Step... (1000/50000 | Loss: 1.7686773538589478, Acc: 0.6487793326377869): 3%|▊ | 1286/50000 [29:40<20:20:20, 1.50s/it]

README.md

@@ -0,0 +1,31 @@
+---
+language: es
+license: CC-BY 4.0
+tags:
+- spanish
+- roberta
+pipeline_tag: fill-mask
+widget:
+- text: "Fui a la librería a comprar un <mask>."
+---
+
+This is a **RoBERTa-base** model trained from scratch in Spanish.
+
+The training dataset is [mc4](https://huggingface.co/datasets/bertin-project/mc4-es-sampled ) subsampling documents to a total of about 50 million examples. Sampling is biased towards average perplexity values (using a Gaussian function), discarding more often documents with very large values (poor quality) of very small values (short, repetitive texts).
+
+This model takes the one using [sequence length 128](https://huggingface.co/bertin-project/bertin-base-stepwise) and trains during 25.000 steps using sequence length 512.
+
+Please see our main [card](https://huggingface.co/bertin-project/bertin-roberta-base-spanish) for more information.
+
+This is part of the
+[Flax/Jax Community Week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104), organised by [HuggingFace](https://huggingface.co/) and TPU usage sponsored by Google.
+
+
+## Team members
+
+- Eduardo González ([edugp](https://huggingface.co/edugp))
+- Javier de la Rosa ([versae](https://huggingface.co/versae))
+- Manu Romero ([mrm8488](https://huggingface.co/))
+- María Grandury ([mariagrandury](https://huggingface.co/))
+- Pablo González de Prado ([Pablogps](https://huggingface.co/Pablogps))
+- Paulo Villegas ([paulo](https://huggingface.co/paulo))

config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "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": "4.9.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

configs/base/config.json

@@ -0,0 +1,25 @@
+{
+    "architectures": [
+      "RobertaForMaskedLM"
+    ],
+    "attention_probs_dropout_prob": 0.1,
+    "bos_token_id": 0,
+    "eos_token_id": 2,
+    "gradient_checkpointing": false,
+    "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": "4.9.0.dev0",
+    "type_vocab_size": 1,
+    "use_cache": true,
+    "vocab_size": 50265
+  }

configs/large/config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.9.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

convert.py

@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+import tempfile
+
+import jax
+from jax import numpy as jnp
+from transformers import AutoTokenizer, FlaxRobertaForMaskedLM, RobertaForMaskedLM
+
+
+def to_f32(t):
+    return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def main():
+    # Saving extra files from config.json and tokenizer.json files
+    tokenizer = AutoTokenizer.from_pretrained("./")
+    tokenizer.save_pretrained("./")
+
+    # Temporary saving bfloat16 Flax model into float32
+    tmp = tempfile.mkdtemp()
+    flax_model = FlaxRobertaForMaskedLM.from_pretrained("./")
+    flax_model.params = to_f32(flax_model.params)
+    flax_model.save_pretrained(tmp)
+    # Converting float32 Flax to PyTorch
+    model = RobertaForMaskedLM.from_pretrained(tmp, from_flax=True)
+    model.save_pretrained("./", save_config=False)
+
+
+if __name__ == "__main__":
+    main()

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a03759b12cc223a5e978aaada836d2e6c39f107a37072fc713c1cd6c8cb58f1
+size 249750019

outputs/checkpoints/checkpoint-1000/config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "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": "4.9.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

outputs/checkpoints/checkpoint-1000/data_collator.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e02a6e9cfa63cb321cac9402efd29841b652999fcbf787800ae050e747b161ee
+size 1471394

outputs/checkpoints/checkpoint-1000/flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a03759b12cc223a5e978aaada836d2e6c39f107a37072fc713c1cd6c8cb58f1
+size 249750019

outputs/checkpoints/checkpoint-1000/optimizer_state.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45abc23402f1ddc533f4b24453f01978f94bc139d5216bf074d6542cec750bab
+size 499500278

outputs/checkpoints/checkpoint-1000/training_args.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dcde99c91fe01c5143995806e1d6595b728cb8ed0a2d9f2f3c5610aeebeb7c2
+size 1871

outputs/checkpoints/checkpoint-1000/training_state.json

@@ -0,0 +1 @@
+{"step": 1001}

outputs/config.json

@@ -0,0 +1,25 @@
+{
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "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": "4.9.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

outputs/data_collator.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e02a6e9cfa63cb321cac9402efd29841b652999fcbf787800ae050e747b161ee
+size 1471394

outputs/events.out.tfevents.1627258355.tablespoon.3000110.3.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c15d6accbbffb760087f0c740d328a1efb90da5cf430f6b59cc62f5450455429
+size 147205

outputs/flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a03759b12cc223a5e978aaada836d2e6c39f107a37072fc713c1cd6c8cb58f1
+size 249750019

outputs/optimizer_state.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45abc23402f1ddc533f4b24453f01978f94bc139d5216bf074d6542cec750bab
+size 499500278

outputs/training_args.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dcde99c91fe01c5143995806e1d6595b728cb8ed0a2d9f2f3c5610aeebeb7c2
+size 1871

outputs/training_state.json

@@ -0,0 +1 @@
+{"step": 1001}

push_to_hub.sh

@@ -0,0 +1,3 @@
+git add -A
+git commit -m "$(sed 's/\r/\n/g' run_stream.512.log | tail -1)"
+git push

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c870186998e7c252130640a265d196f785ac49cace7f1f7085dfd0cf139aa6
+size 498858859

