}, "output_dense": {"bias": 0.01129150390625, "kernel": 0.248046875}}, "final_layer_norm": {"bias": 0.032958984375, "scale": 0.045654296875}, "layer_norm": {"bias": 0.03369140625, "scale": 0.0341796875}}, "35": {"attention": {"k_proj": {"bias": 0.0001678466796875, "kernel": 0.2041015625}, "out_proj": {"bias": 0.0098876953125, "kernel": 0.28515625}, "q_proj": {"bias": 0.009765625, "kernel": 0.2060546875}, "v_proj": {"bias": 0.01483154296875, "kernel": 0.255859375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0150146484375, "kernel": 0.279296875}, "output_dense": {"bias": 0.0098876953125, "kernel": 0.2080078125}}, "final_layer_norm": {"bias": 0.02734375, "scale": 0.0301513671875}, "layer_norm": {"bias": 0.03466796875, "scale": 0.04345703125}}, "36": {"attention": {"k_proj": {"bias": 0.0001220703125, "kernel": 0.1455078125}, "out_proj": {"bias": 0.00909423828125, "kernel": 0.2177734375}, "q_proj": {"bias": 0.007049560546875, "kernel": 0.140625}, "v_proj": {"bias": 0.013427734375, "kernel": 0.2021484375}}, "feed_forward": {"intermediate_dense": {"bias": 0.01251220703125, "kernel": 0.234375}, "output_dense": {"bias": 0.0089111328125, "kernel": 0.1845703125}}, "final_layer_norm": {"bias": 0.021484375, "scale": 0.0230712890625}, "layer_norm": {"bias": 0.0277099609375, "scale": 0.0267333984375}}, "37": {"attention": {"k_proj": {"bias": 0.0001087188720703125, "kernel": 0.158203125}, "out_proj": {"bias": 0.0081787109375, "kernel": 0.212890625}, "q_proj": {"bias": 0.007659912109375, "kernel": 0.150390625}, "v_proj": {"bias": 0.01300048828125, "kernel": 0.197265625}}, "feed_forward": {"intermediate_dense": {"bias": 0.01141357421875, "kernel": 0.21484375}, "output_dense": {"bias": 0.00799560546875, "kernel": 0.16796875}}, "final_layer_norm": {"bias": 0.020263671875, "scale": 0.01806640625}, "layer_norm": {"bias": 0.0283203125, "scale": 0.044677734375}}, "38": {"attention": {"k_proj": {"bias": 0.0001354217529296875, "kernel": 0.1357421875}, "out_proj": {"bias": 0.007354736328125, "kernel": 0.1845703125}, "q_proj": {"bias": 0.0059814453125, "kernel": 0.12890625}, "v_proj": {"bias": 0.011474609375, "kernel": 0.1796875}}, "feed_forward": {"intermediate_dense": {"bias": 0.01025390625, "kernel": 0.193359375}, "output_dense": {"bias": 0.0072021484375, "kernel": 0.154296875}}, "final_layer_norm": {"bias": 0.0189208984375, "scale": 0.01513671875}, "layer_norm": {"bias": 0.024658203125, "scale": 0.0286865234375}}, "39": {"attention": {"k_proj": {"bias": 7.677078247070312e-05, "kernel": 0.1259765625}, "out_proj": {"bias": 0.00640869140625, "kernel": 0.1708984375}, "q_proj": {"bias": 0.00567626953125, "kernel": 0.12109375}, "v_proj": {"bias": 0.01031494140625, "kernel": 0.169921875}}, "feed_forward": {"intermediate_dense": {"bias": 0.00860595703125, "kernel": 0.169921875}, "output_dense": {"bias": 0.006195068359375, "kernel": 0.150390625}}, "final_layer_norm": {"bias": 0.01611328125, "scale": 0.01422119140625}, "layer_norm": {"bias": 0.0240478515625, "scale": 0.0244140625}}, "4": {"attention": {"k_proj": {"bias": 0.0002346038818359375, "kernel": 0.283203125}, "out_proj": {"bias": 0.03759765625, "kernel": 0.796875}, "q_proj": {"bias": 0.0179443359375, "kernel": 0.279296875}, "v_proj": {"bias": 0.058349609375, "kernel": 0.734375}}, "feed_forward": {"intermediate_dense": {"bias": 0.05078125, "kernel": 0.8203125}, "output_dense": {"bias": 0.034423828125, "kernel": 0.62890625}}, "final_layer_norm": {"bias": 0.0888671875, "scale": 0.09814453125}, "layer_norm": {"bias": 0.0908203125, "scale": 0.08544921875}}, "40": {"attention": {"k_proj": {"bias": 4.482269287109375e-05, "kernel": 0.1181640625}, "out_proj": {"bias": 0.006072998046875, "kernel": 0.138671875}, "q_proj": {"bias": 0.004791259765625, "kernel": 0.09765625}, "v_proj": {"bias": 0.00885009765625, "kernel": 0.1328125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0081787109375, "kernel": 0.1513671875}, "output_dense": {"bias": 0.005859375, "kernel": 0.1298828125}}, "final_layer_norm": {"bias": 0.016845703125, "scale": 0.01806640625}, "layer_norm": {"bias": 0.018310546875, "scale": 0.0198974609375}}, "41": {"attention": {"k_proj": {"bias": 6.246566772460938e-05, "kernel": 0.11181640625}, "out_proj": {"bias": 0.00543212890625, "kernel": 0.1640625}, "q_proj": {"bias": 0.004302978515625, "kernel": 0.1015625}, "v_proj": {"bias": 0.0091552734375, "kernel": 0.173828125}}, "feed_forward": {"intermediate_dense": {"bias": 0.006805419921875, "kernel": 0.1376953125}, "output_dense": {"bias": 0.00531005859375, "kernel": 0.119140625}}, "final_layer_norm": {"bias": 0.0142822265625, "scale": 0.01708984375}, "layer_norm": {"bias": 0.019287109375, "scale": 0.0267333984375}}, "42": {"attention": {"k_proj": {"bias": 2.0384788513183594e-05, "kernel": 0.0400390625}, "out_proj": {"bias": 0.00537109375, "kernel": 0.0908203125}, "q_proj": {"bias": 0.001922607421875, "kernel": 0.039306640625}, "v_proj": {"bias": 0.006591796875, "kernel": 0.08642578125}}, "feed_forward": {"intermediate_dense": {"bias": 0.00604248046875, "kernel": 0.1259765625}, "output_dense": {"bias": 0.00537109375, "kernel": 0.1015625}}, "final_layer_norm": {"bias": 0.01263427734375, "scale": 0.014892578125}, "layer_norm": {"bias": 0.01220703125, "scale": 0.02099609375}}, "43": {"attention": {"k_proj": {"bias": 6.854534149169922e-06, "kernel": 0.019287109375}, "out_proj": {"bias": 0.00543212890625, "kernel": 0.0634765625}, "q_proj": {"bias": 0.0009918212890625, "kernel": 0.01953125}, "v_proj": {"bias": 