Saving weights and logs of step 1000

.gitattributes

@@ -25,3 +25,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220114_234119-1zya86oe/files/output.log filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220114_234119-1zya86oe/logs/debug-internal.log filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220114_234119-1zya86oe/run-1zya86oe.wandb filter=lfs diff=lfs merge=lfs -text

config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": "./",
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.16.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

eval_results.json

@@ -0,0 +1,5 @@
+{
+    "eval_accuracy": 0.6885889057979672,
+    "eval_loss": 1.430497475427537,
+    "eval_perplexity": 4.180778509252052
+}

events.out.tfevents.1642203685.t1v-n-eedfb410-w-0.10537.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d263549b78d9bb80e115caaffac01f4b141c31e954b64fb7a2ee59c5e4138641
+size 19208160

events.out.tfevents.1642204242.t1v-n-eedfb410-w-0.profile-empty

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ac614ecef2709e4ed2bc443ce4ade10122a22097363c5eb86dfadf8e74fa7c5
+size 40

events.out.tfevents.1642608722.t1v-n-eedfb410-w-0.1271442.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d194308b26bfc127e2d922f51dc9ca9914119f3b530144d116457498072ad97
+size 147136

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a12b2311a330a0e74118143c453bf8ad2490d7109372302e7cf9b878f65e181
+size 498796983

run_mlm_flax.py

@@ -0,0 +1,815 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 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 library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=fill-mask
+"""
+import json
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    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"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[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)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+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 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 main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level=logging.INFO,
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            # Enable Weight&Biases
+            import wandb
+            wandb.init(
+                entity='versae',
+                project='roberta-base-ncc',
+                sync_tensorboard=False,
+            )
+            wandb.config.update(training_args)
+            wandb.config.update(model_args)
+            wandb.config.update(data_args)
+
+            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."
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # 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()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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 FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = 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,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    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()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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 = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            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)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # 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 step {cur_step}", blocking=False)
+
+    # Eval after training
+    if training_args.do_eval:
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for _, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda metric: jnp.sum(metric).item(), eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        try:
+            perplexity = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            perplexity = float("inf")
+        eval_metrics["perplexity"] = perplexity
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

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 @@
+{"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>", "add_prefix_space": false, "errors": "replace", "sep_token": "</s>", "cls_token": "<s>", "pad_token": "<pad>", "mask_token": "<mask>", "trim_offsets": true, "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "RobertaTokenizer"}

train.128.sh

@@ -0,0 +1,25 @@
+python run_mlm_flax.py \
+    --output_dir="./" \
+    --model_type="roberta" \
+    --config_name="roberta-base" \
+    --tokenizer_name="NbAiLab/nb-roberta-base" \
+    --dataset_name="NbAiLab/NCC" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="232" \
+    --per_device_eval_batch_size="232" \
+    --pad_to_max_length \
+    --learning_rate="6e-4" \
+    --warmup_steps="10000" \
+    --overwrite_output_dir \
+    --num_train_epochs="3" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --logging_steps="1000" \
+    --save_steps="1000" \
+    --eval_steps="1000" \
+    --do_train \
+    --do_eval \
+    --dtype="bfloat16" \
+    --push_to_hub

train.512.sh

@@ -0,0 +1,26 @@
+python run_mlm_flax.py \
+    --output_dir="./" \
+    --model_type="roberta" \
+    --model_name_or_path="./" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --dataset_name="NbAiLab/NCC" \
+    --max_seq_length="512" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="46" \
+    --per_device_eval_batch_size="46" \
+    --pad_to_max_length \
+    --learning_rate="6e-4" \
+    --warmup_steps="1000" \
+    --overwrite_output_dir \
+    --num_train_epochs="3" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --logging_steps="1000" \
+    --save_steps="1000" \
+    --eval_steps="1000" \
+    --do_train \
+    --do_eval \
+    --dtype="bfloat16" \
+    --push_to_hub

wandb/debug-internal.log

@@ -0,0 +1 @@
+run-20220119_161158-274aad95/logs/debug-internal.log

wandb/debug.log

@@ -0,0 +1 @@
+run-20220119_161158-274aad95/logs/debug.log

wandb/latest-run

@@ -0,0 +1 @@
+run-20220119_161158-274aad95

wandb/run-20220114_212855-32qdb4k5/files/code/run_mlm_flax.py

@@ -0,0 +1,815 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 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 library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=fill-mask
+"""
+import json
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    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"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[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)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+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 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 main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level=logging.INFO,
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            # Enable Weight&Biases
+            import wandb
+            wandb.init(
+                entity='versae',
+                project='roberta-base-ncc',
+                sync_tensorboard=False,
+            )
+            wandb.config.update(training_args)
+            wandb.config.update(model_args)
+            wandb.config.update(data_args)
+
+            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."
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # 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()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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 FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = 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,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    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()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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 = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            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)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # 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 step {cur_step}", blocking=False)
+
+    # Eval after training
+    if training_args.do_eval:
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for _, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda metric: jnp.sum(metric).item(), eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        try:
+            perplexity = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            perplexity = float("inf")
+        eval_metrics["perplexity"] = perplexity
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

wandb/run-20220114_212855-32qdb4k5/files/config.yaml

@@ -0,0 +1,152 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_mlm_flax.py
+    framework: huggingface
+    huggingface_version: 4.16.0.dev0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1642195735
+    t:
+      1:
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 2
+      - 3
+      - 11
+      - 12
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.16.0.dev0
+      8:
+      - 5
+adafactor:
+  desc: null
+  value: false
+adam_beta1:
+  desc: null
+  value: 0.9
+adam_beta2:
+  desc: null
+  value: 0.98
+adam_epsilon:
+  desc: null
+  value: 1.0e-06
+cache_dir:
+  desc: null
+  value: null
+config_name:
+  desc: null
+  value: roberta-base
+dataset_config_name:
+  desc: null
+  value: null
+dataset_name:
+  desc: null
+  value: NbAiLab/NCC
+do_eval:
+  desc: null
+  value: true
+do_train:
+  desc: null
+  value: true
+dtype:
+  desc: null
+  value: bfloat16
+eval_steps:
+  desc: null
+  value: 1000
+hub_model_id:
+  desc: null
+  value: null
+hub_token:
+  desc: null
+  value: null
+learning_rate:
+  desc: null
+  value: 0.0006
+line_by_line:
+  desc: null
+  value: false
+logging_steps:
+  desc: null
+  value: 1000
+max_seq_length:
+  desc: null
+  value: 128
+mlm_probability:
+  desc: null
+  value: 0.15
+model_name_or_path:
+  desc: null
+  value: null
+model_type:
+  desc: null
+  value: roberta
+num_train_epochs:
+  desc: null
+  value: 3.0
+output_dir:
+  desc: null
+  value: ./
+overwrite_cache:
+  desc: null
+  value: false
+overwrite_output_dir:
+  desc: null
+  value: true
+pad_to_max_length:
+  desc: null
+  value: true
+per_device_eval_batch_size:
+  desc: null
+  value: 250
+per_device_train_batch_size:
+  desc: null
+  value: 250
+preprocessing_num_workers:
+  desc: null
+  value: null
+push_to_hub:
+  desc: null
+  value: true
+save_steps:
+  desc: null
+  value: 1000
+seed:
+  desc: null
+  value: 42
+tokenizer_name:
+  desc: null
+  value: NbAiLab/nb-roberta-base
+train_file:
+  desc: null
+  value: null
+train_ref_file:
+  desc: null
+  value: null
+use_fast_tokenizer:
+  desc: null
+  value: true
+validation_file:
+  desc: null
+  value: null
+validation_ref_file:
+  desc: null
+  value: null
+validation_split_percentage:
+  desc: null
+  value: 5
+warmup_steps:
+  desc: null
+  value: 10000
+weight_decay:
+  desc: null
+  value: 0.01

wandb/run-20220114_212855-32qdb4k5/files/output.log

@@ -0,0 +1,43 @@
+2022-01-14 21:29:01.798913: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libcuda.so.1'; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory
+2022-01-14 21:29:01.798960: W tensorflow/stream_executor/cuda/cuda_driver.cc:269] failed call to cuInit: UNKNOWN ERROR (303)
+Epoch ... (1/3):   0%|                                                                                                                                                                                                                                              | 0/3 [00:00<?, ?it/s]
+Training...:   0%|                                                                                                                                                                                                                                              | 0/39919 [02:17<?, ?it/s]
+Epoch ... (1/3):   0%|                                                                                                                                                                                                                                              | 0/3 [03:05<?, ?it/s]
+Traceback (most recent call last):
+  File "run_mlm_flax.py", line 815, in <module>
+    main()
+  File "run_mlm_flax.py", line 723, in main
+    state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/traceback_util.py", line 162, in reraise_with_filtered_traceback
+    return fun(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/api.py", line 2058, in cache_miss
+    out_tree, out_flat = f_pmapped_(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/api.py", line 1934, in f_pmapped
+    out = pxla.xla_pmap(
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1727, in bind
+    return call_bind(self, fun, *args, **params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1652, in call_bind
+    outs = primitive.process(top_trace, fun, tracers, params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1730, in process
+    return trace.process_map(self, fun, tracers, params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 633, in process_call
+    return primitive.impl(f, *tracers, **params)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 778, in xla_pmap_impl
+    return compiled_fun(*args)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/profiler.py", line 206, in wrapper
+    return func(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 1502, in execute_replicated
+    out_bufs = compiled.execute_sharded_on_local_devices(input_bufs)
+jax._src.traceback_util.UnfilteredStackTrace: RuntimeError: RESOURCE_EXHAUSTED: Attempting to reserve 12.83G at the bottom of memory. That was not possible. There are 13.18G free, 0B reserved, and 12.71G reservable.: while running replica 0 and partition 0 of a replicated computation (other replicas may have failed as well).
+The stack trace below excludes JAX-internal frames.
+The preceding is the original exception that occurred, unmodified.
+--------------------
+The above exception was the direct cause of the following exception:
+Traceback (most recent call last):
+  File "run_mlm_flax.py", line 815, in <module>
+    main()
+  File "run_mlm_flax.py", line 723, in main
+    state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 1502, in execute_replicated
+    out_bufs = compiled.execute_sharded_on_local_devices(input_bufs)
+RuntimeError: RESOURCE_EXHAUSTED: Attempting to reserve 12.83G at the bottom of memory. That was not possible. There are 13.18G free, 0B reserved, and 12.71G reservable.: while running replica 0 and partition 0 of a replicated computation (other replicas may have failed as well).

