初始化项目，由ModelHub XC社区提供模型

Model: flax-community/gpt2-medium-persian
Source: Original Platform

.gitattributes

@@ -0,0 +1,17 @@
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tar.gz filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,43 @@
+---
+language: fa
+tags:
+- text-generation
+widget:
+ - text: "در یک اتفاق شگفت انگیز، پژوهشگران"
+ - text: "گرفتگی بینی در کودکان و به‌خصوص نوزادان باعث می‌شود"
+ - text: "امیدواریم نوروز امسال سالی"
+---
+
+# GPT2 Medium 4 Persian
+> This is part of the
+[Flax/Jax Community Week](https://discuss.huggingface.co/t/pretrain-gpt2-from-scratch-in-persian/7560), organized by [HuggingFace](https://huggingface.co/) and TPU usage sponsored by Google.
+
+
+## Team Members
+- [Mehrdad Farahani](huggingface.co/m3hrdadfi)
+- [Saied Alimoradi](https://discuss.huggingface.co/u/saied)
+- [M. Reza Zerehpoosh](huggingface.co/ironcladgeek)
+- [Hooman Sedghamiz](https://discuss.huggingface.co/u/hooman650)
+- [Mazeyar Moeini Feizabadi](https://discuss.huggingface.co/u/mazy1998)
+
+## Dataset
+We used  [Oscar](https://huggingface.co/datasets/oscar) dataset, which is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus.
+
+## How To Use
+You can use this model directly with a pipeline for text generation.
+
+```python
+from transformers import pipeline, AutoTokenizer, GPT2LMHeadModel
+tokenizer = AutoTokenizer.from_pretrained('flax-community/gpt2-medium-persian')
+model = GPT2LMHeadModel.from_pretrained('flax-community/gpt2-medium-persian')
+generator = pipeline('text-generation', model, tokenizer=tokenizer, config={'max_length':100})
+generated_text = generator('در یک اتفاق شگفت انگیز، پژوهشگران')
+```
+For using Tensorflow import TFGPT2LMHeadModel instead of GPT2LMHeadModel.
+## Demo
+
+... SOON
+
+## Evaluation
+
+... SOON

config.json

@@ -0,0 +1,41 @@
+{
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 5,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 5,
+  "gradient_checkpointing": false,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 1024,
+  "n_head": 16,
+  "n_inner": null,
+  "n_layer": 24,
+  "n_positions": 1024,
+  "n_special": 0,
+  "predict_special_tokens": true,
+  "resid_pdrop": 0.1,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "min_length": 32,
+      "max_length": 256,
+      "top_k": 50,
+      "top_p": 0.95
+    }
+  },
+  "transformers_version": "4.9.0.dev0",
+  "use_cache": true,
+  "vocab_size": 50000
+}

events.out.tfevents.1625933109.t1v-n-a95a71e5-w-0.273188.3.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaab8c1f0762b6d480d6ddef6fbf05107db528987d67c8474792dd5795375a45
+size 8545979

events.out.tfevents.1626033670.t1v-n-a95a71e5-w-0.441100.3.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7663ff60a0b876debfd15a5101a510a59721b473d85bf2e605a223328e96e047
+size 48307053

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:022e3c438752023ce7d26b0bb58fbc27b1cf86b534aa8fbb6119713dba36fa6e
+size 1418249630

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cae1f3c7d3627c1b1ce2ebc6991542e88781e525b4ef12041c25900aea411d12
+size 1443523865

src/convert_flax_to_pytorch.py

@@ -0,0 +1,31 @@
+import torch
+import numpy as np
+
+import jax
+import jax.numpy as jnp
+
+from transformers import AutoTokenizer
+from transformers import FlaxGPT2LMHeadModel
+from transformers import GPT2LMHeadModel
+
+tokenizer = AutoTokenizer.from_pretrained("../")
+tokenizer.pad_token = tokenizer.eos_token
+
+model_fx = FlaxGPT2LMHeadModel.from_pretrained("../")
+
+# def to_f32(t):
+#     return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
+
+# model_fx.params = to_f32(model_fx.params)
+# model_fx.save_pretrained("./fx")
+
+model_pt = GPT2LMHeadModel.from_pretrained("../", from_flax=True)
+model_pt.save_pretrained("./pt")
+
+input_ids = np.asarray(2 * [128 * [0]], dtype=np.int32)
+input_ids_pt = torch.tensor(input_ids)
+
+logits_pt = model_pt(input_ids_pt).logits
+print(logits_pt)
+logits_fx = model_fx(input_ids).logits
+print(logits_fx)

src/convert_flax_to_tf.py

@@ -0,0 +1,24 @@
+import torch
+import numpy as np
+
+import jax
+import jax.numpy as jnp
+
+from transformers import AutoTokenizer
+from transformers import GPT2LMHeadModel
+from transformers import TFGPT2LMHeadModel
+
+tokenizer = AutoTokenizer.from_pretrained("../")
+tokenizer.pad_token = tokenizer.eos_token
+
+model_pt = GPT2LMHeadModel.from_pretrained("./pt")
+model_tf = TFGPT2LMHeadModel.from_pretrained("./pt", from_pt=True)
+model_tf.save_pretrained("./tf")
+
+input_ids = np.asarray(2 * [128 * [0]], dtype=np.int32)
+input_ids_pt = torch.tensor(input_ids)
+
+logits_pt = model_pt(input_ids_pt).logits
+print(logits_pt)
+logits_tf = model_tf(input_ids).logits
+print(logits_tf)

