Skip to content

WINDOWS_SETUP.md

Latest commit

 

History

History
152 lines (122 loc) · 3.49 KB

WINDOWS_SETUP.md

File metadata and controls

152 lines (122 loc) · 3.49 KB

Windows Setup Guide for Shielded RecRL

This guide will help you set up the Shielded RecRL project on Windows without using bash scripts.

Prerequisites

  1. Python 3.8+: Download from python.org
  2. Git: Download from git-scm.com
  3. CUDA Toolkit (optional, for GPU support): Download from nvidia.com

Option 1: Automated Setup (Recommended)

Using PowerShell

# Run PowerShell as Administrator if needed
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
.\setup_windows.ps1

Using Batch File

# Double-click setup_windows.bat or run in Command Prompt
setup_windows.bat

Option 2: Manual Setup

Step 1: Create Project Structure

mkdir data
mkdir data\raw
mkdir data\proc
mkdir data\_checksums
mkdir checkpoints
mkdir logs
mkdir experiments
mkdir docs

Step 2: Create .gitignore

Create a file named .gitignore with the following content:

*.pyc
__pycache__/
checkpoints/
logs/
*.pt
*.pth
*.bin
.env
.venv/
venv/
.idea/
.vscode/
*.log

Step 3: Initialize Git Repository

git init -b main
git add .gitignore
git commit -m "Init repo"

Step 4: Install Python Dependencies

pip install -r requirements.txt

Note: Some packages might require Microsoft Visual C++ Build Tools. If you encounter errors:

  1. Install Visual Studio Build Tools
  2. Or use pre-compiled wheels: pip install --only-binary=all -r requirements.txt

Step 5: Verify Installation

python gpu_test.py

Running the Project

1. Download Datasets

cd code\dataset
python download_datasets.py --dataset books

2. Preprocess Data

python preprocess.py --dataset books

3. Train Ranking Model

cd ..\ranker
python train_sasrec.py --dataset books --epochs 50

4. Initialize Language Model

cd ..\explainer
python init_lora.py --dataset books --int8

5. Compute Projection Basis

cd ..\projection
python basis.py --ranker_ckpt ..\checkpoints\sasrec_books.pt --output ..\checkpoints\basis_books.pt

6. Run Training

cd ..\trainer
python run_recrl.py --dataset books --cfg ..\..\experiments\recrl_default.yaml

Troubleshooting

Common Issues

  1. CUDA not available: Install CUDA Toolkit and PyTorch with CUDA support
  2. Memory errors: Use --int8 flag for 8-bit quantization
  3. Import errors: Ensure all dependencies are installed with pip install -r requirements.txt
  4. Git errors: Make sure Git is in your PATH

GPU Support

To enable GPU support on Windows:

  1. Install NVIDIA drivers
  2. Install CUDA Toolkit
  3. Install PyTorch with CUDA: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Alternative: Use WSL2

If you prefer a Linux environment:

  1. Install WSL2 from Microsoft Store
  2. Install Ubuntu on WSL2
  3. Follow the original bash setup instructions

Next Steps

After setup, you can:

  1. Run experiments with different datasets (books, ml25m, steam)
  2. Modify training parameters in experiments/recrl_default.yaml
  3. Analyze results in the logs/ directory
  4. Generate evaluation reports with code/eval/aggregate_main.py

Support

If you encounter issues:

  1. Check the troubleshooting section above
  2. Verify your Python and CUDA versions
  3. Ensure all dependencies are correctly installed
  4. Check the project's GitHub issues page