wandb/run-20220114_212855-32qdb4k5/files/requirements.txt

@@ -0,0 +1,122 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+charset-normalizer==2.0.10
+chex==0.1.0
+click==8.0.3
+clu==0.0.6
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==1.17.1.dev0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.3.6
+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-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
+huggingface-hub==0.2.1
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.2.26
+jaxlib==0.1.75
+jedi==0.18.1
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+libtpu-nightly==0.1.dev20211208
+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
+numpy==1.22.0
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.0
+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==20.0.2
+pkg-resources==0.0.0
+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
+pygments==2.11.1
+pyparsing==3.0.6
+python-dateutil==2.8.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+rsa==4.8
+sacremoses==0.0.46
+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
+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
+termcolor==1.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.16.0.dev0
+typing-extensions==3.10.0.2
+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-20220114_212855-32qdb4k5/files/wandb-metadata.json

@@ -0,0 +1,47 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-01-14T21:28:58.974844",
+    "startedAt": "2022-01-14T21:28:55.397355",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--output_dir=./",
+        "--model_type=roberta",
+        "--config_name=roberta-base",
+        "--tokenizer_name=NbAiLab/nb-roberta-base",
+        "--dataset_name=NbAiLab/NCC",
+        "--max_seq_length=128",
+        "--weight_decay=0.01",
+        "--per_device_train_batch_size=250",
+        "--per_device_eval_batch_size=250",
+        "--pad_to_max_length",
+        "--learning_rate=6e-4",
+        "--warmup_steps=10000",
+        "--overwrite_output_dir",
+        "--num_train_epochs=3",
+        "--adam_beta1=0.9",
+        "--adam_beta2=0.98",
+        "--adam_epsilon=1e-6",
+        "--logging_steps=1000",
+        "--save_steps=1000",
+        "--eval_steps=1000",
+        "--do_train",
+        "--do_eval",
+        "--dtype=bfloat16",
+        "--push_to_hub"
+    ],
+    "state": "running",
+    "program": "run_mlm_flax.py",
+    "codePath": "run_mlm_flax.py",
+    "git": {
+        "remote": "https://huggingface.co/versae/roberta-base-ncc",
+        "commit": "502df078f73cf93ca9380fcac1c9b9c7598a445f"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/roberta-base-ncc",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220114_212855-32qdb4k5/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb": {"runtime": 200}}