src/create_config.py

@@ -0,0 +1,74 @@
+import ast
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple
+
+from transformers import (
+    HfArgumentParser,
+    AutoConfig
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ConfigArguments:
+    """
+    Arguments to which config we are going to set up.
+    """
+    output_dir: str = field(
+        default=".",
+        metadata={"help": "The output directory where the config will be written."},
+    )
+    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."
+        },
+    )
+    params: Optional[str] = field(
+        default=None,
+        metadata={"help": "Custom configuration for the specific `name_or_path`"}
+    )
+
+    def __post_init__(self):
+        if self.params:
+            try:
+                self.params = ast.literal_eval(self.params)
+            except Exception as e:
+                print(f"Your custom parameters do not acceptable due to {e}")
+
+
+def main():
+    parser = HfArgumentParser([ConfigArguments])
+    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.
+        config_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))[0]
+    else:
+        config_args = parser.parse_args_into_dataclasses()[0]
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO)
+
+    logger.info(f"Setting up configuration {config_args.name_or_path} with extra params {config_args.params}")
+
+    if config_args.params and isinstance(config_args.params, dict):
+        config = AutoConfig.from_pretrained(config_args.name_or_path, **config_args.params)
+    else:
+        config = AutoConfig.from_pretrained(config_args.name_or_path)
+
+    logger.info(f"Your configuration saved here {config_args.output_dir}/config.json")
+    config.save_pretrained(config_args.output_dir)
+
+
+if __name__ == '__main__':
+    main()

src/create_dataset.py

@@ -0,0 +1,136 @@
+import ast
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+import pandas as pd
+from sklearn.model_selection import train_test_split
+from tqdm import tqdm
+from typing import Dict, List, Optional, Tuple
+from datasets import load_dataset
+from transformers import (
+    HfArgumentParser,
+)
+from data_utils import (
+    filter_by_lang_regex,
+    filter_by_num_tokens,
+    filter_by_num_sents,
+    filter_by_adv,
+    normalizer
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class DataArguments:
+    """
+    Arguments to which dataset we are going to set up.
+    """
+    output_dir: str = field(
+        default=".",
+        metadata={"help": "The output directory where the config will be written."},
+    )
+    dataset_name: str = field(
+        default=None,
+        metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+def main():
+    parser = HfArgumentParser([DataArguments])
+    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.
+        data_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))[0]
+    else:
+        data_args = parser.parse_args_into_dataclasses()[0]
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO)
+    logger.info(f"Preparing the dataset")
+    if data_args.dataset_name is not None:
+        dataset = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            cache_dir=data_args.cache_dir,
+            split="train"
+        )
+    else:
+        data_files = {"train": data_args.train_file}
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+
+        dataset = load_dataset(
+            extension,
+            data_files=data_files,
+            delimiter="\t",
+            cache_dir=data_args.cache_dir,
+        )
+    
+    logger.info(f"dataset: {dataset}")
+
+    def data_preparation(item_dict):
+        if "text" not in item_dict:
+            return None
+
+        text = item_dict["text"]
+
+        status = filter_by_lang_regex(text, ratio=0.75)
+        if not status:
+            return None
+
+        status = filter_by_num_tokens(text, gt=64)
+        if not status:
+            return None
+        
+        status = filter_by_num_sents(text, gt=2)
+        if not status:
+            return None
+        
+        status = filter_by_adv(text, ratio=50)
+        if not status:
+            return None
+
+        text = normalizer(text)
+        return {"text": text}
+
+    data_dict = []
+    for item in tqdm(dataset, position=0, total=len(dataset)):
+        item = data_preparation(item)
+
+        if item:
+            data_dict.append(item)
+
+    data_df = pd.DataFrame(data_dict)
+
+    logger.info(f"Preparation - [before] consists of {len(dataset)} records!")
+    logger.info(f"Preparation - [after]  consists of {len(data_df)} records!")
+
+    train, test = train_test_split(data_df, test_size=0.01, random_state=101)
+
+    train = train.reset_index(drop=True)
+    test = test.reset_index(drop=True)
+
+    logger.info(f"Preparation of [train] set consists of {len(train)} records!")
+    logger.info(f"Preparation of [test] set consists of {len(test)} records!")
+
+    os.makedirs(data_args.output_dir, exist_ok=True)
+    train.to_csv(os.path.join(data_args.output_dir, "train.csv"), sep="\t", encoding="utf-8", index=False)
+    test.to_csv(os.path.join(data_args.output_dir, "test.csv"), sep="\t", encoding="utf-8", index=False)
+    logger.info(f"Data saved here {data_args.output_dir}")
+
+if __name__ == '__main__':
+    main()

src/data_utils.py

@@ -0,0 +1,42 @@
+from hazm import word_tokenize
+from hazm import sent_tokenize
+import re
+import six
+
+from normalizer import normalize
+
+persian_regex = "0-9۰۱۲۳۴۵۶۷۸۹ءآئابتثجحخدذرزسشصضطظعغفقلمنهوپچژکگیە\u200c"
+
+
+def filter_by_lang_regex(text, ratio=0.7, regex="0-9۰۱۲۳۴۵۶۷۸۹ءآئابتثجحخدذرزسشصضطظعغفقلمنهوپچژکگیە\u200c"):
+    candidate_text = re.sub(r"[^" + regex + "]+", " ", six.ensure_str(text)).replace(" ", "")
+    text = text.replace(" ", "")
+
+    return (len(candidate_text) / len(text)) > ratio
+
+
+def filter_by_num_tokens(text, gt=64):
+    return len(word_tokenize(text)) > gt
+
+
+def filter_by_num_sents(text, gt=2):
+    return len(sent_tokenize(text)) > gt
+
+
+def filter_by_adv(text, ratio=50):
+    comma = text.split(",")
+    colon = re.findall(r"""(?:([^\W]+):([^\W]+))""", text)
+    virgool = text.split("،")
+    length_add = len(comma) + len(colon) + len(virgool)
+
+    return length_add < ratio
+
+
+def normalizer(text, do_lowercase=False):
+    text = normalize(text)
+
+    if do_lowercase:
+        text = text.lower()
+
+    return text
+

