Code accompanying the NeurIPS 2020 oral paper

tl;dr Our approach distills the residual information of one model with respect to another's and thereby enables translation between fixed off-the-shelf expert models such as BERT and BigGAN without having to modify or finetune them.


A suitable conda environment named net2net can be created
and activated with:

conda env create -f environment.yaml
conda activate net2net


  • CelebA: Create a symlink 'data/CelebA' pointing to a folder which contains the following files:
    ├── identity_CelebA.txt
    ├── img_align_celeba
    ├── list_attr_celeba.txt
    └── list_eval_partition.txt
    These files can be obtained here.
  • CelebA-HQ: Create a symlink data/celebahq pointing to a folder containing
    the .npy files of CelebA-HQ (instructions to obtain them can be found in
    the PGGAN repository).
  • FFHQ: Create a symlink data/ffhq pointing to the images1024x1024 folder
    obtained from the FFHQ repository.
  • Anime Faces: First download the face images from the Anime Crop dataset and then apply
    the preprocessing of FFHQ to those images. We only keep images
    where the underlying dlib face recognition model recognizes
    a face. Finally, create a symlink data/anime which contains the processed anime face images.
  • Oil Portraits: Download here.
    Unpack the content and place the files in data/portraits.

ML4Creativity Demo

We include a streamlit demo, which utilizes our
approach to demonstrate biases of datasets and their creative applications.
More information can be found in our paper A Note on Data Biases in Generative
from the Machine Learning for Creativity and Design at NeurIPS 2020. Download the models from

and place them into logs. Run the demo with

streamlit run


Our code uses Pytorch-Lightning and thus natively supports
things like 16-bit precision, multi-GPU training and gradient accumulation. Training details for any model need to be specified in a dedicated .yaml file.
In general, such a config file is structured as follows:

  base_learning_rate: 4.5e-6
  target: <path/to/lightning/module>
  target: translation.DataModuleFromConfig
    batch_size: ...
    num_workers: ...
      target: <path/to/train/dataset>
      target: <path/to/validation/dataset>

Any Pytorch-Lightning model specified under is then trained on the specified data
by running the command:

python --base <path/to/yaml> -t --gpus 0,

All available Pytorch-Lightning trainer arguments can be added via the command line, e.g. run

python --base <path/to/yaml> -t --gpus 0,1,2,3 --precision 16 --accumulate_grad_batches 2

to train a model on 4 GPUs using 16-bit precision and a 2-step gradient accumulation.
More details are provided in the examples below.

Training a cINN

Training a cINN for network-to-network translation usually utilizes the Lighnting Module net2net.models.flows.flow.Net2NetFlow
and makes a few further assumptions on the configuration file and model interface:

  base_learning_rate: 4.5e-6
  target: net2net.models.flows.flow.Net2NetFlow
      target: <path/to/cinn>

      target: <path/to/network1>

      target: <path/to/network2>

Here, the entries under flow_config specifies the architecture and parameters of the conditional INN;
cond_stage_config specifies the first network whose representation is to be translated into another network
specified by first_stage_config. Our model net2net.models.flows.flow.Net2NetFlow expects that the first
network has a .encode() method which produces the representation of interest, while the second network should
have an encode() and a decode() method, such that both of them applied sequentially produce the networks output. This allows for a modular combination of arbitrary models of interest. For more details, see the examples below.

Training a cINN - Superresolution

Training details for a cINN to concatenate two autoencoders from different image scales for stochastic
superresolution are specified in configs/translation/faces32-to-256.yaml.

To train a model for translating from 32 x 32 images to 256 x 256 images on GPU 0, run

python --base configs/translation/faces32-to-faces256.yaml -t --gpus 0, 

and specify any additional training commands as described above. Note that this setup requires two
pretrained autoencoder models, one on 32 x 32 images and the other on 256 x 256. If you want to
train them yourself on a combination of FFHQ and CelebA-HQ, run

python --base configs/autoencoder/faces32.yaml -t --gpus <n>, 

for the 32 x 32 images; and

python --base configs/autoencoder/faces256.yaml -t --gpus <n>, 

for the model on 256 x 256 images. After training, adopt the corresponding model paths in configs/translation/faces32-to-faces256.yaml. Additionally, we provide weights of pretrained autoencoders for both settings:
Weights 32x32; Weights256x256.
To run the training as described above, put them into
logs/2020-09-16T16-23-39_FacesXL256z128/checkpoints/last.ckpt, respectively.

Training a cINN - Unpaired Translation

All training scenarios for unpaired translation are specified in the configs in configs/creativity.
We provide code and pretrained autoencoder models for three different translation tasks:

  • AnimePhotography; see configs/creativity/anime_photography_256.yaml.
    Download autoencoder checkpoint (Download Anime+Photography) and place into logs/2020-09-30T21-40-22_AnimeAndFHQ/checkpoints/epoch=000007.ckpt.
  • Oil-PortraitPhotography; see configs/creativity/portraits_photography_256.yaml
    Download autoencoder checkpoint (Download Portrait+Photography) and place into logs/2020-09-29T23-47-10_PortraitsAndFFHQ/checkpoints/epoch=000004.ckpt.
  • FFHQCelebA-HQCelebA; see configs/creativity/celeba_celebahq_ffhq_256.yaml
    Download autoencoder checkpoint (Download FFHQ+CelebAHQ+CelebA) and place into logs/2020-09-16T16-23-39_FacesXL256z128/checkpoints/last.ckpt.
    Note that this is the same autoencoder checkpoint as for the stochastic superresolution experiment.

To train a cINN on one of these unpaired transfer tasks using the first GPU, simply run

python --base configs/translation/<task-of-interest>.yaml -t --gpus 0, 

where <task-of-interest>.yaml is one of portraits_photography_256.yaml, celeba_celebahq_ffhq_256.yaml
or anime_photography_256.yaml. Providing additional arguments to the pytorch-lightning
trainer object is also possible and described above.


      title={Network-to-Network Translation with Conditional Invertible Neural Networks},
      author={Robin Rombach and Patrick Esser and Björn Ommer},