wandb/run-20220114_212855-32qdb4k5/logs/debug-internal.log

@@ -0,0 +1,189 @@
+2022-01-14 21:28:56,265 INFO    MainThread:8253 [internal.py:wandb_internal():87] W&B internal server running at pid: 8253, started at: 2022-01-14 21:28:56.265129
+2022-01-14 21:28:56,268 DEBUG   SenderThread:8253 [sender.py:send():234] send: header
+2022-01-14 21:28:56,268 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: check_version
+2022-01-14 21:28:56,268 INFO    WriterThread:8253 [datastore.py:open_for_write():77] open: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/run-32qdb4k5.wandb
+2022-01-14 21:28:56,268 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: check_version
+2022-01-14 21:28:56,352 DEBUG   SenderThread:8253 [sender.py:send():234] send: run
+2022-01-14 21:28:56,515 INFO    SenderThread:8253 [dir_watcher.py:__init__():169] watching files in: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files
+2022-01-14 21:28:56,515 INFO    SenderThread:8253 [sender.py:_start_run_threads():804] run started: 32qdb4k5 with start time 1642195735
+2022-01-14 21:28:56,515 DEBUG   SenderThread:8253 [sender.py:send():234] send: summary
+2022-01-14 21:28:56,515 INFO    SenderThread:8253 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 21:28:56,515 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: run_start
+2022-01-14 21:28:57,561 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-summary.json
+2022-01-14 21:28:58,974 DEBUG   HandlerThread:8253 [meta.py:__init__():40] meta init
+2022-01-14 21:28:58,974 DEBUG   HandlerThread:8253 [meta.py:__init__():54] meta init done
+2022-01-14 21:28:58,974 DEBUG   HandlerThread:8253 [meta.py:probe():214] probe
+2022-01-14 21:28:58,975 DEBUG   HandlerThread:8253 [meta.py:_setup_git():204] setup git
+2022-01-14 21:28:59,006 DEBUG   HandlerThread:8253 [meta.py:_setup_git():211] setup git done
+2022-01-14 21:28:59,006 DEBUG   HandlerThread:8253 [meta.py:_save_code():92] save code
+2022-01-14 21:28:59,018 DEBUG   HandlerThread:8253 [meta.py:_save_code():113] save code done
+2022-01-14 21:28:59,018 DEBUG   HandlerThread:8253 [meta.py:_save_patches():130] save patches
+2022-01-14 21:28:59,561 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/code/run_mlm_flax.py
+2022-01-14 21:28:59,562 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/code
+2022-01-14 21:29:01,562 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:29:03,563 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:29:03,563 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/diff.patch
+2022-01-14 21:29:08,274 ERROR   HandlerThread:8253 [meta.py:_save_patches():171] Error generating diff: Command '['git', 'diff', '--submodule=diff', 'HEAD']' timed out after 5 seconds
+2022-01-14 21:29:08,274 DEBUG   HandlerThread:8253 [meta.py:_save_patches():172] save patches done
+2022-01-14 21:29:08,274 DEBUG   HandlerThread:8253 [meta.py:_save_pip():58] save pip
+2022-01-14 21:29:08,275 DEBUG   HandlerThread:8253 [meta.py:_save_pip():72] save pip done
+2022-01-14 21:29:08,275 DEBUG   HandlerThread:8253 [meta.py:probe():252] probe done
+2022-01-14 21:29:08,283 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:29:08,284 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:29:08,566 DEBUG   SenderThread:8253 [sender.py:send():234] send: config
+2022-01-14 21:29:08,566 DEBUG   SenderThread:8253 [sender.py:send():234] send: config
+2022-01-14 21:29:08,566 DEBUG   SenderThread:8253 [sender.py:send():234] send: config
+2022-01-14 21:29:08,567 DEBUG   SenderThread:8253 [sender.py:send():234] send: files
+2022-01-14 21:29:08,567 INFO    SenderThread:8253 [sender.py:_save_file():939] saving file wandb-metadata.json with policy now
+2022-01-14 21:29:08,567 INFO    SenderThread:8253 [sender.py:_save_file():939] saving file code/run_mlm_flax.py with policy now
+2022-01-14 21:29:08,571 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/requirements.txt
+2022-01-14 21:29:08,571 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-metadata.json
+2022-01-14 21:29:09,032 INFO    Thread-12 :8253 [upload_job.py:push():137] Uploaded file /tmp/tmpdg54qv_0wandb/w1tibuxq-code/run_mlm_flax.py
+2022-01-14 21:29:09,069 INFO    Thread-11 :8253 [upload_job.py:push():137] Uploaded file /tmp/tmpdg54qv_0wandb/35h4ryp5-wandb-metadata.json
+2022-01-14 21:29:09,571 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:29:21,520 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:29:21,521 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:29:27,049 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:29:27,579 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/config.yaml
+2022-01-14 21:29:36,656 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:29:36,657 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:29:51,794 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:29:51,795 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:29:57,119 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:29:58,591 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:30:06,945 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:30:06,945 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:30:22,126 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:30:22,127 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:30:27,189 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:30:37,330 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:30:37,330 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:30:52,532 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:30:52,532 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:30:57,257 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:31:07,691 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:31:07,692 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:31:22,944 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:31:22,945 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:31:27,323 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:31:38,085 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:31:38,086 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:31:53,231 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:31:53,231 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:31:57,395 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:32:08,366 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 21:32:08,367 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 21:32:16,649 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:32:16,893 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:16,893 DEBUG   SenderThread:8253 [sender.py:send():234] send: telemetry
+2022-01-14 21:32:16,893 DEBUG   SenderThread:8253 [sender.py:send():234] send: exit
+2022-01-14 21:32:16,893 INFO    SenderThread:8253 [sender.py:send_exit():366] handling exit code: 1
+2022-01-14 21:32:16,894 INFO    SenderThread:8253 [sender.py:send_exit():368] handling runtime: 200
+2022-01-14 21:32:16,894 INFO    SenderThread:8253 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 21:32:16,894 INFO    SenderThread:8253 [sender.py:send_exit():374] send defer
+2022-01-14 21:32:16,894 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:16,895 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:16,895 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 0
+2022-01-14 21:32:16,895 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:16,895 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 0
+2022-01-14 21:32:16,895 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 1
+2022-01-14 21:32:16,896 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:16,896 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 1
+2022-01-14 21:32:16,941 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:16,941 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 1
+2022-01-14 21:32:16,941 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 2
+2022-01-14 21:32:16,941 DEBUG   SenderThread:8253 [sender.py:send():234] send: stats
+2022-01-14 21:32:16,942 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:16,942 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 2
+2022-01-14 21:32:16,942 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:16,942 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 2
+2022-01-14 21:32:16,942 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 3
+2022-01-14 21:32:16,942 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:16,942 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 3
+2022-01-14 21:32:16,943 DEBUG   SenderThread:8253 [sender.py:send():234] send: summary
+2022-01-14 21:32:16,943 INFO    SenderThread:8253 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 21:32:16,943 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:16,943 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 3
+2022-01-14 21:32:16,943 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 4
+2022-01-14 21:32:16,943 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:16,943 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 4
+2022-01-14 21:32:16,944 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:16,944 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 4
+2022-01-14 21:32:16,997 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:17,650 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:32:17,650 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-summary.json
+2022-01-14 21:32:17,685 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 5
+2022-01-14 21:32:17,686 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:17,686 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:17,686 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 5
+2022-01-14 21:32:17,686 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:17,686 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 5
+2022-01-14 21:32:17,687 INFO    SenderThread:8253 [dir_watcher.py:finish():283] shutting down directory watcher
+2022-01-14 21:32:17,787 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:18,650 INFO    Thread-8  :8253 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/config.yaml
+2022-01-14 21:32:18,651 INFO    SenderThread:8253 [dir_watcher.py:finish():313] scan: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files
+2022-01-14 21:32:18,651 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/config.yaml config.yaml
+2022-01-14 21:32:18,651 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/diff.patch diff.patch
+2022-01-14 21:32:18,651 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/requirements.txt requirements.txt
+2022-01-14 21:32:18,652 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log output.log
+2022-01-14 21:32:18,652 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-summary.json wandb-summary.json
+2022-01-14 21:32:18,652 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-metadata.json wandb-metadata.json
+2022-01-14 21:32:18,656 INFO    SenderThread:8253 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/code/run_mlm_flax.py code/run_mlm_flax.py
+2022-01-14 21:32:18,656 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 6
+2022-01-14 21:32:18,656 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:18,657 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:18,657 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 6
+2022-01-14 21:32:18,662 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:18,665 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 6
+2022-01-14 21:32:18,665 INFO    SenderThread:8253 [file_pusher.py:finish():177] shutting down file pusher
+2022-01-14 21:32:18,757 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:18,758 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:18,859 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:18,860 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:18,961 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:18,962 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,063 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,064 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,139 INFO    Thread-15 :8253 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/output.log
+2022-01-14 21:32:19,148 INFO    Thread-14 :8253 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/requirements.txt
+2022-01-14 21:32:19,165 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,165 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,171 INFO    Thread-13 :8253 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/config.yaml
+2022-01-14 21:32:19,267 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,267 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,288 INFO    Thread-16 :8253 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/files/wandb-summary.json
+2022-01-14 21:32:19,370 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,370 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,472 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,472 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,489 INFO    Thread-7  :8253 [sender.py:transition_state():387] send defer: 7
+2022-01-14 21:32:19,489 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:19,490 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 7
+2022-01-14 21:32:19,490 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:19,490 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 7
+2022-01-14 21:32:19,573 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:19,915 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 8
+2022-01-14 21:32:19,916 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:19,916 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:19,916 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 8
+2022-01-14 21:32:19,917 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:19,917 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 8
+2022-01-14 21:32:19,917 INFO    SenderThread:8253 [sender.py:transition_state():387] send defer: 9
+2022-01-14 21:32:19,917 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 21:32:19,917 INFO    HandlerThread:8253 [handler.py:handle_request_defer():147] handle defer: 9
+2022-01-14 21:32:19,918 DEBUG   SenderThread:8253 [sender.py:send():234] send: final
+2022-01-14 21:32:19,918 DEBUG   SenderThread:8253 [sender.py:send():234] send: footer
+2022-01-14 21:32:19,918 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: defer
+2022-01-14 21:32:19,918 INFO    SenderThread:8253 [sender.py:send_request_defer():383] handle sender defer: 9
+2022-01-14 21:32:20,017 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 21:32:20,018 DEBUG   SenderThread:8253 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 21:32:20,018 INFO    SenderThread:8253 [file_pusher.py:join():182] waiting for file pusher
+2022-01-14 21:32:20,278 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: get_summary
+2022-01-14 21:32:20,278 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: sampled_history
+2022-01-14 21:32:20,279 DEBUG   HandlerThread:8253 [handler.py:handle_request():130] handle_request: shutdown
+2022-01-14 21:32:20,279 INFO    HandlerThread:8253 [handler.py:finish():731] shutting down handler
+2022-01-14 21:32:20,918 INFO    WriterThread:8253 [datastore.py:close():281] close: /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/run-32qdb4k5.wandb
+2022-01-14 21:32:21,277 INFO    SenderThread:8253 [sender.py:finish():1070] shutting down sender
+2022-01-14 21:32:21,277 INFO    SenderThread:8253 [file_pusher.py:finish():177] shutting down file pusher
+2022-01-14 21:32:21,277 INFO    SenderThread:8253 [file_pusher.py:join():182] waiting for file pusher
+2022-01-14 21:32:21,279 INFO    MainThread:8253 [internal.py:handle_exit():77] Internal process exited

wandb/run-20220114_212855-32qdb4k5/logs/debug.log

@@ -0,0 +1,150 @@
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_setup.py:_flush():71] setting env: {}
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_setup.py:_flush():71] setting login settings: {}
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_init.py:_log_setup():371] Logging user logs to /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/logs/debug.log
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_init.py:_log_setup():372] Logging internal logs to /data/roberta-base-ncc/wandb/run-20220114_212855-32qdb4k5/logs/debug-internal.log
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_init.py:init():404] calling init triggers
+2022-01-14 21:28:55,408 INFO    MainThread:5000 [wandb_init.py:init():409] wandb.init called with sweep_config: {}
+config: {}
+2022-01-14 21:28:55,409 INFO    MainThread:5000 [wandb_init.py:init():460] starting backend
+2022-01-14 21:28:55,409 INFO    MainThread:5000 [backend.py:_multiprocessing_setup():99] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
+2022-01-14 21:28:55,461 INFO    MainThread:5000 [backend.py:ensure_launched():216] starting backend process...
+2022-01-14 21:28:55,488 INFO    MainThread:5000 [backend.py:ensure_launched():221] started backend process with pid: 8253
+2022-01-14 21:28:55,490 INFO    MainThread:5000 [wandb_init.py:init():469] backend started and connected
+2022-01-14 21:28:55,501 INFO    MainThread:5000 [wandb_init.py:init():533] updated telemetry
+2022-01-14 21:28:55,563 INFO    MainThread:5000 [wandb_init.py:init():563] communicating current version
+2022-01-14 21:28:56,351 INFO    MainThread:5000 [wandb_init.py:init():568] got version response 
+2022-01-14 21:28:56,351 INFO    MainThread:5000 [wandb_init.py:init():578] communicating run to backend with 30 second timeout
+2022-01-14 21:28:56,515 INFO    MainThread:5000 [wandb_init.py:init():606] starting run threads in backend
+2022-01-14 21:29:01,520 INFO    MainThread:5000 [wandb_run.py:_console_start():1810] atexit reg
+2022-01-14 21:29:01,520 INFO    MainThread:5000 [wandb_run.py:_redirect():1684] redirect: SettingsConsole.REDIRECT
+2022-01-14 21:29:01,520 INFO    MainThread:5000 [wandb_run.py:_redirect():1689] Redirecting console.
+2022-01-14 21:29:01,523 INFO    MainThread:5000 [wandb_run.py:_redirect():1745] Redirects installed.
+2022-01-14 21:29:01,523 INFO    MainThread:5000 [wandb_init.py:init():633] run started, returning control to user process
+2022-01-14 21:29:01,523 INFO    MainThread:5000 [wandb_run.py:_config_callback():956] config_cb None None {'output_dir': './', 'overwrite_output_dir': True, 'do_train': True, 'do_eval': True, 'per_device_train_batch_size': 250, 'per_device_eval_batch_size': 250, 'learning_rate': 0.0006, 'weight_decay': 0.01, 'adam_beta1': 0.9, 'adam_beta2': 0.98, 'adam_epsilon': 1e-06, 'adafactor': False, 'num_train_epochs': 3.0, 'warmup_steps': 10000, 'logging_steps': 1000, 'save_steps': 1000, 'eval_steps': 1000, 'seed': 42, 'push_to_hub': True, 'hub_model_id': None, 'hub_token': None}
+2022-01-14 21:29:01,524 INFO    MainThread:5000 [wandb_run.py:_config_callback():956] config_cb None None {'model_name_or_path': None, 'model_type': 'roberta', 'config_name': 'roberta-base', 'tokenizer_name': 'NbAiLab/nb-roberta-base', 'cache_dir': None, 'use_fast_tokenizer': True, 'dtype': 'bfloat16'}
+2022-01-14 21:29:01,524 INFO    MainThread:5000 [wandb_run.py:_config_callback():956] config_cb None None {'dataset_name': 'NbAiLab/NCC', 'dataset_config_name': None, 'train_file': None, 'validation_file': None, 'train_ref_file': None, 'validation_ref_file': None, 'overwrite_cache': False, 'validation_split_percentage': 5, 'max_seq_length': 128, 'preprocessing_num_workers': None, 'mlm_probability': 0.15, 'pad_to_max_length': True, 'line_by_line': False}
+2022-01-14 21:32:14,189 INFO    MainThread:5000 [wandb_run.py:_atexit_cleanup():1780] got exitcode: 1
+2022-01-14 21:32:14,192 INFO    MainThread:5000 [wandb_run.py:_restore():1752] restore
+2022-01-14 21:32:16,895 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-14 21:32:17,686 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-14 21:32:18,657 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:18,759 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:18,861 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45903
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:18,962 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45903
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,064 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,166 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,269 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,371 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,473 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:19,917 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+
+2022-01-14 21:32:20,277 INFO    MainThread:5000 [wandb_run.py:_wait_for_finish():1912] got exit ret: done: true
+exit_result {
+}
+file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45931
+  total_bytes: 45931
+}
+local_info {
+}
+
+2022-01-14 21:32:23,445 INFO    MainThread:5000 [wandb_run.py:_append_files():2180] logging synced files