src/dictionary.py

@@ -0,0 +1,139 @@
+characters = {
+    "ك": "ک",
+    "دِ": "د",
+    "بِ": "ب",
+    "زِ": "ز",
+    "ذِ": "ذ",
+    "شِ": "ش",
+    "سِ": "س",
+    "ى": "ی",
+    "ي": "ی",
+    "ؤ": "و",
+    "ے": "ی",
+    "ۀ": "ه",
+    "ﭘ": "پ",
+    "ﮐ": "ک",
+    "ﯽ": "ی",
+    "ﺎ": "ا",
+    "ﺑ": "ب",
+    "ﺘ": "ت",
+    "ﺧ": "خ",
+    "ﺩ": "د",
+    "ﺱ": "س",
+    "ﻀ": "ض",
+    "ﻌ": "ع",
+    "ﻟ": "ل",
+    "ﻡ": "م",
+    "ﻢ": "م",
+    "ﻪ": "ه",
+    "ﻮ": "و",
+    "ﺍ": "ا",
+    "ة": "ه",
+    "ﯾ": "ی",
+    "ﯿ": "ی",
+    "ﺒ": "ب",
+    "ﺖ": "ت",
+    "ﺪ": "د",
+    "ﺮ": "ر",
+    "ﺴ": "س",
+    "ﺷ": "ش",
+    "ﺸ": "ش",
+    "ﻋ": "ع",
+    "ﻤ": "م",
+    "ﻥ": "ن",
+    "ﻧ": "ن",
+    "ﻭ": "و",
+    "ﺭ": "ر",
+    "ﮔ": "گ",
+    "إ": "ا",
+    "ٕ": " ",
+    "ھ": "ه",
+    "...": ".",
+    "…": ".",
+    "-": " - ",
+    "هٔ": "ه",
+    "ﻯ": "ی",
+    "ﻛ": "ک",
+    "ﭼ": "چ",
+    "ﺓ": "ه",
+    "ﻴ": "ی",
+    "ﻊ": "ع",
+    "ﮬ": "ه",
+    "ﺟ": "ج",
+    "ﺳ": "س",
+    "ﻦ": "ن",
+    "ﺬ": "ذ",
+    "ﺋ": "ئ",
+    "ﷲ": "لله",
+    "ﺞ": "ج",
+    "ﺙ": "ث",
+    "ﻗ": "ق",
+    "ﮪ": "ه",
+    "ﺰ": "ز",
+    "ﯼ": "ی",
+    "ٺ": "ت",
+    "ﺻ": "ص",
+    "ﻂ": "ط",
+    "ﻣ": "م",
+    "ﻈ": "ظ",
+    "ﺐ": "ب",
+    "ﻍ": "غ",
+    "ݸ": "و",
+    "ﻨ": "ن",
+    "ﻝ": "ل",
+    "ﻩ": "ه",
+    "ﻲ": "ی",
+    "ﻐ": "غ",
+    "ﺲ": "س",
+    "ﺁ": "آ",
+    "ڔ": "ر",
+    "ﺫ": "ذ",
+    "ﭻ": "چ",
+    "ﺠ": "ج",
+    "ﯙ": "و",
+    "ﮏ": "ک",
+    "ﺣ": "ح",
+    "ﺝ": "ج",
+    "ﺼ": "ص",
+    "ﻳ": "ی",
+    "ﻘ": "ق",
+    "ﺨ": "خ",
+    "ﻔ": "ف",
+    "ﻎ": "غ",
+    "ئ": "ی",
+    "ﻓ": "ف",
+    "ﻕ": "ق",
+    "ﮋ": "ژ",
+    "ﺗ": "ت",
+    "ﻁ": "ط",
+    "ﺯ": "ز",
+    "ﮕ": "گ",
+    "ﺌ": "ئ",
+    "ﺵ": "ش",
+    "ۮ": "د",
+    "ﻫ": "ه",
+    "ﻬ": "ه",
+    "ﻏ": "غ",
+    "ﻰ": "ی",
+    "﷼": "ریال",
+    "ﺿ": "ض",
+    "ﺛ": "ث",
+    "ݐ": "پ",
+    "ﺏ": "ب",
+    "ﭙ": "پ",
+    "ﭽ": "چ",
+    "ﺜ": "ث",
+    "ﻃ": "ط",
+    "ۂ": "ه",
+    "ﻑ": "ف",
+    "ﺕ": "ت",
+    "ﻞ": "ل",
+}
+
+special_tokens = {}
+
+words_map = {
+    "Leave a comment": "",
+    "[…]": "",
+    "[.]": "",
+}

