user8614 / hezar

The all-in-one AI library for Persian, supporting a wide variety of tasks and modalities!

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

You can also install the latest version from the source. Clone the repo and execute the following commands:

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:

• Getting Started
• Quick Tour
• Tutorials
• Developer Guides
• Contribution
• Reference API

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 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)

{'labels': ['positive'], 'probs': [0.812910258769989]}

• Sequence Labeling (POS, NER, etc.)

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

• Language Modeling

from hezar import Model

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

{'filled_texts': ['سلام بچه ها حالتون چطوره'], 'filled_tokens': [' چطوره']}

• Speech Recognition

from hezar import Model

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

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

• Image to Text (OCR)

from hezar import Model
# OCR with TrOCR
model = Model.load("hezarai/trocr-fa-v1")
texts = model.predict(["examples/assets/ocr_example.jpg"])
print(f"TrOCR Output: {texts}")

# OCR with CRNN
model = Model.load("hezarai/crnn-base-fa-64x256")
texts = model.predict("examples/assets/ocr_example.jpg")
print(f"CRNN Output: {texts}")

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

• Image to Text (Image Captioning)

from hezar import Model

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

{'texts': ['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 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 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 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'}]

Datasets

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

from hezar 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
...

Training

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

from hezar import (
    BertSequenceLabeling,
    BertSequenceLabelingConfig,
    TrainerConfig,
    Trainer,
    Dataset,
    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(
    device="cuda",
    init_weights_from=base_model_path,
    batch_size=8,
    num_epochs=5,
    checkpoints_dir="checkpoints/",
    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 trainers. Refer to the notebooks to see the examples!

Going Deeper

Hezar's primary focus is on providing ready to use models (implementations & pretrained weights) for different casual tasks without 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

This is a really heavy project to be maintained by a couple of developers. The idea isn't novel at all but actually doing it is really difficult hence being the only one in the whole history of the Persian open source! So any contribution is appreciated ❤️