up

config.json

@@ -0,0 +1,273 @@
+{
+  "architectures": [
+    "SpeechEncoderDecoderModel"
+  ],
+  "decoder": {
+    "_name_or_path": "facebook/bart-large-cnn",
+    "_num_labels": 3,
+    "activation_dropout": 0.0,
+    "activation_function": "gelu",
+    "add_cross_attention": true,
+    "add_final_layer_norm": false,
+    "architectures": [
+      "BartForConditionalGeneration"
+    ],
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "bos_token_id": 0,
+    "chunk_size_feed_forward": 0,
+    "classif_dropout": 0.0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": null,
+    "d_model": 1024,
+    "decoder_attention_heads": 16,
+    "decoder_ffn_dim": 4096,
+    "decoder_layerdrop": 0.0,
+    "decoder_layers": 12,
+    "decoder_start_token_id": 2,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": true,
+    "encoder_attention_heads": 16,
+    "encoder_ffn_dim": 4096,
+    "encoder_layerdrop": 0.0,
+    "encoder_layers": 12,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "finetuning_task": null,
+    "force_bos_token_to_be_generated": true,
+    "forced_bos_token_id": 0,
+    "forced_eos_token_id": 2,
+    "gradient_checkpointing": false,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1",
+      "2": "LABEL_2"
+    },
+    "init_std": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1,
+      "LABEL_2": 2
+    },
+    "length_penalty": 2.0,
+    "max_length": 142,
+    "max_position_embeddings": 1024,
+    "min_length": 56,
+    "model_type": "bart",
+    "no_repeat_ngram_size": 3,
+    "normalize_before": false,
+    "num_beam_groups": 1,
+    "num_beams": 4,
+    "num_hidden_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_past": true,
+    "output_scores": false,
+    "pad_token_id": 1,
+    "prefix": " ",
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "scale_embedding": false,
+    "sep_token_id": null,
+    "task_specific_params": {
+      "summarization": {
+        "early_stopping": true,
+        "length_penalty": 2.0,
+        "max_length": 142,
+        "min_length": 56,
+        "no_repeat_ngram_size": 3,
+        "num_beams": 4
+      }
+    },
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "transformers_version": "4.18.0.dev0",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 50264
+  },
+  "decoder_start_token_id": 0,
+  "encoder": {
+    "_name_or_path": "facebook/wav2vec2-large-lv60",
+    "activation_dropout": 0.1,
+    "adapter_kernel_size": 3,
+    "adapter_stride": 2,
+    "add_adapter": true,
+    "add_cross_attention": false,
+    "apply_spec_augment": true,
+    "architectures": [
+      "Wav2Vec2ForPreTraining"
+    ],
+    "attention_dropout": 0.1,
+    "bad_words_ids": null,
+    "bos_token_id": 1,
+    "chunk_size_feed_forward": 0,
+    "classifier_proj_size": 256,
+    "codevector_dim": 768,
+    "contrastive_logits_temperature": 0.1,
+    "conv_bias": true,
+    "conv_dim": [
+      512,
+      512,
+      512,
+      512,
+      512,
+      512,
+      512
+    ],
+    "conv_kernel": [
+      10,
+      3,
+      3,
+      3,
+      3,
+      2,
+      2
+    ],
+    "conv_stride": [
+      5,
+      2,
+      2,
+      2,
+      2,
+      2,
+      2
+    ],
+    "cross_attention_hidden_size": null,
+    "ctc_loss_reduction": "sum",
+    "ctc_zero_infinity": false,
+    "decoder_start_token_id": null,
+    "diversity_loss_weight": 0.1,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "do_stable_layer_norm": true,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "feat_extract_activation": "gelu",
+    "feat_extract_dropout": 0.0,
+    "feat_extract_norm": "layer",
+    "feat_proj_dropout": 0.0,
+    "feat_quantizer_dropout": 0.0,
+    "final_dropout": 0.0,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "gradient_checkpointing": false,
+    "hidden_act": "gelu",
+    "hidden_dropout": 0.1,
+    "hidden_dropout_prob": 0.1,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_eps": 1e-05,
+    "layerdrop": 0.0,
+    "length_penalty": 1.0,
+    "mask_feature_length": 10,
+    "mask_feature_min_masks": 0,
+    "mask_feature_prob": 0.0,
+    "mask_time_length": 10,
+    "mask_time_min_masks": 2,
+    "mask_time_prob": 0.1,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "wav2vec2",
+    "no_repeat_ngram_size": 0,
+    "num_adapter_layers": 3,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_codevector_groups": 2,
+    "num_codevectors_per_group": 320,
+    "num_conv_pos_embedding_groups": 16,
+    "num_conv_pos_embeddings": 128,
+    "num_feat_extract_layers": 7,
+    "num_hidden_layers": 24,
+    "num_negatives": 100,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_size": 1024,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 0,
+    "prefix": null,
+    "problem_type": null,
+    "proj_codevector_dim": 768,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "task_specific_params": null,
+    "tdnn_dilation": [
+      1,
+      2,
+      3,
+      1,
+      1
+    ],
+    "tdnn_dim": [
+      512,
+      512,
+      512,
+      512,
+      1500
+    ],
+    "tdnn_kernel": [
+      5,
+      3,
+      3,
+      1,
+      1
+    ],
+    "temperature": 1.0,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "transformers_version": "4.18.0.dev0",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_weighted_layer_sum": false,
+    "vocab_size": 32,
+    "xvector_output_dim": 512
+  },
+  "eos_token_id": 2,
+  "is_encoder_decoder": true,
+  "max_length": 40,
+  "model_type": "speech-encoder-decoder",
+  "pad_token_id": 1,
+  "processor_class": "Wav2Vec2Processor",
+  "transformers_version": null,
+  "use_cache": false
+}