src/normalizer.py

@@ -0,0 +1,138 @@
+import hazm
+import re
+import string
+
+from regexes.currency import CURRENCY_REGEX
+from regexes.email import EMAIL_REGEX
+from regexes.latin import LATIN_REGEX
+from regexes.latin import LATIN_REGEX, LATIN_WITH_SPECIAL_REGEX
+from regexes.number import NUMBERS_REGEX
+from regexes.phone import PHONE_REGEX
+from regexes.quote import DOUBLE_QUOTE_REGEX, SINGLE_QUOTE_REGEX
+from regexes.url import URL_REGEX
+from regexes.persian import PERSIAN_REGEX
+from regexes.punk import PUNK_REGEX
+import dictionary
+
+allowed_char = string.ascii_letters + string.digits + ':/@_-. '
+
+
+def make_trans(list_a, list_b):
+    return dict((ord(a), b) for a, b in zip(list_a, list_b))
+
+
+def multiple_replace(text, chars_to_mapping):
+    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
+    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))
+
+
+def remove_adv_by_tag_name(text, tag_name):
+    found = text.find(tag_name)
+
+    if found > 0:
+        text = text[:found]
+
+    return text
+
+
+def clean_url(text):
+    # removing html tags
+    text = re.sub('<.*?>', '', text)
+
+    # removing normal(without space urls)
+    text = re.sub(r'(?:(?:http|https):\/\/)?([-a-zA-Z0-9.]{2,256}\.[a-z]{2,4})\b(?:\/[-a-zA-Z0-9@:%_\+.~#?&//=]*)?', "",
+                  text)
+
+    # removing urls that contains space
+    result = ''
+    for char in text:
+        if char in allowed_char:
+            result += char
+    result = result.replace('  ', '')
+    result = result.split(':')
+    for phrase in result:
+        p = phrase
+        if '//' in p:
+            if ('https :' + p) in text:
+                text = text.replace('https :' + p, '')
+            elif ('http :' + p) in text:
+                text = text.replace('http :' + p, '')
+        elif '@' in p:
+            if p in text:
+                text = text.replace(p, '')
+
+    return text
+
+
+ar2fa_digits = make_trans("٠١٢٣٤٥٦٧٨٩٪", "۰۱۲۳۴۵۶۷۸۹٪")
+fa2en_digits = make_trans("۰۱۲۳۴۵۶۷۸۹٪", "0123456789%")
+normalizer = hazm.Normalizer(persian_numbers=True, punctuation_spacing=False)
+
+
+def normalize(text, zwnj="\u200c", tokenized=False):
+    text = text.replace("\n", " ").replace("\t", " ")
+    text = re.sub(r"\u200c+", "\u200c", text)
+    text = text.replace('ـ', '')
+    text = normalizer.normalize(text)
+
+    if len(dictionary.characters) > 0:
+        text = multiple_replace(text, dictionary.characters)
+
+    if len(dictionary.words_map) > 0:
+        text = multiple_replace(text, dictionary.words_map)
+
+    text = text.translate(ar2fa_digits)
+    text = text.translate(fa2en_digits)
+
+    text = SINGLE_QUOTE_REGEX.sub("'", text)
+    text = DOUBLE_QUOTE_REGEX.sub('"', text)
+    text = CURRENCY_REGEX.sub(r" \1 ", text)
+    text = clean_url(text)
+    text = remove_adv_by_tag_name(text, tag_name="برچسب ها :")
+    text = URL_REGEX.sub(" ", text)
+    text = EMAIL_REGEX.sub(" ", text)
+    text = PHONE_REGEX.sub(r" \1 ", text)
+    text = NUMBERS_REGEX.sub(r" \1 ", text)
+    text = LATIN_REGEX.sub(r" \1 ", text)
+    # text = PUNK_REGEX.sub(r" \1 ", text)  # must be remained the same!
+
+    # Allow only english and persian characters
+    text = re.sub(PERSIAN_REGEX, " ", text)
+
+    text = text.replace(f" {zwnj} ", f"{zwnj}")
+    text = text.replace(f"{zwnj} ", f"{zwnj}")
+    text = text.replace(f" {zwnj}", f"{zwnj}")
+
+    if len(dictionary.special_tokens) > 0:
+        text = multiple_replace(text, dictionary.special_tokens)
+
+    tokens = []
+    for token in text.split():
+        token = token.strip()
+        if token:
+            if token.startswith(zwnj) and token.endswith(zwnj):
+                token = token[1:-1]
+            if token.startswith(zwnj):
+                token = token[1:]
+            elif token.endswith(zwnj):
+                token = token[:-1]
+            else:
+                token = token
+
+            tokens.append(token)
+
+    if tokenized:
+        return tokens
+
+    return " ".join(tokens)
+
+
+
+if __name__ == '__main__':
+    import textwrap
+
+    # input_text = " «هفتاد سی» "
+    # input_text = normalize(input_text)
+    # input_text = DOUBLE_QUOTE_REGEX.sub('"', input_text)
+    # print(textwrap.fill(input_text))
+    # print(normalize(input_text, tokenized=True))

src/regexes/currency.py

@@ -0,0 +1,23 @@
+import re
+
+CURRENCIES = {
+    "$": "USD",
+    "zł": "PLN",
+    "£": "GBP",
+    "¥": "JPY",
+    "฿": "THB",
+    "₡": "CRC",
+    "₦": "NGN",
+    "₩": "KRW",
+    "₪": "ILS",
+    "₫": "VND",
+    "€": "EUR",
+    "₱": "PHP",
+    "₲": "PYG",
+    "₴": "UAH",
+    "₹": "INR",
+    "﷼": "IRR",
+}
+CURRENCY_REGEX = re.compile(
+    "({})+".format("|".join(re.escape(c) for c in CURRENCIES.keys()))
+)

src/regexes/email.py

@@ -0,0 +1,6 @@
+import re
+
+EMAIL_REGEX = re.compile(
+    r"(?:^|(?<=[^\w@.)]))([\w+-](\.(?!\.))?)*?[\w+-](@|[(<{\[]at[)>}\]])(?:(?:[a-z\\u00a1-\\uffff0-9]-?)*[a-z\\u00a1-\\uffff0-9]+)(?:\.(?:[a-z\\u00a1-\\uffff0-9]-?)*[a-z\\u00a1-\\uffff0-9]+)*(?:\.(?:[a-z\\u00a1-\\uffff]{2,}))",
+    flags=re.IGNORECASE | re.UNICODE,
+)

src/regexes/latin.py

@@ -0,0 +1,13 @@
+import re
+
+LATIN_WITH_SPECIAL_REGEX = re.compile(
+    r"(\b(?!URL|EMAIL|PHONE|NUMBER|CUR|LATIN\b)[0-9a-zA-Z]+)"
+)
+
+LATIN_REGEX = re.compile(
+    r"([0-9a-zA-Z]+)"
+)
+
+LATIN_SPACES_REGEX = re.compile(
+    r"([0-9a-zA-Z])"
+)

