Net2net - Network-to-Network Translation with Conditional Invertible Neural Networks

Overview

Net2Net

Code accompanying the NeurIPS 2020 oral paper

Network-to-Network Translation with Conditional Invertible Neural Networks
Robin Rombach*, Patrick Esser*, Björn Ommer
* equal contribution

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.

teaser arXiv | BibTeX | Project Page

News Dec 19th, 2020: added SBERT-to-BigGAN, SBERT-to-BigBiGAN and SBERT-to-AE (COCO)

Requirements

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

conda env create -f environment.yaml
conda activate net2net

Datasets

  • 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. It consists of 18k oil portraits, which were obtained by running dlib on a subset of the WikiArt dataset dataset, kindly provided by A Style-Aware Content Loss for Real-time HD Style Transfer.
  • COCO: Create a symlink data/coco containing the images from the 2017 split in train2017 and val2017, and their annotations in annotations. Files can be obtained from the COCO webpage.

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 Models 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 ml4cad.py

Training

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:

model:
  base_learning_rate: 4.5e-6
  target: 
   
    
  params:
    ...
data:
  target: translation.DataModuleFromConfig
  params:
    batch_size: ...
    num_workers: ...
    train:
      target: 
    
     
      params:
        ...
    validation:
      target: 
     
      
      params:
        ...

     
    
   

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

python translation.py --base 
   
     -t --gpus 0,

   

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

python translation.py --base 
   
     -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:

model:
  base_learning_rate: 4.5e-6
  target: net2net.models.flows.flow.Net2NetFlow
  params:
    flow_config:
      target: 
   
    
      params:
        ...

    cond_stage_config:
      target: 
    
     
      params:
        ...

    first_stage_config:
      target: 
     
      
      params:
        ...

     
    
   

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

superres 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 translation.py --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 translation.py --base configs/autoencoder/faces32.yaml -t --gpus 
   
    , 

   

for the 32 x 32 images; and

python translation.py --base configs/autoencoder/faces256.yaml -t --gpus 
   
    , 

   

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/2020-10-16T17-11-42_FacesFQ32x32/checkpoints/last.ckptand logs/2020-09-16T16-23-39_FacesXL256z128/checkpoints/last.ckpt, respectively.

Training a cINN - Unpaired Translation

superres 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 translation.py --base configs/creativity/
   
    .yaml -t --gpus 0,

   

where .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 as described above.

In our framework, unpaired translation between domains is formulated as a translation between expert 1, a model which can infer the domain a given image belongs to, and expert 2, a model which can synthesize images of each domain. In the examples provided, we assume that the domain label comes with the dataset and provide the net2net.modules.labels.model.Labelator module, which simply returns a one hot encoding of this label. However, one could also use a classification model which infers the domain label from the image itself. For expert 2, our examples use an autoencoder trained jointly on all domains, which is easily achieved by concatenating datasets together. The provided net2net.data.base.ConcatDatasetWithIndex concatenates datasets and returns the corresponding dataset label for each example, which can then be used by the Labelator class for the translation. The training configurations for the autoencoders used in the creativity experiments are included in configs/autoencoder/anime_photography_256.yaml, configs/autoencoder/celeba_celebahq_ffhq_256.yaml and configs/autoencoder/portraits_photography_256.yaml.

Unpaired Translation on Custom Datasets

Create pytorch datasets for each of your domains, create a concatenated dataset with ConcatDatasetWithIndex (follow the example in net2net.data.faces.CCFQTrain), train an autoencoder on the concatenated dataset (adjust the data section in configs/autoencoder/celeba_celebahq_ffhq_256.yaml) and finally train a net2net translation model between a Labelator and your autoencoder (adjust the sections data and first_stage_config in configs/creativity/celeba_celebahq_ffhq_256.yaml). You can then also add your new model to the available modes in the ml4cad.py demo to visualize the results.

Training a cINN - Text-to-Image

texttoimage We provide code to obtain a text-to-image model by translating between a text model (SBERT) and an image decoder. To show the flexibility of our approach, we include code for three different decoders: BigGAN, as described in the paper, BigBiGAN, which is only available as a tensorflow model and thus nicely shows how our approach can work with black-box experts, and an autoencoder.

SBERT-to-BigGAN

Train with

python translation.py --base configs/translation/sbert-to-biggan256.yaml -t --gpus 0,

When running it for the first time, the required models will be downloaded automatically.

SBERT-to-BigBiGAN

Since BigBiGAN is only available on tensorflow-hub, this example has an additional dependency on tensorflow. A suitable environment is provided in env_bigbigan.yaml, and you will need COCO for training. You can then start training with

python translation.py --base configs/translation/sbert-to-bigbigan.yaml -t --gpus 0,

Note that the BigBiGAN class is just a naive wrapper, which converts pytorch tensors to numpy arrays, feeds them to the tensorflow graph and again converts the result to pytorch tensors. It does not require gradients of the expert model and serves as a good example on how to use black-box experts.

SBERT-to-AE

Similarly to the other examples, you can also train your own autoencoder on COCO with

python translation.py --base configs/autoencoder/coco256.yaml -t --gpus 0,

or download a pre-trained one, and translate to it by running

python translation.py --base configs/translation/sbert-to-ae-coco256.yaml -t --gpus 0,

Shout-outs

Thanks to everyone who makes their code and models available.

BibTeX

@misc{rombach2020networktonetwork,
      title={Network-to-Network Translation with Conditional Invertible Neural Networks},
      author={Robin Rombach and Patrick Esser and Björn Ommer},
      year={2020},
      eprint={2005.13580},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}
@misc{esser2020note,
      title={A Note on Data Biases in Generative Models}, 
      author={Patrick Esser and Robin Rombach and Björn Ommer},
      year={2020},
      eprint={2012.02516},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}
Owner
CompVis Heidelberg
Computer Vision research group at the Ruprecht-Karls-University Heidelberg
CompVis Heidelberg
Official implementations of PSENet, PAN and PAN++.

