DaGAN Example¶

This example demonstrates a GAN-like vision model. The nnscaler trainer assumes there is only one end-to-end module that needs to be parallelized. However, GAN-like models always have both a generator and a discriminator. Here, you will learn how to run your code without the nnscaler trainer, and how to parallelize, synchronize, and update modules during training.

In this example, both GeneratorFullModel and DiscriminatorFullModel contain the same keypoint detector, generator, and discriminator modules. A module cannot be parallelized multiple times, so keypoint detector, generator, and discriminator must be parallelized separately. Separate synchronization and updates are also needed during training.

# clone nnScaler & CVPR2022-DaGAN repositories
git clone --recursive https://github.com/microsoft/nnscaler
cd nnscaler/examples/vision/dagan
git clone https://github.com/harlanhong/CVPR2022-DaGAN.git

# Install dependent packages
mv CVPR2022-DaGAN CVPR2022_DaGAN
pip install --ignore-installed -r CVPR2022_DaGAN/requirements.txt
export PYTHONPATH=$PYTHONPATH:CVPR2022_DaGAN

# Run
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc-per-node=4 --master_port=12348 run.py \
    --config vox-adv-256.yaml --name DaGAN --batchsize 8 --kp_num 15 --generator DepthAwareGenerator