src/regexes/number.py

@@ -0,0 +1,5 @@
+import re
+
+NUMBERS_REGEX = re.compile(
+    r"(?:^|(?<=[^\w,.]))[+–-]?(([1-9]\d{0,2}(,\d{3})+(\.\d*)?)|([1-9]\d{0,2}([ .]\d{3})+(,\d*)?)|(\d*?[.,]\d+)|\d+)(?:$|(?=\b))"
+)

src/regexes/persian.py

@@ -0,0 +1,19 @@
+import re
+
+
+PERSIAN_ALPHA = "ءآئابتثجحخدذرزسشصضطظعغفقلمنهوپچژکگیە"  # noqa: E501
+PERSIAN_DIGIT = "۰۱۲۳۴۵۶۷۸۹"
+
+
+ZWNJ = "\u200c"
+PUNK = '\!\"\#\$\%\&\'\(\)\*\+\,\-\.\/\:\;\<\=\>\?\@\[\]\^\_\`\{\|\}\~\«\»\؟\:\×\٬\٫\﷼\٪\،'
+
+PERSIAN = (
+    "a-zA-Z0-9" +
+    PERSIAN_ALPHA +
+    PERSIAN_DIGIT +
+    ZWNJ +
+    PUNK
+)
+
+PERSIAN_REGEX = r"[^" + PERSIAN + "+]"

src/regexes/phone.py

@@ -0,0 +1,6 @@
+import re
+
+
+PHONE_REGEX = re.compile(
+    r"((?:^|(?<=[^\w)]))(((\+?[01])|(\+\d{2}))[ .-]?)?(\(?\d{3,4}\)?/?[ .-]?)?(\d{3}[ .-]?\d{4})(\s?(?:ext\.?|[#x-])\s?\d{2,6})?(?:$|(?=\W)))|\+?\d{4,5}[ .-/]\d{6,9}"
+)

src/regexes/punk.py

@@ -0,0 +1,5 @@
+import re
+
+PUNK_REGEX = re.compile(
+    r"([\!\"\#\$\%\&\'\(\)\*\+\,\-\.\/\:\;\=\?\@\[\\\]\^\_\`\{\|\}\~\«\»\⸮\؟\،\٬\٫\؛])"
+)

src/regexes/quote.py

@@ -0,0 +1,25 @@
+import re
+
+
+strange_double_quotes = [
+    "«",
+    "‹",
+    "»",
+    "›",
+    "„",
+    "“",
+    "‟",
+    "”",
+    "❝",
+    "❞",
+    "❮",
+    "❯",
+    "〝",
+    "〞",
+    "〟",
+    "＂",
+]
+strange_single_quotes = ["‘", "‛", "’", "❛", "❜", "`", "´", "‘", "’"]
+
+DOUBLE_QUOTE_REGEX = re.compile("|".join(strange_double_quotes))
+SINGLE_QUOTE_REGEX = re.compile("|".join(strange_single_quotes))

src/regexes/url.py

@@ -0,0 +1,38 @@
+import re
+
+URL_REGEX = re.compile(
+    r"(?:^|(?<![\w\/\.]))"
+    # protocol identifier
+    # r"(?:(?:https?|ftp)://)"  <-- alt?
+    r"(?:(?:https?:\/\/|ftp:\/\/|www\d{0,3}\.))"
+    # user:pass authentication
+    r"(?:\S+(?::\S*)?@)?" r"(?:"
+    # IP address exclusion
+    # private & local networks
+    r"(?!(?:10|127)(?:\.\d{1,3}){3})"
+    r"(?!(?:169\.254|192\.168)(?:\.\d{1,3}){2})"
+    r"(?!172\.(?:1[6-9]|2\d|3[0-1])(?:\.\d{1,3}){2})"
+    # IP address dotted notation octets
+    # excludes loopback network 0.0.0.0
+    # excludes reserved space >= 224.0.0.0
+    # excludes network & broadcast addresses
+    # (first & last IP address of each class)
+    r"(?:[1-9]\d?|1\d\d|2[01]\d|22[0-3])"
+    r"(?:\.(?:1?\d{1,2}|2[0-4]\d|25[0-5])){2}"
+    r"(?:\.(?:[1-9]\d?|1\d\d|2[0-4]\d|25[0-4]))"
+    r"|"
+    # host name
+    r"(?:(?:[a-z\\u00a1-\\uffff0-9]-?)*[a-z\\u00a1-\\uffff0-9]+)"
+    # domain name
+    r"(?:\.(?:[a-z\\u00a1-\\uffff0-9]-?)*[a-z\\u00a1-\\uffff0-9]+)*"
+    # TLD identifier
+    r"(?:\.(?:[a-z\\u00a1-\\uffff]{2,}))" r"|" r"(?:(localhost))" r")"
+    # port number
+    r"(?::\d{2,5})?"
+    # resource path
+    r"(?:\/[^\)\]\}\s]*)?",
+    # r"(?:$|(?![\w?!+&\/\)]))",
+    # @jfilter: I removed the line above from the regex because I don't understand what it is used for, maybe it was useful?
+    # But I made sure that it does not include ), ] and } in the URL.
+    flags=re.UNICODE | re.IGNORECASE,
+)

src/requirements.txt

@@ -0,0 +1,7 @@
+datasets >= 1.1.3
+jax>=0.2.8
+jaxlib>=0.1.59
+flax>=0.3.4
+optax>=0.0.8
+hazm
+tensorboard