wandb/run-20220114_221533-24dma583/files/code/run_mlm_flax.py

@@ -0,0 +1,815 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 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 library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=fill-mask
+"""
+import json
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    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"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[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)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+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 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 main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level=logging.INFO,
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            # Enable Weight&Biases
+            import wandb
+            wandb.init(
+                entity='versae',
+                project='roberta-base-ncc',
+                sync_tensorboard=False,
+            )
+            wandb.config.update(training_args)
+            wandb.config.update(model_args)
+            wandb.config.update(data_args)
+
+            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."
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # 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()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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 FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = 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,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    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()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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 = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            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)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # 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 step {cur_step}", blocking=False)
+
+    # Eval after training
+    if training_args.do_eval:
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for _, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda metric: jnp.sum(metric).item(), eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        try:
+            perplexity = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            perplexity = float("inf")
+        eval_metrics["perplexity"] = perplexity
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

wandb/run-20220114_221533-24dma583/files/config.yaml

@@ -0,0 +1,152 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_mlm_flax.py
+    framework: huggingface
+    huggingface_version: 4.16.0.dev0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1642198533
+    t:
+      1:
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 2
+      - 3
+      - 11
+      - 12
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.16.0.dev0
+      8:
+      - 5
+adafactor:
+  desc: null
+  value: false
+adam_beta1:
+  desc: null
+  value: 0.9
+adam_beta2:
+  desc: null
+  value: 0.98
+adam_epsilon:
+  desc: null
+  value: 1.0e-06
+cache_dir:
+  desc: null
+  value: null
+config_name:
+  desc: null
+  value: roberta-base
+dataset_config_name:
+  desc: null
+  value: null
+dataset_name:
+  desc: null
+  value: NbAiLab/NCC
+do_eval:
+  desc: null
+  value: true
+do_train:
+  desc: null
+  value: true
+dtype:
+  desc: null
+  value: bfloat16
+eval_steps:
+  desc: null
+  value: 1000
+hub_model_id:
+  desc: null
+  value: null
+hub_token:
+  desc: null
+  value: null
+learning_rate:
+  desc: null
+  value: 0.0006
+line_by_line:
+  desc: null
+  value: false
+logging_steps:
+  desc: null
+  value: 1000
+max_seq_length:
+  desc: null
+  value: 128
+mlm_probability:
+  desc: null
+  value: 0.15
+model_name_or_path:
+  desc: null
+  value: null
+model_type:
+  desc: null
+  value: roberta
+num_train_epochs:
+  desc: null
+  value: 3.0
+output_dir:
+  desc: null
+  value: ./
+overwrite_cache:
+  desc: null
+  value: false
+overwrite_output_dir:
+  desc: null
+  value: true
+pad_to_max_length:
+  desc: null
+  value: true
+per_device_eval_batch_size:
+  desc: null
+  value: 250
+per_device_train_batch_size:
+  desc: null
+  value: 250
+preprocessing_num_workers:
+  desc: null
+  value: null
+push_to_hub:
+  desc: null
+  value: true
+save_steps:
+  desc: null
+  value: 1000
+seed:
+  desc: null
+  value: 42
+tokenizer_name:
+  desc: null
+  value: NbAiLab/nb-roberta-base
+train_file:
+  desc: null
+  value: null
+train_ref_file:
+  desc: null
+  value: null
+use_fast_tokenizer:
+  desc: null
+  value: true
+validation_file:
+  desc: null
+  value: null
+validation_ref_file:
+  desc: null
+  value: null
+validation_split_percentage:
+  desc: null
+  value: 5
+warmup_steps:
+  desc: null
+  value: 10000
+weight_decay:
+  desc: null
+  value: 0.01

wandb/run-20220114_221533-24dma583/files/output.log

@@ -0,0 +1,43 @@
+2022-01-14 22:15:40.254500: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libcuda.so.1'; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory
+2022-01-14 22:15:40.254546: W tensorflow/stream_executor/cuda/cuda_driver.cc:269] failed call to cuInit: UNKNOWN ERROR (303)
+Epoch ... (1/3):   0%|                                                                                                          | 0/3 [00:00<?, ?it/s]
+Training...:   0%|                                                                                                          | 0/39919 [02:25<?, ?it/s]
+Epoch ... (1/3):   0%|                                                                                                          | 0/3 [03:13<?, ?it/s]
+Traceback (most recent call last):
+  File "run_mlm_flax.py", line 815, in <module>
+    main()
+  File "run_mlm_flax.py", line 723, in main
+    state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/traceback_util.py", line 162, in reraise_with_filtered_traceback
+    return fun(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/api.py", line 2058, in cache_miss
+    out_tree, out_flat = f_pmapped_(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/api.py", line 1934, in f_pmapped
+    out = pxla.xla_pmap(
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1727, in bind
+    return call_bind(self, fun, *args, **params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1652, in call_bind
+    outs = primitive.process(top_trace, fun, tracers, params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 1730, in process
+    return trace.process_map(self, fun, tracers, params)
+  File "/data/flax/lib/python3.8/site-packages/jax/core.py", line 633, in process_call
+    return primitive.impl(f, *tracers, **params)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 778, in xla_pmap_impl
+    return compiled_fun(*args)
+  File "/data/flax/lib/python3.8/site-packages/jax/_src/profiler.py", line 206, in wrapper
+    return func(*args, **kwargs)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 1502, in execute_replicated
+    out_bufs = compiled.execute_sharded_on_local_devices(input_bufs)
+jax._src.traceback_util.UnfilteredStackTrace: RuntimeError: RESOURCE_EXHAUSTED: Attempting to reserve 12.83G at the bottom of memory. That was not possible. There are 13.18G free, 0B reserved, and 12.71G reservable.: while running replica 0 and partition 0 of a replicated computation (other replicas may have failed as well).
+The stack trace below excludes JAX-internal frames.
+The preceding is the original exception that occurred, unmodified.
+--------------------
+The above exception was the direct cause of the following exception:
+Traceback (most recent call last):
+  File "run_mlm_flax.py", line 815, in <module>
+    main()
+  File "run_mlm_flax.py", line 723, in main
+    state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+  File "/data/flax/lib/python3.8/site-packages/jax/interpreters/pxla.py", line 1502, in execute_replicated
+    out_bufs = compiled.execute_sharded_on_local_devices(input_bufs)
+RuntimeError: RESOURCE_EXHAUSTED: Attempting to reserve 12.83G at the bottom of memory. That was not possible. There are 13.18G free, 0B reserved, and 12.71G reservable.: while running replica 0 and partition 0 of a replicated computation (other replicas may have failed as well).

wandb/run-20220114_221533-24dma583/files/requirements.txt

@@ -0,0 +1,122 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+charset-normalizer==2.0.10
+chex==0.1.0
+click==8.0.3
+clu==0.0.6
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==1.17.1.dev0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.3.6
+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-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
+huggingface-hub==0.2.1
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.2.26
+jaxlib==0.1.75
+jedi==0.18.1
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+libtpu-nightly==0.1.dev20211208
+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
+numpy==1.22.0
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.0
+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==20.0.2
+pkg-resources==0.0.0
+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
+pygments==2.11.1
+pyparsing==3.0.6
+python-dateutil==2.8.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+rsa==4.8
+sacremoses==0.0.46
+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
+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
+termcolor==1.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.16.0.dev0
+typing-extensions==3.10.0.2
+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-20220114_221533-24dma583/files/wandb-metadata.json

