Official repository for ABC-GAN

Overview

ABC-GAN

The work represented in this repository is the result of a 14 week semesterthesis on photo-realistic image generation using generative adversarial networks at ETH zurich.

Additional Experiments

There is a second branch called loose_encoder in which you will find another experiment conducted during the thesis. Unfortunately we didn't had enough time to make an in-depth analysis of the results. The loose encoder can be added to any other experiment with a simple flag and in specific datasets such as CelebA results in astonishing improvement of generated image quality. To the best of our knowledge there is no GAN out there resulting in similar realistic images on celebA with a 64 by 64 output resolution. (Check the other branch for some samples)

Prerequisites

The code has been built and tested using the following packages (and versions)

  • Pillow (4.0.0)
  • scikit-learn (0.18.1)
  • scipy (0.18.1)
  • numpy (1.12.1)
  • tensorflow-gpu (1.1.0rc1)

Usage

Make sure you have a running tensorflow setup.

We added some special flags to have a better overview of the different experiments. One thing we added is a folder_suffix which will be appended to all checkpoint, log and samples folders.

Here are some examples to train the ABC-GAN with datasets:

  • python main.py --dataset celebA --folder_suffix=_abcgan_1_GpD_o64 --input_height=128 --output_height=64 --GpD_ratio=1 --blur_strategy=3x3 --epoch=8 --train --crop True
  • python main.py --dataset lsun --folder_suffix=_abcgan_lsun_controller_blur_3x3_o128 --input_fname_pattern=*.webp --input_height=256 --output_height=128 --GpD_ratio=-1 --blur_strategy=3x3 --epoch=20 --batch-size=64 --train --crop True
  • python main.py --dataset cifar10 --folder_suffix=_abcgan_cifar_3GpD_regressive_hyperbolic_o32 --input_height=32 --input_fname_pattern=*.png --output_height=32 --blur_strategy=reg_hyp --epoch=100 --train --crop True

Datasets

The following datasets have been used:

  • CelebA (> 200'000 celebrity faces, 178 px lower side)
  • Cifar10 (60'000 pictures of 10 categories, 32x32 px)
  • LSUN (> 3 Mio images of bedrooms 256 px lower side)
  • ImageNet subset (ImageNet subset, 64 by 64 pixels)

The easiest way to include the datasets is by having all images in one folder. Using such a dataset can be done by just changing the input_fname_pattern to the correct file ending and specifying the folder name with dataset. (The folder with the dataset has to be in the subfolder data)

Folder structure used for our experiments:

  • ABC-GAN
    • data
      • celebA
      • cifar10
      • lsun
      • train_64x64
    • download.py
    • LICENSE
    • main.py
    • model.py
    • ops.py
    • utils.py
    • README.md
    • report

train_64x64 is referring to the ImageNet subset. We used the same one as used in Improved Wasserstein GAN

Special Case LSUN

Since the LSUN dataset comes in a hierarchical structure with many files it makes sense to just use a reference file with the respective paths to the files. The best way to do that is:

  1. Inside the abc-gan/data folder create a subfolder lsun
  2. Extract the downloaded lsun dataset here (we used bedroom_train_lmdb.zip)
  3. Make a list of all files appearing in the extracted lsun_train folder
  4. Name this file lsun_images

The lsun_images file should have a structure such as:

lsun_train/f/f/f/f/f/b/fffffbb9225d069b7f47e464bdd75e6eff82b61c.webp
lsun_train/f/f/f/f/f/6/fffff6cd254f0ead6191f3003519f6805e1e6619.webp
lsun_train/f/f/f/f/f/5/fffff548f9109fc3be2d71088f8e202ea78ac620.webp
lsun_train/f/f/f/f/f/a/fffffa900959150cb53ac851b355ec4adbc22e4e.webp
lsun_train/f/f/f/f/8/0/ffff80a1dc7e7d790ccd46f2fdd4dcfca929d2c3.webp
...

In order to use LSUN just again change the input_fname_pattern and switch the dataset to lsun. We hard coded the special case of lsun such that we will use the reference file to get the paths to the images.

Results

Adaptive Controller

One of the most important points during training of GANs is balancing the discriminator against the generator. If one of the two dominates the other a mode collapse can occur. Many people started playing around with the ratio between discriminator and generator. And others used thresholds to determine if one has to train the discriminator or the generator.

In our work, we implemented a simple controller to get rid of this manual tuning. The output of the controller is a probability of either training the discriminator or the generator for one iteration. The controller gives you the following benefits:

  • Reduced training time (up to a factor of 5)
  • Reuse the same network for different datasets (The controller automatically adapts to other datasets so you don't have to tune the ratio between D and G anymore)
  • In some cases, the controller also improves stability during training

Controller

Controller architecture:

Note: The controller input is calculated using the two losses of the discriminator (loss for real and for fake images).

controller

The controller tries to keep the avg. value always at a reference point. The output of the controller is a probability of training either the discriminator or the generator.

