hezar_pi: A Python repository from mostafapiran

The all-in-one AI library for Persian

Hezar (meaning thousand in Persian) is a multipurpose AI library built to make AI easy for the Persian community!

Hezar is a library that:

brings together all the best works in AI for Persian
makes using AI models as easy as a couple of lines of code
seamlessly integrates with Hugging Face Hub for all of its models
has a highly developer-friendly interface
has a task-based model interface which is more convenient for general users.
is packed with additional tools like word embeddings, tokenizers, feature extractors, etc.
comes with a lot of supplementary ML tools for deployment, benchmarking, optimization, etc.
and more!

Installation

Hezar is available on PyPI and can be installed with pip:

pip install hezar

Note that Hezar is a collection of models and tools, hence having different installation variants:

pip install hezar[all]  # For a full installation
pip install hezar[nlp]  # For NLP
pip install hezar[vision]  # For computer vision models
pip install hezar[audio]  # For audio and speech
pip install hezar[embeddings]  # For word embedding models

You can also install the latest version from the source:

git clone https://github.com/hezarai/hezar.git
pip install ./hezar

Documentation

Explore Hezar to learn more on the docs page or explore the key concepts:

Quick Tour

Models

There's a bunch of ready to use trained models for different tasks on the Hub!

🤗Hugging Face Hub Page: https://huggingface.co/hezarai

Let's walk you through some examples!

Text Classification (sentiment analysis, categorization, etc)

from hezar.models import Model

example = ["هزار، کتابخانه‌ای کامل برای به کارگیری آسان هوش مصنوعی"]
model = Model.load("hezarai/bert-fa-sentiment-dksf")
outputs = model.predict(example)
print(outputs)

[[{'label': 'positive', 'score': 0.812910258769989}]]

Sequence Labeling (POS, NER, etc.)

from hezar.models import Model

pos_model = Model.load("hezarai/bert-fa-pos-lscp-500k")  # Part-of-speech
ner_model = Model.load("hezarai/bert-fa-ner-arman")  # Named entity recognition
inputs = ["شرکت هوش مصنوعی هزار"]
pos_outputs = pos_model.predict(inputs)
ner_outputs = ner_model.predict(inputs)
print(f"POS: {pos_outputs}")
print(f"NER: {ner_outputs}")

POS: [[{'token': 'شرکت', 'label': 'Ne'}, {'token': 'هوش', 'label': 'Ne'}, {'token': 'مصنوعی', 'label': 'AJe'}, {'token': 'هزار', 'label': 'NUM'}]]
NER: [[{'token': 'شرکت', 'label': 'B-org'}, {'token': 'هوش', 'label': 'I-org'}, {'token': 'مصنوعی', 'label': 'I-org'}, {'token': 'هزار', 'label': 'I-org'}]]

Language Modeling (Mask Filling)

from hezar.models import Model

roberta_mlm = Model.load("hezarai/roberta-fa-mlm")
inputs = ["سلام بچه ها حالتون <mask>"]
outputs = roberta_mlm.predict(inputs, top_k=1)
print(outputs)

[[{'token': 'چطوره', 'sequence': 'سلام بچه ها حالتون چطوره', 'token_id': 34505, 'score': 0.2230483442544937}]]

Speech Recognition

from hezar.models import Model

whisper = Model.load("hezarai/whisper-small-fa")
transcripts = whisper.predict("examples/assets/speech_example.mp3")
print(transcripts)

[{'text': 'و این تنها محدود به محیط کار نیست'}]

Image to Text (OCR)

from hezar.models import Model
# OCR with TrOCR
model = Model.load("hezarai/trocr-base-fa-v2")
texts = model.predict(["examples/assets/ocr_example.jpg"])
print(f"TrOCR Output: {texts}")

# OCR with CRNN
model = Model.load("hezarai/crnn-fa-printed-96-long")
texts = model.predict("examples/assets/ocr_example.jpg")
print(f"CRNN Output: {texts}")

TrOCR Output: [{'text': 'چه میشه کرد، باید صبر کنیم'}]
CRNN Output: [{'text': 'چه میشه کرد، باید صبر کنیم'}]

Image to Text (License Plate Recognition)

from hezar.models import Model

model = Model.load("hezarai/crnn-fa-64x256-license-plate-recognition")
plate_text = model.predict("assets/license_plate_ocr_example.jpg")
print(plate_text)  # Persian text of mixed numbers and characters might not show correctly in the console

[{'text': '۵۷س۷۷۹۷۷'}]

Image to Text (Image Captioning)

from hezar.models import Model

model = Model.load("hezarai/vit-roberta-fa-image-captioning-flickr30k")
texts = model.predict("examples/assets/image_captioning_example.jpg")
print(texts)

[{'text': 'سگی با توپ تنیس در دهانش می دود.'}]

We constantly keep working on adding and training new models and this section will hopefully be expanding over time ;)

Word Embeddings

FastText

from hezar.embeddings import Embedding

fasttext = Embedding.load("hezarai/fasttext-fa-300")
most_similar = fasttext.most_similar("هزار")
print(most_similar)

[{'score': 0.7579, 'word': 'میلیون'},
 {'score': 0.6943, 'word': '21هزار'},
 {'score': 0.6861, 'word': 'میلیارد'},
 {'score': 0.6825, 'word': '26هزار'},
 {'score': 0.6803, 'word': '٣هزار'}]

Word2Vec (Skip-gram)

from hezar.embeddings import Embedding

word2vec = Embedding.load("hezarai/word2vec-skipgram-fa-wikipedia")
most_similar = word2vec.most_similar("هزار")
print(most_similar)