@@ -0,0 +1,47 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-01-14T22:15:37.284889",
+    "startedAt": "2022-01-14T22:15:33.798491",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--output_dir=./",
+        "--model_type=roberta",
+        "--config_name=roberta-base",
+        "--tokenizer_name=NbAiLab/nb-roberta-base",
+        "--dataset_name=NbAiLab/NCC",
+        "--max_seq_length=128",
+        "--weight_decay=0.01",
+        "--per_device_train_batch_size=250",
+        "--per_device_eval_batch_size=250",
+        "--pad_to_max_length",
+        "--learning_rate=6e-4",
+        "--warmup_steps=10000",
+        "--overwrite_output_dir",
+        "--num_train_epochs=3",
+        "--adam_beta1=0.9",
+        "--adam_beta2=0.98",
+        "--adam_epsilon=1e-6",
+        "--logging_steps=1000",
+        "--save_steps=1000",
+        "--eval_steps=1000",
+        "--do_train",
+        "--do_eval",
+        "--dtype=bfloat16",
+        "--push_to_hub"
+    ],
+    "state": "running",
+    "program": "run_mlm_flax.py",
+    "codePath": "run_mlm_flax.py",
+    "git": {
+        "remote": "https://huggingface.co/versae/roberta-base-ncc",
+        "commit": "502df078f73cf93ca9380fcac1c9b9c7598a445f"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/roberta-base-ncc",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220114_221533-24dma583/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb": {"runtime": 208}}

wandb/run-20220114_221533-24dma583/logs/debug-internal.log

@@ -0,0 +1,187 @@
+2022-01-14 22:15:34,709 INFO    MainThread:7834 [internal.py:wandb_internal():87] W&B internal server running at pid: 7834, started at: 2022-01-14 22:15:34.709583
+2022-01-14 22:15:34,711 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: check_version
+2022-01-14 22:15:34,712 INFO    WriterThread:7834 [datastore.py:open_for_write():77] open: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/run-24dma583.wandb
+2022-01-14 22:15:34,712 DEBUG   SenderThread:7834 [sender.py:send():234] send: header
+2022-01-14 22:15:34,712 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: check_version
+2022-01-14 22:15:34,785 DEBUG   SenderThread:7834 [sender.py:send():234] send: run
+2022-01-14 22:15:34,980 INFO    SenderThread:7834 [dir_watcher.py:__init__():169] watching files in: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files
+2022-01-14 22:15:34,980 INFO    SenderThread:7834 [sender.py:_start_run_threads():804] run started: 24dma583 with start time 1642198533
+2022-01-14 22:15:34,980 DEBUG   SenderThread:7834 [sender.py:send():234] send: summary
+2022-01-14 22:15:34,980 INFO    SenderThread:7834 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 22:15:34,981 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: run_start
+2022-01-14 22:15:35,985 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-summary.json
+2022-01-14 22:15:37,284 DEBUG   HandlerThread:7834 [meta.py:__init__():40] meta init
+2022-01-14 22:15:37,284 DEBUG   HandlerThread:7834 [meta.py:__init__():54] meta init done
+2022-01-14 22:15:37,284 DEBUG   HandlerThread:7834 [meta.py:probe():214] probe
+2022-01-14 22:15:37,286 DEBUG   HandlerThread:7834 [meta.py:_setup_git():204] setup git
+2022-01-14 22:15:37,315 DEBUG   HandlerThread:7834 [meta.py:_setup_git():211] setup git done
+2022-01-14 22:15:37,315 DEBUG   HandlerThread:7834 [meta.py:_save_code():92] save code
+2022-01-14 22:15:37,326 DEBUG   HandlerThread:7834 [meta.py:_save_code():113] save code done
+2022-01-14 22:15:37,326 DEBUG   HandlerThread:7834 [meta.py:_save_patches():130] save patches
+2022-01-14 22:15:37,985 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/code/run_mlm_flax.py
+2022-01-14 22:15:37,986 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/code
+2022-01-14 22:15:39,986 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:15:40,986 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/diff.patch
+2022-01-14 22:15:42,987 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:15:45,607 ERROR   HandlerThread:7834 [meta.py:_save_patches():171] Error generating diff: Command '['git', 'diff', '--submodule=diff', 'HEAD']' timed out after 5 seconds
+2022-01-14 22:15:45,607 DEBUG   HandlerThread:7834 [meta.py:_save_patches():172] save patches done
+2022-01-14 22:15:45,607 DEBUG   HandlerThread:7834 [meta.py:_save_pip():58] save pip
+2022-01-14 22:15:45,607 DEBUG   HandlerThread:7834 [meta.py:_save_pip():72] save pip done
+2022-01-14 22:15:45,608 DEBUG   HandlerThread:7834 [meta.py:probe():252] probe done
+2022-01-14 22:15:45,643 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:15:45,643 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:15:45,786 DEBUG   SenderThread:7834 [sender.py:send():234] send: config
+2022-01-14 22:15:45,787 DEBUG   SenderThread:7834 [sender.py:send():234] send: config
+2022-01-14 22:15:45,787 DEBUG   SenderThread:7834 [sender.py:send():234] send: config
+2022-01-14 22:15:45,787 DEBUG   SenderThread:7834 [sender.py:send():234] send: files
+2022-01-14 22:15:45,787 INFO    SenderThread:7834 [sender.py:_save_file():939] saving file wandb-metadata.json with policy now
+2022-01-14 22:15:45,789 INFO    SenderThread:7834 [sender.py:_save_file():939] saving file code/run_mlm_flax.py with policy now
+2022-01-14 22:15:45,988 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-metadata.json
+2022-01-14 22:15:45,989 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_created():217] file/dir created: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/requirements.txt
+2022-01-14 22:15:46,312 INFO    Thread-12 :7834 [upload_job.py:push():137] Uploaded file /tmp/tmpxqv1l1fswandb/2juok80v-code/run_mlm_flax.py
+2022-01-14 22:15:46,330 INFO    Thread-11 :7834 [upload_job.py:push():137] Uploaded file /tmp/tmpxqv1l1fswandb/xnc44171-wandb-metadata.json
+2022-01-14 22:15:50,990 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:15:59,991 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:15:59,991 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:16:05,368 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:16:05,997 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/config.yaml
+2022-01-14 22:16:15,132 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:16:15,132 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:16:30,272 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:16:30,272 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:16:35,439 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:16:39,009 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:16:45,408 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:16:45,408 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:17:00,601 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:17:00,601 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:17:05,512 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:17:15,756 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:17:15,756 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:17:30,970 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:17:30,971 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:17:35,586 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:17:46,135 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:17:46,136 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:18:01,309 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:18:01,309 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:18:05,663 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:18:16,458 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:18:16,458 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:18:31,596 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:18:31,597 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:18:35,742 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:18:46,731 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: stop_status
+2022-01-14 22:18:46,732 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: stop_status
+2022-01-14 22:19:03,067 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:19:03,953 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:03,953 DEBUG   SenderThread:7834 [sender.py:send():234] send: telemetry
+2022-01-14 22:19:03,953 DEBUG   SenderThread:7834 [sender.py:send():234] send: exit
+2022-01-14 22:19:03,953 INFO    SenderThread:7834 [sender.py:send_exit():366] handling exit code: 1
+2022-01-14 22:19:03,954 INFO    SenderThread:7834 [sender.py:send_exit():368] handling runtime: 208
+2022-01-14 22:19:03,954 INFO    SenderThread:7834 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 22:19:03,954 INFO    SenderThread:7834 [sender.py:send_exit():374] send defer
+2022-01-14 22:19:03,954 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:03,955 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:03,955 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 0
+2022-01-14 22:19:03,955 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:03,955 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 0
+2022-01-14 22:19:03,955 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 1
+2022-01-14 22:19:03,955 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:03,955 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 1
+2022-01-14 22:19:04,011 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:04,011 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 1
+2022-01-14 22:19:04,011 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 2
+2022-01-14 22:19:04,011 DEBUG   SenderThread:7834 [sender.py:send():234] send: stats
+2022-01-14 22:19:04,012 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:04,012 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 2
+2022-01-14 22:19:04,012 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:04,012 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 2
+2022-01-14 22:19:04,012 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 3
+2022-01-14 22:19:04,012 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:04,012 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 3
+2022-01-14 22:19:04,012 DEBUG   SenderThread:7834 [sender.py:send():234] send: summary
+2022-01-14 22:19:04,013 INFO    SenderThread:7834 [sender.py:_save_file():939] saving file wandb-summary.json with policy end
+2022-01-14 22:19:04,013 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:04,013 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 3
+2022-01-14 22:19:04,013 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 4
+2022-01-14 22:19:04,013 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:04,013 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 4
+2022-01-14 22:19:04,013 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:04,013 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 4
+2022-01-14 22:19:04,057 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:04,068 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:19:04,068 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-summary.json
+2022-01-14 22:19:04,198 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 5
+2022-01-14 22:19:04,198 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:04,198 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:04,198 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 5
+2022-01-14 22:19:04,199 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:04,199 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 5
+2022-01-14 22:19:04,199 INFO    SenderThread:7834 [dir_watcher.py:finish():283] shutting down directory watcher
+2022-01-14 22:19:04,300 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,068 INFO    Thread-8  :7834 [dir_watcher.py:_on_file_modified():230] file/dir modified: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/config.yaml
+2022-01-14 22:19:05,069 INFO    SenderThread:7834 [dir_watcher.py:finish():313] scan: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files
+2022-01-14 22:19:05,069 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/config.yaml config.yaml
+2022-01-14 22:19:05,069 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/diff.patch diff.patch
+2022-01-14 22:19:05,069 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/requirements.txt requirements.txt
+2022-01-14 22:19:05,069 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log output.log
+2022-01-14 22:19:05,070 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-summary.json wandb-summary.json
+2022-01-14 22:19:05,070 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-metadata.json wandb-metadata.json
+2022-01-14 22:19:05,072 INFO    SenderThread:7834 [dir_watcher.py:finish():327] scan save: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/code/run_mlm_flax.py code/run_mlm_flax.py
+2022-01-14 22:19:05,073 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 6
+2022-01-14 22:19:05,073 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,081 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:05,081 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 6
+2022-01-14 22:19:05,081 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:05,081 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 6
+2022-01-14 22:19:05,081 INFO    SenderThread:7834 [file_pusher.py:finish():177] shutting down file pusher
+2022-01-14 22:19:05,183 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,183 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,285 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,285 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,387 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,387 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,488 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,489 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,539 INFO    Thread-13 :7834 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/config.yaml
+2022-01-14 22:19:05,556 INFO    Thread-15 :7834 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/output.log
+2022-01-14 22:19:05,561 INFO    Thread-14 :7834 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/requirements.txt
+2022-01-14 22:19:05,590 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,590 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,599 INFO    Thread-16 :7834 [upload_job.py:push():137] Uploaded file /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/files/wandb-summary.json
+2022-01-14 22:19:05,692 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,692 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,794 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:05,794 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:05,799 INFO    Thread-7  :7834 [sender.py:transition_state():387] send defer: 7
+2022-01-14 22:19:05,800 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:05,800 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 7
+2022-01-14 22:19:05,800 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:05,800 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 7
+2022-01-14 22:19:05,896 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:06,218 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 8
+2022-01-14 22:19:06,218 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:06,218 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:06,218 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 8
+2022-01-14 22:19:06,219 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:06,219 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 8
+2022-01-14 22:19:06,219 INFO    SenderThread:7834 [sender.py:transition_state():387] send defer: 9
+2022-01-14 22:19:06,219 DEBUG   SenderThread:7834 [sender.py:send():234] send: final
+2022-01-14 22:19:06,219 DEBUG   SenderThread:7834 [sender.py:send():234] send: footer
+2022-01-14 22:19:06,219 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: defer
+2022-01-14 22:19:06,220 INFO    HandlerThread:7834 [handler.py:handle_request_defer():147] handle defer: 9
+2022-01-14 22:19:06,220 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: defer
+2022-01-14 22:19:06,220 INFO    SenderThread:7834 [sender.py:send_request_defer():383] handle sender defer: 9
+2022-01-14 22:19:06,320 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: poll_exit
+2022-01-14 22:19:06,320 DEBUG   SenderThread:7834 [sender.py:send_request():248] send_request: poll_exit
+2022-01-14 22:19:06,320 INFO    SenderThread:7834 [file_pusher.py:join():182] waiting for file pusher
+2022-01-14 22:19:06,598 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: get_summary
+2022-01-14 22:19:06,618 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: sampled_history
+2022-01-14 22:19:06,619 DEBUG   HandlerThread:7834 [handler.py:handle_request():130] handle_request: shutdown
+2022-01-14 22:19:06,619 INFO    HandlerThread:7834 [handler.py:finish():731] shutting down handler
+2022-01-14 22:19:07,220 INFO    WriterThread:7834 [datastore.py:close():281] close: /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/run-24dma583.wandb
+2022-01-14 22:19:07,575 INFO    SenderThread:7834 [sender.py:finish():1070] shutting down sender
+2022-01-14 22:19:07,575 INFO    SenderThread:7834 [file_pusher.py:finish():177] shutting down file pusher
+2022-01-14 22:19:07,576 INFO    SenderThread:7834 [file_pusher.py:join():182] waiting for file pusher
+2022-01-14 22:19:07,578 INFO    MainThread:7834 [internal.py:handle_exit():77] Internal process exited