Training behaviour

Training curve showing discriminator vs generator training iterations

Note: Without a controller and a fixed ratio between discriminator and generator updates we would see two straight lines

training curve G against D Through the controller, the training of the networks adapts itself. In the beginning, the generator is getting trained more often but after around 5k steps the discriminator takes over. As known from GAN theory we actually want the discriminator to dominate the generator. And without a controller, this is very hard to achieve without changing the loss function (Wasserstein Loss, Cramer loss etc.)

Convergence speed comparison of DCGAN with different GpD ratios and our controller

Note: One iteration is either training once the discriminator or the generator

convergence comparison with and without controller Comparison of convergence speed. GpD: Generator per Discriminator training iterations. (e.g. 3 GpD means we train the generator 3 times per discriminator)

Adaptive Blur

GANs still have trouble with stability, image quality and output resolution. We implemented an adaptive blur filter to assist the discriminator and therefore improve overall results. We figured out that the discriminator has issues with details in images. To overcome this issue we just blur all images before they reach the discriminator. So in the end, the discriminator either sees blurred images from the generator or blurred images of the dataset. Using a fixed blur such as a 3 by 3 Gaussian kernel as we used in our experiments, has the side effect of additional noise in the output image. Since the generator has not to care about the details (since his output will be blurred anyway) he can add noise. To mitigate this effect, we added an adaptive blur which changes over training time. In the beginning, we have a strong blur such that the GAN can focus on improving the base structure of the output. In the end of the training, we have almost no blur, such that the GAN can now focus on the details.

The blur gives you the following benefits:

  • Improved stability during training. (We encountered several times the case that without blur the network was not able to converge at all)
  • Improved image quality. (Despite the noise the output images look much more realistic)
  • You can increase the resolution (We were able to use DCGAN for generating images at 256 by 256 pixels using CelebA or LSUN)

DCGAN vs DCGAN with Blur (DCGAN+B)

Image showing plain DCGAN without and with blur (DCGAN+B)

DCGAN without and with blur The resulting images look much better with blur than without. They have more details but also noise.

Comparison of different blurring strategies

We compare two regressive blur kernel (e.g. make sigma of a Gaussian blur smaller during training)

Comparison of blurring strategies The hyperbolic decreasing Gaussian blur is best when it comes to reducing the noise in the images.

ABC-GAN

Combination of the Adaptive Blur and Controller GAN.

We conducted different experiments using various datasets such as LSUN, CIFAR10 and CelebA. Some of the resulting images have been downscaled by a factor of two in order to reduce noise. (Since on some screens and also printed the noise looks very annoying.)

Note: In some samples, you will see a green bar at the top. The bar is the actual input value of the controller on a per image basis. If the reference value is for example 0.25 (like in our code) this means that on average over one batch we want the green bar to be around 25%.

LSUN bedrooms dataset

ABC-GAN with a fixed Gaussian Blur kernel of 3x3 and output resolution of 256 by 256 pixels

Randomly sampled batch. Downscaled to 128 by 128 pixels to reduce noise.

lsun using ABC-GAN with fixed kernel and downscaled to 128 by 128 pixels We seem to reach the limit of DCGAN with this experiment. The same experiment without Blur failed. The images look not very realistic but one still sees that some of them look close to bedrooms.

CIFAR10

Comparison of results gathered using CIFAR10

CIFAR10 comparison our experiments

Comparison of our best results with other works

CIFAR10 comparison our work and others

Publications

Acknowledgement

  • Thanks, Prof. Luc Van Gool for the semester thesis at the Computer Vision Lab D-ITET at ETH Zurich
  • Thanks for supervising this thesis Eirikur Agustsson and Radu Timofte
  • This work has been based on the DCGAN implementation found on GitHub

Author

Igor Susmelj / @igorsusmelj

Owner
IgorSusmelj
Co-founder at Lightly Degree from ETH Zurich with a focus on embedded computing and machine learning.
IgorSusmelj
Implementations of the algorithms in the paper Approximative Algorithms for Multi-Marginal Optimal Transport and Free-Support Wasserstein Barycenters

Implementations of the algorithms in the paper Approximative Algorithms for Multi-Marginal Optimal Transport and Free-Support Wasserstein Barycenters

Johannes von Lindheim 3 Oct 29, 2022
Indonesian Car License Plate Character Recognition using Tensorflow, Keras and OpenCV.