run_mlm_flax_stream.py

@@ -0,0 +1,832 @@
+#!/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=masked-lm
+"""
+import logging
+import json
+import os
+import shutil
+import sys
+import tempfile
+import time
+from collections import defaultdict
+from dataclasses import dataclass, field
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+import joblib
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import datasets
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import kenlm  # pip install https://github.com/kpu/kenlm/archive/master.zip
+import optax
+from flax import jax_utils, traverse_util
+from flax.serialization import from_bytes, to_bytes
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+    FlaxRobertaForMaskedLM,
+    RobertaForMaskedLM,
+)
+
+
+if datasets.__version__ <= "1.8.0":
+    raise ValueError("Make sure to upgrade `datasets` to a version >= 1.9.0 to use dataset streaming")
+
+
+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 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."},
+    )
+    text_column_name: str = field(
+        default="text", metadata={"help": "The name of the column to retrieve the training text."}
+    )
+    shuffle_buffer_size: int = field(
+        default=10000, metadata={"help": "The number of examples to pre-load for shuffling."}
+    )
+    num_train_steps: int = field(default=50000, metadata={"help": "The number of training steps."})
+    num_eval_samples: int = field(default=50000, metadata={"help": "The number of samples to be used for evaluation"})
+
+    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", "jsonl", "txt", "gz"], "`train_file` should be a csv, a json (lines) or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "jsonl", "txt", "gz"], "`validation_file` should be a csv, a json (lines) 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[jnp.ndarray, jnp.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
+
+
+@dataclass
+class SamplingArguments:
+    """
+    Arguments pertaining to how to perform sampling of the dataset.
+    """
+
+    perplexity_model: Optional[str] = field(
+        default="./es.arpa.bin", metadata={"help": "Path to KenLM model to use to get perplexity values."}
+    )
+    sampling_method: Optional[str] = field(
+        default=None, metadata={"help": "Sample using a 'step' or 'gaussian' perplexity function per document, or 'random'."}
+    )
+    sampling_factor: Optional[float]  = field(
+        default=None, metadata={"help": "Sampling factor. Integers for step function, decimals for gaussian."}
+    )
+    boundaries: Optional[str] = field(
+        default="536394.99320948,662247.50212365,919250.87225178", metadata={"help": "Quartile boundaries"}
+    )
+
+    def __post_init__(self):
+        self.boundaries = [float(q.strip()) for q in self.boundaries.split(",")]
+
+
+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 advance_iter_and_group_samples(train_iterator, num_samples, max_seq_length):
+    """
+    The training iterator is advanced so that after groupifying the samples,
+    `num_samples` of length `max_seq_length` are returned.
+    """
+    num_total_tokens = max_seq_length * num_samples
+    samples = defaultdict(list)
+
+    i = 0
+    while i < num_total_tokens:
+        tokenized_samples = next(train_iterator)
+        i += len(tokenized_samples["input_ids"])
+
+        # concatenate tokenized samples to list
+        samples = {k: samples[k] + tokenized_samples[k] for k in tokenized_samples.keys()}
+
+    # Concatenated tokens are split to lists of length `max_seq_length`.
+    # Note that remainedr of % max_seq_length are thrown away.
+    def group_texts(examples):
+        result = {
+            k: [t[i : i + max_seq_length] for i in range(0, num_total_tokens, max_seq_length)]
+            for k, t in examples.items()
+        }
+        return result
+
+    grouped_samples = group_texts(samples)
+    return grouped_samples
+
+
+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 save_checkpoint_files(state, data_collator, training_args, save_dir):
+    unreplicated_state = jax_utils.unreplicate(state)
+    with open(os.path.join(save_dir, "optimizer_state.msgpack"), "wb") as f:
+        f.write(to_bytes(unreplicated_state.opt_state))
+    joblib.dump(training_args, os.path.join(save_dir, "training_args.joblib"))
+    joblib.dump(data_collator, os.path.join(save_dir, "data_collator.joblib"))
+    with open(os.path.join(save_dir, "training_state.json"), "w") as f:
+        json.dump({"step": unreplicated_state.step.item()}, f)
+
+
+def restore_checkpoint(save_dir, state):
+    logger.info(f"Restoring checkpoint from {save_dir}")
+    with open(os.path.join(save_dir, "flax_model.msgpack"), "rb") as f:
+        params = from_bytes(state.params, f.read())
+
+    with open(os.path.join(save_dir, "optimizer_state.msgpack"), "rb") as f:
+        opt_state = from_bytes(state.opt_state, f.read())
+
+    args = joblib.load(os.path.join(save_dir, "training_args.joblib"))
+    data_collator = joblib.load(os.path.join(save_dir, "data_collator.joblib"))
+
+    with open(os.path.join(save_dir, "training_state.json"), "r") as f:
+        training_state = json.load(f)
+    step = training_state["step"]
+
+    return params, opt_state, step, args, data_collator
+
+
+def rotate_checkpoints(path, max_checkpoints=5):
+    paths = sorted(Path(path).iterdir(), key=os.path.getmtime)[::-1]
+    if len(paths) > max_checkpoints:
+        for path_to_delete in paths[max_checkpoints:]:
+            try:
+                shutil.rmtree(path_to_delete)
+            except OSError:
+                os.remove(path_to_delete)
+
+
+def to_f32(t):
+    return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def convert(output_dir, destination_dir="./"):
+    shutil.copyfile(Path(output_dir) / "flax_model.msgpack", Path(destination_dir) / "flax_model.msgpack")
+    shutil.copyfile(Path(output_dir) / "config.json", Path(destination_dir) / "config.json")
+    # Saving extra files from config.json and tokenizer.json files
+    tokenizer = AutoTokenizer.from_pretrained(destination_dir)
+    tokenizer.save_pretrained(destination_dir)
+
+    # Temporary saving bfloat16 Flax model into float32
+    tmp = tempfile.mkdtemp()
+    flax_model = FlaxRobertaForMaskedLM.from_pretrained(destination_dir)
+    flax_model.params = to_f32(flax_model.params)
+    flax_model.save_pretrained(tmp)
+    # Converting float32 Flax to PyTorch
+    model = RobertaForMaskedLM.from_pretrained(tmp, from_flax=True)
+    model.save_pretrained(destination_dir, save_config=False)
+
+
+if __name__ == "__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, SamplingArguments))
+    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, sampling_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args, sampling_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="INFO",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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).
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        filepaths = {}
+        if data_args.train_file:
+            filepaths["train"] = data_args.train_file
+        if data_args.validation_file:
+            filepaths["validation"] = data_args.validation_file
+        try:
+            dataset = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                cache_dir=model_args.cache_dir,
+                streaming=True,
+                split="train",
+                sampling_method=sampling_args.sampling_method,
+                sampling_factor=sampling_args.sampling_factor,
+                boundaries=sampling_args.boundaries,
+                perplexity_model=sampling_args.perplexity_model,
+                seed=training_args.seed,
+                data_files=filepaths,
+            )
+        except Exception as exc:
+            logger.warning(
+                f"Unable to load local dataset with perplexity sampling support. Using huggingface.co/datasets/{data_args.dataset_name}: {exc}"
+            )
+            dataset = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                cache_dir=model_args.cache_dir,
+                streaming=True,
+                split="train",
+            )
+
+    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."
+        )
+
+    # 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[data_args.text_column_name],
+            return_special_tokens_mask=True
+        )
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+    )
+
+    shuffle_seed = training_args.seed
+    tokenized_datasets = tokenized_datasets.shuffle(buffer_size=data_args.shuffle_buffer_size, seed=shuffle_seed)
+
+    # 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='wandb',
+                project='hf-flax-bertin-roberta-es',
+                sync_tensorboard=True,
+            )
+            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()
+
+    # define number steps per stream epoch
+    num_train_steps = data_args.num_train_steps
+
+    # 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
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+        mask=decay_mask_fn,
+    )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
+    saved_step = -1
+    if model_args.model_name_or_path and "checkpoint" in model_args.model_name_or_path:
+        params, opt_state, saved_step, args, data_collator = restore_checkpoint(model_args.model_name_or_path, state)
+        # Create learning rate schedule
+        warmup_fn = optax.linear_schedule(
+            init_value=0.0, end_value=args.learning_rate, transition_steps=args.warmup_steps
+        )
+        decay_fn = optax.linear_schedule(
+            init_value=args.learning_rate,
+            end_value=0,
+            transition_steps=data_args.num_train_steps - args.warmup_steps,
+        )
+        linear_decay_lr_schedule_fn = optax.join_schedules(
+            schedules=[warmup_fn, decay_fn], boundaries=[args.warmup_steps]
+        )
+        # create adam optimizer
+        adamw = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=args.weight_decay,
+            mask=decay_mask_fn,
+        )
+        state = train_state.TrainState(
+            step=saved_step,
+            apply_fn=model.__call__,
+            params=params,
+            tx=adamw,
+            opt_state=opt_state,
+        )
+        # self.args = args
+        # data_collator = data_collator
+        # scheduler_fn = args.learning_rate
+        model.params = params
+
+
+    # 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
+    train_start = time.time()
+    train_metrics = []
+    eval_metrics = []
+
+    training_iter = iter(tokenized_datasets)
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+    eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
+
+    last_desc = ""
+    steps = tqdm(range(num_train_steps), desc="Training...", position=0)
+    for step in range(num_train_steps):
+        if step < saved_step:
+            steps.update(1)
+            continue
+        # ======================== Training ================================
+        try:
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
+        except StopIteration:
+            # Once the end of the dataset stream is reached, the training iterator
+            # is reinitialized and reshuffled and a new eval dataset is randomely chosen.
+            shuffle_seed += 1
+            tokenized_datasets.set_epoch(shuffle_seed)
+
+            training_iter = iter(tokenized_datasets)
+
+            eval_dataset = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
+
+        # process input samples
+        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)
+
+        if step % training_args.logging_steps == 0 and step > 0:
+            steps.write(
+                f"Step... ({step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+            )
+            train_time += time.time() - train_start
+            if has_tensorboard and jax.process_index() == 0:
+                write_train_metric(summary_writer, train_metrics, train_time, step)
+            train_metrics = []
+
+        # ======================== Evaluating ==============================
+        if step % training_args.eval_steps == 0 and step > 0:
+            eval_samples_idx = jnp.arange(data_args.num_eval_samples)
+            eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+            for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=1)):
+                # process input samples
+                batch_eval_samples = {k: [v[idx] for idx in batch_idx] for k, v in eval_samples.items()}
+                model_inputs = data_collator(batch_eval_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
+            steps.desc = f"Step... ({step}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+            last_desc = steps.desc
+
+            if has_tensorboard and jax.process_index() == 0:
+                write_eval_metric(summary_writer, eval_metrics, step)
+            eval_metrics = []
+
+        # save checkpoint after eval_steps
+        if step % training_args.save_steps == 0 and step > 0 and jax.process_index() == 0:
+            logger.info(f"Saving checkpoint at {step} steps")
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=False,
+            )
+            save_checkpoint_files(state, data_collator, training_args, training_args.output_dir)
+            checkpoints_dir = Path(training_args.output_dir) / "checkpoints" / f"checkpoint-{step}"
+            checkpoints_dir.mkdir(parents=True, exist_ok=True)
+            model.save_pretrained(checkpoints_dir, params=params)
+            save_checkpoint_files(state, data_collator, training_args, checkpoints_dir)
+            rotate_checkpoints(
+                Path(training_args.output_dir) / "checkpoints",
+                max_checkpoints=training_args.save_total_limit
+            )
+            convert(training_args.output_dir, "./")
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=training_args.push_to_hub,
+                commit_message=last_desc,
+            )
+
+        # update tqdm bar
+        steps.update(1)
+
+    if jax.process_index() == 0:
+        logger.info(f"Saving checkpoint at {step} steps")
+        params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+        model.save_pretrained(
+            training_args.output_dir,
+            params=params,
+            push_to_hub=False,
+        )
+        save_checkpoint_files(state, data_collator, training_args, training_args.output_dir)
+        checkpoints_dir = Path(training_args.output_dir) / "checkpoints" / f"checkpoint-{step}"
+        checkpoints_dir.mkdir(parents=True, exist_ok=True)
+        model.save_pretrained(checkpoints_dir, params=params)
+        save_checkpoint_files(state, data_collator, training_args, checkpoints_dir)
+        convert(training_args.output_dir, "./")
+        model.save_pretrained(
+            training_args.output_dir,
+            params=params,
+            push_to_hub=training_args.push_to_hub,
+            commit_message=last_desc or "Saving model after training",
+        )