wandb/run-20220114_221533-24dma583/logs/debug.log

@@ -0,0 +1,141 @@
+2022-01-14 22:15:33,825 INFO    MainThread:4503 [wandb_setup.py:_flush():71] setting env: {}
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_setup.py:_flush():71] setting login settings: {}
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_init.py:_log_setup():371] Logging user logs to /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/logs/debug.log
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_init.py:_log_setup():372] Logging internal logs to /data/roberta-base-ncc/wandb/run-20220114_221533-24dma583/logs/debug-internal.log
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_init.py:init():404] calling init triggers
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_init.py:init():409] wandb.init called with sweep_config: {}
+config: {}
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [wandb_init.py:init():460] starting backend
+2022-01-14 22:15:33,826 INFO    MainThread:4503 [backend.py:_multiprocessing_setup():99] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
+2022-01-14 22:15:33,871 INFO    MainThread:4503 [backend.py:ensure_launched():216] starting backend process...
+2022-01-14 22:15:33,898 INFO    MainThread:4503 [backend.py:ensure_launched():221] started backend process with pid: 7834
+2022-01-14 22:15:33,900 INFO    MainThread:4503 [wandb_init.py:init():469] backend started and connected
+2022-01-14 22:15:33,911 INFO    MainThread:4503 [wandb_init.py:init():533] updated telemetry
+2022-01-14 22:15:33,976 INFO    MainThread:4503 [wandb_init.py:init():563] communicating current version
+2022-01-14 22:15:34,784 INFO    MainThread:4503 [wandb_init.py:init():568] got version response 
+2022-01-14 22:15:34,784 INFO    MainThread:4503 [wandb_init.py:init():578] communicating run to backend with 30 second timeout
+2022-01-14 22:15:34,980 INFO    MainThread:4503 [wandb_init.py:init():606] starting run threads in backend
+2022-01-14 22:15:39,985 INFO    MainThread:4503 [wandb_run.py:_console_start():1810] atexit reg
+2022-01-14 22:15:39,985 INFO    MainThread:4503 [wandb_run.py:_redirect():1684] redirect: SettingsConsole.REDIRECT
+2022-01-14 22:15:39,986 INFO    MainThread:4503 [wandb_run.py:_redirect():1689] Redirecting console.
+2022-01-14 22:15:39,988 INFO    MainThread:4503 [wandb_run.py:_redirect():1745] Redirects installed.
+2022-01-14 22:15:39,988 INFO    MainThread:4503 [wandb_init.py:init():633] run started, returning control to user process
+2022-01-14 22:15:39,989 INFO    MainThread:4503 [wandb_run.py:_config_callback():956] config_cb None None {'output_dir': './', 'overwrite_output_dir': True, 'do_train': True, 'do_eval': True, 'per_device_train_batch_size': 250, 'per_device_eval_batch_size': 250, 'learning_rate': 0.0006, 'weight_decay': 0.01, 'adam_beta1': 0.9, 'adam_beta2': 0.98, 'adam_epsilon': 1e-06, 'adafactor': False, 'num_train_epochs': 3.0, 'warmup_steps': 10000, 'logging_steps': 1000, 'save_steps': 1000, 'eval_steps': 1000, 'seed': 42, 'push_to_hub': True, 'hub_model_id': None, 'hub_token': None}
+2022-01-14 22:15:39,989 INFO    MainThread:4503 [wandb_run.py:_config_callback():956] config_cb None None {'model_name_or_path': None, 'model_type': 'roberta', 'config_name': 'roberta-base', 'tokenizer_name': 'NbAiLab/nb-roberta-base', 'cache_dir': None, 'use_fast_tokenizer': True, 'dtype': 'bfloat16'}
+2022-01-14 22:15:39,990 INFO    MainThread:4503 [wandb_run.py:_config_callback():956] config_cb None None {'dataset_name': 'NbAiLab/NCC', 'dataset_config_name': None, 'train_file': None, 'validation_file': None, 'train_ref_file': None, 'validation_ref_file': None, 'overwrite_cache': False, 'validation_split_percentage': 5, 'max_seq_length': 128, 'preprocessing_num_workers': None, 'mlm_probability': 0.15, 'pad_to_max_length': True, 'line_by_line': False}
+2022-01-14 22:19:01,641 INFO    MainThread:4503 [wandb_run.py:_atexit_cleanup():1780] got exitcode: 1
+2022-01-14 22:19:01,645 INFO    MainThread:4503 [wandb_run.py:_restore():1752] restore
+2022-01-14 22:19:03,955 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-14 22:19:04,199 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-14 22:19:05,082 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,184 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,286 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,387 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,489 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,591 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,693 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:05,795 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:06,219 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+
+2022-01-14 22:19:06,576 INFO    MainThread:4503 [wandb_run.py:_wait_for_finish():1912] got exit ret: done: true
+exit_result {
+}
+file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 45535
+  total_bytes: 45535
+}
+local_info {
+}
+
+2022-01-14 22:19:09,886 INFO    MainThread:4503 [wandb_run.py:_append_files():2180] logging synced files

