Chinese Minority Costume Classification (CMC-Classification)

This repository contains the official implementation of the paper "Ensemble Learning for Fine-Grained Chinese Minority Costume Classification: A Comprehensive Empirical Study".

The project provides an automated framework to train, evaluate, and compare various deep learning architectures (CNNs, Vision Transformers, and Hybrid models) on the detailed classification of costumes from five Chinese ethnic groups.

📖 Overview

• Goal: Fine-grained classification of minority costumes.
• Classes: 5 (Hui, Zhuang, Man, Yao, and Bai).
• Dataset: 3,000 high-resolution images (600 per class), split 80/20 for training and validation.
• Performance: The proposed Ensemble method achieves 94.50% accuracy, outperforming individual baselines.

🚀 Key Features

The core script auto_baseline_runner.py is a complete pipeline that handles:

1. Automated Training: Runs multiple models (ResNet, EfficientNet, ViT, etc.) in sequence.
2. Ablation Studies: Can test training strategies like "Training from Scratch" vs. "Transfer Learning".
3. Analytics: Automatically generates confusion matrices, accuracy plots, and comparison tables (CSV/LaTeX).
4. Ensembling: Supports voting-based fusion of top-performing models.

🛠️ Installation

1. Clone the Repository

git clone https://github.com/TiandaSun/CMC-Classification.git
cd CMC-Classification

2. Install Dependencies

This project requires PyTorch and timm (PyTorch Image Models).

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install timm numpy pandas matplotlib seaborn scikit-learn tqdm

(Note: Adjust the PyTorch installation command based on your CUDA version).

💻 Usage

Data Preparation

Organize your dataset in standard ImageFolder format:

dataset/
├── Hui/
│   ├── image_01.jpg
│   └── ...
├── Zhuang/
├── Man/
├── Yao/
└── Bai/

Running the Baseline Suite

The auto_baseline_runner.py script is the main entry point.

1. Basic Run (All Models)

Train and evaluate all supported models (ResNet, VGG, EfficientNet, ViT, Swin, etc.):

python auto_baseline_runner.py --data_dir ./path/to/dataset --output_dir ./results

2. Quick Test (Essential Models Only)

Run only the 5 core models to save time:

python auto_baseline_runner.py --data_dir ./dataset --essential_only

3. Run Ablation Studies

Include experiments like training from scratch (no pre-trained weights):

python auto_baseline_runner.py --data_dir ./dataset --ablations

Command Line Arguments

Argument	Default	Description
--data_dir	./dataset	Path to the dataset root directory.
--output_dir	./baseline_results	Folder to save logs, plots, and checkpoints.
--batch_size	32	Batch size for training.
--max_epochs	100	Maximum epochs per model.
--patience	15	Early stopping patience (epochs).
--essential_only	False	Run only the core 5 baseline models.
--ablations	False	Run additional ablation study experiments.

📊 Models & Results

The code supports the following architectures via the timm library:

Category	Models
Traditional CNNs	ResNet50, VGG16
Efficient CNNs	EfficientNet-B0, MobileNetV2
Transformers	ViT-Small, DeiT-Tiny, DeiT-Small
Hybrid / Modern	ConvNeXt-Tiny, Swin-Tiny

Performance Snapshot

(Based on paper results on the validation set)

Model	Accuracy
Ensemble (Voting)	94.50%
ConvNeXt-Tiny	93.17%
VGG16	92.00%
ResNet50	90.17%
DeiT-Small	83.67%