run_stream.128.sh

@@ -0,0 +1,27 @@
+# From https://arxiv.org/pdf/1907.11692.pdf for base model
+python -c "import jax; print('TPUs', jax.device_count())"
+python ./run_mlm_flax_stream.py \
+    --output_dir="./outputs" \
+    --model_type="roberta" \
+    --config_name="./configs/base" \
+    --tokenizer_name="./configs/base" \
+    --dataset_name="./mc4" \
+    --dataset_config_name="es" \
+    --train_file="../mc4-es-train-50M-random.jsonl" \
+    --max_seq_length="128" \
+    --pad_to_max_length  \
+    --per_device_train_batch_size="256" \
+    --per_device_eval_batch_size="256" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --learning_rate="6e-4" \
+    --weight_decay="0.01" \
+    --save_steps="10000" \
+    --save_total_limit="5" \
+    --warmup_steps="24000" \
+    --overwrite_output_dir \
+    --num_train_steps="250000" \
+    --eval_steps="10000" \
+    --dtype="bfloat16" \
+    --logging_steps="500" 2>&1 | tee run_stream.log

run_stream.512.sh

@@ -0,0 +1,27 @@
+# From https://arxiv.org/pdf/1907.11692.pdf for base model
+python -c "import jax; print('TPUs', jax.device_count())"
+python ./run_mlm_flax_stream.py \
+    --model_name_or_path="bertin-project/bertin-base-stepwise" \
+    --output_dir="./outputs" \
+    --model_type="roberta" \
+    --config_name="./configs/base" \
+    --tokenizer_name="./configs/base" \
+    --dataset_name="bertin-project/mc4-es-sampled" \
+    --dataset_config_name="stepwise" \
+    --max_seq_length="512" \
+    --pad_to_max_length  \
+    --per_device_train_batch_size="48" \
+    --per_device_eval_batch_size="48" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --learning_rate="6e-4" \
+    --weight_decay="0.01" \
+    --save_steps="1000" \
+    --save_total_limit="5" \
+    --warmup_steps="500" \
+    --overwrite_output_dir \
+    --num_train_steps="50000" \
+    --eval_steps="1000" \
+    --dtype="bfloat16" \
+    --logging_steps="500" 2>&1 | tee run_stream.512.log

