DORAEMON: Deep Object Recognition And Embedding Model Of Networks

🚀 Quick Start

Installation Guide

# Create and activate environment
python -m venv doraemon
source doraemon/bin/activate

# Install Doraemon
pip install doraemon-torch

# If you need to install in editable mode (for development)
pip install -e .

📢 What's New

• 🎁 2025.03.16: Doraemon v0.1.0 released
• 🎁 2024.10.01: Content-Based Image Retrieval (CBIR): Training on a real Amazon product dataset with a complete pipeline for training, end-to-end validation, and visualization. Please check ImageRetrieval.md
• 🎁 2024.04.01: Face Recognition: Based on a cleaned MS-Celeb-1M-v1c with over 70,000 IDs and 3.6 million images, validated with LFW. Includes loss functions like ArcFace, CircleLoss, and MagFace.
• 🎁 2023.06.01: Image Classification (IC): Given the Oxford-IIIT Pet dataset. Supports different learning rates for different layers, hard example mining, multi-label and single-label training, bad case analysis, GradCAM visualization, automatic labeling to aid semi-supervised training, and category-specific data augmentation. Refer to ImageClassification.md

✨ Highlights

• Optimization Algorithms: Various optimization techniques to enhance model training efficiency, including SGD, Adam, and SAM (Sharpness-Aware Minimization).

• Data Augmentation: A variety of data augmentation techniques to improve model robustness, such as CutOut, Color-Jitter, and Copy-Paste etc.

• Regularization: Techniques to prevent overfitting and improve model generalization, including Label Smoothing, OHEM, Focal Loss, and Mixup.

• Visualization: Integrated visualization tool to understand model decision-making, featuring GradCAM.

• Personalized Data Augmentation: Apply exclusive data augmentation to specific classes with Class-Specific Augmentation.

• Personalized Hyperparameter Tuning: Apply different learning rates to specific layers using Layer-Specific Learning Rates.

🚀 Deployment API

Doraemon offers incredibly simple yet powerful deployment options:

• Local API Inference: Deploy models with just a single weight file (*.pt) - one command setup for high-performance local inference
• Seamless HuggingFace Integration: Effortlessly deploy to the Huggingface ecosystem with full support for:
  • AutoModel.from_pretrained()
  • AutoProcessor.from_pretrained()
  • And all standard Hugging Face API interfaces

For detailed deployment instructions and ready-to-use examples, see our Deployment Guide.

📚 Tutorials

For detailed guidance on specific tasks, please refer to the following resources:

• Image Classification: If you are working on image classification tasks, please refer to Doc: Image Classification.

• Image Retrieval: For image retrieval tasks, please refer to Doc: Image Retrieval.

• Face Recognition: Stay tuned.

📊 Datasets

Doraemon integrates the following datasets, allowing users to quickly start training:

• Image Retrieval: Available at Ecommerce Product
• Face Recognition: Available at MS-Celeb-1M-v1c
• Image Classification: Available at Oxford-IIIT Pet

🧩 Supported Models

Doraemon now supports 1000+ models through integration with Timm:

• All models from timm.list_models(pretrained=True)
• Including CLIP, SigLIP, DeiT, BEiT, MAE, EVA, DINO and more

Model Performance Benchmarks can help you select the most suitable model by comparing:

• Inference speed
• Training efficiency
• Accuracy across different datasets
• Parameter count vs performance trade-offs

For detailed benchmark results, see @huggingface/pytorch-image-models#1933