[{'score': 0.7885, 'word': 'چهارهزار'},
 {'score': 0.7788, 'word': '۱۰هزار'},
 {'score': 0.7727, 'word': 'دویست'},
 {'score': 0.7679, 'word': 'میلیون'},
 {'score': 0.7602, 'word': 'پانصد'}]

Word2Vec (CBOW)

from hezar.embeddings import Embedding

word2vec = Embedding.load("hezarai/word2vec-cbow-fa-wikipedia")
most_similar = word2vec.most_similar("هزار")
print(most_similar)

[{'score': 0.7407, 'word': 'دویست'},
 {'score': 0.7400, 'word': 'میلیون'},
 {'score': 0.7326, 'word': 'صد'},
 {'score': 0.7276, 'word': 'پانصد'},
 {'score': 0.7011, 'word': 'سیصد'}]

For a full guide on the embeddings module, see the embeddings tutorial.

Datasets

You can load any of the datasets on the Hub like below:

from hezar.data import Dataset

sentiment_dataset = Dataset.load("hezarai/sentiment-dksf")  # A TextClassificationDataset instance
lscp_dataset = Dataset.load("hezarai/lscp-pos-500k")  # A SequenceLabelingDataset instance
xlsum_dataset = Dataset.load("hezarai/xlsum-fa")  # A TextSummarizationDataset instance
alpr_ocr_dataset = Dataset.load("hezarai/persian-license-plate-v1")  # An OCRDataset instance
...

The returned dataset objects from load() are PyTorch Dataset wrappers for specific tasks and can be used by a data loader out-of-the-box!

You can also load Hezar's datasets using 🤗Datasets:

from datasets import load_dataset

dataset = load_dataset("hezarai/sentiment-dksf")

For a full guide on Hezar's datasets, see the datasets tutorial.

Training

Hezar makes it super easy to train models using out-of-the-box models and datasets provided in the library.

from hezar.models import BertSequenceLabeling, BertSequenceLabelingConfig
from hezar.data import Dataset
from hezar.trainer import Trainer, TrainerConfig
from hezar.preprocessors import Preprocessor

base_model_path = "hezarai/bert-base-fa"
dataset_path = "hezarai/lscp-pos-500k"

train_dataset = Dataset.load(dataset_path, split="train", tokenizer_path=base_model_path)
eval_dataset = Dataset.load(dataset_path, split="test", tokenizer_path=base_model_path)

model = BertSequenceLabeling(BertSequenceLabelingConfig(id2label=train_dataset.config.id2label))
preprocessor = Preprocessor.load(base_model_path)

train_config = TrainerConfig(
    output_dir="bert-fa-pos-lscp-500k",
    task="sequence_labeling",
    device="cuda",
    init_weights_from=base_model_path,
    batch_size=8,
    num_epochs=5,
    metrics=["seqeval"],
)

trainer = Trainer(
    config=train_config,
    model=model,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    data_collator=train_dataset.data_collator,
    preprocessor=preprocessor,
)
trainer.train()

trainer.push_to_hub("bert-fa-pos-lscp-500k")  # push model, config, preprocessor, trainer files and configs

You can actually go way deeper with the Trainer. See more details here.

Going Deeper

Hezar's primary focus is on providing ready to use models (implementations & pretrained weights) for different casual tasks not by reinventing the wheel, but by being built on top of PyTorch, 🤗Transformers, 🤗Tokenizers, 🤗Datasets, Scikit-learn, Gensim, etc. Besides, it's deeply integrated with the 🤗Hugging Face Hub and almost any module e.g, models, datasets, preprocessors, trainers, etc. can be uploaded to or downloaded from the Hub!

More specifically, here's a simple summary of the core modules in Hezar:

Models: Every model is a hezar.models.Model instance which is in fact, a PyTorch nn.Module wrapper with extra features for saving, loading, exporting, etc.
Datasets: Every dataset is a hezar.data.Dataset instance which is a PyTorch Dataset implemented specifically for each task that can load the data files from the Hugging Face Hub.
Preprocessors: All preprocessors are preferably backed by a robust library like Tokenizers, pillow, etc.
Embeddings: All embeddings are developed on top of Gensim and can be easily loaded from the Hub and used in just 2 lines of code!
Trainer: Trainer is the base class for training almost any model in Hezar or even your own custom models backed by Hezar. The Trainer comes with a lot of features and is also exportable to the Hub!
Metrics: Metrics are also another configurable and portable modules backed by Scikit-learn, seqeval, etc. and can be easily used in the trainers!

For more info, check the tutorials

Contribution

Maintaining this robust project is quite a task for our small team of developers. Although the idea isn't super new, putting it into action has been really tough. In fact, Hezar stands as the sole project of its kind in the entire history of Persian open source.

We could really use your help. Any contribution, big or small, would mean a lot to us. So, if you're interested, let's team up and make Hezar even better together! ❤️

Don't forget to check out our contribution guidelines in CONTRIBUTING.md before diving in. Your support is much appreciated!

Citation

If you found this project useful in your work or research please cite it by using this BibTeX entry:

@misc{hezar2023,
  title =        {Hezar: The all-in-one AI library for Persian},
  author =       {Aryan Shekarlaban & Pooya Mohammadi Kazaj},
  publisher =    {GitHub},
  howpublished = {\url{https://github.com/hezarai/hezar}},
  year =         {2023}
}

mostafapiran/hezar_pi