create_model.py

@@ -0,0 +1,30 @@
+import jax.numpy as jnp
+from transformers import AutoFeatureExtractor, AutoTokenizer, FlaxSpeechEncoderDecoderModel
+
+
+encoder_id = "facebook/wav2vec2-large-lv60"
+decoder_id = "facebook/bart-large-cnn"
+
+model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)
+
+model.config.encoder.feat_proj_dropout = 0.0
+model.config.encoder.final_dropout = 0.0
+model.config.encoder.mask_time_prob = 0.1
+model.config.decoder_start_token_id = model.config.decoder.bos_token_id
+model.config.pad_token_id = model.config.decoder.pad_token_id
+model.config.eos_token_id = model.config.decoder.eos_token_id
+model.config.max_length = 40
+model.config.num_beams = 1
+model.config.encoder.layerdrop = 0.0
+model.config.use_cache = False
+model.config.processor_class = "Wav2Vec2Processor"
+
+# check if generation works
+out = model.generate(jnp.ones((1, 2000)))
+
+model.save_pretrained("./")
+
+feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
+feature_extractor.save_pretrained("./")
+tokenizer = AutoTokenizer.from_pretrained(decoder_id)
+tokenizer.save_pretrained("./")

preprocessor_config.json

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

run_flax_speech_recognition_seq2seq.py

@@ -0,0 +1,897 @@
+#!/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 sequence to sequence 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 os
+import sys
+import time
+from dataclasses import field
+from functools import partial
+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
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from huggingface_hub import Repository
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoProcessor,
+    AutoTokenizer,
+    FlaxAutoModelForSpeechSeq2Seq,
+    HfArgumentParser,
+    Seq2SeqTrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.file_utils import get_full_repo_name
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.17.0.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."}
+    )
+
+
+@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)."},
+    )
+    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."
+        },
+    )
+    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_target_length: Optional[int] = field(
+        default=128,
+        metadata={
+            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    min_target_length: Optional[int] = field(
+        default=0,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+            "than this will be filtered."
+        },
+    )
+    pad_input_to_multiple_of: Optional[int] = field(
+        default=None,
+        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="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    do_lower_case: bool = field(
+        default=True,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def shift_tokens_right(label_ids: np.array, decoder_start_token_id: int) -> np.ndarray:
+    """
+    Shift label ids one token to the right.
+    """
+    shifted_label_ids = np.zeros_like(label_ids)
+    shifted_label_ids[:, 1:] = label_ids[:, :-1]
+    shifted_label_ids[:, 0] = decoder_start_token_id
+
+    return shifted_label_ids
+
+
+@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 (`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).
+        max_target_length (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` 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
+    decoder_start_token_id: int
+    input_padding: Union[bool, str] = "max_length"
+    target_padding: Union[bool, str] = "max_length"
+    max_input_length: Optional[float] = None
+    max_target_length: Optional[int] = None
+    pad_input_to_multiple_of: Optional[int] = None
+    pad_target_to_multiple_of: Optional[int] = 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_target_length,
+            padding=self.target_padding,
+            pad_to_multiple_of=self.pad_target_to_multiple_of,
+            return_tensors="np",
+        )
+
+        # if bos token is appended in previous tokenization step,
+        # cut bos token here as it's append later anyways
+        labels = labels_batch["input_ids"]
+        if (labels[:, 0] == self.decoder_start_token_id).all().item():
+            labels = labels[:, 1:]
+            labels_batch.attention_mask = labels_batch.attention_mask[:, 1:]
+
+        decoder_input_ids = shift_tokens_right(labels, self.decoder_start_token_id)
+
+        # replace padding with -100 to ignore loss correctly
+        labels = np.ma.array(labels, mask=np.not_equal(labels_batch.attention_mask, 1))
+        labels = labels.filled(fill_value=-100)
+
+        batch["inputs"] = batch.pop("input_values")
+        batch["labels"] = labels
+        batch["decoder_input_ids"] = decoder_input_ids
+        # decoder_attention_mask known to give issues with nan's
+        # remove decoder_attention_mask as an arg for the time being - handled by the causal mask in XXXForCausalLM
+        # batch["decoder_attention_mask"] = labels_batch.attention_mask
+
+        return batch
+
+
+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):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    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 generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    num_samples = len(samples_idx)
+    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
+    batch_idx = np.split(samples_idx, sections_split)
+    return batch_idx
+
+
+def main():
+    # 1. Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # 2. Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    # 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()
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    logger.info(f"JAX devices: {jax.device_count()}")
