InterFaceGAN Documentation

When loading pre-trained networks, you may encounter different types of snapshots:

• _G: An instantaneous snapshot of the generator. Primarily used for resuming a previous training run.
• _D: An instantaneous snapshot of the discriminator. Primarily used for resuming a previous training run.
• Gs: The long-term average of the generator. This version yields higher-quality results than the instantaneous snapshot and is preferred for generation tasks.

Follow these steps to verify that you can correctly import and run a pre-trained CelebA-HQ network:

1. Add the Progressive GAN repository to your PYTHONPATH.
2. Install dependencies: pip install -r requirements-pip.txt.
3. Download import_example.py from the official Google Drive resources.
4. Download karras2018iclr-celebahq-1024x1024.pkl and place it in the same directory as the script.
5. Run the script: python import_example.py.

If successful, the script will generate 10 PNG images (img0.png through img9.png) matching the reference images.

To train a StyleGAN network:

1. Edit train.py to specify your dataset and training configuration.
2. Run the training script: python train.py.
3. Results are saved to results/<ID>-<DESCRIPTION>.

Note: Training is computationally intensive and may take days or weeks. The default configuration is optimized for 8 GPUs. Training with fewer GPUs may produce different results compared to official benchmarks.

To run the TensorFlow implementation of Progressive GAN, ensure your environment meets the following requirements:

• OS: Linux (strongly recommended) or Windows.
• Python: 64-bit Python 3.6 with numpy 1.13.3 or newer (Anaconda3 recommended).
• GPU: One or more high-end NVIDIA Pascal or Volta GPUs with 16GB of DRAM (e.g., NVIDIA DGX-1 with 8 Tesla V100 GPUs).
• Drivers/Toolkits: NVIDIA driver 391.25 or newer, CUDA toolkit 9.0 or newer, and cuDNN 7.1.2 or newer.
• Dependencies: Install additional packages via pip install -r requirements-pip.txt.

Use run_metrics.py to evaluate metrics such as Fréchet Inception Distance (fid50k) and Perceptual Path Length (ppl_zfull, ppl_wfull, etc.).

By default, the script evaluates the pre-trained FFHQ generator and writes results to a new directory under results. You can customize the evaluation by editing run_metrics.py.

StyleGAN training requires datasets stored as multi-resolution TFRecords. Each dataset directory should contain the same image data in several resolutions.

By default, scripts look for datasets at datasets/<NAME>/<NAME>-<RESOLUTION>.tfrecords. You can modify these paths in config.py using data_dir.

Use dataset_tool.py to convert various image formats/libraries into the required TFRecord format.

To run the official TensorFlow implementation of StyleGAN, ensure your environment meets the following specifications:

• OS: Linux (strongly recommended) or Windows.
• Python: 64-bit Python 3.6 (Anaconda3 with numpy 1.14.3+ is recommended).
• TensorFlow: 1.10.0 or newer with GPU support.
• GPU: One or more high-end NVIDIA GPUs with at least 11GB of DRAM (e.g., NVIDIA Tesla V100).
• Drivers/Toolkits: NVIDIA driver 391.35+, CUDA toolkit 9.0+, and cuDNN 7.3.1+.

Pre-trained networks are stored as pickle files. When loading, the file yields three instances of dnnlib.tflib.Network:

• _G: Instantaneous snapshot of the generator (useful for resuming training).
• _D: Instantaneous snapshot of the discriminator (useful for resuming training).
• Gs: Long-term average of the generator (yields higher-quality results; typically used for image generation).

Requirements:

• The dnnlib source directory must be in your PYTHONPATH.
• A tf.Session must be set as default, initialized via dnnlib.tflib.init_tf().

import pickle
import dnnlib.tflib

url = 'https://drive.google.com/uc?id=1MEGjdvVpUsu1jB4zrXZN7Y4kBBOzizDQ' # karras2019stylegan-ffhq-1024x1024.pkl
with dnnlib.util.open_url(url, cache_dir=config.cache_dir) as f:
    _G, _D, Gs = pickle.load(f)

To edit facial attributes using InterFaceGAN, you need a pre-trained model, a boundary file (representing the semantic direction), and a latent code.

Prerequisite: Download the pre-trained ProgressiveGAN model on CelebA-HQ dataset and place it in the .models/pretrain/ folder.

Run the edit.py script with the following arguments:

• -m: The model name (e.g., pggan_celebahq).
• -b: Path to the boundary .npy file.
• -n: Number of latent codes to use for editing.
• -o: Output directory for the results.

LATENT_CODE_NUM=10
python edit.py \
    -m pggan_celebahq \
    -b boundaries/pggan_celebahq_smile_boundary.npy \
    -n "$LATENT_CODE_NUM" \
    -o results/pggan_celebahq_smile_editing

Pre-trained networks are available as pickled instances of dnnlib.tflib.Network. You can download specific models for different datasets via the following links:

• Flickr-Faces-HQ (1024x1024): stylegan-ffhq-1024x1024.pkl
• CelebA-HQ (1024x1024): stylegan-celebahq-1024x1024.pkl
• LSUN Bedrooms (256x256): stylegan-bedrooms-256x256.pkl
• LSUN Cars (512x384): stylegan-cars-512x384.pkl
• LSUN Cats (256x256): stylegan-cats-256x256.pkl