0.005859375, "kernel": 0.0625}}, "feed_forward": {"intermediate_dense": {"bias": 0.005950927734375, "kernel": 0.12890625}, "output_dense": {"bias": 0.005523681640625, "kernel": 0.09912109375}}, "final_layer_norm": {"bias": 0.011474609375, "scale": 0.01495361328125}, "layer_norm": {"bias": 0.0107421875, "scale": 0.0145263671875}}, "44": {"attention": {"k_proj": {"bias": 6.377696990966797e-06, "kernel": 0.02197265625}, "out_proj": {"bias": 0.005615234375, "kernel": 0.068359375}, "q_proj": {"bias": 0.00116729736328125, "kernel": 0.0225830078125}, "v_proj": {"bias": 0.006256103515625, "kernel": 0.068359375}}, "feed_forward": {"intermediate_dense": {"bias": 0.005523681640625, "kernel": 0.130859375}, "output_dense": {"bias": 0.005767822265625, "kernel": 0.0947265625}}, "final_layer_norm": {"bias": 0.01025390625, "scale": 0.01220703125}, "layer_norm": {"bias": 0.01318359375, "scale": 0.01171875}}, "45": {"attention": {"k_proj": {"bias": 8.881092071533203e-06, "kernel": 0.021240234375}, "out_proj": {"bias": 0.00579833984375, "kernel": 0.06884765625}, "q_proj": {"bias": 0.001190185546875, "kernel": 0.020751953125}, "v_proj": {"bias": 0.007049560546875, "kernel": 0.0732421875}}, "feed_forward": {"intermediate_dense": {"bias": 0.00543212890625, "kernel": 0.126953125}, "output_dense": {"bias": 0.00592041015625, "kernel": 0.0927734375}}, "final_layer_norm": {"bias": 0.01025390625, "scale": 0.013916015625}, "layer_norm": {"bias": 0.0179443359375, "scale": 0.021484375}}, "46": {"attention": {"k_proj": {"bias": 1.049041748046875e-05, "kernel": 0.02490234375}, "out_proj": {"bias": 0.0059814453125, "kernel": 0.0693359375}, "q_proj": {"bias": 0.0014495849609375, "kernel": 0.0230712890625}, "v_proj": {"bias": 0.00830078125, "kernel": 0.0849609375}}, "feed_forward": {"intermediate_dense": {"bias": 0.004913330078125, "kernel": 0.103515625}, "output_dense": {"bias": 0.00604248046875, "kernel": 0.0830078125}}, "final_layer_norm": {"bias": 0.0118408203125, "scale": 0.02001953125}, "layer_norm": {"bias": 0.02294921875, "scale": 0.01806640625}}, "47": {"attention": {"k_proj": {"bias": 2.2411346435546875e-05, "kernel": 0.027099609375}, "out_proj": {"bias": 0.006103515625, "kernel": 0.0615234375}, "q_proj": {"bias": 0.00177001953125, "kernel": 0.02294921875}, "v_proj": {"bias": 0.010009765625, "kernel": 0.09375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0045166015625, "kernel": 0.076171875}, "output_dense": {"bias": 0.0062255859375, "kernel": 0.07470703125}}, "final_layer_norm": {"bias": 0.0125732421875, "scale": 0.019775390625}, "layer_norm": {"bias": 0.032470703125, "scale": 0.0284423828125}}, "5": {"attention": {"k_proj": {"bias": 0.000118255615234375, "kernel": 0.251953125}, "out_proj": {"bias": 0.03662109375, "kernel": 0.515625}, "q_proj": {"bias": 0.0137939453125, "kernel": 0.2470703125}, "v_proj": {"bias": 0.056640625, "kernel": 0.52734375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0498046875, "kernel": 0.7890625}, "output_dense": {"bias": 0.03466796875, "kernel": 0.6015625}}, "final_layer_norm": {"bias": 0.087890625, "scale": 0.07373046875}, "layer_norm": {"bias": 0.0888671875, "scale": 0.10546875}}, "6": {"attention": {"k_proj": {"bias": 0.00016021728515625, "kernel": 0.296875}, "out_proj": {"bias": 0.03466796875, "kernel": 0.6953125}, "q_proj": {"bias": 0.0169677734375, "kernel": 0.28515625}, "v_proj": {"bias": 0.05908203125, "kernel": 0.734375}}, "feed_forward": {"intermediate_dense": {"bias": 0.0478515625, "kernel": 0.7890625}, "output_dense": {"bias": 0.032958984375, "kernel": 0.57421875}}, "final_layer_norm": {"bias": 0.0849609375, "scale": 0.08447265625}, "layer_norm": {"bias": 0.0927734375, "scale": 0.11865234375}}, "7": {"attention": {"k_proj": {"bias": 0.0002079010009765625, "kernel": 0.30078125}, "out_proj": {"bias": 0.03271484375, "kernel": 0.6796875}, "q_proj": {"bias": 0.018798828125, "kernel": 0.29296875}, "v_proj": {"bias": 0.05419921875, "kernel": 0.7265625}}, "feed_forward": {"intermediate_dense": {"bias": 0.048583984375, "kernel": 0.8125}, "output_dense": {"bias": 0.03125, "kernel": 0.5859375}}, "final_layer_norm": {"bias": 0.08984375, "scale": 0.09033203125}, "layer_norm": {"bias": 0.0966796875, "scale": 0.107421875}}, "8": {"attention": {"k_proj": {"bias": 0.0001468658447265625, "kernel": 0.27734375}, "out_proj": {"bias": 0.03076171875, "kernel": 0.58984375}, "q_proj": {"bias": 0.016357421875, "kernel": 0.275390625}, "v_proj": {"bias": 0.052001953125, "kernel": 0.62109375}}, "feed_forward": {"intermediate_dense": {"bias": 0.046630859375, "kernel": 0.7890625}, "output_dense": {"bias": 0.0294189453125, "kernel": 0.58984375}}, "final_layer_norm": {"bias": 0.080078125, "scale": 0.0830078125}, "layer_norm": {"bias": 0.08642578125, "scale": 0.12451171875}}, "9": {"attention": {"k_proj": {"bias": 0.00026702880859375, "kernel": 0.369140625}, "out_proj": {"bias": 0.0269775390625, "kernel": 0.77734375}, "q_proj": {"bias": 0.01904296875, "kernel": 0.36328125}, "v_proj": {"bias": 0.044921875, "kernel": 0.78125}}, "feed_forward": {"intermediate_dense": {"bias": 0.0390625, "kernel": 0.671875}, "output_dense": {"bias": 0.0263671875, "kernel": 0.53125}}, "final_layer_norm": {"bias": 0.06884765625, "scale": 0.068359375}, "layer_norm": {"bias": 0.08349609375, "scale": 0.0927734375}}}, "pos_conv_embed": {"conv": {"bias": 0.1298828125, "weight_g": 0.1240234375, "weight_v": 