Fader Networks: Manipulating Images by Sliding Attributes - NIPS 2017

Overview

FaderNetworks

PyTorch implementation of Fader Networks (NIPS 2017).

Fader Networks can generate different realistic versions of images by modifying attributes such as gender or age group. They can swap multiple attributes at a time, and continuously interpolate between each attribute value. In this repository we provide the code to reproduce the results presented in the paper, as well as trained models.

Single-attribute swap

Below are some examples of different attribute swaps:

Multi-attributes swap

The Fader Networks are also designed to disentangle multiple attributes at a time:

Model

The main branch of the model (Inference Model), is an autoencoder of images. Given an image x and an attribute y (e.g. male/female), the decoder is trained to reconstruct the image from the latent state E(x) and y. The other branch (Adversarial Component), is composed of a discriminator trained to predict the attribute from the latent state. The encoder of the Inference Model is trained not only to reconstruct the image, but also to fool the discriminator, by removing from E(x) the information related to the attribute. As a result, the decoder needs to consider y to properly reconstruct the image. During training, the model is trained using real attribute values, but at test time, y can be manipulated to generate variations of the original image.

Dependencies

Installation

Simply clone the repository:

git clone https://github.com/facebookresearch/FaderNetworks.git
cd FaderNetworks

Dataset

Download the aligned and cropped CelebA dataset from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html. Extract all images and move them to the data/img_align_celeba/ folder. There should be 202599 images. The dataset also provides a file list_attr_celeba.txt containing the list of the 40 attributes associated with each image. Move it to data/. Then simply run:

cd data
./preprocess.py

It will resize images, and create 2 files: images_256_256.pth and attributes.pth. The first one contains a tensor of size (202599, 3, 256, 256) containing the concatenation of all resized images. Note that you can update the image size in preprocess.py to work with different resolutions. The second file is a pre-processed version of the attributes.

Pretrained models

You can download pretrained classifiers and Fader Networks by running:

cd models
./download.sh

Train your own models

Train a classifier

To train your own model you first need to train a classifier to let the model evaluate the swap quality during the training. Training a good classifier is relatively simple for most attributes, and a good model can be trained in a few minutes. We provide a trained classifier for all attributes in models/classifier256.pth. Note that the classifier does not need to be state-of-the-art, it is not used during the training process, but is just here to monitor the swap quality. If you want to train your own classifier, you can run classifier.py, using the following parameters:

python classifier.py

# Main parameters
--img_sz 256                  # image size
--img_fm 3                    # number of feature maps
--attr "*"                    # attributes list. "*" for all attributes

# Network architecture
--init_fm 32                  # number of feature maps in the first layer
--max_fm 512                  # maximum number of feature maps
--hid_dim 512                 # hidden layer size

# Training parameters
--v_flip False                # randomly flip images vertically (data augmentation)
--h_flip True                 # randomly flip images horizontally (data augmentation)
--batch_size 32               # batch size
--optimizer "adam,lr=0.0002"  # optimizer
--clip_grad_norm 5            # clip gradient L2 norm
--n_epochs 1000               # number of epochs
--epoch_size 50000            # number of images per epoch

# Reload
--reload ""                   # reload a trained classifier
--debug False                 # debug mode (if True, load a small subset of the dataset)

Train a Fader Network

You can train a Fader Network with train.py. The autoencoder can receive feedback from:

  • The image reconstruction loss
  • The latent discriminator loss
  • The PatchGAN discriminator loss
  • The classifier loss

In the paper, only the first two losses are used, but the two others could improve the results further. You can tune the impact of each of these losses with the lambda_ae, lambda_lat_dis, lambda_ptc_dis, and lambda_clf_dis coefficients. Below is a complete list of all parameters:

# Main parameters
--img_sz 256                      # image size
--img_fm 3                        # number of feature maps
--attr "Male"                     # attributes list. "*" for all attributes

# Networks architecture
--instance_norm False             # use instance normalization instead of batch normalization
--init_fm 32                      # number of feature maps in the first layer
--max_fm 512                      # maximum number of feature maps
--n_layers 6                      # number of layers in the encoder / decoder
--n_skip 0                        # number of skip connections
--deconv_method "convtranspose"   # deconvolution method
--hid_dim 512                     # hidden layer size
--dec_dropout 0                   # dropout in the decoder
--lat_dis_dropout 0.3             # dropout in the latent discriminator