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>", "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "RobertaTokenizer"}

wandb/debug-internal.log

@@ -0,0 +1 @@
+run-20210726_001233-17u6inbn/logs/debug-internal.log

wandb/debug.log

@@ -0,0 +1 @@
+run-20210726_001233-17u6inbn/logs/debug.log

wandb/latest-run

@@ -0,0 +1 @@
+run-20210726_001233-17u6inbn

wandb/run-20210726_001233-17u6inbn/files/code/run_mlm_flax_stream.py

@@ -0,0 +1,832 @@
+#!/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=masked-lm
+"""
+import logging
+import json
+import os
+import shutil
+import sys
+import tempfile
+import time
+from collections import defaultdict
+from dataclasses import dataclass, field
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+import joblib
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import datasets
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import kenlm  # pip install https://github.com/kpu/kenlm/archive/master.zip
+import optax
+from flax import jax_utils, traverse_util
+from flax.serialization import from_bytes, to_bytes
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+    FlaxRobertaForMaskedLM,
+    RobertaForMaskedLM,
+)
+
+
+if datasets.__version__ <= "1.8.0":
+    raise ValueError("Make sure to upgrade `datasets` to a version >= 1.9.0 to use dataset streaming")
+
+
+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 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."},
+    )
+    text_column_name: str = field(
+        default="text", metadata={"help": "The name of the column to retrieve the training text."}
+    )
+    shuffle_buffer_size: int = field(
+        default=10000, metadata={"help": "The number of examples to pre-load for shuffling."}
+    )
+    num_train_steps: int = field(default=50000, metadata={"help": "The number of training steps."})
+    num_eval_samples: int = field(default=50000, metadata={"help": "The number of samples to be used for evaluation"})
+
+    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", "jsonl", "txt", "gz"], "`train_file` should be a csv, a json (lines) or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "jsonl", "txt", "gz"], "`validation_file` should be a csv, a json (lines) 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[jnp.ndarray, jnp.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
+
+
+@dataclass
+class SamplingArguments:
+    """
+    Arguments pertaining to how to perform sampling of the dataset.
+    """
+
+    perplexity_model: Optional[str] = field(
+        default="./es.arpa.bin", metadata={"help": "Path to KenLM model to use to get perplexity values."}
+    )
+    sampling_method: Optional[str] = field(
+        default=None, metadata={"help": "Sample using a 'step' or 'gaussian' perplexity function per document, or 'random'."}
+    )
+    sampling_factor: Optional[float]  = field(
+        default=None, metadata={"help": "Sampling factor. Integers for step function, decimals for gaussian."}
+    )
+    boundaries: Optional[str] = field(
+        default="536394.99320948,662247.50212365,919250.87225178", metadata={"help": "Quartile boundaries"}
+    )
+
+    def __post_init__(self):
+        self.boundaries = [float(q.strip()) for q in self.boundaries.split(",")]
+
+
+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 advance_iter_and_group_samples(train_iterator, num_samples, max_seq_length):
+    """
+    The training iterator is advanced so that after groupifying the samples,
+    `num_samples` of length `max_seq_length` are returned.
+    """
+    num_total_tokens = max_seq_length * num_samples
+    samples = defaultdict(list)
+
+    i = 0
+    while i < num_total_tokens:
+        tokenized_samples = next(train_iterator)
+        i += len(tokenized_samples["input_ids"])
+
+        # concatenate tokenized samples to list
+        samples = {k: samples[k] + tokenized_samples[k] for k in tokenized_samples.keys()}
+
+    # Concatenated tokens are split to lists of length `max_seq_length`.
+    # Note that remainedr of % max_seq_length are thrown away.
+    def group_texts(examples):
+        result = {
+            k: [t[i : i + max_seq_length] for i in range(0, num_total_tokens, max_seq_length)]
+            for k, t in examples.items()
+        }
+        return result
+
+    grouped_samples = group_texts(samples)
+    return grouped_samples
+
+
+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 save_checkpoint_files(state, data_collator, training_args, save_dir):
+    unreplicated_state = jax_utils.unreplicate(state)
+    with open(os.path.join(save_dir, "optimizer_state.msgpack"), "wb") as f:
+        f.write(to_bytes(unreplicated_state.opt_state))
+    joblib.dump(training_args, os.path.join(save_dir, "training_args.joblib"))
+    joblib.dump(data_collator, os.path.join(save_dir, "data_collator.joblib"))
+    with open(os.path.join(save_dir, "training_state.json"), "w") as f:
+        json.dump({"step": unreplicated_state.step.item()}, f)
+
+
+def restore_checkpoint(save_dir, state):
+    logger.info(f"Restoring checkpoint from {save_dir}")
+    with open(os.path.join(save_dir, "flax_model.msgpack"), "rb") as f:
+        params = from_bytes(state.params, f.read())
+
+    with open(os.path.join(save_dir, "optimizer_state.msgpack"), "rb") as f:
+        opt_state = from_bytes(state.opt_state, f.read())
+
+    args = joblib.load(os.path.join(save_dir, "training_args.joblib"))
+    data_collator = joblib.load(os.path.join(save_dir, "data_collator.joblib"))
+
+    with open(os.path.join(save_dir, "training_state.json"), "r") as f:
+        training_state = json.load(f)
+    step = training_state["step"]
+
+    return params, opt_state, step, args, data_collator
+
+
+def rotate_checkpoints(path, max_checkpoints=5):
+    paths = sorted(Path(path).iterdir(), key=os.path.getmtime)[::-1]
+    if len(paths) > max_checkpoints:
+        for path_to_delete in paths[max_checkpoints:]:
+            try:
+                shutil.rmtree(path_to_delete)
+            except OSError:
+                os.remove(path_to_delete)
+
+
+def to_f32(t):
+    return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+
+def convert(output_dir, destination_dir="./"):
+    shutil.copyfile(Path(output_dir) / "flax_model.msgpack", Path(destination_dir) / "flax_model.msgpack")
+    shutil.copyfile(Path(output_dir) / "config.json", Path(destination_dir) / "config.json")
+    # Saving extra files from config.json and tokenizer.json files
+    tokenizer = AutoTokenizer.from_pretrained(destination_dir)
+    tokenizer.save_pretrained(destination_dir)
+
+    # Temporary saving bfloat16 Flax model into float32
+    tmp = tempfile.mkdtemp()
+    flax_model = FlaxRobertaForMaskedLM.from_pretrained(destination_dir)
+    flax_model.params = to_f32(flax_model.params)
+    flax_model.save_pretrained(tmp)
+    # Converting float32 Flax to PyTorch
+    model = RobertaForMaskedLM.from_pretrained(tmp, from_flax=True)
+    model.save_pretrained(destination_dir, save_config=False)
+
+
+if __name__ == "__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, SamplingArguments))
+    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, sampling_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args, sampling_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="INFO",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 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).
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        filepaths = {}
+        if data_args.train_file:
+            filepaths["train"] = data_args.train_file
+        if data_args.validation_file:
+            filepaths["validation"] = data_args.validation_file
+        try:
+            dataset = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                cache_dir=model_args.cache_dir,