src/run.sh

@@ -0,0 +1,69 @@
+#!/bin/bash
+
+export LC_ALL=C.UTF-8
+export LANG=C.UTF-8
+
+export MODEL_NAME_OR_PATH=/home/m3hrdadfi/code/gpt2-medium-persian
+export OUTPUT_DIR=/home/m3hrdadfi/code/gpt2-medium-persian
+# export MODEL_TYPE=gpt2
+# export CONFIG_NAME=/home/m3hrdadfi/code/gpt2-medium-persian
+# export TOKENIZER_NAME=/home/m3hrdadfi/code/gpt2-medium-persian
+
+export TRAIN_FILE=/home/m3hrdadfi/data/train-fixed.csv
+export VALIDATION_FILE=/home/m3hrdadfi/data/test-fixed.csv
+export TEST_FILE=/home/m3hrdadfi/code/data/test-fixed.csv
+# export DATASET_NAME=oscar
+# export DATASET_CONFIG_NAME=unshuffled_deduplicated_fa
+export MAX_SEQUENCE_LENGTH=512
+
+#export MAX_TRAIN_SAMPLE=5000
+#export MAX_EVAL_SAMPLES=5000
+
+export PER_DEVICE_TRAIN_BATCH_SIZE=16
+export PER_DEVICE_EVAL_BATCH_SIZE=16
+export NUM_TRAIN_EPOCHS=9.0
+export LEARNING_RATE=8e-4
+export WARMUP_STEPS=5000
+export LOGGING_STEPS=500
+export EVAL_STEPS=2500
+export SAVE_STEPS=2500
+
+python src/run_clm_flax.py \
+    --output_dir="$OUTPUT_DIR"  \
+    --model_name_or_path="$MODEL_NAME_OR_PATH" \
+    --train_file="$TRAIN_FILE" \
+    --validation_file="$VALIDATION_FILE" \
+    --block_size=$MAX_SEQUENCE_LENGTH \
+    --per_device_train_batch_size=$PER_DEVICE_TRAIN_BATCH_SIZE \
+    --per_device_eval_batch_size=$PER_DEVICE_EVAL_BATCH_SIZE \
+    --num_train_epochs=$NUM_TRAIN_EPOCHS \
+    --learning_rate=$LEARNING_RATE \
+    --warmup_steps=$WARMUP_STEPS \
+    --logging_step=$LOGGING_STEPS \
+    --eval_steps=$EVAL_STEPS \
+    --save_steps=$SAVE_STEPS \
+    --do_train \
+    --do_eval \
+    --overwrite_output_dir \
+    --push_to_hub
+
+# python src/run_clm_flax.py \
+#     --output_dir="$OUTPUT_DIR"  \
+#     --model_type="$MODEL_TYPE" \
+#     --config_name="$CONFIG_NAME" \
+#     --tokenizer_name="$TOKENIZER_NAME" \
+#     --dataset_name="$DATASET_NAME" \
+#     --dataset_config_name="$DATASET_CONFIG_NAME" \
+#     --block_size=$MAX_SEQUENCE_LENGTH \
+#     --per_device_train_batch_size=$PER_DEVICE_TRAIN_BATCH_SIZE \
+#     --per_device_eval_batch_size=$PER_DEVICE_EVAL_BATCH_SIZE \
+#     --num_train_epochs=$NUM_TRAIN_EPOCHS \
+#     --learning_rate=$LEARNING_RATE \
+#     --warmup_steps=$WARMUP_STEPS \
+#     --logging_step=$LOGGING_STEPS \
+#     --eval_steps=$EVAL_STEPS \
+#     --save_steps=$SAVE_STEPS \
+#     --do_train \
+#     --do_eval \
+#     --overwrite_output_dir \
+#     --push_to_hub

src/run_clm_flax.py

@@ -0,0 +1,709 @@
+#!/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.
+"""
+Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) 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=causal-lm
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+import jax
+from jax import lax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import checkpoints
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForCausalLM,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.testing_utils import CaptureLogger
+
+from data_utils import (
+    filter_by_lang_regex,
+    filter_by_num_tokens,
+    filter_by_num_sents,
+    filter_by_adv,
+    normalizer
+)
+
+print(jax.devices())
+
+logger = logging.getLogger(__name__)
+
+# 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_CAUSAL_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)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+                    "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+                    "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=1,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+                    "The training dataset will be truncated in block of this size for training. "
+                    "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+    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."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+# 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)
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+        train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def 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."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    checkpoints_dir = os.path.join(training_args.output_dir, "checkpoints")
+    os.makedirs(checkpoints_dir, exist_ok=True)
+
+    #  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.
+        raw_dataset = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            cache_dir=model_args.cache_dir,
+            keep_in_memory=False
+        )
+
+        if "validation" not in raw_dataset.keys():
+            raw_dataset["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,
+            )
+            raw_dataset["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"
+
+        raw_dataset = load_dataset(
+            extension,
+            data_files=data_files,
+            delimiter="\t",
+            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.
+    # logger.info("Preprocessing the dataset")
+    # dataset = raw_dataset.filter(lambda example: filter_by_lang_regex(example["text"], ratio=0.75))
+    # dataset = dataset.filter(lambda example: filter_by_num_tokens(example["text"], gt=64))
+    # dataset = dataset.filter(lambda example: filter_by_num_sents(example["text"], gt=2))
+    # dataset = dataset.filter(lambda example: filter_by_adv(example["text"], ratio=50))
+    # dataset = dataset.map(normalizer)
+    # logger.info(f"Preprocessed dataset kept {len(dataset)} out of {len(raw_dataset)}")
+    dataset = raw_dataset
+    logger.info(f"dataset: {dataset}")
+
+    # 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."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            config=config,
+            seed=training_args.seed,
+            dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForCausalLM.from_config(
+            config,
+            seed=training_args.seed,
+            dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+    logger.info(f"text_column_name: {text_column_name}")
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    tokenized_datasets = dataset.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,
+    )
+
+    if data_args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > config.max_position_embeddings:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+            )
+            block_size = 1024
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warning(
+                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+            )
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    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.
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i: i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        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
+
+    lm_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # 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))
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # 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:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
+
+    def loss_fn(logits, labels):
+        shift_logits = logits[..., :-1, :]
+        shift_labels = labels[..., 1:]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (len(train_dataset) // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    logger.info(f"*** Writing training summary after {cur_step} steps ***")
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0 and training_args.do_eval:
+                logger.info(f"*** Evaluation after {cur_step} steps ***")
+
+                eval_metrics = []
+                eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+                eval_steps = len(eval_dataset) // eval_batch_size
+                for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+                    # Model forward
+                    batch = next(eval_loader)
+                    metrics = p_eval_step(state.params, batch)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                try:
+                    eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+                except OverflowError:
+                    eval_metrics["perplexity"] = float("inf")
+
+                # Print metrics and update progress bar
+                desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+                epochs.write(desc)
+                epochs.desc = desc
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    logger.info(f"*** Writing evaluation summary after {cur_step} steps ***")
+                    # cur_step = epoch * (len(train_dataset) // train_batch_size)
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                logger.info(f"*** Saving checkpoints after {cur_step} steps ***")
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(unreplicate(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 {cur_step}",
+                    )
+
+                    if not os.path.exists(os.path.join(training_args.output_dir, "tokenizer.json")):
+                        logger.info(f"*** Saving tokenizer ***")
+                        tokenizer.save_pretrained(
+                            training_args.output_dir,
+                            push_to_hub=training_args.push_to_hub,
+                            commit_message=f"Saving tokenizer",
+                        )
+
+
+if __name__ == "__main__":
+    main()