0.953125}}}, "feature_extractor": {"conv_layers": {"0": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "1": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "2": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "3": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "4": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "5": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}, "6": {"conv": {"bias": 0.0, "kernel": 0.0}, "layer_norm": {"bias": 0.0, "scale": 0.0}}}}, "feature_projection": {"layer_norm": {"bias": 0.34765625, "scale": 0.45703125}, "projection": {"bias": 0.169921875, "kernel": 2.96875}}, "masked_spec_embed": 0.0}}, "layer_param_norm/": {"lm_head": {"bias": 0.0668909028172493, "kernel": 5.564761161804199}, "wav2vec2": {"encoder": {"layer_norm": {"bias": 0.9411773681640625, "scale": 22.735441207885742}, "layers": {"0": {"attention": {"k_proj": {"bias": 0.1572660654783249, "kernel": 26.265689849853516}, "out_proj": {"bias": 1.540636420249939, "kernel": 25.332489013671875}, "q_proj": {"bias": 1.345062494277954, "kernel": 26.540752410888672}, "v_proj": {"bias": 0.34915056824684143, "kernel": 25.96399688720703}}, "feed_forward": {"intermediate_dense": {"bias": 1.7922512292861938, "kernel": 96.14887237548828}, "output_dense": {"bias": 1.0332554578781128, "kernel": 91.98905181884766}}, "final_layer_norm": {"bias": 1.3027501106262207, "scale": 19.902679443359375}, "layer_norm": {"bias": 3.2321395874023438, "scale": 15.86198616027832}}, "1": {"attention": {"k_proj": {"bias": 0.14874659478664398, "kernel": 39.97101593017578}, "out_proj": {"bias": 1.3095301389694214, "kernel": 41.88142776489258}, "q_proj": {"bias": 2.8878695964813232, "kernel": 39.850341796875}, "v_proj": {"bias": 0.2854360044002533, "kernel": 40.33198547363281}}, "feed_forward": {"intermediate_dense": {"bias": 1.6540393829345703, "kernel": 94.21043395996094}, "output_dense": {"bias": 0.8036744594573975, "kernel": 84.77149963378906}}, "final_layer_norm": {"bias": 1.2088514566421509, "scale": 18.529014587402344}, "layer_norm": {"bias": 1.8125758171081543, "scale": 18.96085548400879}}, "10": {"attention": {"k_proj": {"bias": 0.1729704886674881, "kernel": 47.594642639160156}, "out_proj": {"bias": 1.2517130374908447, "kernel": 50.368370056152344}, "q_proj": {"bias": 2.4282305240631104, "kernel": 47.5654411315918}, "v_proj": {"bias": 0.31612780690193176, "kernel": 50.5357780456543}}, "feed_forward": {"intermediate_dense": {"bias": 1.6828261613845825, "kernel": 98.35198974609375}, "output_dense": {"bias": 0.560376763343811, "kernel": 92.19532775878906}}, "final_layer_norm": {"bias": 2.2686586380004883, "scale": 20.38600730895996}, "layer_norm": {"bias": 1.7384451627731323, "scale": 22.122407913208008}}, "11": {"attention": {"k_proj": {"bias": 0.20578038692474365, "kernel": 47.44072723388672}, "out_proj": {"bias": 1.1011972427368164, "kernel": 49.64116287231445}, "q_proj": {"bias": 2.4754388332366943, "kernel": 47.175193786621094}, "v_proj": {"bias": 0.37477394938468933, "kernel": 50.14057922363281}}, "feed_forward": {"intermediate_dense": {"bias": 1.7321875095367432, "kernel": 99.11766052246094}, "output_dense": {"bias": 0.5437204837799072, "kernel": 93.87814331054688}}, "final_layer_norm": {"bias": 2.2578558921813965, "scale": 20.41146469116211}, "layer_norm": {"bias": 1.7143844366073608, "scale": 22.52808952331543}}, "12": {"attention": {"k_proj": {"bias": 0.1765490174293518, "kernel": 48.007171630859375}, "out_proj": {"bias": 1.0823392868041992, "kernel": 49.907554626464844}, "q_proj": {"bias": 2.351224899291992, "kernel": 47.764991760253906}, "v_proj": {"bias": 0.356992244720459, "kernel": 50.29539489746094}}, "feed_forward": {"intermediate_dense": {"bias": 1.7750531435012817, "kernel": 99.95246887207031}, "output_dense": {"bias": 0.5311572551727295, "kernel": 95.42961120605469}}, "final_layer_norm": {"bias": 2.2109756469726562, "scale": 20.368785858154297}, "layer_norm": {"bias": 1.7906843423843384, "scale": 23.005416870117188}}, "13": {"attention": {"k_proj": {"bias": 0.18811266124248505, "kernel": 49.8721923828125}, "out_proj": {"bias": 1.076762318611145, "kernel": 49.606414794921875}, "q_proj": {"bias": 2.322277307510376, "kernel": 49.74319076538086}, "v_proj": {"bias": 0.3691728115081787, "kernel": 49.76460647583008}}, "feed_forward": {"intermediate_dense": {"bias": 1.829641342163086, "kernel": 100.58248138427734}, "output_dense": {"bias": 0.5479011535644531, "kernel": 95.87431335449219}}, "final_layer_norm": {"bias": 2.109241485595703, "scale": 20.520580291748047}, "layer_norm": {"bias": 1.9025626182556152, "scale": 23.28399085998535}}, "14": {"attention": {"k_proj": {"bias": 0.22546514868736267, "kernel": 50.01702117919922}, "out_proj": {"bias": 1.2384552955627441, "kernel": 47.94068908691406}, "q_proj": {"bias": 2.382194995880127, "kernel": 50.0910758972168}, "v_proj": {"bias": 0.35839876532554626, "kernel": 47.508934020996094}}, "feed_forward": {"intermediate_dense": {"bias": 1.86549711227417, "kernel": 101.19500732421875}, "output_dense": {"bias": 0.5690709948539734, "kernel": 97.23810577392578}}, "final_layer_norm": {"bias": 2.2292520999908447, "scale": 20.66695213317871}, "layer_norm": {"bias": 2.0422415733337402, "scale": 23.39774513244629}}, "15": {"attention": {"k_proj": {"bias": 0.18015369772911072, "kernel": 50.14893341064453}, "out_proj": {"bias": 1.2943248748779297, "kernel": 48.63526916503906}, "q_proj": {"bias": 2.523224353790283, "kernel": 50.23434829711914}, "v_proj": {"bias": 0.4106982350349426, "kernel": 48.305335998535156}}, "feed_forward": {"intermediate_dense": {"bias": 