+                streaming=True,
+                split="train",
+                sampling_method=sampling_args.sampling_method,
+                sampling_factor=sampling_args.sampling_factor,
+                boundaries=sampling_args.boundaries,
+                perplexity_model=sampling_args.perplexity_model,
+                seed=training_args.seed,
+                data_files=filepaths,
+            )
+        except Exception as exc:
+            logger.warning(
+                f"Unable to load local dataset with perplexity sampling support. Using huggingface.co/datasets/{data_args.dataset_name}: {exc}"
+            )
+            dataset = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                cache_dir=model_args.cache_dir,
+                streaming=True,
+                split="train",
+            )
+
+    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."
+        )
+
+    # 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[data_args.text_column_name],
+            return_special_tokens_mask=True
+        )
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+    )
+
+    shuffle_seed = training_args.seed
+    tokenized_datasets = tokenized_datasets.shuffle(buffer_size=data_args.shuffle_buffer_size, seed=shuffle_seed)
+
+    # 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='wandb',
+                project='hf-flax-bertin-roberta-es',
+                sync_tensorboard=True,
+            )
+            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()
+
+    # define number steps per stream epoch
+    num_train_steps = data_args.num_train_steps
+
+    # 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
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+        mask=decay_mask_fn,
+    )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
+    saved_step = -1
+    if model_args.model_name_or_path and "checkpoint" in model_args.model_name_or_path:
+        params, opt_state, saved_step, args, data_collator = restore_checkpoint(model_args.model_name_or_path, state)
+        # Create learning rate schedule
+        warmup_fn = optax.linear_schedule(
+            init_value=0.0, end_value=args.learning_rate, transition_steps=args.warmup_steps
+        )
+        decay_fn = optax.linear_schedule(
+            init_value=args.learning_rate,
+            end_value=0,
+            transition_steps=data_args.num_train_steps - args.warmup_steps,
+        )
+        linear_decay_lr_schedule_fn = optax.join_schedules(
+            schedules=[warmup_fn, decay_fn], boundaries=[args.warmup_steps]
+        )
+        # create adam optimizer
+        adamw = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=args.weight_decay,
+            mask=decay_mask_fn,
+        )
+        state = train_state.TrainState(
+            step=saved_step,
+            apply_fn=model.__call__,
+            params=params,
+            tx=adamw,
+            opt_state=opt_state,
+        )
+        # self.args = args
+        # data_collator = data_collator
+        # scheduler_fn = args.learning_rate
+        model.params = params
+
+
+    # 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
+    train_start = time.time()
+    train_metrics = []
+    eval_metrics = []
+
+    training_iter = iter(tokenized_datasets)
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+    eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
+
+    last_desc = ""
+    steps = tqdm(range(num_train_steps), desc="Training...", position=0)
+    for step in range(num_train_steps):
+        if step < saved_step:
+            steps.update(1)
+            continue
+        # ======================== Training ================================
+        try:
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
+        except StopIteration:
+            # Once the end of the dataset stream is reached, the training iterator
+            # is reinitialized and reshuffled and a new eval dataset is randomely chosen.
+            shuffle_seed += 1
+            tokenized_datasets.set_epoch(shuffle_seed)
+
+            training_iter = iter(tokenized_datasets)
+
+            eval_dataset = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
+
+        # process input samples
+        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)
+
+        if step % training_args.logging_steps == 0 and step > 0:
+            steps.write(
+                f"Step... ({step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+            )
+            train_time += time.time() - train_start
+            if has_tensorboard and jax.process_index() == 0:
+                write_train_metric(summary_writer, train_metrics, train_time, step)
+            train_metrics = []
+
+        # ======================== Evaluating ==============================
+        if step % training_args.eval_steps == 0 and step > 0:
+            eval_samples_idx = jnp.arange(data_args.num_eval_samples)
+            eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+            for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=1)):
+                # process input samples
+                batch_eval_samples = {k: [v[idx] for idx in batch_idx] for k, v in eval_samples.items()}
+                model_inputs = data_collator(batch_eval_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
+            steps.desc = f"Step... ({step}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+            last_desc = steps.desc
+
+            if has_tensorboard and jax.process_index() == 0:
+                write_eval_metric(summary_writer, eval_metrics, step)
+            eval_metrics = []
+
+        # save checkpoint after eval_steps
+        if step % training_args.save_steps == 0 and step > 0 and jax.process_index() == 0:
+            logger.info(f"Saving checkpoint at {step} steps")
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=False,
+            )
+            save_checkpoint_files(state, data_collator, training_args, training_args.output_dir)
+            checkpoints_dir = Path(training_args.output_dir) / "checkpoints" / f"checkpoint-{step}"
+            checkpoints_dir.mkdir(parents=True, exist_ok=True)
+            model.save_pretrained(checkpoints_dir, params=params)
+            save_checkpoint_files(state, data_collator, training_args, checkpoints_dir)
+            rotate_checkpoints(
+                Path(training_args.output_dir) / "checkpoints",
+                max_checkpoints=training_args.save_total_limit
+            )
+            convert(training_args.output_dir, "./")
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=training_args.push_to_hub,
+                commit_message=last_desc,
+            )
+
+        # update tqdm bar
+        steps.update(1)
+
+    if jax.process_index() == 0:
+        logger.info(f"Saving checkpoint at {step} steps")
+        params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+        model.save_pretrained(
+            training_args.output_dir,
+            params=params,
+            push_to_hub=False,
+        )
+        save_checkpoint_files(state, data_collator, training_args, training_args.output_dir)
+        checkpoints_dir = Path(training_args.output_dir) / "checkpoints" / f"checkpoint-{step}"
+        checkpoints_dir.mkdir(parents=True, exist_ok=True)
+        model.save_pretrained(checkpoints_dir, params=params)
+        save_checkpoint_files(state, data_collator, training_args, checkpoints_dir)
+        convert(training_args.output_dir, "./")
+        model.save_pretrained(
+            training_args.output_dir,
+            params=params,
+            push_to_hub=training_args.push_to_hub,
+            commit_message=last_desc or "Saving model after training",
+        )