Monopol Indonesian Car License Plate (Indonesia Mobil Nomor Polisi) Character Recognition using Tensorflow, Keras and OpenCV. Background This applicat

Jayaku Briliantio 3 Apr 07, 2022
Official code for the CVPR 2021 paper "How Well Do Self-Supervised Models Transfer?"

How Well Do Self-Supervised Models Transfer? This repository hosts the code for the experiments in the CVPR 2021 paper How Well Do Self-Supervised Mod

Linus Ericsson 157 Dec 16, 2022
Real-time Joint Semantic Reasoning for Autonomous Driving

MultiNet MultiNet is able to jointly perform road segmentation, car detection and street classification. The model achieves real-time speed and state-

Marvin Teichmann 518 Dec 12, 2022
Categorizing comments on YouTube into different categories.

Youtube Comments Categorization This repo is for categorizing comments on a youtube video into different categories. negative (grievances, complaints,

Rhitik 5 Nov 26, 2022
Tgbox-bench - Simple TGBOX upload speed benchmark

TGBOX Benchmark This script will benchmark upload speed to TGBOX storage. Build

Non 1 Jan 09, 2022
Code for "Unsupervised State Representation Learning in Atari"

Unsupervised State Representation Learning in Atari Ankesh Anand*, Evan Racah*, Sherjil Ozair*, Yoshua Bengio, Marc-Alexandre Côté, R Devon Hjelm This

Mila 217 Jan 03, 2023
The openspoor package is intended to allow easy transformation between different geographical and topological systems commonly used in Dutch Railway

Openspoor The openspoor package is intended to allow easy transformation between different geographical and topological systems commonly used in Dutch

7 Aug 22, 2022
Codes for NeurIPS 2021 paper "On the Equivalence between Neural Network and Support Vector Machine".

On the Equivalence between Neural Network and Support Vector Machine Codes for NeurIPS 2021 paper "On the Equivalence between Neural Network and Suppo

Leslie 8 Oct 25, 2022
Reference models and tools for Cloud TPUs.

Cloud TPUs This repository is a collection of reference models and tools used with Cloud TPUs. The fastest way to get started training a model on a Cl

5k Jan 05, 2023
PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)

About PyTorch 1.2.0 Now the master branch supports PyTorch 1.2.0 by default. Due to the serious version problem (especially torch.utils.data.dataloade

Sanghyun Son 2.1k Jan 01, 2023
Pytorch Implementation of rpautrat/SuperPoint

SuperPoint-Pytorch (A Pure Pytorch Implementation) SuperPoint: Self-Supervised Interest Point Detection and Description Thanks This work is based on:

76 Dec 27, 2022
Weakly Supervised Dense Event Captioning in Videos, i.e. generating multiple sentence descriptions for a video in a weakly-supervised manner.

WSDEC This is the official repo for our NeurIPS paper Weakly Supervised Dense Event Captioning in Videos. Description Repo directories ./: global conf

Melon(Xuguang Duan) 96 Nov 01, 2022
CoReNet is a technique for joint multi-object 3D reconstruction from a single RGB image.

CoReNet CoReNet is a technique for joint multi-object 3D reconstruction from a single RGB image. It produces coherent reconstructions, where all objec

Google Research 80 Dec 25, 2022
Code for the paper "On the Power of Edge Independent Graph Models"

Edge Independent Graph Models Code for the paper: "On the Power of Edge Independent Graph Models" Sudhanshu Chanpuriya, Cameron Musco, Konstantinos So

Konstantinos Sotiropoulos 0 Oct 26, 2021
This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR)

CEDR This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR) introduced in the following paper

phoenix 3 Feb 27, 2022
Everything you need to know about NumPy( Creating Arrays, Indexing, Math,Statistics,Reshaping).

Everything you need to know about NumPy( Creating Arrays, Indexing, Math,Statistics,Reshaping).

1 Feb 14, 2022
Python/Rust implementations and notes from Proofs Arguments and Zero Knowledge

What is this? This is where I'll be collecting resources related to the Study Group on Dr. Justin Thaler's Proofs Arguments And Zero Knowledge Book. T

Thor 66 Jan 04, 2023
Automatic tool focused on deriving metallicities of open clusters

metalcode Automatic tool focused on deriving metallicities of open clusters. Based on the method described in Pöhnl & Paunzen (2010, https://ui.adsabs

2 Dec 13, 2021
This program creates a formatted excel file which highlights the undervalued stock according to Graham's number.

Over-and-Undervalued-Stocks Of Nepse Using Graham's Number Scrap the latest data using different websites and creates a formatted excel file that high

6 May 03, 2022