1.8729243278503418, "kernel": 101.01643371582031}, "output_dense": {"bias": 0.7231183052062988, "kernel": 97.94534301757812}}, "final_layer_norm": {"bias": 2.1691672801971436, "scale": 20.768878936767578}, "layer_norm": {"bias": 2.275395154953003, "scale": 23.74152374267578}}, "16": {"attention": {"k_proj": {"bias": 0.1596199870109558, "kernel": 50.00080871582031}, "out_proj": {"bias": 1.217071771621704, "kernel": 48.02516555786133}, "q_proj": {"bias": 2.6123976707458496, "kernel": 49.893272399902344}, "v_proj": {"bias": 0.3556521236896515, "kernel": 47.71862030029297}}, "feed_forward": {"intermediate_dense": {"bias": 1.8646655082702637, "kernel": 101.64906311035156}, "output_dense": {"bias": 0.745720386505127, "kernel": 98.78334045410156}}, "final_layer_norm": {"bias": 2.2145400047302246, "scale": 21.247108459472656}, "layer_norm": {"bias": 2.196000814437866, "scale": 22.552139282226562}}, "17": {"attention": {"k_proj": {"bias": 0.15246188640594482, "kernel": 50.159278869628906}, "out_proj": {"bias": 1.1550219058990479, "kernel": 47.34062194824219}, "q_proj": {"bias": 2.700528144836426, "kernel": 50.237125396728516}, "v_proj": {"bias": 0.3945621848106384, "kernel": 47.01970672607422}}, "feed_forward": {"intermediate_dense": {"bias": 1.875906229019165, "kernel": 102.71083068847656}, "output_dense": {"bias": 0.763410210609436, "kernel": 99.16596984863281}}, "final_layer_norm": {"bias": 2.297891616821289, "scale": 21.794700622558594}, "layer_norm": {"bias": 2.0964274406433105, "scale": 21.963212966918945}}, "18": {"attention": {"k_proj": {"bias": 0.17540258169174194, "kernel": 50.5605583190918}, "out_proj": {"bias": 1.258694052696228, "kernel": 48.3692626953125}, "q_proj": {"bias": 2.578209400177002, "kernel": 50.937294006347656}, "v_proj": {"bias": 0.42091283202171326, "kernel": 47.91288757324219}}, "feed_forward": {"intermediate_dense": {"bias": 1.912561297416687, "kernel": 103.00012969970703}, "output_dense": {"bias": 0.8791664838790894, "kernel": 100.81649780273438}}, "final_layer_norm": {"bias": 2.39349102973938, "scale": 21.778369903564453}, "layer_norm": {"bias": 2.2818918228149414, "scale": 23.874561309814453}}, "19": {"attention": {"k_proj": {"bias": 0.14555136859416962, "kernel": 49.727783203125}, "out_proj": {"bias": 1.2327790260314941, "kernel": 48.223941802978516}, "q_proj": {"bias": 2.850454330444336, "kernel": 50.13501739501953}, "v_proj": {"bias": 0.3850022852420807, "kernel": 47.482051849365234}}, "feed_forward": {"intermediate_dense": {"bias": 1.962997555732727, "kernel": 103.55958557128906}, "output_dense": {"bias": 0.9461472034454346, "kernel": 101.71952819824219}}, "final_layer_norm": {"bias": 2.3411974906921387, "scale": 22.13956069946289}, "layer_norm": {"bias": 2.1903624534606934, "scale": 22.96629524230957}}, "2": {"attention": {"k_proj": {"bias": 0.16273045539855957, "kernel": 46.281925201416016}, "out_proj": {"bias": 1.224226713180542, "kernel": 44.15675735473633}, "q_proj": {"bias": 3.0906779766082764, "kernel": 46.02598571777344}, "v_proj": {"bias": 0.3187757134437561, "kernel": 44.112770080566406}}, "feed_forward": {"intermediate_dense": {"bias": 1.6842701435089111, "kernel": 99.14990234375}, "output_dense": {"bias": 0.6949976086616516, "kernel": 87.93550872802734}}, "final_layer_norm": {"bias": 1.4942196607589722, "scale": 21.074359893798828}, "layer_norm": {"bias": 1.6854312419891357, "scale": 21.76892852783203}}, "20": {"attention": {"k_proj": {"bias": 0.13836346566677094, "kernel": 49.7047004699707}, "out_proj": {"bias": 1.2614378929138184, "kernel": 47.5956916809082}, "q_proj": {"bias": 2.763533115386963, "kernel": 50.45436096191406}, "v_proj": {"bias": 0.35550954937934875, "kernel": 46.5172004699707}}, "feed_forward": {"intermediate_dense": {"bias": 1.9570611715316772, "kernel": 104.79476165771484}, "output_dense": {"bias": 1.0612233877182007, "kernel": 102.35823822021484}}, "final_layer_norm": {"bias": 2.3444409370422363, "scale": 23.082225799560547}, "layer_norm": {"bias": 2.165036201477051, "scale": 23.06981086730957}}, "21": {"attention": {"k_proj": {"bias": 0.15504375100135803, "kernel": 50.163543701171875}, "out_proj": {"bias": 1.3001418113708496, "kernel": 47.61540222167969}, "q_proj": {"bias": 2.7271313667297363, "kernel": 50.96747589111328}, "v_proj": {"bias": 0.4037085771560669, "kernel": 46.739601135253906}}, "feed_forward": {"intermediate_dense": {"bias": 1.9970109462738037, "kernel": 104.97686004638672}, "output_dense": {"bias": 1.140713095664978, "kernel": 102.67676544189453}}, "final_layer_norm": {"bias": 2.370567798614502, "scale": 22.737632751464844}, "layer_norm": {"bias": 2.2212581634521484, "scale": 23.29868507385254}}, "22": {"attention": {"k_proj": {"bias": 0.16382987797260284, "kernel": 50.56926727294922}, "out_proj": {"bias": 1.2247910499572754, "kernel": 47.1660041809082}, "q_proj": {"bias": 2.7855143547058105, "kernel": 50.92293930053711}, "v_proj": {"bias": 0.35865116119384766, "kernel": 47.0435791015625}}, "feed_forward": {"intermediate_dense": {"bias": 1.9342015981674194, "kernel": 105.3548583984375}, "output_dense": {"bias": 1.1557631492614746, "kernel": 101.9817886352539}}, "final_layer_norm": {"bias": 2.264819622039795, "scale": 22.244813919067383}, "layer_norm": {"bias": 2.229525566101074, "scale": 22.353832244873047}}, "23": {"attention": {"k_proj": {"bias": 0.21246662735939026, "kernel": 51.59439468383789}, "out_proj": {"bias": 1.364210844039917, "kernel": 48.22317886352539}, "q_proj": {"bias": 2.6520018577575684, "kernel": 