wandb/run-20210726_001233-17u6inbn/files/config.yaml

@@ -0,0 +1,324 @@
+wandb_version: 1
+
+__cached__setup_devices:
+  desc: null
+  value: cpu
+_n_gpu:
+  desc: null
+  value: 0
+_wandb:
+  desc: null
+  value:
+    cli_version: 0.10.33
+    code_path: code/run_mlm_flax_stream.py
+    framework: huggingface
+    huggingface_version: 4.9.0.dev0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    python_version: 3.8.10
+    t:
+      1:
+      - 1
+      - 3
+      - 11
+      4: 3.8.10
+      5: 0.10.33
+      6: 4.9.0.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: ./configs/base
+dataloader_drop_last:
+  desc: null
+  value: false
+dataloader_num_workers:
+  desc: null
+  value: 0
+dataloader_pin_memory:
+  desc: null
+  value: true
+dataset_config_name:
+  desc: null
+  value: stepwise
+dataset_name:
+  desc: null
+  value: bertin-project/mc4-es-sampled
+ddp_find_unused_parameters:
+  desc: null
+  value: null
+debug:
+  desc: null
+  value: []
+deepspeed:
+  desc: null
+  value: null
+disable_tqdm:
+  desc: null
+  value: false
+do_eval:
+  desc: null
+  value: false
+do_predict:
+  desc: null
+  value: false
+do_train:
+  desc: null
+  value: false
+dtype:
+  desc: null
+  value: bfloat16
+eval_accumulation_steps:
+  desc: null
+  value: null
+eval_steps:
+  desc: null
+  value: 1000
+evaluation_strategy:
+  desc: null
+  value: IntervalStrategy.NO
+fp16:
+  desc: null
+  value: false
+fp16_backend:
+  desc: null
+  value: auto
+fp16_full_eval:
+  desc: null
+  value: false
+fp16_opt_level:
+  desc: null
+  value: O1
+gradient_accumulation_steps:
+  desc: null
+  value: 1
+greater_is_better:
+  desc: null
+  value: null
+group_by_length:
+  desc: null
+  value: false
+ignore_data_skip:
+  desc: null
+  value: false
+label_names:
+  desc: null
+  value: null
+label_smoothing_factor:
+  desc: null
+  value: 0.0
+learning_rate:
+  desc: null
+  value: 0.0006
+length_column_name:
+  desc: null
+  value: length
+line_by_line:
+  desc: null
+  value: false
+load_best_model_at_end:
+  desc: null
+  value: false
+local_rank:
+  desc: null
+  value: -1
+log_level:
+  desc: null
+  value: -1
+log_level_replica:
+  desc: null
+  value: -1
+log_on_each_node:
+  desc: null
+  value: true
+logging_dir:
+  desc: null
+  value: ./outputs/runs/Jul26_00-12-25_tablespoon
+logging_first_step:
+  desc: null
+  value: false
+logging_steps:
+  desc: null
+  value: 500
+logging_strategy:
+  desc: null
+  value: IntervalStrategy.STEPS
+lr_scheduler_type:
+  desc: null
+  value: SchedulerType.LINEAR
+max_grad_norm:
+  desc: null
+  value: 1.0
+max_seq_length:
+  desc: null
+  value: 512
+max_steps:
+  desc: null
+  value: -1
+metric_for_best_model:
+  desc: null
+  value: null
+mlm_probability:
+  desc: null
+  value: 0.15
+model_name_or_path:
+  desc: null
+  value: bertin-project/bertin-base-stepwise
+model_type:
+  desc: null
+  value: roberta
+mp_parameters:
+  desc: null
+  value: ''
+no_cuda:
+  desc: null
+  value: false
+num_eval_samples:
+  desc: null
+  value: 50000
+num_train_epochs:
+  desc: null
+  value: 3.0
+num_train_steps:
+  desc: null
+  value: 50000
+output_dir:
+  desc: null
+  value: ./outputs
+overwrite_cache:
+  desc: null
+  value: false
+overwrite_output_dir:
+  desc: null
+  value: true
+pad_to_max_length:
+  desc: null
+  value: true
+past_index:
+  desc: null
+  value: -1
+per_device_eval_batch_size:
+  desc: null
+  value: 48
+per_device_train_batch_size:
+  desc: null
+  value: 48
+per_gpu_eval_batch_size:
+  desc: null
+  value: null
+per_gpu_train_batch_size:
+  desc: null
+  value: null
+prediction_loss_only:
+  desc: null
+  value: false
+preprocessing_num_workers:
+  desc: null
+  value: null
+push_to_hub:
+  desc: null
+  value: false
+push_to_hub_model_id:
+  desc: null
+  value: outputs
+push_to_hub_organization:
+  desc: null
+  value: null
+push_to_hub_token:
+  desc: null
+  value: null
+remove_unused_columns:
+  desc: null
+  value: true
+report_to:
+  desc: null
+  value:
+  - tensorboard
+  - wandb
+resume_from_checkpoint:
+  desc: null
+  value: null
+run_name:
+  desc: null
+  value: ./outputs
+save_on_each_node:
+  desc: null
+  value: false
+save_steps:
+  desc: null
+  value: 1000
+save_strategy:
+  desc: null
+  value: IntervalStrategy.STEPS
+save_total_limit:
+  desc: null
+  value: 5
+seed:
+  desc: null
+  value: 42
+sharded_ddp:
+  desc: null
+  value: []
+shuffle_buffer_size:
+  desc: null
+  value: 10000
+skip_memory_metrics:
+  desc: null
+  value: true
+text_column_name:
+  desc: null
+  value: text
+tokenizer_name:
+  desc: null
+  value: ./configs/base
+tpu_metrics_debug:
+  desc: null
+  value: false
+tpu_num_cores:
+  desc: null
+  value: null
+train_file:
+  desc: null
+  value: null
+train_ref_file:
+  desc: null
+  value: null
+use_fast_tokenizer:
+  desc: null
+  value: true
+use_legacy_prediction_loop:
+  desc: null
+  value: false
+validation_file:
+  desc: null
+  value: null
+validation_ref_file:
+  desc: null
+  value: null
+validation_split_percentage:
+  desc: null
+  value: 5
+warmup_ratio:
+  desc: null
+  value: 0.0
+warmup_steps:
+  desc: null
+  value: 500
+weight_decay:
+  desc: null
+  value: 0.01

wandb/run-20210726_001233-17u6inbn/files/events.out.tfevents.1627258355.tablespoon.3000110.3.v2

@@ -0,0 +1 @@
+/var/hf/experiment-base-exp-512seq-stepwise/outputs/events.out.tfevents.1627258355.tablespoon.3000110.3.v2

