Skip to content

Latest commit

 

History

History
97 lines (74 loc) · 4.3 KB

README.md

File metadata and controls

97 lines (74 loc) · 4.3 KB
Raw

Hack-A-Sol 2024-25: AI/ML Track PS-1

Project Overview

This project is built for Hack-A-Sol 2024-25, hosted by IIIT NR. The focus is to apply Large Events Models (LEMs) for sequential event prediction in soccer, using the Wyscout dataset. LEMs are used to simulate matches, predict player performance, and fine-tune models for various contextual scenarios, such as different teams and player behaviors.

The project includes two sets of notebooks:

  1. Notebooks satisfying the hackathon's problem statement (PS).
  2. Notebooks inside the "ScaledUP" directory that represent upscaled versions, fine-tuned on more data than required by the problem statement.

Directory Structure

LargeEventsModel/
│
├── data/
│   ├── wyscout/
│   │   ├── csv/
│   │   │   ├── events/ (CSV Files)
│   │   ├── json/
│   │       ├── events/ (JSON Files)
│   │       ├── matches/ (Match Data)
├── lib/
│   ├── data_utils.py        # Helper functions for handling Wyscout data
│   ├── model_utils.py       # Utility functions for LEMs model building
│   ├── simulator.py         # Simulation functions using LEMs
├── models/
│   ├── lem/
│       ├── LEMv3_MODEL_DATA_TORCH.pth  # Pre-trained LEM data model
│       ├── LEMv3_MODEL_TYPE_TORCH.pth  # Pre-trained LEM type model
│       ├── LEMv4_MODEL_ACC_TORCH.pth   # Pre-trained accuracy model
├── ScaledUP/                # Upscaled versions, fine-tuned on more data than expected by PS
├── Dockerfile               # Docker setup for environment consistency
├── requirements.txt         # Python dependencies
└── README.md                # This file

Key Files

1. Notebooks for Hackathon Problem Statement (PS):

  • 011_Train_Model_Type.ipynb: Train the LEM to predict event types.
  • 012_Train_Model_Accuracy.ipynb: Train the model to estimate the accuracy of events.
  • 013_Train_Model_Data.ipynb: Predict event location and time intervals between events.
  • 021_Learning_State_Values.ipynb: Reinforcement learning for learning state values.
  • 022_Valuing_Actions_From_States.ipynb: Estimating value of actions, key for decision-making.
  • 023_Finetuning_Framework_Lib.ipynb: Fine-tune LEMs based on new data for better predictions.

2. Upscaled Notebooks (ScaledUP Directory):

The notebooks inside ScaledUP/ are designed to handle more data than required by the problem statement, offering more advanced, fine-tuned models:

  • ScaledUP/011_Train_Model_Type.ipynb
  • ScaledUP/022_Valuing_Actions_From_States.ipynb
  • ScaledUP/013_Train_Model_Data.ipynb

These notebooks are fine-tuned on larger datasets for more comprehensive insights into event sequences and player performance.

Technologies and Frameworks

  • Python for core development
  • PyTorch for deep learning model training and inference
  • NumPy and Pandas for data manipulation
  • Jupyter Notebooks for experimentation and prototyping
  • Docker for creating reproducible environments

Setup Instructions

1. Install Dependencies

Ensure you have Python 3.8+ installed. You can install all necessary dependencies using requirements.txt:

pip install -r requirements.txt

2. Dataset

The project uses Wyscout soccer data stored in data/wyscout/csv and data/wyscout/json. The dataset contains detailed match events and metadata for various leagues.

3. Running the Models

  • Event Type Prediction: Run 011_Train_Model_Type.ipynb for the basic problem statement or use the ScaledUP version for fine-tuning on more data.
  • Accuracy Prediction: Run 012_Train_Model_Accuracy.ipynb.
  • Event Data Prediction: Use 013_Train_Model_Data.ipynb.

4. Simulation

Simulate matches and predict sequences of game events using simulator.py, which generates and updates the context based on LEMs’ predictions.

Acknowledgements

This project is inspired by Large Language Models (LLMs) and adapted for sequential event prediction in sports, particularly soccer. Special thanks to IIIT NR and Hack-A-Sol organizers for providing a platform to showcase AI in sports analytics.

Enjoy the code!