51.69462585449219}, "v_proj": {"bias": 0.5232762098312378, "kernel": 48.86384582519531}}, "feed_forward": {"intermediate_dense": {"bias": 1.9091339111328125, "kernel": 105.17338562011719}, "output_dense": {"bias": 1.1327968835830688, "kernel": 102.7879638671875}}, "final_layer_norm": {"bias": 2.5350351333618164, "scale": 22.22771644592285}, "layer_norm": {"bias": 2.701894760131836, "scale": 23.687578201293945}}, "24": {"attention": {"k_proj": {"bias": 0.18005745112895966, "kernel": 50.28105926513672}, "out_proj": {"bias": 1.421328067779541, "kernel": 50.1175537109375}, "q_proj": {"bias": 2.7797858715057373, "kernel": 50.268226623535156}, "v_proj": {"bias": 0.481330931186676, "kernel": 50.19598388671875}}, "feed_forward": {"intermediate_dense": {"bias": 2.0205435752868652, "kernel": 104.63185119628906}, "output_dense": {"bias": 1.171934962272644, "kernel": 105.64089965820312}}, "final_layer_norm": {"bias": 2.61605167388916, "scale": 22.27845001220703}, "layer_norm": {"bias": 2.4518842697143555, "scale": 23.27135467529297}}, "25": {"attention": {"k_proj": {"bias": 0.1678708791732788, "kernel": 50.65785217285156}, "out_proj": {"bias": 1.2310528755187988, "kernel": 48.01655197143555}, "q_proj": {"bias": 2.8710856437683105, "kernel": 50.43537139892578}, "v_proj": {"bias": 0.5626442432403564, "kernel": 48.55765914916992}}, "feed_forward": {"intermediate_dense": {"bias": 1.9219928979873657, "kernel": 104.88790130615234}, "output_dense": {"bias": 1.0354797840118408, "kernel": 105.56619262695312}}, "final_layer_norm": {"bias": 2.343381881713867, "scale": 22.79658317565918}, "layer_norm": {"bias": 2.5976107120513916, "scale": 22.297609329223633}}, "26": {"attention": {"k_proj": {"bias": 0.18550752103328705, "kernel": 50.914405822753906}, "out_proj": {"bias": 1.1548503637313843, "kernel": 48.805328369140625}, "q_proj": {"bias": 2.851818561553955, "kernel": 50.67176818847656}, "v_proj": {"bias": 0.478745698928833, "kernel": 49.415489196777344}}, "feed_forward": {"intermediate_dense": {"bias": 2.024549961090088, "kernel": 104.31269836425781}, "output_dense": {"bias": 0.9997164607048035, "kernel": 102.69577026367188}}, "final_layer_norm": {"bias": 1.9872686862945557, "scale": 21.646724700927734}, "layer_norm": {"bias": 2.4832143783569336, "scale": 22.715808868408203}}, "27": {"attention": {"k_proj": {"bias": 0.41229814291000366, "kernel": 51.72254943847656}, "out_proj": {"bias": 1.3984119892120361, "kernel": 50.182186126708984}, "q_proj": {"bias": 2.616152048110962, "kernel": 51.59402847290039}, "v_proj": {"bias": 0.5837572813034058, "kernel": 50.64431381225586}}, "feed_forward": {"intermediate_dense": {"bias": 2.182586669921875, "kernel": 102.61892700195312}, "output_dense": {"bias": 0.8772892951965332, "kernel": 102.38434600830078}}, "final_layer_norm": {"bias": 2.2664685249328613, "scale": 20.938594818115234}, "layer_norm": {"bias": 2.558232307434082, "scale": 23.56865692138672}}, "28": {"attention": {"k_proj": {"bias": 0.4503825604915619, "kernel": 52.60496139526367}, "out_proj": {"bias": 1.418035864830017, "kernel": 50.870330810546875}, "q_proj": {"bias": 2.7903497219085693, "kernel": 52.25389862060547}, "v_proj": {"bias": 0.4630447328090668, "kernel": 51.217445373535156}}, "feed_forward": {"intermediate_dense": {"bias": 2.125363349914551, "kernel": 102.62171936035156}, "output_dense": {"bias": 0.7697524428367615, "kernel": 104.55252075195312}}, "final_layer_norm": {"bias": 2.181206703186035, "scale": 21.27033233642578}, "layer_norm": {"bias": 2.0507431030273438, "scale": 24.352882385253906}}, "29": {"attention": {"k_proj": {"bias": 0.17947150766849518, "kernel": 49.040016174316406}, "out_proj": {"bias": 1.3947250843048096, "kernel": 53.36112976074219}, "q_proj": {"bias": 2.733522415161133, "kernel": 48.861534118652344}, "v_proj": {"bias": 0.4220423698425293, "kernel": 53.27291488647461}}, "feed_forward": {"intermediate_dense": {"bias": 2.115380048751831, "kernel": 103.29501342773438}, "output_dense": {"bias": 0.8926779627799988, "kernel": 108.8265380859375}}, "final_layer_norm": {"bias": 2.3777034282684326, "scale": 22.42627716064453}, "layer_norm": {"bias": 2.17183256149292, "scale": 25.36501693725586}}, "3": {"attention": {"k_proj": {"bias": 0.20873260498046875, "kernel": 50.3943977355957}, "out_proj": {"bias": 1.3884034156799316, "kernel": 46.793052673339844}, "q_proj": {"bias": 2.741361618041992, "kernel": 50.61634063720703}, "v_proj": {"bias": 0.3030884861946106, "kernel": 47.16839599609375}}, "feed_forward": {"intermediate_dense": {"bias": 1.6966179609298706, "kernel": 100.75701141357422}, "output_dense": {"bias": 0.6564837694168091, "kernel": 90.81097412109375}}, "final_layer_norm": {"bias": 1.773003339767456, "scale": 21.166318893432617}, "layer_norm": {"bias": 1.8634135723114014, "scale": 23.474411010742188}}, "30": {"attention": {"k_proj": {"bias": 0.3080509305000305, "kernel": 50.9363899230957}, "out_proj": {"bias": 1.1957952976226807, "kernel": 49.6925163269043}, "q_proj": {"bias": 2.8136203289031982, "kernel": 51.00883102416992}, "v_proj": {"bias": 0.4762360453605652, "kernel": 50.050315856933594}}, "feed_forward": {"intermediate_dense": {"bias": 2.048686981201172, "kernel": 103.77497863769531}, "output_dense": {"bias": 0.8379020690917969, "kernel": 107.83690643310547}}, "final_layer_norm": {"bias": 2.2059359550476074, "scale": 23.55939483642578}, "layer_norm": {"bias": 2.2876853942871094, "scale": 25.086034774780273}}, "31": {"attention": {"k_proj": {"bias": 