README.md

metadata

title: AST Dashboard
emoji: 🐢
colorFrom: yellow
colorTo: indigo
sdk: gradio
sdk_version: 6.0.2
app_file: app.py
pinned: false

AST Training Dashboard - HuggingFace Space

Interactive dashboard for training models with Adaptive Sparse Training (AST).

Features

• 🚀 Live Training: Watch your model train in real-time
• 📊 Energy Tracking: See energy savings as you train
• 🎯 Model Card Generation: Auto-generate HuggingFace model cards
• ⚡ 60-70% Energy Savings: Train faster with minimal accuracy loss

Quick Start

Deploy to HuggingFace Spaces

1. Create new Space at https://huggingface.co/spaces
2. Choose Gradio as SDK
3. Upload files from this directory:
   • app.py
   • requirements.txt
   • README.md
4. Space will auto-deploy!

Run Locally

cd hf_space
pip install -r requirements.txt
python app.py

Then open http://localhost:7860

Usage

1. Select Model: Choose from ResNet18, EfficientNet, MobileNet
2. Set Activation Rate: Lower = more energy savings (0.35 recommended)
3. Choose Epochs: 30-50 epochs for good results
4. Start Training: Click "Start Training" and watch live metrics
5. Get Model Card: Copy auto-generated card for HuggingFace Hub

Example Results

Training ResNet18 on CIFAR-10 with AST (activation_rate=0.35):

• Accuracy: 92.1% (vs 92.3% baseline)
• Energy Savings: 65%
• Training Time: 2.8h (vs 7.2h baseline)

About AST

Adaptive Sparse Training (AST) automatically selects the most important training samples per batch, reducing compute by 60-70% while maintaining accuracy.

How it works:

1. Computes significance score (loss + entropy) for each sample
2. PI controller dynamically adjusts selection threshold
3. Only backpropagates through "hard" samples
4. Result: Same accuracy, way less compute

Links

• 📦 PyPI Package
• 🐙 GitHub
• 📚 Full Documentation

Citation

@software{adaptive_sparse_training,
    title={Adaptive Sparse Training},
    author={Idiakhoa, Oluwafemi},
    year={2024},
    url={https://github.com/oluwafemidiakhoa/adaptive-sparse-training}
}