src/run_clm_flax_with_ckpts.py

@@ -0,0 +1,700 @@
+#!/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.
+"""
+Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) 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=causal-lm
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+import jax
+from jax import lax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import checkpoints
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForCausalLM,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.testing_utils import CaptureLogger
+
+logger = logging.getLogger(__name__)
+
+# 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_CAUSAL_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)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+                    "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+                    "value if set."
+        },
+    )
+    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"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+                    "The training dataset will be truncated in block of this size for training. "
+                    "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+    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."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+# 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)
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+        train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def restore_checkpoint(state, workdir):
+    return checkpoints.restore_checkpoint(workdir, state)
+
+
+def save_checkpoint(state, workdir):
+    if jax.process_index() == 0:
+        # get train state from the first replica
+        state = jax.device_get(jax.tree_map(lambda x: x[0], state))
+        step = int(state.step)
+        checkpoints.save_checkpoint(workdir, state, step, keep=3)
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+            os.path.exists(training_args.output_dir)
+            and os.listdir(training_args.output_dir)
+            and training_args.do_train
+            and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    checkpoints_dir = os.path.join(training_args.output_dir, "checkpoints")
+    os.makedirs(checkpoints_dir, exist_ok=True)
+
+    #  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.
+        dataset = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            cache_dir=model_args.cache_dir,
+            keep_in_memory=False
+        )
+
+        if "validation" not in dataset.keys():
+            dataset["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,
+            )
+            dataset["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"
+
+        dataset = load_dataset(
+            extension,
+            data_files=data_files,
+            delimiter="\t",
+            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."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            config=config,
+            seed=training_args.seed,
+            dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForCausalLM.from_config(
+            config,
+            seed=training_args.seed,
+            dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    tokenized_datasets = dataset.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,
+    )
+
+    if data_args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > config.max_position_embeddings:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+            )
+            block_size = 1024
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warning(
+                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+            )
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    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 // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i: i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        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
+
+    lm_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # 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))
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+
+    # total_train_steps = steps_per_epoch * num_epochs
+    if training_args.max_steps == -1:
+        total_train_steps = steps_per_epoch * num_epochs
+    else:
+        total_train_steps = training_args.max_steps
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # 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:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "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 = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+
+    # Restore states
+    state = restore_checkpoint(state, checkpoints_dir)
+    step_offset = int(state.step)  # step_offset > 0 if restarting from checkpoint
+    epoch_offset = int(num_epochs - ((total_train_steps - step_offset) / steps_per_epoch))
+
+    def loss_fn(logits, labels):
+        shift_logits = logits[..., :-1, :]
+        shift_labels = labels[..., 1:]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    if step_offset > 0:
+        logger.info("  Continuing training from checkpoint")
+        logger.info(f"  Continuing training from epoch {epoch_offset}")
+        logger.info(f"  Continuing training from global step {step_offset}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(epoch_offset, num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        num_steps = abs(step_offset - (steps_per_epoch * (epoch + 1)))
+
+        # train
+        for step in tqdm(range(num_steps), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (len(train_dataset) // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.save_steps and cur_step > 0:
+                save_checkpoint(state, checkpoints_dir)
+
+        # ======================== Evaluating ==============================
+        eval_metrics = []
+        eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+        eval_steps = len(eval_dataset) // eval_batch_size
+        for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+            # Model forward
+            batch = next(eval_loader)
+            metrics = p_eval_step(state.params, batch)
+            eval_metrics.append(metrics)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+
+        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+        try:
+            eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+        except OverflowError:
+            eval_metrics["perplexity"] = float("inf")
+
+        # Print metrics and update progress bar
+        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+        epochs.write(desc)
+        epochs.desc = desc
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(train_dataset) // train_batch_size)
+            write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(unreplicate(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}",
+            )
+
+
+if __name__ == "__main__":
+    main()

src/run_config.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export LC_ALL=C.UTF-8
+export LANG=C.UTF-8
+
+# export OUTPUT_DIR=./
+export OUTPUT_DIR=/home/saied/code/gpt2-medium-persian
+export NAME_OR_PATH=gpt2-medium
+
+python src/create_config.py \
+    --output_dir="$OUTPUT_DIR"  \
+    --name_or_path="$NAME_OR_PATH" \
+    --params='{"vocab_size": 50000,"bos_token_id": 5,"eos_token_id": 5}'

src/run_dataset.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export LC_ALL=C.UTF-8
+export LANG=C.UTF-8
+
+export OUTPUT_DIR=/home/m3hrdadfi/data/
+export DATASET_NAME=oscar
+export DATASET_CONFIG_NAME=unshuffled_deduplicated_fa
+
+python src/create_dataset.py \
+    --output_dir="$OUTPUT_DIR"  \
+    --dataset_name="$DATASET_NAME" \
+    --dataset_config_name="$DATASET_CONFIG_NAME"

