Updated code to work with streaming version

.gitignore

@@ -0,0 +1 @@
+__pycache__

.vscode/launch.json

@@ -0,0 +1,42 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Python: StreamingFile",
+            "type": "python",
+            "request": "launch",
+            "program": "${file}",
+            "args": [
+                "--output_dir","roberta_mc4_sentence_piece",
+                "--model_type","roberta",
+                "--config_name","roberta_mc4_sentence_piece",
+                "--tokenizer_name","roberta_mc4_sentence_piece",
+                "--dataset_name","mc4",
+                "--dataset_config_name","hi",
+                "--max_seq_length","256",
+                "--per_device_train_batch_size","128",
+                "--per_device_eval_batch_size","128",
+                "--learning_rate","3e-4",
+                "--warmup_steps","1000",
+                "--overwrite_output_dir",
+                "--adam_beta1","0.9",
+                "--adam_beta2","0.98",
+                "--num_train_steps","10000",
+                "--num_eval_samples","5000",
+                "--logging_steps","250",
+                "--eval_steps","1000"
+            ],
+            "console": "integratedTerminal"
+        },
+        {
+            "name": "Python: Current File",
+            "type": "python",
+            "request": "launch",
+            "program": "${file}",
+            "console": "integratedTerminal"
+        }
+    ]
+}

events.out.tfevents.1625416432.t1v-n-9df4ce0e-w-0.447041.3.v2

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:10ad97376c517e5e19a89e76161722dd57fe0da7a5aa8bb2b16eb3234749e607
-size 40

events.out.tfevents.1625418057.t1v-n-9df4ce0e-w-0.452509.3.v2

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e6c4826820ff2695d2564e7e0bd9b2d1b181883ba8fd1e89493d0379a4425454
-size 41580506

flax_model.msgpack

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:904705cb48bfd59a5d3aed1f30746844238567349459d7e3e09aca20dbfb1e35
-size 498796983

pytorch_model.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4887e2e3b00807be6e343d2aa0abd4d3fd8eb844d48d55ef39efa22b0e1ba0b1
-size 498877970

roberta_mc4_sentence_piece/config.json

@@ -0,0 +1,26 @@
+{
+  "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",
+  "torch_dtype": "float32",
+  "transformers_version": "4.6.1",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

run_mlm_flax.py

@@ -40,6 +40,7 @@ import jax.numpy as jnp
 import optax
 from flax import jax_utils, traverse_util
 from flax.training import train_state
+from flax.serialization import from_bytes, to_bytes
 from flax.training.common_utils import get_metrics, onehot, shard
 from transformers import (
     CONFIG_MAPPING,
@@ -56,22 +57,6 @@ from transformers import (
 )
 
 
-# Cache the result
-has_tensorboard = is_tensorboard_available()
-if has_tensorboard:
-    try:
-        from flax.metrics.tensorboard import SummaryWriter
-    except ImportError as ie:
-        has_tensorboard = False
-        print(f"Unable to display metrics through TensorBoard because some package are not installed: {ie}")
-
-else:
-    print(
-        "Unable to display metrics through TensorBoard because the package is not installed: "
-        "Please run pip install tensorboard to enable."
-    )
-
-
 MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
 MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
 
@@ -186,6 +171,7 @@ class DataTrainingArguments:
                 assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
 
 
+
 @flax.struct.dataclass
 class FlaxDataCollatorForLanguageModeling:
     """
@@ -269,7 +255,7 @@ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndar
     return batch_idx
 
 
-def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
+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)
@@ -278,6 +264,8 @@ def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
         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)
 
@@ -315,16 +303,11 @@ if __name__ == "__main__":
 
     # 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.
-    training_args.seed = 42
     set_seed(training_args.seed)
 
     # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
@@ -363,6 +346,19 @@ if __name__ == "__main__":
         if extension == "txt":
             extension = "text"
         datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+        if data_args.validation_file is None:
+            datasets["validation"] = load_dataset(
+                extension, data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension, data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+            print(datasets)
+
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
 