News (2021/11/03) Paddle implementation of PAN, see Paddle-PANet. Thanks @simplify23. (2021/04/08) PSENet and PAN are included in MMOCR. Introduction

395 Dec 14, 2022
PyTorch implementation of Towards Accurate Alignment in Real-time 3D Hand-Mesh Reconstruction (ICCV 2021).

Towards Accurate Alignment in Real-time 3D Hand-Mesh Reconstruction Introduction This is official PyTorch implementation of Towards Accurate Alignment

TANG Xiao 96 Dec 27, 2022
Biomarker identification for COVID-19 Severity in BALF cells Single-cell RNA-seq data

scBALF Covid-19 dataset Analysis Here is the Github page that has the codes for the bioinformatics pipeline described in the paper COVID-Datathon: Bio

Nami Niyakan 2 May 21, 2022
Image segmentation with private İstanbul Dataset

Image Segmentation This repo was created for academic research and test result. Repo will update after academic article online. This repo contains wei

İrem KÖMÜRCÜ 9 Dec 11, 2022
Official Pytorch implementation for 2021 ICCV paper "Learning Motion Priors for 4D Human Body Capture in 3D Scenes" and trained models / data

Learning Motion Priors for 4D Human Body Capture in 3D Scenes (LEMO) Official Pytorch implementation for 2021 ICCV (oral) paper "Learning Motion Prior

165 Dec 19, 2022
Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue

Realtime Unsupervised Depth Estimation from an Image This is the caffe implementation of our paper "Unsupervised CNN for single view depth estimation:

Ravi Garg 227 Nov 28, 2022
NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size

NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size Xuanyi Dong, Lu Liu, Katarzyna Musial, Bogdan Gabrys in IEEE Transactions o

D-X-Y 137 Dec 20, 2022
It's a implement of this paper:Relation extraction via Multi-Level attention CNNs

Relation Classification via Multi-Level Attention CNNs It's a implement of this paper:Relation Classification via Multi-Level Attention CNNs. Training

Aybss 2 Nov 04, 2022
Code for the IJCAI 2021 paper "Structure Guided Lane Detection"

SGNet Project for the IJCAI 2021 paper "Structure Guided Lane Detection" Abstract Recently, lane detection has made great progress with the rapid deve

Jinming Su 27 Dec 08, 2022
scAR (single-cell Ambient Remover) is a package for data denoising in single-cell omics.

scAR scAR (single cell Ambient Remover) is a package for denoising multiple single cell omics data. It can be used for multiple tasks, such as, sgRNA

19 Nov 28, 2022
Advancing mathematics by guiding human intuition with AI

Advancing mathematics by guiding human intuition with AI This repo contains two colab notebooks which accompany the paper, available online at https:/

DeepMind 315 Dec 26, 2022
Implementation of STAM (Space Time Attention Model), a pure and simple attention model that reaches SOTA for video classification

STAM - Pytorch Implementation of STAM (Space Time Attention Model), yet another pure and simple SOTA attention model that bests all previous models in

Phil Wang 109 Dec 28, 2022
Grad2Task: Improved Few-shot Text Classification Using Gradients for Task Representation

Grad2Task: Improved Few-shot Text Classification Using Gradients for Task Representation Prerequisites This repo is built upon a local copy of transfo

Jixuan Wang 10 Sep 28, 2022
ELSED: Enhanced Line SEgment Drawing

ELSED: Enhanced Line SEgment Drawing This repository contains the source code of ELSED: Enhanced Line SEgment Drawing the fastest line segment detecto

Iago Suárez 125 Dec 31, 2022
Statistical-Rethinking-with-Python-and-PyMC3 - Python/PyMC3 port of the examples in " Statistical Rethinking A Bayesian Course with Examples in R and Stan" by Richard McElreath

Statistical Rethinking with Python and PyMC3 This repository has been deprecated in favour of this one, please check that repository for updates, for

Osvaldo Martin 786 Dec 29, 2022
PyTorch implementation of "Representing Shape Collections with Alignment-Aware Linear Models" paper.

deep-linear-shapes PyTorch implementation of "Representing Shape Collections with Alignment-Aware Linear Models" paper. If you find this code useful i

Romain Loiseau 27 Sep 24, 2022
PyKale is a PyTorch library for multimodal learning and transfer learning as well as deep learning and dimensionality reduction on graphs, images, texts, and videos

PyKale is a PyTorch library for multimodal learning and transfer learning as well as deep learning and dimensionality reduction on graphs, images, texts, and videos. By adopting a unified pipeline-ba

PyKale 370 Dec 27, 2022
"Learning and Analyzing Generation Order for Undirected Sequence Models" in Findings of EMNLP, 2021

undirected-generation-dev This repo contains the source code of the models described in the following paper "Learning and Analyzing Generation Order f

Yichen Jiang 0 Mar 25, 2022
[TOG 2021] PyTorch implementation for the paper: SofGAN: A Portrait Image Generator with Dynamic Styling.

This repository contains the official PyTorch implementation for the paper: SofGAN: A Portrait Image Generator with Dynamic Styling. We propose a SofGAN image generator to decouple the latent space o

Anpei Chen 694 Dec 23, 2022
Large Scale Multi-Illuminant (LSMI) Dataset for Developing White Balance Algorithm under Mixed Illumination

Large Scale Multi-Illuminant (LSMI) Dataset for Developing White Balance Algorithm under Mixed Illumination (ICCV 2021) Dataset License This work is l

DongYoung Kim 33 Jan 04, 2023