wandb/run-20210726_001233-17u6inbn/files/output.log

@@ -0,0 +1,823 @@
+2021-07-26 00:12:35.575266: 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
+2021-07-26 00:12:35.575304: W tensorflow/stream_executor/cuda/cuda_driver.cc:326] failed call to cuInit: UNKNOWN ERROR (303)
+[00:12:36] - INFO - filelock - Lock 139656499698272 acquired on /home/versae/.cache/huggingface/transformers/27b7e968d2908b27f8c1df265c2dc08aef61be0f25bdc735df4df552829968fd.04a8293889c44bb7f31a5ee6212b8aa0b690121444e9c7ce1616fbe2a461ebba.lock
+
+
+
+Downloading: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 250M/250M [00:06<00:00, 35.8MB/s]
+[00:12:43] - INFO - filelock - Lock 139656499698272 released on /home/versae/.cache/huggingface/transformers/27b7e968d2908b27f8c1df265c2dc08aef61be0f25bdc735df4df552829968fd.04a8293889c44bb7f31a5ee6212b8aa0b690121444e9c7ce1616fbe2a461ebba.lock
+/var/hf/venv/lib/python3.8/site-packages/jax/lib/xla_bridge.py:386: UserWarning: jax.host_count has been renamed to jax.process_count. This alias will eventually be removed; please update your code.
+  warnings.warn(
+/var/hf/venv/lib/python3.8/site-packages/jax/lib/xla_bridge.py:373: UserWarning: jax.host_id has been renamed to jax.process_index. This alias will eventually be removed; please update your code.
+  warnings.warn(
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+Training...:   2%|█▊                                                                                          | 1000/50000 [22:19<17:30:45,  1.29s/it]
+Step... (500 | Loss: 1.8920137882232666, Learning Rate: 0.0006000000284984708)
+Training...:   2%|█▊                                                                                          | 1000/50000 [22:21<17:30:45,  1.29s/it]
+
+
+
+
+
+
+
+
+
+
+
+
+
+[02:30:54] - INFO - __main__ - Saving checkpoint at 1000 steps██████████████████████████████████████████████████████| 130/130 [00:31<00:00,  4.59it/s]
+/var/hf/transformers-orig/src/transformers/modeling_flax_pytorch_utils.py:201: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at  /pytorch/torch/csrc/utils/tensor_numpy.cpp:180.)
+  pt_model_dict[flax_key] = torch.from_numpy(flax_tensor)
+All Flax model weights were used when initializing RobertaForMaskedLM.
+Some weights of RobertaForMaskedLM were not initialized from the Flax model and are newly initialized: ['lm_head.decoder.weight', 'roberta.embeddings.position_ids', 'lm_head.decoder.bias']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

wandb/run-20210726_001233-17u6inbn/files/requirements.txt

@@ -0,0 +1,108 @@
+absl-py==0.13.0
+aiohttp==3.7.4.post0
+astunparse==1.6.3
+async-timeout==3.0.1
+attrs==21.2.0
+backcall==0.2.0
+cachetools==4.2.2
+certifi==2021.5.30
+chardet==4.0.0
+chex==0.0.8
+click==8.0.1
+configparser==5.0.2
+cycler==0.10.0
+datasets==1.9.1.dev0
+decorator==5.0.9
+dill==0.3.4
+dm-tree==0.1.6
+docker-pycreds==0.4.0
+filelock==3.0.12
+flatbuffers==1.12
+flax==0.3.4
+fsspec==2021.6.1
+gast==0.4.0
+gitdb==4.0.7
+gitpython==3.1.18
+google-auth-oauthlib==0.4.4
+google-auth==1.32.1
+google-pasta==0.2.0
+grpcio==1.38.1
+h5py==3.1.0
+huggingface-hub==0.0.12
+idna==2.10
+ipython-genutils==0.2.0
+ipython==7.25.0
+jax==0.2.17
+jaxlib==0.1.68
+jedi==0.18.0
+joblib==1.0.1
+kenlm==0.0.0
+keras-nightly==2.5.0.dev2021032900
+keras-preprocessing==1.1.2
+kiwisolver==1.3.1
+libtpu-nightly==0.1.dev20210615
+markdown==3.3.4
+matplotlib-inline==0.1.2
+matplotlib==3.4.2
+msgpack==1.0.2
+multidict==5.1.0
+multiprocess==0.70.12.2
+numpy==1.21.0
+oauthlib==3.1.1
+opt-einsum==3.3.0
+optax==0.0.9
+packaging==21.0
+pandas==1.3.0
+parso==0.8.2
+pathtools==0.1.2
+pexpect==4.8.0
+pickleshare==0.7.5
+pillow==8.3.1
+pip==20.0.2
+pkg-resources==0.0.0
+promise==2.3
+prompt-toolkit==3.0.19
+protobuf==3.17.3
+psutil==5.8.0
+ptyprocess==0.7.0
+pyarrow==4.0.1
+pyasn1-modules==0.2.8
+pyasn1==0.4.8
+pygments==2.9.0
+pyparsing==2.4.7
+python-dateutil==2.8.1
+pytz==2021.1
+pyyaml==5.4.1
+regex==2021.7.6
+requests-oauthlib==1.3.0
+requests==2.25.1
+rsa==4.7.2
+sacremoses==0.0.45
+scipy==1.7.0
+sentry-sdk==1.3.0
+setuptools==44.0.0
+shortuuid==1.0.1
+six==1.15.0
+smmap==4.0.0
+subprocess32==3.5.4
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.0
+tensorboard==2.5.0
+tensorflow-estimator==2.5.0
+tensorflow==2.5.0
+termcolor==1.1.0
+tokenizers==0.10.3
+toolz==0.11.1
+torch==1.9.0
+tqdm==4.61.2
+traitlets==5.0.5
+transformers==4.9.0.dev0
+typing-extensions==3.7.4.3
+urllib3==1.26.6
+wandb==0.10.33
+wcwidth==0.2.5
+werkzeug==2.0.1
+wheel==0.36.2
+wrapt==1.12.1
+xxhash==2.0.2
+yarl==1.6.3

wandb/run-20210726_001233-17u6inbn/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": "2021-07-26T00:12:35.406409",
+    "startedAt": "2021-07-26T00:12:33.305928",
+    "docker": null,
+    "cpu_count": 96,
+    "cuda": null,
+    "args": [
+        "--model_name_or_path=bertin-project/bertin-base-stepwise",
+        "--output_dir=./outputs",
+        "--model_type=roberta",
+        "--config_name=./configs/base",
+        "--tokenizer_name=./configs/base",
+        "--dataset_name=bertin-project/mc4-es-sampled",
+        "--dataset_config_name=stepwise",
+        "--max_seq_length=512",
+        "--pad_to_max_length",
+        "--per_device_train_batch_size=48",
+        "--per_device_eval_batch_size=48",
+        "--adam_beta1=0.9",
+        "--adam_beta2=0.98",
+        "--adam_epsilon=1e-6",
+        "--learning_rate=6e-4",
+        "--weight_decay=0.01",
+        "--save_steps=1000",
+        "--save_total_limit=5",
+        "--warmup_steps=500",
+        "--overwrite_output_dir",
+        "--num_train_steps=50000",
+        "--eval_steps=1000",
+        "--dtype=bfloat16",
+        "--logging_steps=500"
+    ],
+    "state": "running",
+    "program": "./run_mlm_flax_stream.py",
+    "codePath": "run_mlm_flax_stream.py",
+    "git": {
+        "remote": "https://huggingface.co/bertin-project/bertin-base-stepwise-exp-512seqlen",
+        "commit": "529e26d977dcd80df13f8ff4dc528756c974c3b3"
+    },
+    "email": "versae@gmail.com",
+    "root": "/var/hf/experiment-base-exp-512seq-stepwise",
+    "host": "tablespoon",
+    "username": "versae",
+    "executable": "/var/hf/venv/bin/python"
+}