@@ -450,7 +446,8 @@ if __name__ == "__main__":
             total_length = len(concatenated_examples[list(examples.keys())[0]])
             # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
             # customize this part to your needs.
-            total_length = (total_length // max_seq_length) * max_seq_length
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
             # Split by chunks of max_len.
             result = {
                 k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
@@ -472,8 +469,22 @@ if __name__ == "__main__":
         )
 
     # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
     if has_tensorboard and jax.process_index() == 0:
-        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
 
     # Data collator
     # This one will take care of randomly masking the tokens.
@@ -483,7 +494,14 @@ if __name__ == "__main__":
     rng = jax.random.PRNGKey(training_args.seed)
     dropout_rngs = jax.random.split(rng, jax.local_device_count())
 
-    model = FlaxAutoModelForMaskedLM.from_config(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))
+    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)
@@ -518,17 +536,24 @@ if __name__ == "__main__":
         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=1e-8,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
 
     # Define gradient update step fn
     def train_step(state, batch, dropout_rng):
@@ -588,7 +613,6 @@ if __name__ == "__main__":
 
     train_time = 0
     epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
-    save_checkpoint=True
     for epoch in epochs:
         # ======================== Training ================================
         train_start = time.time()
@@ -603,7 +627,7 @@ if __name__ == "__main__":
         train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
 
         # Gather the indexes for creating the batch and do a training step
-        for i, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
+        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
             samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
             model_inputs = data_collator(samples, pad_to_multiple_of=16)
 
@@ -611,58 +635,63 @@ if __name__ == "__main__":
             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 save_checkpoint and (train_metric['loss'] < 1.).all():
-                params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
-                model.save_pretrained(
-                '/home/khandelia1000/checkpoints/',
-                params=params,
-                push_to_hub=False
-                )
-                save_checkpoint = False
 
-        train_time += time.time() - train_start
+            cur_step = epoch * (num_train_samples // train_batch_size) + step
 
-        epochs.write(
-            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
-        )
-
-        # ======================== Evaluating ==============================
-        num_eval_samples = len(tokenized_datasets["validation"])
-        eval_samples_idx = jnp.arange(num_eval_samples)
-        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = jax_utils.unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
 
-        eval_metrics = []
-        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
-            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
-            model_inputs = data_collator(samples, pad_to_multiple_of=16)
-
-            # Model forward
-            model_inputs = shard(model_inputs.data)
-            metrics = p_eval_step(state.params, model_inputs)
-            eval_metrics.append(metrics)
-
-        # normalize eval metrics
-        eval_metrics = get_metrics(eval_metrics)
-        eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
-        eval_normalizer = eval_metrics.pop("normalizer")
-        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
-
-        # Update progress bar
-        epochs.desc = (
-            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
-        )
+                epochs.write(
+                    f"Step... ({cur_step} | Train Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
 
-        # Save metrics
-        if has_tensorboard and jax.process_index() == 0:
-            cur_step = epoch * (len(tokenized_datasets["train"]) // train_batch_size)
-            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
-
-        # save checkpoint after each epoch and push checkpoint to the hub
-        if jax.process_index() == 0:
-            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
-            model.save_pretrained(
-                training_args.output_dir,
-                params=params,
-                push_to_hub=training_args.push_to_hub,
-                commit_message=f"Saving weights and logs of epoch {epoch+1}",
-            )
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Val Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    step_output_dir = f"checkpoint_{cur_step}"
+                    os.mkdir(step_output_dir)
+                    print(f"Saving weights, optimizer state and logs of step {cur_step} at {step_output_dir}")
+                    params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                    model.save_pretrained(
+                        step_output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )
+                    with open("opt_state.msgpack", "wb") as f:
+                        f.write(to_bytes(state.opt_state))