# Training parameters
--n_lat_dis 1                     # number of latent discriminator training steps
--n_ptc_dis 0                     # number of PatchGAN discriminator training steps
--n_clf_dis 0                     # number of classifier training steps
--smooth_label 0.2                # smooth discriminator labels
--lambda_ae 1                     # autoencoder loss coefficient
--lambda_lat_dis 0.0001           # latent discriminator loss coefficient
--lambda_ptc_dis 0                # PatchGAN discriminator loss coefficient
--lambda_clf_dis 0                # classifier loss coefficient
--lambda_schedule 500000          # lambda scheduling (0 to disable)
--v_flip False                    # randomly flip images vertically (data augmentation)
--h_flip True                     # randomly flip images horizontally (data augmentation)
--batch_size 32                   # batch size
--ae_optimizer "adam,lr=0.0002"   # autoencoder optimizer
--dis_optimizer "adam,lr=0.0002"  # discriminator optimizer
--clip_grad_norm 5                # clip gradient L2 norm
--n_epochs 1000                   # number of epochs
--epoch_size 50000                # number of images per epoch

# Reload
--ae_reload ""                    # reload pretrained autoencoder
--lat_dis_reload ""               # reload pretrained latent discriminator
--ptc_dis_reload ""               # reload pretrained PatchGAN discriminator
--clf_dis_reload ""               # reload pretrained classifier
--eval_clf ""                     # evaluation classifier (trained with classifier.py)
--debug False                     # debug mode (if True, load a small subset of the dataset)

Generate interpolations

Given a trained model, you can use it to swap attributes of images in the dataset. Below are examples using the pretrained models:

# Narrow Eyes
python interpolate.py --model_path models/narrow_eyes.pth --n_images 10 --n_interpolations 10 --alpha_min 10.0 --alpha_max 10.0 --output_path narrow_eyes.png

# Eyeglasses
python interpolate.py --model_path models/eyeglasses.pth --n_images 10 --n_interpolations 10 --alpha_min 2.0 --alpha_max 2.0 --output_path eyeglasses.png

# Age
python interpolate.py --model_path models/young.pth --n_images 10 --n_interpolations 10 --alpha_min 10.0 --alpha_max 10.0 --output_path young.png

# Gender
python interpolate.py --model_path models/male.pth --n_images 10 --n_interpolations 10 --alpha_min 2.0 --alpha_max 2.0 --output_path male.png

# Pointy nose
python interpolate.py --model_path models/pointy_nose.pth --n_images 10 --n_interpolations 10 --alpha_min 10.0 --alpha_max 10.0 --output_path pointy_nose.png

These commands will generate images with 10 rows of 12 columns with the interpolated images. The first column corresponds to the original image, the second is the reconstructed image (without alteration of the attribute), and the remaining ones correspond to the interpolated images. alpha_min and alpha_max represent the range of the interpolation. Values superior to 1 represent generations over the True / False range of the boolean attribute in the model. Note that the variations of some attributes may only be noticeable for high values of alphas. For instance, for the "eyeglasses" or "gender" attributes, alpha_max=2 is usually enough, while for the "age" or "narrow eyes" attributes, it is better to go up to alpha_max=10.

References

If you find this code useful, please consider citing:

Fader Networks: Manipulating Images by Sliding Attributes - G. Lample, N. Zeghidour, N. Usunier, A. Bordes, L. Denoyer, M'A. Ranzato

@inproceedings{lample2017fader,
  title={Fader Networks: Manipulating Images by Sliding Attributes},
  author={Lample, Guillaume and Zeghidour, Neil and Usunier, Nicolas and Bordes, Antoine and DENOYER, Ludovic and others},
  booktitle={Advances in Neural Information Processing Systems},
  pages={5963--5972},
  year={2017}
}

Contact: [email protected], [email protected]

Owner
Facebook Research
Facebook Research
A simple PyTorch Implementation of Generative Adversarial Networks, focusing on anime face drawing.

AnimeGAN A simple PyTorch Implementation of Generative Adversarial Networks, focusing on anime face drawing. Randomly Generated Images The images are

Jie Lei 雷杰 1.2k Jan 03, 2023
A collection of models for image<->text generation in ACM MM 2021.