wandb/run-20220114_234119-1zya86oe/files/code/run_mlm_flax.py

@@ -0,0 +1,815 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 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 library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=fill-mask
+"""
+import json
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    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"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[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)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+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 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 main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level=logging.INFO,
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            # Enable Weight&Biases
+            import wandb
+            wandb.init(
+                entity='versae',
+                project='roberta-base-ncc',
+                sync_tensorboard=False,
+            )
+            wandb.config.update(training_args)
+            wandb.config.update(model_args)
+            wandb.config.update(data_args)
+
+            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."
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # 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()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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 FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = 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,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    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()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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 = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            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)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # 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 step {cur_step}", blocking=False)
+
+    # Eval after training
+    if training_args.do_eval:
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for _, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda metric: jnp.sum(metric).item(), eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        try:
+            perplexity = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            perplexity = float("inf")
+        eval_metrics["perplexity"] = perplexity
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

wandb/run-20220114_234119-1zya86oe/files/config.yaml

@@ -0,0 +1,152 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_mlm_flax.py
+    framework: huggingface
+    huggingface_version: 4.16.0.dev0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1642203679
+    t:
+      1:
+      - 2
+      - 3
+      - 11
+      - 12
+      2:
+      - 2
+      - 3
+      - 11
+      - 12
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.16.0.dev0
+      8:
+      - 5
+adafactor:
+  desc: null
+  value: false
+adam_beta1:
+  desc: null
+  value: 0.9
+adam_beta2:
+  desc: null
+  value: 0.98
+adam_epsilon:
+  desc: null
+  value: 1.0e-06
+cache_dir:
+  desc: null
+  value: null
+config_name:
+  desc: null
+  value: roberta-base
+dataset_config_name:
+  desc: null
+  value: null
+dataset_name:
+  desc: null
+  value: NbAiLab/NCC
+do_eval:
+  desc: null
+  value: true
+do_train:
+  desc: null
+  value: true
+dtype:
+  desc: null
+  value: bfloat16
+eval_steps:
+  desc: null
+  value: 1000
+hub_model_id:
+  desc: null
+  value: null
+hub_token:
+  desc: null
+  value: null
+learning_rate:
+  desc: null
+  value: 0.0006
+line_by_line:
+  desc: null
+  value: false
+logging_steps:
+  desc: null
+  value: 1000
+max_seq_length:
+  desc: null
+  value: 128
+mlm_probability:
+  desc: null
+  value: 0.15
+model_name_or_path:
+  desc: null
+  value: null
+model_type:
+  desc: null
+  value: roberta
+num_train_epochs:
+  desc: null
+  value: 3.0
+output_dir:
+  desc: null
+  value: ./
+overwrite_cache:
+  desc: null
+  value: false
+overwrite_output_dir:
+  desc: null
+  value: true
+pad_to_max_length:
+  desc: null
+  value: true
+per_device_eval_batch_size:
+  desc: null
+  value: 232
+per_device_train_batch_size:
+  desc: null
+  value: 232
+preprocessing_num_workers:
+  desc: null
+  value: null
+push_to_hub:
+  desc: null
+  value: true
+save_steps:
+  desc: null
+  value: 1000
+seed:
+  desc: null
+  value: 42
+tokenizer_name:
+  desc: null
+  value: NbAiLab/nb-roberta-base
+train_file:
+  desc: null
+  value: null
+train_ref_file:
+  desc: null
+  value: null
+use_fast_tokenizer:
+  desc: null
+  value: true
+validation_file:
+  desc: null
+  value: null
+validation_ref_file:
+  desc: null
+  value: null
+validation_split_percentage:
+  desc: null
+  value: 5
+warmup_steps:
+  desc: null
+  value: 10000
+weight_decay:
+  desc: null
+  value: 0.01

wandb/run-20220114_234119-1zya86oe/files/output.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85f905cb3060152eb20f5227d157a7605e33c762b80a6b8f4792e5791a1acd2b
+size 26403055

wandb/run-20220114_234119-1zya86oe/files/requirements.txt

@@ -0,0 +1,122 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+astunparse==1.6.3
+async-timeout==4.0.2
+attrs==21.4.0
+backcall==0.2.0
+cachetools==4.2.4
+certifi==2021.10.8
+charset-normalizer==2.0.10
+chex==0.1.0
+click==8.0.3
+clu==0.0.6
+configparser==5.2.0
+contextlib2==21.6.0
+cycler==0.11.0
+datasets==1.17.1.dev0
+decorator==5.1.0
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+filelock==3.4.2
+flatbuffers==2.0
+flax==0.3.6
+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-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
+huggingface-hub==0.2.1
+idna==3.3
+importlib-metadata==4.10.0
+importlib-resources==5.4.0
+ipython==7.31.0
+jax==0.2.26
+jaxlib==0.1.75
+jedi==0.18.1
+joblib==1.1.0
+keras-preprocessing==1.1.2
+keras==2.7.0
+kiwisolver==1.3.2
+libclang==12.0.0
+libtpu-nightly==0.1.dev20211208
+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
+numpy==1.22.0
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.1.0
+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==20.0.2
+pkg-resources==0.0.0
+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
+pygments==2.11.1
+pyparsing==3.0.6
+python-dateutil==2.8.2
+pytz==2021.3
+pyyaml==6.0
+regex==2021.11.10
+requests-oauthlib==1.3.0
+requests==2.27.0
+rsa==4.8
+sacremoses==0.0.46
+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
+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
+termcolor==1.1.0
+tokenizers==0.11.2
+toolz==0.11.2
+tqdm==4.62.3
+traitlets==5.1.1
+transformers==4.16.0.dev0
+typing-extensions==3.10.0.2
+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-20220114_234119-1zya86oe/files/wandb-metadata.json

@@ -0,0 +1,47 @@
+{
+    "os": "Linux-5.4.0-1043-gcp-x86_64-with-glibc2.29",
+    "python": "3.8.10",
+    "heartbeatAt": "2022-01-14T23:41:22.464389",
+    "startedAt": "2022-01-14T23:41:19.296699",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--output_dir=./",
+        "--model_type=roberta",
+        "--config_name=roberta-base",
+        "--tokenizer_name=NbAiLab/nb-roberta-base",
+        "--dataset_name=NbAiLab/NCC",
+        "--max_seq_length=128",
+        "--weight_decay=0.01",
+        "--per_device_train_batch_size=232",
+        "--per_device_eval_batch_size=232",
+        "--pad_to_max_length",
+        "--learning_rate=6e-4",
+        "--warmup_steps=10000",
+        "--overwrite_output_dir",
+        "--num_train_epochs=3",
+        "--adam_beta1=0.9",
+        "--adam_beta2=0.98",
+        "--adam_epsilon=1e-6",
+        "--logging_steps=1000",
+        "--save_steps=1000",
+        "--eval_steps=1000",
+        "--do_train",
+        "--do_eval",
+        "--dtype=bfloat16",
+        "--push_to_hub"
+    ],
+    "state": "running",
+    "program": "run_mlm_flax.py",
+    "codePath": "run_mlm_flax.py",
+    "git": {
+        "remote": "https://huggingface.co/versae/roberta-base-ncc",
+        "commit": "502df078f73cf93ca9380fcac1c9b9c7598a445f"
+    },
+    "email": "versae@gmail.com",
+    "root": "/data/roberta-base-ncc",
+    "host": "t1v-n-eedfb410-w-0",
+    "username": "javierr",
+    "executable": "/data/flax/bin/python"
+}

wandb/run-20220114_234119-1zya86oe/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb": {"runtime": 356856}}

wandb/run-20220114_234119-1zya86oe/logs/debug-internal.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0a61912984f47a6ce815f00cc68ecb84206e92936ff60cd6573447587dd00aa
+size 38335870