0.3784014582633972, "kernel": 49.39188003540039}, "out_proj": {"bias": 1.119596242904663, "kernel": 50.44926834106445}, "q_proj": {"bias": 2.5993165969848633, "kernel": 49.48455810546875}, "v_proj": {"bias": 0.522638738155365, "kernel": 50.58903884887695}}, "feed_forward": {"intermediate_dense": {"bias": 2.1306161880493164, "kernel": 102.43810272216797}, "output_dense": {"bias": 1.0122932195663452, "kernel": 105.18507385253906}}, "final_layer_norm": {"bias": 2.1017799377441406, "scale": 23.437347412109375}, "layer_norm": {"bias": 2.281771183013916, "scale": 24.799633026123047}}, "32": {"attention": {"k_proj": {"bias": 0.25974252820014954, "kernel": 48.31093215942383}, "out_proj": {"bias": 1.1245567798614502, "kernel": 49.66279602050781}, "q_proj": {"bias": 2.827341079711914, "kernel": 48.274757385253906}, "v_proj": {"bias": 0.39349639415740967, "kernel": 49.95775604248047}}, "feed_forward": {"intermediate_dense": {"bias": 2.0681939125061035, "kernel": 101.28511810302734}, "output_dense": {"bias": 1.0686500072479248, "kernel": 104.53141784667969}}, "final_layer_norm": {"bias": 2.0827746391296387, "scale": 23.84016227722168}, "layer_norm": {"bias": 2.262401580810547, "scale": 25.117748260498047}}, "33": {"attention": {"k_proj": {"bias": 0.26482468843460083, "kernel": 48.2266960144043}, "out_proj": {"bias": 1.1604676246643066, "kernel": 49.49713897705078}, "q_proj": {"bias": 2.972139835357666, "kernel": 48.19778060913086}, "v_proj": {"bias": 0.41569650173187256, "kernel": 49.77226257324219}}, "feed_forward": {"intermediate_dense": {"bias": 2.076352834701538, "kernel": 99.64717864990234}, "output_dense": {"bias": 1.0581281185150146, "kernel": 103.25572204589844}}, "final_layer_norm": {"bias": 1.994544506072998, "scale": 23.598308563232422}, "layer_norm": {"bias": 2.439363956451416, "scale": 25.341171264648438}}, "34": {"attention": {"k_proj": {"bias": 0.2755166292190552, "kernel": 47.42335510253906}, "out_proj": {"bias": 1.4115952253341675, "kernel": 50.949275970458984}, "q_proj": {"bias": 2.8844351768493652, "kernel": 47.45287322998047}, "v_proj": {"bias": 0.38429465889930725, "kernel": 50.88615417480469}}, "feed_forward": {"intermediate_dense": {"bias": 2.164264678955078, "kernel": 98.45794677734375}, "output_dense": {"bias": 0.9942412972450256, "kernel": 102.51951599121094}}, "final_layer_norm": {"bias": 1.9257347583770752, "scale": 23.230693817138672}, "layer_norm": {"bias": 2.5131359100341797, "scale": 25.724273681640625}}, "35": {"attention": {"k_proj": {"bias": 0.3880418837070465, "kernel": 49.17444610595703}, "out_proj": {"bias": 1.338890552520752, "kernel": 49.785396575927734}, "q_proj": {"bias": 2.5939016342163086, "kernel": 49.47700500488281}, "v_proj": {"bias": 0.4775533974170685, "kernel": 49.64286422729492}}, "feed_forward": {"intermediate_dense": {"bias": 2.2472829818725586, "kernel": 97.02595520019531}, "output_dense": {"bias": 0.8886467218399048, "kernel": 101.20555877685547}}, "final_layer_norm": {"bias": 1.9989571571350098, "scale": 23.346912384033203}, "layer_norm": {"bias": 2.3170738220214844, "scale": 26.273502349853516}}, "36": {"attention": {"k_proj": {"bias": 0.23954293131828308, "kernel": 46.442291259765625}, "out_proj": {"bias": 1.3719561100006104, "kernel": 51.094085693359375}, "q_proj": {"bias": 2.6886184215545654, "kernel": 46.40093994140625}, "v_proj": {"bias": 0.3625168800354004, "kernel": 51.2959098815918}}, "feed_forward": {"intermediate_dense": {"bias": 2.129176378250122, "kernel": 96.0584716796875}, "output_dense": {"bias": 0.9140334129333496, "kernel": 100.84213256835938}}, "final_layer_norm": {"bias": 1.6761391162872314, "scale": 23.85495376586914}, "layer_norm": {"bias": 2.0316162109375, "scale": 25.767738342285156}}, "37": {"attention": {"k_proj": {"bias": 0.5490036010742188, "kernel": 45.51734924316406}, "out_proj": {"bias": 1.6235731840133667, "kernel": 51.08349609375}, "q_proj": {"bias": 2.3980464935302734, "kernel": 45.56562805175781}, "v_proj": {"bias": 0.35822078585624695, "kernel": 50.9661979675293}}, "feed_forward": {"intermediate_dense": {"bias": 2.023285388946533, "kernel": 95.24488830566406}, "output_dense": {"bias": 0.9175524115562439, "kernel": 100.54832458496094}}, "final_layer_norm": {"bias": 1.4775093793869019, "scale": 24.243749618530273}, "layer_norm": {"bias": 1.9990556240081787, "scale": 25.78732681274414}}, "38": {"attention": {"k_proj": {"bias": 0.6277763843536377, "kernel": 43.699893951416016}, "out_proj": {"bias": 1.3179919719696045, "kernel": 50.56129837036133}, "q_proj": {"bias": 2.3240532875061035, "kernel": 43.691986083984375}, "v_proj": {"bias": 0.4102995991706848, "kernel": 50.447662353515625}}, "feed_forward": {"intermediate_dense": {"bias": 1.9623992443084717, "kernel": 93.24853515625}, "output_dense": {"bias": 0.9016005992889404, "kernel": 98.77943420410156}}, "final_layer_norm": {"bias": 1.512964129447937, "scale": 24.965984344482422}, "layer_norm": {"bias": 2.1805269718170166, "scale": 26.572105407714844}}, "39": {"attention": {"k_proj": {"bias": 0.6627014875411987, "kernel": 43.5059814453125}, "out_proj": {"bias": 1.6037161350250244, "kernel": 50.41535186767578}, "q_proj": {"bias": 2.112717628479004, "kernel": 43.87850570678711}, "v_proj": {"bias": 0.390166699886322, "kernel": 50.09760665893555}}, "feed_forward": {"intermediate_dense": {"bias": 1.9482412338256836, "kernel": 91.50100708007812}, "output_dense": {"bias": 0.9774587750434875, "kernel": 99.10246276855469}}, "final_layer_norm": {"bias": 1.6250338554382324, "scale": 