wandb/run-20210726_001233-17u6inbn/files/wandb-summary.json

@@ -0,0 +1 @@
+{"global_step": 1000, "_timestamp": 1627266616.074129, "train_time": 15844.4716796875, "train_learning_rate": 0.0005939393886364996, "_step": 1994, "train_loss": 1.8408620357513428}

wandb/run-20210726_001233-17u6inbn/logs/debug.log

@@ -0,0 +1,27 @@
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_setup.py:_flush():69] setting env: {}
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_setup.py:_flush():69] setting login settings: {}
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_init.py:_log_setup():337] Logging user logs to /var/hf/experiment-base-exp-512seq-stepwise/wandb/run-20210726_001233-17u6inbn/logs/debug.log
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_init.py:_log_setup():338] Logging internal logs to /var/hf/experiment-base-exp-512seq-stepwise/wandb/run-20210726_001233-17u6inbn/logs/debug-internal.log
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_init.py:init():370] calling init triggers
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_init.py:init():375] wandb.init called with sweep_config: {}
+config: {}
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [wandb_init.py:init():419] starting backend
+2021-07-26 00:12:33,307 INFO    MainThread:3000110 [backend.py:_multiprocessing_setup():70] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
+2021-07-26 00:12:33,351 INFO    MainThread:3000110 [backend.py:ensure_launched():135] starting backend process...
+2021-07-26 00:12:33,394 INFO    MainThread:3000110 [backend.py:ensure_launched():139] started backend process with pid: 3001431
+2021-07-26 00:12:33,396 INFO    MainThread:3000110 [wandb_init.py:init():424] backend started and connected
+2021-07-26 00:12:33,399 INFO    MainThread:3000110 [wandb_init.py:init():472] updated telemetry
+2021-07-26 00:12:33,400 INFO    MainThread:3000110 [wandb_init.py:init():491] communicating current version
+2021-07-26 00:12:34,050 INFO    MainThread:3000110 [wandb_init.py:init():496] got version response upgrade_message: "wandb version 0.11.0 is available!  To upgrade, please run:\n $ pip install wandb --upgrade"
+
+2021-07-26 00:12:34,050 INFO    MainThread:3000110 [wandb_init.py:init():504] communicating run to backend with 30 second timeout
+2021-07-26 00:12:34,261 INFO    MainThread:3000110 [wandb_init.py:init():529] starting run threads in backend
+2021-07-26 00:12:35,502 INFO    MainThread:3000110 [wandb_run.py:_console_start():1623] atexit reg
+2021-07-26 00:12:35,502 INFO    MainThread:3000110 [wandb_run.py:_redirect():1497] redirect: SettingsConsole.REDIRECT
+2021-07-26 00:12:35,503 INFO    MainThread:3000110 [wandb_run.py:_redirect():1502] Redirecting console.
+2021-07-26 00:12:35,505 INFO    MainThread:3000110 [wandb_run.py:_redirect():1558] Redirects installed.
+2021-07-26 00:12:35,505 INFO    MainThread:3000110 [wandb_init.py:init():554] run started, returning control to user process
+2021-07-26 00:12:35,506 INFO    MainThread:3000110 [wandb_run.py:_config_callback():872] config_cb None None {'output_dir': './outputs', 'overwrite_output_dir': True, 'do_train': False, 'do_eval': False, 'do_predict': False, 'evaluation_strategy': 'IntervalStrategy.NO', 'prediction_loss_only': False, 'per_device_train_batch_size': 48, 'per_device_eval_batch_size': 48, 'per_gpu_train_batch_size': None, 'per_gpu_eval_batch_size': None, 'gradient_accumulation_steps': 1, 'eval_accumulation_steps': None, 'learning_rate': 0.0006, 'weight_decay': 0.01, 'adam_beta1': 0.9, 'adam_beta2': 0.98, 'adam_epsilon': 1e-06, 'max_grad_norm': 1.0, 'num_train_epochs': 3.0, 'max_steps': -1, 'lr_scheduler_type': 'SchedulerType.LINEAR', 'warmup_ratio': 0.0, 'warmup_steps': 500, 'log_level': -1, 'log_level_replica': -1, 'log_on_each_node': True, 'logging_dir': './outputs/runs/Jul26_00-12-25_tablespoon', 'logging_strategy': 'IntervalStrategy.STEPS', 'logging_first_step': False, 'logging_steps': 500, 'save_strategy': 'IntervalStrategy.STEPS', 'save_steps': 1000, 'save_total_limit': 5, 'save_on_each_node': False, 'no_cuda': False, 'seed': 42, 'fp16': False, 'fp16_opt_level': 'O1', 'fp16_backend': 'auto', 'fp16_full_eval': False, 'local_rank': -1, 'tpu_num_cores': None, 'tpu_metrics_debug': False, 'debug': [], 'dataloader_drop_last': False, 'eval_steps': 1000, 'dataloader_num_workers': 0, 'past_index': -1, 'run_name': './outputs', 'disable_tqdm': False, 'remove_unused_columns': True, 'label_names': None, 'load_best_model_at_end': False, 'metric_for_best_model': None, 'greater_is_better': None, 'ignore_data_skip': False, 'sharded_ddp': [], 'deepspeed': None, 'label_smoothing_factor': 0.0, 'adafactor': False, 'group_by_length': False, 'length_column_name': 'length', 'report_to': ['tensorboard', 'wandb'], 'ddp_find_unused_parameters': None, 'dataloader_pin_memory': True, 'skip_memory_metrics': True, 'use_legacy_prediction_loop': False, 'push_to_hub': False, 'resume_from_checkpoint': None, 'push_to_hub_model_id': 'outputs', 'push_to_hub_organization': None, 'push_to_hub_token': None, 'mp_parameters': '', '_n_gpu': 0, '__cached__setup_devices': 'cpu'}
+2021-07-26 00:12:35,507 INFO    MainThread:3000110 [wandb_run.py:_config_callback():872] config_cb None None {'model_name_or_path': 'bertin-project/bertin-base-stepwise', 'model_type': 'roberta', 'config_name': './configs/base', 'tokenizer_name': './configs/base', 'cache_dir': None, 'use_fast_tokenizer': True, 'dtype': 'bfloat16'}
+2021-07-26 00:12:35,508 INFO    MainThread:3000110 [wandb_run.py:_config_callback():872] config_cb None None {'dataset_name': 'bertin-project/mc4-es-sampled', 'dataset_config_name': 'stepwise', 'train_file': None, 'validation_file': None, 'train_ref_file': None, 'validation_ref_file': None, 'overwrite_cache': False, 'validation_split_percentage': 5, 'max_seq_length': 512, 'preprocessing_num_workers': None, 'mlm_probability': 0.15, 'pad_to_max_length': True, 'line_by_line': False, 'text_column_name': 'text', 'shuffle_buffer_size': 10000, 'num_train_steps': 50000, 'num_eval_samples': 50000}
+2021-07-26 00:12:35,587 INFO    MainThread:3000110 [wandb_run.py:_tensorboard_callback():943] tensorboard callback: outputs, None