wandb/run-20220114_234119-1zya86oe/logs/debug.log

@@ -0,0 +1,168 @@
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_setup.py:_flush():71] setting env: {}
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_setup.py:_flush():71] setting login settings: {}
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_init.py:_log_setup():371] Logging user logs to /data/roberta-base-ncc/wandb/run-20220114_234119-1zya86oe/logs/debug.log
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_init.py:_log_setup():372] Logging internal logs to /data/roberta-base-ncc/wandb/run-20220114_234119-1zya86oe/logs/debug-internal.log
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_init.py:init():404] calling init triggers
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_init.py:init():409] wandb.init called with sweep_config: {}
+config: {}
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [wandb_init.py:init():460] starting backend
+2022-01-14 23:41:19,298 INFO    MainThread:10537 [backend.py:_multiprocessing_setup():99] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
+2022-01-14 23:41:19,328 INFO    MainThread:10537 [backend.py:ensure_launched():216] starting backend process...
+2022-01-14 23:41:19,355 INFO    MainThread:10537 [backend.py:ensure_launched():221] started backend process with pid: 11989
+2022-01-14 23:41:19,357 INFO    MainThread:10537 [wandb_init.py:init():469] backend started and connected
+2022-01-14 23:41:19,369 INFO    MainThread:10537 [wandb_init.py:init():533] updated telemetry
+2022-01-14 23:41:19,437 INFO    MainThread:10537 [wandb_init.py:init():563] communicating current version
+2022-01-14 23:41:20,144 INFO    MainThread:10537 [wandb_init.py:init():568] got version response 
+2022-01-14 23:41:20,145 INFO    MainThread:10537 [wandb_init.py:init():578] communicating run to backend with 30 second timeout
+2022-01-14 23:41:20,323 INFO    MainThread:10537 [wandb_init.py:init():606] starting run threads in backend
+2022-01-14 23:41:25,327 INFO    MainThread:10537 [wandb_run.py:_console_start():1810] atexit reg
+2022-01-14 23:41:25,328 INFO    MainThread:10537 [wandb_run.py:_redirect():1684] redirect: SettingsConsole.REDIRECT
+2022-01-14 23:41:25,328 INFO    MainThread:10537 [wandb_run.py:_redirect():1689] Redirecting console.
+2022-01-14 23:41:25,330 INFO    MainThread:10537 [wandb_run.py:_redirect():1745] Redirects installed.
+2022-01-14 23:41:25,331 INFO    MainThread:10537 [wandb_init.py:init():633] run started, returning control to user process
+2022-01-14 23:41:25,331 INFO    MainThread:10537 [wandb_run.py:_config_callback():956] config_cb None None {'output_dir': './', 'overwrite_output_dir': True, 'do_train': True, 'do_eval': True, 'per_device_train_batch_size': 232, 'per_device_eval_batch_size': 232, 'learning_rate': 0.0006, 'weight_decay': 0.01, 'adam_beta1': 0.9, 'adam_beta2': 0.98, 'adam_epsilon': 1e-06, 'adafactor': False, 'num_train_epochs': 3.0, 'warmup_steps': 10000, 'logging_steps': 1000, 'save_steps': 1000, 'eval_steps': 1000, 'seed': 42, 'push_to_hub': True, 'hub_model_id': None, 'hub_token': None}
+2022-01-14 23:41:25,332 INFO    MainThread:10537 [wandb_run.py:_config_callback():956] config_cb None None {'model_name_or_path': None, 'model_type': 'roberta', 'config_name': 'roberta-base', 'tokenizer_name': 'NbAiLab/nb-roberta-base', 'cache_dir': None, 'use_fast_tokenizer': True, 'dtype': 'bfloat16'}
+2022-01-14 23:41:25,332 INFO    MainThread:10537 [wandb_run.py:_config_callback():956] config_cb None None {'dataset_name': 'NbAiLab/NCC', 'dataset_config_name': None, 'train_file': None, 'validation_file': None, 'train_ref_file': None, 'validation_ref_file': None, 'overwrite_cache': False, 'validation_split_percentage': 5, 'max_seq_length': 128, 'preprocessing_num_workers': None, 'mlm_probability': 0.15, 'pad_to_max_length': True, 'line_by_line': False}
+2022-01-19 02:48:53,379 INFO    MainThread:10537 [wandb_run.py:_atexit_cleanup():1780] got exitcode: 0
+2022-01-19 02:48:53,381 INFO    MainThread:10537 [wandb_run.py:_restore():1752] restore
+2022-01-19 02:48:56,346 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-19 02:48:56,559 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 1
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 37446
+}
+
+2022-01-19 02:48:56,919 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,021 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 37446
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,123 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 10068910
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,225 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 13042606
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,327 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 21463982
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,429 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,531 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,633 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,735 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,837 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:57,939 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:58,457 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+
+2022-01-19 02:48:58,818 INFO    MainThread:10537 [wandb_run.py:_wait_for_finish():1912] got exit ret: done: true
+exit_result {
+}
+file_counts {
+  wandb_count: 5
+  other_count: 1
+}
+pusher_stats {
+  uploaded_bytes: 26444957
+  total_bytes: 26444957
+}
+local_info {
+}
+
+2022-01-19 02:49:00,429 INFO    MainThread:10537 [wandb_run.py:_append_files():2180] logging synced files

wandb/run-20220114_234119-1zya86oe/run-1zya86oe.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f93d599b590506c08952e582e6ab64f317f0ac3488bf1937504d605ab8ecf5b
+size 118429569

wandb/run-20220119_161158-274aad95/files/code/run_mlm_flax.py

@@ -0,0 +1,815 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 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 library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=fill-mask
+"""
+import json
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from enum import Enum
+from itertools import chain
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from huggingface_hub import Repository
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.file_utils import get_full_repo_name
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class TrainingArguments:
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
+    )
+    overwrite_output_dir: bool = field(
+        default=False,
+        metadata={
+            "help": (
+                "Overwrite the content of the output directory. "
+                "Use this to continue training if output_dir points to a checkpoint directory."
+            )
+        },
+    )
+    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
+    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
+    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
+    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
+    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
+    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
+    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
+    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
+    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
+    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
+    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
+    push_to_hub: bool = field(
+        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
+    )
+    hub_model_id: str = field(
+        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
+    )
+    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+
+    def __post_init__(self):
+        if self.output_dir is not None:
+            self.output_dir = os.path.expanduser(self.output_dir)
+
+    def to_dict(self):
+        """
+        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
+        the token values by removing their value.
+        """
+        d = asdict(self)
+        for k, v in d.items():
+            if isinstance(v, Enum):
+                d[k] = v.value
+            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
+                d[k] = [x.value for x in v]
+            if k.endswith("_token"):
+                d[k] = f"<{k.upper()}>"
+        return d
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    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"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[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)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+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 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 main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level=logging.INFO,
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            # Enable Weight&Biases
+            import wandb
+            wandb.init(
+                entity='versae',
+                project='roberta-base-ncc',
+                sync_tensorboard=False,
+            )
+            wandb.config.update(training_args)
+            wandb.config.update(model_args)
+            wandb.config.update(data_args)
+
+            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."
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # 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()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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 FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = 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,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    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()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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 = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            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)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # 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 step {cur_step}", blocking=False)
+
+    # Eval after training
+    if training_args.do_eval:
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for _, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(lambda metric: jnp.sum(metric).item(), eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        try:
+            perplexity = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            perplexity = float("inf")
+        eval_metrics["perplexity"] = perplexity
+
+        if jax.process_index() == 0:
+            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
+            path = os.path.join(training_args.output_dir, "eval_results.json")
+            with open(path, "w") as f:
+                json.dump(eval_metrics, f, indent=4, sort_keys=True)
+
+
+if __name__ == "__main__":
+    main()

wandb/run-20220119_161158-274aad95/files/config.yaml

@@ -0,0 +1,147 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.12.9
+    code_path: code/run_mlm_flax.py
+    framework: huggingface
+    huggingface_version: 4.16.0.dev0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    start_time: 1642608719
+    t:
+      1:
+      - 2
+      - 3
+      - 11
+      - 12
+      4: 3.8.10
+      5: 0.12.9
+      6: 4.16.0.dev0
+      8:
+      - 5
+adafactor:
+  desc: null
+  value: false
+adam_beta1:
+  desc: null
+  value: 0.9
+adam_beta2:
+  desc: null
+  value: 0.98
+adam_epsilon:
+  desc: null
+  value: 1.0e-06
+cache_dir:
+  desc: null
+  value: null
+config_name:
+  desc: null
+  value: ./
+dataset_config_name:
+  desc: null
+  value: null
+dataset_name:
+  desc: null
+  value: NbAiLab/NCC
+do_eval:
+  desc: null
+  value: true
+do_train:
+  desc: null
+  value: true
+dtype:
+  desc: null
+  value: bfloat16
+eval_steps:
+  desc: null
+  value: 1000
+hub_model_id:
+  desc: null
+  value: null
+hub_token:
+  desc: null
+  value: null
+learning_rate:
+  desc: null
+  value: 0.0006
+line_by_line:
+  desc: null
+  value: false
+logging_steps:
+  desc: null
+  value: 1000
+max_seq_length:
+  desc: null
+  value: 512
+mlm_probability:
+  desc: null
+  value: 0.15
+model_name_or_path:
+  desc: null
+  value: ./
+model_type:
+  desc: null
+  value: roberta
+num_train_epochs:
+  desc: null
+  value: 3.0
+output_dir:
+  desc: null
+  value: ./
+overwrite_cache:
+  desc: null
+  value: false
+overwrite_output_dir:
+  desc: null
+  value: true
+pad_to_max_length:
+  desc: null
+  value: true
+per_device_eval_batch_size:
+  desc: null
+  value: 46
+per_device_train_batch_size:
+  desc: null
+  value: 46
+preprocessing_num_workers:
+  desc: null
+  value: null
+push_to_hub:
+  desc: null
+  value: true
+save_steps:
+  desc: null
+  value: 1000
+seed:
+  desc: null
+  value: 42
+tokenizer_name:
+  desc: null
+  value: ./
+train_file:
+  desc: null
+  value: null
+train_ref_file:
+  desc: null
+  value: null
+use_fast_tokenizer:
+  desc: null
+  value: true
+validation_file:
+  desc: null
+  value: null
+validation_ref_file:
+  desc: null
+  value: null
+validation_split_percentage:
+  desc: null
+  value: 5
+warmup_steps:
+  desc: null
+  value: 1000
+weight_decay:
+  desc: null
+  value: 0.01