An Efficient Library for Training Deep Models

This repository provides an efficient PyTorch-based library for training deep models.


Make sure your Python >= 3.7, CUDA version >= 11.1, and CUDNN version >= 7.6.5.

  1. Install package requirements via conda:

    conda create -n <ENV_NAME> python=3.7  # create virtual environment with Python 3.7
    conda activate <ENV_NAME>
    pip install -r requirements/minimal.txt -f
  2. To use video visualizer (optional), please also install ffmpeg.

    • Ubuntu: sudo apt-get install ffmpeg.
    • MacOS: brew install ffmpeg.
  3. To reduce memory footprint (optional), you can switch to either jemalloc (recommended) or tcmalloc rather than your default memory allocator.

    • jemalloc (recommended):
      • Ubuntu: sudo apt-get install libjemalloc
    • tcmalloc:
      • Ubuntu: sudo apt-get install google-perftools
  4. (optional) To speed up data loading on NVIDIA GPUs, you can install DALI, together with dill to pickle python objects. It is optional to also install CuPy for some customized operations if needed:

    pip install --extra-index-url --upgrade nvidia-dali-<CUDA_VERSION>
    pip install dill
    pip install cupy  # optional, installation can be slow

    For example, on CUDA 11.1, DALI can be installed via:

    pip install --extra-index-url --upgrade nvidia-dali-cuda110  # CUDA 11.1 compatible
    pip install dill
    pip install cupy  # optional, installation can be slow

Quick Demo

Train StyleGAN2 on FFHQ in Resolution of 256×256

In your Terminal, run:

./scripts/training_demos/ <NUM_GPUS> <PATH_TO_DATA> [OPTIONS]


  • <NUM_GPUS> refers to the number of GPUs. Setting <NUM_GPUS> as 1 helps launch a training job on single-GPU platforms.

  • <PATH_TO_DATA> refers to the path of FFHQ dataset (in resolution of 256×256) with zip format. If running on local machines, a soft link of the data will be created under the data folder of the working directory to save disk space.

  • [OPTIONS] refers to any additional option to pass. Detailed instructions on available options can be shown via ./scripts/training_demos/ <NUM_GPUS> <PATH_TO_DATA> --help.

This demo script uses stylegan2_ffhq256 as the default value of job_name, which is particularly used to identify experiments. Concretely, a directory with name job_name will be created under the root working directory (with is set as work_dirs/ by default). To prevent overwriting previous experiments, an exception will be raised to interrupt the training if the job_name directory has already existed. To change the job name, please use --job_name=<NEW_JOB_NAME> option.

More Demos

Please find more training demos under ./scripts/training_demos/.

Inspect Training Results

Besides using TensorBoard to track the training process, the raw results (e.g., training losses and running time) are saved in JSON format. They can be easily inspected with the following script

import json

file_name = '<PATH_TO_WORK_DIR>/log.json'

data_entries = []
with open(file_name, 'r') as f:
    for line in f:
        data_entry = json.loads(line)

# An example of data entry
# {"Loss/D Fake": 0.4833524551040682, "Loss/D Real": 0.4966000154727226, "Loss/G": 1.1439273656869773, "Learning Rate/Discriminator": 0.002352941082790494, "Learning Rate/Generator": 0.0020000000949949026, "data time": 0.0036810599267482758, "iter time": 0.24490128830075264, "run time": 66108.140625}

Convert Pre-trained Models

See Model Conversion for details.

Prepare Datasets

See Dataset Preparation for details.


See Contributing Guide for details.


The project is under MIT License.


This repository originates from GenForce, with all modules carefully optimized to make it more flexible and robust for distributed training. On top of GenForce where only StyleGAN training is provided, this repository also supports training StyleGAN2 and StyleGAN3, both of which are fully reproduced. Any new method is welcome to merge into this repository! Please refer to the Develop section.


The main contributors are listed as follows.

Member Contribution
Yujun Shen Refactor and optimize the entire codebase and reproduce start-of-the-art approaches.
Zhiyi Zhang Contribute to a number of sub-modules and functions, especially dataset related.
Dingdong Yang Contribute to DALI data loading acceleration.
Yinghao Xu Originally contribute to runner and loss functions in GenForce.
Ceyuan Yang Originally contribute to data loader in GenForce.
Jiapeng Zhu Originally contribute to evaluation metrics in GenForce.


We open source this library to the community to facilitate the research. If you do like our work and use the codebase for your projects, please cite our work as follows.

  title =        {Hammer: An Efficient Toolkit for Training Deep Models},
  author =       {Shen, Yujun and Zhang, Zhiyi and Yang, Dingdong and Xu, Yinghao and Yang, Ceyuan and Zhu, Jiapeng},
  howpublished = {\url{}},
  year =         {2022}