run_mlm_flax_old.py

@@ -0,0 +1,668 @@
+#!/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 os
+import sys
+import time
+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.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+
+
+# Cache the result
+has_tensorboard = is_tensorboard_available()
+if has_tensorboard:
+    try:
+        from flax.metrics.tensorboard import SummaryWriter
+    except ImportError as ie:
+        has_tensorboard = False
+        print(f"Unable to display metrics through TensorBoard because some package are not installed: {ie}")
+
+else:
+    print(
+        "Unable to display metrics through TensorBoard because the package is not installed: "
+        "Please run pip install tensorboard to enable."
+    )
+
+
+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."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[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
+
+
+def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    num_samples = len(samples_idx)
+    samples_to_remove = num_samples % batch_size
+
+    if samples_to_remove != 0:
+        samples_idx = samples_idx[:-samples_to_remove]
+    sections_split = num_samples // batch_size
+    batch_idx = np.split(samples_idx, sections_split)
+    return batch_idx
+
+
+def write_metric(summary_writer, train_metrics, eval_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)
+
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+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))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level="NOTSET",
+        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.
+    training_args.seed = 42
+    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).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples,
+                return_special_tokens_mask=True,
+                padding=padding,
+                truncation=True,
+                max_length=max_seq_length,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            input_columns=[text_column_name],
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Enable tensorboard only on the master node
+    if has_tensorboard and jax.process_index() == 0:
+        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+
+    # 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())
+
+    model = FlaxAutoModelForMaskedLM.from_config(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=1e-8,
+        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)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    save_checkpoint=True
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
+        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
+
+        # Gather the indexes for creating the batch and do a training step
+        for i, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
+            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            train_metrics.append(train_metric)
+            if save_checkpoint and (train_metric['loss'] < 1.).all():
+                params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                model.save_pretrained(
+                '/home/khandelia1000/checkpoints/',
+                params=params,
+                push_to_hub=False
+                )
+                save_checkpoint = False
+
+        train_time += time.time() - train_start
+
+        epochs.write(
+            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+        )
+
+        # ======================== Evaluating ==============================
+        num_eval_samples = len(tokenized_datasets["validation"])
+        eval_samples_idx = jnp.arange(num_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            metrics = p_eval_step(state.params, model_inputs)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+        eval_normalizer = eval_metrics.pop("normalizer")
+        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+        # Update progress bar
+        epochs.desc = (
+            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+        )
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(tokenized_datasets["train"]) // train_batch_size)
+            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=training_args.push_to_hub,
+                commit_message=f"Saving weights and logs of epoch {epoch+1}",
+            )

run_mlm_flax_stream.py

@@ -0,0 +1,640 @@
+#!/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 os
+import sys
+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.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+from utils import keep_devnagri
+
+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 optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+
+
+# 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", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]]) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, 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
+
+
+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
+    doc_count = 0
+    while i < num_total_tokens:
+        tokenized_samples = next(train_iterator)
+        i += len(tokenized_samples["input_ids"])
+        doc_count += 1
+
+        # 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 doc_count, 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)
+
+
+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))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        level="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.
+        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)
+
+    cleaned_dataset = dataset.map(
+        keep_devnagri,
+        batched=False,
+    )
+    tokenized_datasets = cleaned_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)
+
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+
+        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+
+    # 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)
+        )
+    if jax.device_count() < 8:
+        print('Number of device as per jax device count is {}. Press Enter to continue'.format(jax.device_count()))
+
+    # 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
+    dataset_doc_count = 18507273
+    num_train_steps = ((dataset_doc_count//train_batch_size) + 1) * num_epochs * 2
+
+    # 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)
+
+    # 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)
+    doc_count, eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
+
+    steps = tqdm(range(num_train_steps), desc="Training...", position=0)
+    docs_progress_bar = tqdm(range(dataset_doc_count * num_epochs), desc="Docs Processed...", position=0)
+    for step in range(num_train_steps):
+        # ======================== Training ================================
+        try:
+            doc_count, 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)
+            doc_count, samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
+            
+
+        # process input samples
+        model_inputs = data_collator(samples)
+
+        # 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)
+
+                # 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 + 1}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+            if has_tensorboard and jax.process_index() == 0:
+                write_eval_metric(summary_writer, eval_metrics, step)
+            eval_metrics = []
+
+            # save checkpoint after each epoch and push checkpoint to the hub
+            if jax.process_index() == 0:
+                params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                model.save_pretrained(
+                    training_args.output_dir,
+                    params=params,
+                    push_to_hub=training_args.push_to_hub,
+                    commit_message=f"Saving weights and logs of step {step+1}",
+                )
+
+        # update tqdm bar
+        docs_progress_bar.update(doc_count)
+        steps.update(1)

