MaiDiffPredictor - maimai DX Chart Difficulty Prediction System

Project Overview

A deep learning system for predicting and analyzing chart difficulty in SEGA's rhythm game maimai DX. Combines Transformer architecture with latent factor modeling to achieve:

• Comprehensive difficulty prediction
• 6-dimensional skill analysis (Technical/Stamina/Accuracy/etc)
• Latent difficulty visualization
• Dynamic difficulty calibration

Key Features

• Multimodal Fusion: Integrates temporal note patterns with player performance statistics
• Contrastive Learning: Enhances robustness through data augmentation
• Interpretable Analysis: Factor loading matrix reveals hidden difficulty dimensions
• Dynamic Weighting: Adaptive fusion of statistical and sequential features
• Production Readiness: Supports multi-GPU training and C# interoperability

Current Limitations (Issue #1)

The current model exhibits excessive reliance on:

• Statistical metrics from diving-fish's dataset
• Chart length/duration
• Note density averages

This leads to suboptimal recognition of:

• Complex rhythm patterns
• Note arrangement complexity
• Stamina-demanding sequences
• Unconventional slide patterns

We welcome community contributions to improve pattern recognition capabilities.

Environment Requirements

• Python 3.8+
• PyTorch 2.0+
• .NET 6.0 SDK
• Key Python packages: scikit-learn, seaborn, tqdm, joblib

Installation

# Clone repository
git clone https://github.com/yourname/MaiDiffPredictor.git

# Install Python dependencies
# pip install -r requirements.txt

# Install .NET SDK (Ubuntu example)
wget https://packages.microsoft.com/config/ubuntu/20.04/packages-microsoft-prod.deb
sudo dpkg -i packages-microsoft-prod.deb
sudo apt-get update && sudo apt-get install dotnet-sdk-6.0

Data Preparation

# Serialize chart data
python serializer.py

# Fetch latest chart stats
python .\preprocess\getDiff.py
python .\preprocess\fixSlideTime.py .\serialized_data\ -o .\serialized_data\
python .\preprocess\enhance_json.py .\serialized_data\ -o .\serialized_data\
python .\preprocess\genInfo.py

Model Training

# Single GPU training
python .\train\ori_train.py

Prediction with Pretrained Model

python .\train\test.py

Project Structure

MaiDiffPredictor/
├── config/             # Project configuration
│   ├── api.json        # API configuration
│   └── train.json      # Training configuration
├── serialized_data/    # Processed chart data
├── info/               # Chart statistics
├── train/              # Core model code
│   ├── model.py        # Model definition & training
│   ├── dataset.py      # Data loading & features 
│   ├── test.py         # Prediction interface
│   └── ...
├── tools/              # C# serialization utilities
├── data/               # Chart data
├── serializer.py       # Serializer for chart data
└── preprocess/         # Data preprocessing scripts

Algorithm Overview

The system processes chart data through three main stages:

1. Feature Extraction

   • Temporal note patterns (hold, slide, break notes)
   • Player performance statistics (clear rates, FC distribution)
   • Density fluctuations

2. Hierarchical Processing

   • Transformer encoder for temporal patterns
   • Statistical feature fusion layer
   • Latent factor decomposition

3. Dynamic Prediction

   • Multi-head difficulty dimensions
   • Adaptive weight generation
   • Final difficulty synthesis

Sample Output

Epoch [10/1000], Loss: 0.5384
Top 5 high error samples:
Song: 786 | Level: 2
  FitDiff: 12.6518 | DS: 12.7
  Loss: 10.0690 | Predicted: 9.50 | Actual: 12.68
Song: 157 | Level: 2
  FitDiff: 12.2823 | DS: 11.6
  Loss: 9.8733 | Predicted: 8.80 | Actual: 11.94
Song: 531 | Level: 1
  FitDiff: 9.1862 | DS: 9.0
  Loss: 8.9822 | Predicted: 12.09 | Actual: 9.09
Song: 212 | Level: 2
  FitDiff: 12.4363 | DS: 12.2
  Loss: 8.7804 | Predicted: 9.35 | Actual: 12.32
Song: 189 | Level: 3
  FitDiff: 12.3803 | DS: 12.7
  Loss: 8.6610 | Predicted: 9.60 | Actual: 12.54
模型已保存: trained_models\model_epoch_10.pth
最好的模型已保存: trained_models/best_model.pth

License

...

---

Note: Ensure maidata.txt files comply with SEGA's secondary creation guidelines. This system is for research purposes only.