src/run_tokenizer.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+export LC_ALL=C.UTF-8
+export LANG=C.UTF-8
+
+export OUTPUT_DIR=/home/saied/code/gpt2-medium-persian
+export DATASET_NAME=oscar
+export DATASET_CONFIG_NAME=unshuffled_deduplicated_fa
+export VOCAB_SIZE=50000
+export MIN_FREQUENCY=2
+export SPECIAL_TOKENS='<s>','<pad>','</s>','<unk>','<mask>','<|endoftext|>','<|startoftext|>','<sep>','<cls>','<nl>','<tab>','<zwnj>','[U1]','[U2]','[U3]','[U4]','[U5]','[U6]','[U7]','[U8]','[U9]','[U10]','[U11]','[U12]','[U13]','[U14]','[U15]','[U16]','[U17]','[U18]','[U19]','[U20]'
+
+
+python src/train_tokenizer.py \
+    --output_dir="$OUTPUT_DIR"  \
+    --dataset_name="$DATASET_NAME" \
+    --dataset_config_name="$DATASET_CONFIG_NAME" \
+    --vocab_size=$VOCAB_SIZE \
+    --min_frequency=$MIN_FREQUENCY \
+    --special_tokens="$SPECIAL_TOKENS"

src/train_tokenizer.py

@@ -0,0 +1,151 @@
+import ast
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple, Union, Any
+
+from datasets import load_dataset
+from tokenizers import ByteLevelBPETokenizer
+from transformers import (
+    HfArgumentParser,
+)
+
+from data_utils import (
+    filter_by_lang_regex,
+    filter_by_num_tokens,
+    filter_by_num_sents,
+    filter_by_adv,
+    normalizer
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TokenizerArguments:
+    """
+    Arguments to which tokenizer we are going to set up.
+    """
+
+    output_dir: str = field(
+        default=".",
+        metadata={"help": "The output directory where the config will be written."},
+    )
+    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)."})
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    special_tokens: Optional[str] = field(
+        default=None,
+        metadata={"help": "The list of special tokens that you want to add in your training."}
+    )
+    vocab_size: Optional[int] = field(
+        default=56000,
+        metadata={"help": "The size of the final vocabulary, including all tokens and alphabet"}
+    )
+    min_frequency: Optional[int] = field(
+        default=2,
+        metadata={"help": "The minimum frequency a pair should have in order to be merged"}
+    )
+    show_progress: Optional[bool] = field(
+        default=True,
+        metadata={"help": "Whether to show progress bars while training"}
+    )
+
+    def __post_init__(self):
+        if self.special_tokens is None:
+            special_tokens = [
+                "<s>", "<pad>", "</s>", "<unk>", "<mask>",
+                "<|endoftext|>", "<|startoftext|>",
+                "<sep>", "<cls>", "<nl>", "<tab>", "<zwnj>"
+            ]
+            special_tokens += [f"[U{i}]" for i in range(1, 21)]
+        else:
+            special_tokens = list(self.special_tokens.split(","))
+
+        self.special_tokens = special_tokens
+        if self.dataset_name is None and self.train_file is None:
+            raise ValueError("Need either a dataset name or a training 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."
+
+
+def main():
+    parser = HfArgumentParser([TokenizerArguments])
+    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.
+        tokenizer_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))[0]
+    else:
+        tokenizer_args = parser.parse_args_into_dataclasses()[0]
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO)
+
+    logger.info(f"Training tokenizer")
+
+    if tokenizer_args.dataset_name is not None:
+        raw_dataset = load_dataset(
+            tokenizer_args.dataset_name,
+            tokenizer_args.dataset_config_name,
+            cache_dir=tokenizer_args.cache_dir,
+            split="train"
+        )
+    else:
+        data_files = {"train": tokenizer_args.train_file}
+        extension = tokenizer_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+
+        raw_dataset = load_dataset(
+            extension,
+            data_files=data_files,
+            delimiter="\t",
+            cache_dir=tokenizer_args.cache_dir,
+        )
+
+    logger.info("Preprocessing the dataset")
+    dataset = raw_dataset.filter(lambda example: filter_by_lang_regex(example["text"], ratio=0.75))
+    dataset = dataset.filter(lambda example: filter_by_num_tokens(example["text"], gt=64))
+    dataset = dataset.filter(lambda example: filter_by_num_sents(example["text"], gt=2))
+    dataset = dataset.filter(lambda example: filter_by_adv(example["text"], ratio=50))
+    dataset = dataset.map(normalizer)
+    logger.info(f"Preprocessed dataset kept {len(dataset)} out of {len(raw_dataset)}")
+
+    tokenizer = ByteLevelBPETokenizer()
+
+    def batch_iterative(batch_size=1000):
+        for i in range(0, len(dataset), batch_size):
+            yield dataset[i: i + batch_size]["text"]
+
+    tokenizer.train_from_iterator(
+        batch_iterative(),
+        vocab_size=tokenizer_args.vocab_size,
+        special_tokens=tokenizer_args.special_tokens,
+        min_frequency=tokenizer_args.min_frequency,
+        show_progress=tokenizer_args.show_progress,
+    )
+
+    logger.info(f"Your tokenizer saved here {tokenizer_args.output_dir}")
+    os.makedirs(tokenizer_args.output_dir, exist_ok=True)
+    tokenizer.save_model(tokenizer_args.output_dir)
+    tokenizer.save(f"{tokenizer_args.output_dir}/tokenizer.json", pretty=True)
+
+
+if __name__ == '__main__':
+    main()

tf_model.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7ac4bf62ab348f0729b3f0aebd539b072bb75a5bcffb7f5ec7778185f305f2
+size 1418594792