+
+    # 3. Detecting last checkpoint and eventually continue from last checkpoint
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # 4. Load dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, split=data_args.train_split_name
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, split=data_args.eval_split_name
+        )
+
+    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 = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+    model = FlaxAutoModelForSpeechSeq2Seq.from_pretrained(
+        model_args.model_name_or_path,
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    # 6. Resample speech dataset if necessary
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    max_input_length = 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_target_length
+    min_target_length = data_args.min_target_length
+    pad_input_to_multiple_of = data_args.pad_input_to_multiple_of
+    pad_target_to_multiple_of = data_args.pad_target_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
+
+    if data_args.max_train_samples is not None:
+        raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    def prepare_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.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]
+        batch["labels"] = tokenizer(input_str).input_ids
+        batch["labels_length"] = len(batch["labels"])
+        return batch
+
+    with training_args.main_process_first(desc="dataset map pre-processing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=data_args.preprocessing_num_workers,
+            desc="preprocess train 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 Metric
+    metric = load_metric("wer")
+
+    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, skip_special_tokens=True)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(padded_ids, skip_special_tokens=True)
+
+        wer = metric.compute(predictions=pred_str, references=label_str)
+
+        return {"wer": wer}
+
+    # 9. Create a single speech processor
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+
+    data_collator = FlaxDataCollatorSpeechSeq2SeqWithPadding(
+        processor=processor,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+        input_padding="max_length",
+        target_padding="max_length",
+        max_input_length=max_input_length,
+        max_target_length=max_target_length,
+        pad_input_to_multiple_of=pad_input_to_multiple_of,
+        pad_target_to_multiple_of=pad_target_to_multiple_of,
+    )
+
+    # 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:
+        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 constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(vectorized_datasets["train"]) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+    gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(vectorized_datasets["train"]),
+        train_batch_size,
+        training_args.num_train_epochs,
+        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 FlaxBart.
+    # For FlaxT5, one should correct the layer norm parameter naming
+    # accordingly - see `run_t5_mlm_flax.py` e.g.
+    # TODO: check param dictionary of encoder and decoder match the layer_norm_params list
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        layer_norm_params = [
+            (name, "scale") for name in ["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)
+
+    # create adam optimizer
+    adamw = 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,
+    )
+
+    # augment adam optimizer to facilitate gradient accumulation (ignore for now)
+    # optim = optax.chain(adamw, optax.apply_every(gradient_accumulation_steps))
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+
+    # label smoothed cross entropy
+    def loss_fn(logits, labels, label_smoothing_factor=0.0):
+        """
+        The label smoothing implementation is adapted from Flax's official example:
+        https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
+        """
+        vocab_size = logits.shape[-1]
+        confidence = 1.0 - label_smoothing_factor
+        low_confidence = (1.0 - confidence) / (vocab_size - 1)
+        normalizing_constant = -(
+            confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
+        )
+        soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence)
+
+        loss = optax.softmax_cross_entropy(logits, soft_labels)
+        loss = loss - normalizing_constant
+
+        # ignore padded tokens from loss, i.e. where labels are not set to -100
+        padding = labels > 0
+        loss = loss * padding
+        loss = loss.sum() / padding.sum()
+        return loss
+
+    # Define gradient update step fn
+    def train_step(state, batch, label_smoothing_factor=0.0):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            outputs = state.apply_fn(
+                **batch,
+                params=params,
+                dropout_rng=dropout_rng,
+                freeze_feature_encoder=model_args.freeze_feature_encoder,
+                return_dict=True,
+                output_attentions=True,
+                output_hidden_states=True,
+                train=True,
+            )
+            encoder_hidden_states = jnp.asarray(outputs.encoder_hidden_states)
+            encoder_outputs = outputs.encoder_last_hidden_state
+            decoder_hidden_states = jnp.asarray(outputs.decoder_hidden_states)
+            logits = outputs.logits
+
+            # check for nan in inputs by taking l2-norm over inputs
+            # a single nan in the inputs will return a nan when normed
