minilin 0.2.1

A universal low-resource deep learning framework - Learn More with Less

MiniLin Framework

Learn More with Less - A universal low-resource deep learning framework

English | 中文 | Русский | Français | العربية

What is MiniLin?

MiniLin is a deep learning framework designed for low-resource scenarios where data is scarce and computational resources are limited. It provides an end-to-end automated workflow for text, image, and audio tasks, with built-in optimization for edge device deployment.

Key Features

• 3-Line Solution: Complete ML pipeline from data to deployment
• Auto Strategy Selection: Automatically chooses optimal training strategy based on data size
• Lightweight Models: Pre-integrated efficient models (DistilBERT, MobileNet, Wav2Vec2)
• Model Compression: Quantization, pruning, and knowledge distillation
• Edge Deployment: Export to ONNX, PyTorch, TFLite
• Multi-Modal Support: Text, image, and audio tasks
• Few-Shot Learning: LoRA, Adapter, and Prompt Tuning
• Data Augmentation: Back-translation, Mixup, CutMix, SpecAugment
• API Deployment: FastAPI REST API server

Installation

Basic Installation

pip install minilin

With Optional Dependencies

# For vision tasks
pip install minilin[vision]

# For audio tasks
pip install minilin[audio]

# For optimization features
pip install minilin[optimization]

# For deployment
pip install minilin[deployment]

# Install everything
pip install minilin[all]

From Source

git clone https://github.com/alltobebetter/minilin.git
cd minilin
pip install -e .

Quick Start

Basic Usage (3 lines!)

from minilin import AutoPipeline

pipeline = AutoPipeline(task="text_classification", data_path="./data")
pipeline.train()
pipeline.deploy(output_path="./model.onnx")

Advanced Usage

from minilin import AutoPipeline

pipeline = AutoPipeline(
    task="text_classification",
    data_path="./data",
    target_device="mobile",
    max_samples=500,
    compression_level="high"
)

analysis = pipeline.analyze_data()
print(f"Strategy: {analysis['recommended_strategy']}")

pipeline.train(epochs=10, batch_size=16, learning_rate=2e-5)

metrics = pipeline.evaluate()
print(f"Accuracy: {metrics['accuracy']:.4f}")

pipeline.deploy(output_path="./model.onnx", quantization="int8")

Advanced Features

Few-Shot Learning with LoRA

from minilin.models import apply_few_shot_method

model = apply_few_shot_method(model, method="lora", r=8, alpha=16)
pipeline.train(max_samples=50, epochs=20)

Knowledge Distillation

from minilin.optimization import KnowledgeDistiller

distiller = KnowledgeDistiller(
    teacher_model=large_model,
    student_model=small_model,
    temperature=3.0,
    alpha=0.5
)

metrics = distiller.distill(train_loader, val_loader, epochs=5)

Multi-Modal Learning

from minilin.models import create_multimodal_model

model = create_multimodal_model(
    text_model_name="distilbert-base-uncased",
    image_model_name="mobilenetv3_small_100",
    audio_model_name="facebook/wav2vec2-base",
    num_classes=10,
    fusion_method="attention"
)

FastAPI Deployment

from minilin.deployment import serve_model

serve_model(
    model_path="./model.onnx",
    task="text_classification",
    host="0.0.0.0",
    port=8000
)

Supported Tasks

Text Tasks

• Text Classification
• Named Entity Recognition (NER)
• Sentiment Analysis

Vision Tasks

• Image Classification
• Image Augmentation (Mixup, CutMix)

Audio Tasks

• Audio Classification
• Audio Augmentation (SpecAugment)

Multi-Modal Tasks

• Text + Image
• Text + Audio
• Text + Image + Audio

Core Modules

Data Layer

• DataAnalyzer: Automatic data analysis and quality assessment
• DataLoader: Support for JSON, JSONL, CSV, TXT, directory formats
• DataAugmenter: Text augmentation (synonym, insertion, deletion, swap)
• BackTranslator: Back-translation (googletrans, DeepL)
• ImageAugmenter: Image augmentation (Mixup, CutMix, transforms)
• AudioAugmenter: Audio augmentation (noise, shift, speed, SpecAugment)

Model Layer

• ModelZoo: Pre-integrated lightweight models
• Trainer: Text model trainer with auto hyperparameter tuning
• ImageTrainer: Image model trainer
• AudioTrainer: Audio model trainer
• MultiModalModel: Multi-modal fusion model
• Few-Shot Learning: LoRA, Adapter, Prompt Tuning

Optimization Layer

• Quantizer: INT8/FP16 quantization
• Pruner: Structured/unstructured pruning
• KnowledgeDistiller: Teacher-Student distillation
• ModelCompressor: Compression orchestration

Deployment Layer

• ModelExporter: Export to ONNX, PyTorch, TFLite
• ModelServer: FastAPI REST API server
• Edge Optimization: Mobile and embedded device support

Performance

• Training Speed: 2-3x faster than standard training
• Model Size: Compressed to 10-20% of original size
• Inference Speed: Real-time on edge devices (>30 FPS)
• Accuracy Loss: <2% after compression

Examples

Check out the examples directory:

• Text Classification
• Sentiment Analysis
• Image Classification
• Audio Classification
• Multi-Modal Learning
• Advanced Features

Configuration

Configuration File

Create ~/.minilin/config.json:

{
  "translation_api_key": "your-api-key",
  "huggingface_token": "your-token",
  "cache_dir": "~/.minilin/cache"
}

Environment Variables

export MINILIN_TRANSLATION_API_KEY="your-api-key"
export MINILIN_HUGGINGFACE_TOKEN="your-token"

Contributing

We welcome contributions! Please see CONTRIBUTING.md for details.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

• GitHub: https://github.com/alltobebetter/minilin
• Email: [e-mail]

---

Made with ❤️ by the MiniLin Team