25.600841522216797}, "layer_norm": {"bias": 2.1454267501831055, "scale": 27.16954803466797}}, "4": {"attention": {"k_proj": {"bias": 0.21773464977741241, "kernel": 52.96697998046875}, "out_proj": {"bias": 1.5798711776733398, "kernel": 48.24322509765625}, "q_proj": {"bias": 2.5302865505218506, "kernel": 53.159481048583984}, "v_proj": {"bias": 0.3523319363594055, "kernel": 48.58223342895508}}, "feed_forward": {"intermediate_dense": {"bias": 1.6854965686798096, "kernel": 100.35298156738281}, "output_dense": {"bias": 0.8151395320892334, "kernel": 92.05111694335938}}, "final_layer_norm": {"bias": 1.8813955783843994, "scale": 20.698898315429688}, "layer_norm": {"bias": 1.9626739025115967, "scale": 23.951351165771484}}, "40": {"attention": {"k_proj": {"bias": 0.5965253114700317, "kernel": 42.784297943115234}, "out_proj": {"bias": 1.54435133934021, "kernel": 49.03851318359375}, "q_proj": {"bias": 2.0332603454589844, "kernel": 43.529693603515625}, "v_proj": {"bias": 0.44385749101638794, "kernel": 48.619354248046875}}, "feed_forward": {"intermediate_dense": {"bias": 1.8044453859329224, "kernel": 89.7060317993164}, "output_dense": {"bias": 1.0290844440460205, "kernel": 96.339111328125}}, "final_layer_norm": {"bias": 1.7928125858306885, "scale": 24.868694305419922}, "layer_norm": {"bias": 2.09533953666687, "scale": 26.69829750061035}}, "41": {"attention": {"k_proj": {"bias": 1.676910638809204, "kernel": 40.18012237548828}, "out_proj": {"bias": 1.3033416271209717, "kernel": 50.59858322143555}, "q_proj": {"bias": 1.720700979232788, "kernel": 40.91082000732422}, "v_proj": {"bias": 0.3962639570236206, "kernel": 49.54127502441406}}, "feed_forward": {"intermediate_dense": {"bias": 1.9326260089874268, "kernel": 86.49552154541016}, "output_dense": {"bias": 1.0626506805419922, "kernel": 95.35970306396484}}, "final_layer_norm": {"bias": 2.287132740020752, "scale": 28.329570770263672}, "layer_norm": {"bias": 2.0972471237182617, "scale": 28.477943420410156}}, "42": {"attention": {"k_proj": {"bias": 0.8007075786590576, "kernel": 36.84203338623047}, "out_proj": {"bias": 1.3564722537994385, "kernel": 44.805450439453125}, "q_proj": {"bias": 1.545430302619934, "kernel": 38.173866271972656}, "v_proj": {"bias": 0.6166673898696899, "kernel": 43.14790344238281}}, "feed_forward": {"intermediate_dense": {"bias": 1.6966503858566284, "kernel": 85.47465515136719}, "output_dense": {"bias": 1.1137490272521973, "kernel": 93.52703857421875}}, "final_layer_norm": {"bias": 1.9977667331695557, "scale": 29.619346618652344}, "layer_norm": {"bias": 1.5640318393707275, "scale": 27.318359375}}, "43": {"attention": {"k_proj": {"bias": 1.2139406204223633, "kernel": 33.34306335449219}, "out_proj": {"bias": 1.3445727825164795, "kernel": 41.18675231933594}, "q_proj": {"bias": 1.3703632354736328, "kernel": 34.14249038696289}, "v_proj": {"bias": 0.5302899479866028, "kernel": 39.08563995361328}}, "feed_forward": {"intermediate_dense": {"bias": 1.7241705656051636, "kernel": 84.66365051269531}, "output_dense": {"bias": 0.875501275062561, "kernel": 91.44625091552734}}, "final_layer_norm": {"bias": 1.9585936069488525, "scale": 31.831100463867188}, "layer_norm": {"bias": 1.6979660987854004, "scale": 25.48971939086914}}, "44": {"attention": {"k_proj": {"bias": 2.492720365524292, "kernel": 33.942161560058594}, "out_proj": {"bias": 1.106978416442871, "kernel": 44.9072265625}, "q_proj": {"bias": 1.3083606958389282, "kernel": 34.30083084106445}, "v_proj": {"bias": 0.38603541254997253, "kernel": 44.0037841796875}}, "feed_forward": {"intermediate_dense": {"bias": 1.799367070198059, "kernel": 83.60717010498047}, "output_dense": {"bias": 0.8168225288391113, "kernel": 89.09458923339844}}, "final_layer_norm": {"bias": 1.9565825462341309, "scale": 34.001800537109375}, "layer_norm": {"bias": 1.5935062170028687, "scale": 25.508739471435547}}, "45": {"attention": {"k_proj": {"bias": 2.0522377490997314, "kernel": 33.76741027832031}, "out_proj": {"bias": 0.9882014989852905, "kernel": 48.50224304199219}, "q_proj": {"bias": 1.3905718326568604, "kernel": 33.92951965332031}, "v_proj": {"bias": 0.4323008954524994, "kernel": 48.66568374633789}}, "feed_forward": {"intermediate_dense": {"bias": 1.914729356765747, "kernel": 80.2913589477539}, "output_dense": {"bias": 0.9462040662765503, "kernel": 84.49946594238281}}, "final_layer_norm": {"bias": 1.6955138444900513, "scale": 32.71482467651367}, "layer_norm": {"bias": 1.5199291706085205, "scale": 24.034631729125977}}, "46": {"attention": {"k_proj": {"bias": 1.5418241024017334, "kernel": 34.94271469116211}, "out_proj": {"bias": 0.7529411315917969, "kernel": 50.925716400146484}, "q_proj": {"bias": 1.5559134483337402, "kernel": 35.040340423583984}, "v_proj": {"bias": 0.3770799934864044, "kernel": 51.6875}}, "feed_forward": {"intermediate_dense": {"bias": 1.9622228145599365, "kernel": 74.68043518066406}, "output_dense": {"bias": 1.1015828847885132, "kernel": 74.83889770507812}}, "final_layer_norm": {"bias": 1.6539316177368164, "scale": 28.228740692138672}, "layer_norm": {"bias": 1.3381633758544922, "scale": 22.947879791259766}}, "47": {"attention": {"k_proj": {"bias": 0.2917669415473938, "kernel": 37.277713775634766}, "out_proj": {"bias": 0.6384203433990479, "kernel": 45.20857620239258}, "q_proj": {"bias": 1.6836624145507812, "kernel": 37.879981994628906}, "v_proj": {"bias": 0.3526262640953064, "kernel": 46.196876525878906}}, "feed_forward": {"intermediate_dense": {"bias": 2.025301933288574, "kernel": 