+            logs = {"inputs": jnp.linalg.norm(batch["inputs"])}
+
+            # check for nan in encoder_hidden_states, encoder_outputs
+            logs["encoder_hidden_states"] = jnp.linalg.norm(
+                encoder_hidden_states.reshape(-1, encoder_hidden_states.shape[0]), axis=0
+            )
+            logs["encoder_outputs"] = jnp.linalg.norm(encoder_outputs)
+
+            # check for nan in decoder_hidden_states, decoder_outputs (logits)
+            logs["decoder_hidden_states"] = jnp.linalg.norm(
+                decoder_hidden_states.reshape(-1, decoder_hidden_states.shape[0]), axis=0
+            )
+            logs["logits"] = jnp.linalg.norm(logits)
+
+            loss = loss_fn(logits, labels, label_smoothing_factor)
+            # normalize loss over gradient accumulation steps (ignore for now)
+            # loss = loss / gradient_accumulation_steps
+            return loss, logs
+
+        grad_fn = jax.value_and_grad(compute_loss, has_aux=True)
+        (loss, logs), grad = grad_fn(state.params)
+        # TODO: compute loss correctly over pmapped axis
+        grad = jax.lax.pmean(grad, "batch")
+
+        # compute gradient norm for monitoring
+        # (re-introduce when no nan's on forward pass, currently meaningless)
+        # grad_norm = jnp.linalg.norm(jax.tree_util.tree_leaves(jax.tree_map(jnp.linalg.norm, grad)))
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        # don't log learning-rate and grad-norm until forward pass returns real-valued numbers
+        metrics = {"loss": loss}
+        metrics.update(logs)
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch, label_smoothing_factor=0.0):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels, label_smoothing_factor)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Define generation function
+    gen_kwargs = {"max_length": training_args.generation_max_length, "num_beams": training_args.generation_num_beams}
+
+    def generate_step(params, batch):
+        model.params = params
+        output_ids = model.generate(batch["inputs"], **gen_kwargs)
+        return output_ids.sequences
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(
+        partial(train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0,)
+    )
+    p_eval_step = jax.pmap(partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch")
+    p_generate_step = jax.pmap(generate_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(vectorized_datasets['train'])}")
+    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"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+        train_metrics = []
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(vectorized_datasets["train"])
+        train_samples_idx = np.random.permutation(np.arange(num_train_samples))
+        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
+
+        # Gather the indexes for creating the batch and do a training step
+        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
+            samples = [vectorized_datasets["train"][int(idx)] for idx in batch_idx]
+            batch = data_collator(samples)
+            batch = shard(batch.data)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (num_train_samples // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = jax_utils.unreplicate(train_metric)
+                train_time += time.time() - train_start
+                # if has_tensorboard and jax.process_index() == 0:
+                #    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                # Log everything
+                metric_desc = " ".join([f"{key}: {value} |" for key, value in train_metric.items()])
+                epochs.write(f"Step... ({cur_step}) | {metric_desc}")
+
+                train_metrics = []
+
+        # epochs.write(
+        #    f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+        # )
+
+        continue
+        # ======================== Evaluating ==============================
+        eval_metrics = []
+        eval_preds = []
+        eval_labels = []
+
+        num_eval_samples = len(vectorized_datasets["eval"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [vectorized_datasets["eval"][int(idx)] for idx in batch_idx]
+            batch = data_collator(samples)
+            batch = shard(batch.data)
+            labels = batch["labels"]
+
+            metrics = p_eval_step(state.params, batch)
+            eval_metrics.append(metrics)
+
+            # generation
+            if training_args.predict_with_generate:
+                generated_ids = p_generate_step(state.params, batch)
+                eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
+                eval_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+        # compute WER metric
+        wer_desc = ""
+        if training_args.predict_with_generate:
+            wer_metric = compute_metrics(eval_preds, eval_labels)
+            eval_metrics.update(wer_metric)
+            wer_desc = " ".join([f"Eval {key}: {value} |" for key, value in wer_metric.items()])
+
+        # Print metrics and update progress bar
+        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {wer_desc})"
+        epochs.write(desc)
+        epochs.desc = desc
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(vectorized_datasets["train"]) // train_batch_size)
+            write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(training_args.output_dir, params=params)
+            tokenizer.save_pretrained(training_args.output_dir)
+            if training_args.push_to_hub:
+                repo.push_to_hub(commit_message=f"Saving weights and logs of epoch {epoch}", blocking=False)
+
+
+if __name__ == "__main__":
+    main()