Bi-directional Image and Text Generation UMT-BITG (image & text generator) Unifying Multimodal Transformer for Bi-directional Image and Text Generatio

Multimedia Research 63 Oct 30, 2022
Python Classes: Medical Insurance Project using Object Oriented Programming Concepts

Medical-Insurance-Project-OOP Python Classes: Medical Insurance Project using Object Oriented Programming Concepts Classes are an incredibly useful pr

Hugo B. 0 Feb 04, 2022
Yolov5-opencv-cpp-python - Example of using ultralytics YOLO V5 with OpenCV 4.5.4, C++ and Python

yolov5-opencv-cpp-python Example of performing inference with ultralytics YOLO V

183 Jan 09, 2023
Official PyTorch implementation of "AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks"

AASIST This repository provides the overall framework for training and evaluating audio anti-spoofing systems proposed in 'AASIST: Audio Anti-Spoofing

Clova AI Research 56 Jan 02, 2023
Video Contrastive Learning with Global Context

Video Contrastive Learning with Global Context (VCLR) This is the official PyTorch implementation of our VCLR paper. Install dependencies environments

143 Dec 26, 2022
(JMLR'19) A Python Toolbox for Scalable Outlier Detection (Anomaly Detection)

Python Outlier Detection (PyOD) Deployment & Documentation & Stats Build Status & Coverage & Maintainability & License PyOD is a comprehensive and sca

Yue Zhao 6.6k Jan 03, 2023
Simple transformer model for CIFAR10

CIFAR-Transformer Simple transformer model for CIFAR10. Reference: https://www.tensorflow.org/text/tutorials/transformer https://github.com/huggingfac

9 Nov 07, 2022
In this project, we develop a face recognize platform based on MTCNN object-detection netcwork and FaceNet self-supervised network.

模式识别大作业——人脸检测与识别平台 本项目是一个简易的人脸检测识别平台,提供了人脸信息录入和人脸识别的功能。前端采用 html+css+js,后端采用 pytorch,

Xuhua Huang 5 Aug 02, 2022
Kaggle | 9th place single model solution for TGS Salt Identification Challenge

UNet for segmenting salt deposits from seismic images with PyTorch. General We, tugstugi and xuyuan, have participated in the Kaggle competition TGS S

Erdene-Ochir Tuguldur 276 Dec 20, 2022
Bunch of different tools which helps visualizing and annotating images for semantic/instance segmentation tasks

Data Framework for Semantic/Instance Segmentation Bunch of different tools which helps visualizing, transforming and annotating images for semantic/in

Bruno Fernandes Carvalho 5 Dec 21, 2022
Can we learn gradients by Hamiltonian Neural Networks?

Can we learn gradients by Hamiltonian Neural Networks? This project was carried out as part of the Optimization for Machine Learning course (CS-439) a

2 Aug 22, 2022
Final project code: Implementing BicycleGAN, for CIS680 FA21 at University of Pennsylvania

680 Final Project: BicycleGAN Haoran Tang Instructions 1. Training To train the network, please run train.py. Change hyper-parameters and folder paths

Haoran Tang 0 Apr 22, 2022
Prototypical python implementation of the trust-region algorithm presented in Sequential Linearization Method for Bound-Constrained Mathematical Programs with Complementarity Constraints by Larson, Leyffer, Kirches, and Manns.

Prototypical python implementation of the trust-region algorithm presented in Sequential Linearization Method for Bound-Constrained Mathematical Programs with Complementarity Constraints by Larson, L

3 Dec 02, 2022
Proximal Backpropagation - a neural network training algorithm that takes implicit instead of explicit gradient steps

Proximal Backpropagation Proximal Backpropagation (ProxProp) is a neural network training algorithm that takes implicit instead of explicit gradient s

Thomas Frerix 40 Dec 17, 2022
Tools for computational pathology

A toolkit for computational pathology and machine learning. View documentation Please cite our paper Installation There are several ways to install Pa

254 Dec 12, 2022
FIRA: Fine-Grained Graph-Based Code Change Representation for Automated Commit Message Generation

FIRA is a learning-based commit message generation approach, which first represents code changes via fine-grained graphs and then learns to generate commit messages automatically.

Van 21 Dec 30, 2022
A new framework, collaborative cascade prediction based on graph neural networks (CCasGNN) to jointly utilize the structural characteristics, sequence features, and user profiles.

CCasGNN A new framework, collaborative cascade prediction based on graph neural networks (CCasGNN) to jointly utilize the structural characteristics,

5 Apr 29, 2022
State of the Art Neural Networks for Generative Deep Learning

pyradox-generative State of the Art Neural Networks for Generative Deep Learning Table of Contents pyradox-generative Table of Contents Installation U

Ritvik Rastogi 8 Sep 29, 2022
Prompt Tuning with Rules

PTR Code and datasets for our paper "PTR: Prompt Tuning with Rules for Text Classification" If you use the code, please cite the following paper: @art

THUNLP 118 Dec 30, 2022