72.08566284179688}, "output_dense": {"bias": 0.6106955409049988, "kernel": 68.43527221679688}}, "final_layer_norm": {"bias": 1.548535704612732, "scale": 23.066944122314453}, "layer_norm": {"bias": 1.059852123260498, "scale": 20.202987670898438}}, "5": {"attention": {"k_proj": {"bias": 0.16882865130901337, "kernel": 48.27493667602539}, "out_proj": {"bias": 1.5571939945220947, "kernel": 49.373023986816406}, "q_proj": {"bias": 2.630668878555298, "kernel": 48.41212463378906}, "v_proj": {"bias": 0.31896352767944336, "kernel": 50.12682342529297}}, "feed_forward": {"intermediate_dense": {"bias": 1.6141834259033203, "kernel": 100.44667053222656}, "output_dense": {"bias": 0.8445441722869873, "kernel": 91.35646057128906}}, "final_layer_norm": {"bias": 2.1409213542938232, "scale": 20.886255264282227}, "layer_norm": {"bias": 2.0031042098999023, "scale": 23.135677337646484}}, "6": {"attention": {"k_proj": {"bias": 0.2727879285812378, "kernel": 50.084373474121094}, "out_proj": {"bias": 1.5454241037368774, "kernel": 48.77831268310547}, "q_proj": {"bias": 2.6858837604522705, "kernel": 50.5609130859375}, "v_proj": {"bias": 0.3207029104232788, "kernel": 49.28154754638672}}, "feed_forward": {"intermediate_dense": {"bias": 1.5875343084335327, "kernel": 99.54606628417969}, "output_dense": {"bias": 0.7026320099830627, "kernel": 90.97244262695312}}, "final_layer_norm": {"bias": 2.438199758529663, "scale": 20.364471435546875}, "layer_norm": {"bias": 1.9867676496505737, "scale": 23.706497192382812}}, "7": {"attention": {"k_proj": {"bias": 0.2606049180030823, "kernel": 49.72987365722656}, "out_proj": {"bias": 1.3674311637878418, "kernel": 49.03389358520508}, "q_proj": {"bias": 2.4531450271606445, "kernel": 50.121490478515625}, "v_proj": {"bias": 0.4057798385620117, "kernel": 48.98875427246094}}, "feed_forward": {"intermediate_dense": {"bias": 1.5941696166992188, "kernel": 99.27568054199219}, "output_dense": {"bias": 0.5401815176010132, "kernel": 90.70722961425781}}, "final_layer_norm": {"bias": 2.292677879333496, "scale": 20.583412170410156}, "layer_norm": {"bias": 1.9037420749664307, "scale": 22.42470932006836}}, "8": {"attention": {"k_proj": {"bias": 0.24493072926998138, "kernel": 49.249691009521484}, "out_proj": {"bias": 1.188157558441162, "kernel": 49.54103088378906}, "q_proj": {"bias": 2.4314587116241455, "kernel": 49.03004837036133}, "v_proj": {"bias": 0.33289897441864014, "kernel": 49.696128845214844}}, "feed_forward": {"intermediate_dense": {"bias": 1.6514670848846436, "kernel": 98.86155700683594}, "output_dense": {"bias": 0.5011764168739319, "kernel": 90.14228057861328}}, "final_layer_norm": {"bias": 2.2437543869018555, "scale": 20.356639862060547}, "layer_norm": {"bias": 1.839092493057251, "scale": 22.84292221069336}}, "9": {"attention": {"k_proj": {"bias": 0.2741130292415619, "kernel": 50.02962112426758}, "out_proj": {"bias": 1.4082090854644775, "kernel": 50.41935348510742}, "q_proj": {"bias": 2.3752646446228027, "kernel": 50.20918655395508}, "v_proj": {"bias": 0.35025060176849365, "kernel": 50.80641174316406}}, "feed_forward": {"intermediate_dense": {"bias": 1.730910062789917, "kernel": 97.50648498535156}, "output_dense": {"bias": 0.6366411447525024, "kernel": 90.64373779296875}}, "final_layer_norm": {"bias": 2.162881851196289, "scale": 19.669281005859375}, "layer_norm": {"bias": 1.9278013706207275, "scale": 24.389545440673828}}}, "pos_conv_embed": {"conv": {"bias": 5.617927551269531, "weight_g": 8.927756309509277, "weight_v": 85.1714096069336}}}, "feature_extractor": {"conv_layers": {"0": {"conv": {"bias": 2.0290679931640625, "kernel": 20.55536460876465}, "layer_norm": {"bias": 4.550922393798828, "scale": 16.167570114135742}}, "1": {"conv": {"bias": 1.7790228128433228, "kernel": 51.24136734008789}, "layer_norm": {"bias": 5.962646961212158, "scale": 23.268157958984375}}, "2": {"conv": {"bias": 1.140576720237732, "kernel": 46.50312042236328}, "layer_norm": {"bias": 4.176670551300049, "scale": 20.370853424072266}}, "3": {"conv": {"bias": 0.6725863218307495, "kernel": 44.397525787353516}, "layer_norm": {"bias": 3.888174533843994, "scale": 17.53795051574707}}, "4": {"conv": {"bias": 0.6373162269592285, "kernel": 41.314056396484375}, "layer_norm": {"bias": 2.385471820831299, "scale": 16.34571647644043}}, "5": {"conv": {"bias": 0.5147221684455872, "kernel": 37.479759216308594}, "layer_norm": {"bias": 2.020900011062622, "scale": 17.064470291137695}}, "6": {"conv": {"bias": 0.4947893023490906, "kernel": 40.64780044555664}, "layer_norm": {"bias": 0.5876954793930054, "scale": 19.058603286743164}}}}, "feature_projection": {"layer_norm": {"bias": 6.331979751586914, "scale": 16.56066131591797}, "projection": {"bias": 1.7012581825256348, "kernel": 35.14557647705078}}, "masked_spec_embed": 11.914372444152832}}, "train/learning_rate": 9.961725299945101e-05, "train/loss": 0.8391348123550415, "train/param_norm": 1192.62841796875, "_runtime": 8064, "_timestamp": 1659748575, "_step": 4900}

wandb/run-20220805_230151-2y71vcu4/logs/debug-internal.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1418c615d6b95002fa3e56523498468b78963e0fc761d193f43d63503270a8d6
+size 676709

wandb/run-20220805_230151-2y71vcu4/logs/debug.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:920d65c1c30ad3dfc773632222a1eab21e7f75a6783383b410e1f6c3b2db661a
+size 2637

wandb/run-20220805_230151-2y71vcu4/run-2y71vcu4.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95295b8d59f586b5d5d84e048c8d5ed02c82c7ecf568b944db54ed22b8ce655f
+size 3936514