run_librispeech.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+JAX_DEFAULT_MATMUL_PRECISION=float32 python run_flax_speech_recognition_seq2seq.py \
+        --dataset_name="librispeech_asr" \
+        --model_name_or_path="./" \
+        --dataset_config_name="clean" \
+        --train_split_name="train.100[:5%]" \
+        --eval_split_name="validation[:5%]" \
+        --output_dir="./" \
+        --preprocessing_num_workers="16" \
+        --length_column_name="input_length" \
+        --overwrite_output_dir \
+        --num_train_epochs="1" \
+        --per_device_train_batch_size="2" \
+        --per_device_eval_batch_size="2" \
+        --logging_steps="1" \
+        --max_duration_in_seconds="10" \
+        --max_target_length="32" \
+        --generation_max_length="40" \
+        --generation_num_beams="1" \
+        --learning_rate="3e-4" \
+        --warmup_steps="500" \
+        --text_column_name="text" \
+        --save_total_limit="1" \
+        --freeze_feature_encoder \
+        --predict_with_generate \
+        --do_lower_case \
+        --do_eval \
+        --do_train 
+

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}

tokenizer_config.json

@@ -0,0 +1 @@
+{"errors": "replace", "bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "<unk>", "pad_token": "<pad>", "mask_token": "<mask>", "add_prefix_space": false, "trim_offsets": true, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "facebook/bart-large-cnn", "tokenizer_class": "BartTokenizer"}