- Install anaconda or miniconda
- Install git, then clone respository:
git clone https://github.com/harskish/tlgan/
- Create environment:
conda create -n tlgan python=3.9
- Activate environment:
conda activate tlgan
- Install dependencies
- On NVIDIA Ampere GPUs (3000 series) or newer:
conda env update -f env_cu11.yml --prune
- On older GPUs:
conda env update -f env_cu10.yml --prune
- On NVIDIA Ampere GPUs (3000 series) or newer:
- Setup submodules:
git submodule update --init --recursive
The networks (based on StyleGAN2) contain custom CUDA kernels for improved performance.
- Install CUDA toolkit (match the version in env_cuXX.yml)
- On Windows: install and open 'x64 Native Tools Command Prompt for VS 2019'
- Visual Studio 2019 Community Edition contains the required tools
The interactive viewers (visualize.py and grid_viz.py) benefit in performance from having access to a version of PyCUDA compiled with OpenGL support
Install the included dependencies:
pip install bin/cuXXX/*
- Install CUDA toolkit (match the version in env_cuXX.yml)
- Download pycuda sources from: https://pypi.org/project/pycuda/#files
- Extract files:
tar -xzf pycuda-VERSION.tar.gz
- Configure:
python configure.py --cuda-enable-gl --cuda-root=/path/to/cuda
- Compile and install:
make install