run_stream.sh

@@ -0,0 +1,20 @@
+python3 -c "import jax; print(jax.devices())"
+./run_mlm_flax_stream.py \
+    --output_dir="${MODEL_DIR}" \
+    --model_type="roberta" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="mc4" \
+    --dataset_config_name="hi" \
+    --max_seq_length="256" \
+    --per_device_train_batch_size="128" \
+    --per_device_eval_batch_size="128" \
+    --learning_rate="3e-4" \
+    --warmup_steps="1000" \
+    --overwrite_output_dir \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --num_train_steps="10000" \
+    --num_eval_samples="5000" \
+    --logging_steps="250" \
+    --eval_steps="1000"

train_tokenizer.py

@@ -1,19 +1,31 @@
 #!/usr/bin/env python3
 from datasets import load_dataset
 from datasets import load_from_disk
-from tokenizers import ByteLevelBPETokenizer
+from tokenizers import ByteLevelBPETokenizer, SentencePieceBPETokenizer
 from tqdm import tqdm
-# load dataset
-# dataset = load_dataset("oscar", "unshuffled_deduplicated_hi", split="train")
 
-dataset = load_from_disk("/home/rtx/work/dk/hf/vo")
+from utils import keep_devnagri
+
+# load dataset
+dataset = load_dataset("mc4", "hi", split="train", streaming=True)
 
 # Instantiate tokenizer
-tokenizer = ByteLevelBPETokenizer(add_prefix_space=True)
+tokenizer = SentencePieceBPETokenizer(add_prefix_space=True)
 
 def batch_iterator(batch_size=100_000):
-    for i in range(0, len(dataset), batch_size):
-        yield dataset[i: i + batch_size]["text"]
+    # total docs: 1,85,07,273
+    text_ls = []
+    
+    for example in dataset:
+        devnagari_text, is_just_punctuation = keep_devnagri(example['text'])
+        if not is_just_punctuation:
+            text_ls.append(devnagari_text)
+        if len(text_ls) == batch_size:
+            yield text_ls
+            text_ls = []
+    if len(text_ls) > 0:
+        yield text_ls
+    
 
 # Customized training
 tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=50, special_tokens=[
@@ -22,8 +34,8 @@ tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=
     "</s>",
     "<unk>",
     "<mask>",
-])
+], )
 
 # Save files to disk
-tokenizer.save("./tokenizer.json")
+tokenizer.save("/home/khandelia1000/tokenizer.json")
 

utils.py

@@ -0,0 +1,34 @@
+import regex as re
+import string
+
+def keep_devnagri(document:str):
+    """
+    Remove all non Devnagri characters from the text.
+    Code adapted from https://huggingface.co/flax-community/roberta-base-mr/blob/64d2c745f264f09c3d5b678a718746b2613887db/mr_clean_text.py
+    
+    @param text: str Text to be cleaned
+    @return: Union[str, bool]
+    """
+    text = document['text']
+    pattern = r'[\p{Devanagari}0-9।\s\.\!]+'
+    
+    # regex pattern for all puntuation symbols
+    punctuation_regex = re.compile("[" + re.escape(string.punctuation) + string.digits + "|" + "]")
+
+    # keep only the text which is in devnagari script
+    cleaned = "".join([tok.group() for tok in re.finditer(pattern, text)])
+
+    # remove any extra space between words
+    cleaned = re.sub(r"[ ]+", " ", cleaned)
+    
+    # identify if the clean text only consists of punctuation
+    is_just_punctuation = len(re.sub(punctuation_regex, "", cleaned)) == 0
+    
+    # to handle the tokenizer as empty string may cause issues
+    # also this only happens for 5 out of 10000 docs, should not
+    # affect the results
+    if is_just_punctuation:
+        document['text'] = " "
+    else:
+        document['text'] = cleaned
+    return document