Dirac-GAN

This repository implements (PyTorch) the Dirac-GAN proposed in the paper "Which Training Methods for GANs do actually Converge?" by Mescheder et al. [1]. The original implementation of the authors can be found here.

This work was done as part of the lecture Deep Generative Models at TU Darmstadt held by Dr. Anirban Mukhopadhyay.

Parts of this implementation are taken from my recent mode collapse example repository.

Standard GAN loss	Non-saturating GAN loss	Wasserstein GAN

Wasserstein GAN loss + GP	Least squares GAN	Hinge GAN

DRAGAN loss

This repository implements the following GAN losses and regularizers.

Method	Generator loss	Discriminator loss
Original GAN loss	$\mathcal{L}^{\text{GAN}}_{D}=-\mathbb{E}_{x\sim p_{d}}[\log(D(x))] - \mathbb{E}_{\hat{x}\sim p_{g}}[\log(1 - D(\hat{x}))]$	$\mathcal{L}_{G}^{\text{GAN}}=\mathbb{E}_{\hat{x}\sim p_{g}}[\log(1 - D(\hat{x}))]$
Non-saturating GAN loss	$\mathcal{L}_{D}^{\text{NSGAN}}=-\mathbb{E}_{x\sim p_{d}}[\log(D(x))] - \mathbb{E}_{\hat{x}\sim p_{g}}[\log(1 - D(\hat{x}))]$	$\mathcal{L}_{G}^{\text{NSGAN}}=-\mathbb{E}_{\hat{x}\sim p_{g}}[\log(D(\hat{x}))]$
Wasserstein GAN loss	$\mathcal{L}_{D}^{\text{WGAN}}=-\mathbb{E}_{x\sim p_{d}}[D(x)] + \mathbb{E}_{\hat{x}\sim p_{g}}[D(\hat{x})]$	$\mathcal{L}_{G}^{\text{WGAN}}=-\mathbb{E}_{\hat{x}\sim p_{g}}[D(\hat{x})]$
Wasserstein GAN loss + grad. pen.	$\mathcal{L}_{D}^{\text{WGANGP}}=\mathcal{L}_{D}^{\text{WGAN}} + \lambda\mathbb{E}_{\hat{x}\sim p_{g}}[(\lvert\lvert\nabla D(\alpha x + (1 - \alpha \hat{x}))\rvert\rvert_{2} - 1)^2]$	$\mathcal{L}_{G}^{\text{WGANGP}}=\mathcal{L}_{G}^{\text{WGAN}}$
Least squares GAN loss	$\mathcal{L}^{\text{LSGAN}}_{D}=-\mathbb{E}_{x\sim p_{d}}[(D(x) - 1)^2] + \mathbb{E}_{\hat{x}\sim p_{g}}[D(\hat{x})^2]$	$\mathcal{L}^{\text{LSGAN}}_{G}=-\mathbb{E}_{\hat{x}\sim p_{g}}[(D(\hat{x} - 1))^2]$
Hinge GAN	$\mathcal{L}^{\text{LSGAN}}_{D}=-\mathbb{E}_{x\sim p_{d}}[\min(0, D(x)-1] - \mathbb{E}_{\hat{x}\sim p_{g}}[\min(0, -D(\hat{x})-1)]$	$\mathcal{L}^{\text{LSGAN}}_{G}=\mathcal{L}^{\text{WGAN}}_{G}$
DRAGAN	$\mathcal{L}_{D}^{\text{DRAGAN}}=\mathcal{L}_{D}^{\text{GAN}} + \lambda\mathbb{E}_{\hat{x}\sim p_{d} + \mathcal{N}(0, c)}[(\lvert\lvert\nabla D(\hat{x})\rvert\rvert_{2} - 1)^2]$	$\mathcal{L}_{G}^{\text{DRAGAN}}=\mathcal{L}_{G}^{\text{GAN}}$

Method	Generator loss
$R_{1}$ regularization	$R_{1}=\frac{\gamma}{2}\mathbb{E}_{x\sim p_{d}}[\lvert\lvert\nabla D(x)\rvert\rvert^2]$
$R_{2}$ regularization	$R_{2}=\frac{\gamma}{2}\mathbb{E}_{\hat{x}\sim p_{g}}[\lvert\lvert\nabla D(x)\rvert\rvert^2]$
$R_{\text{LC}}$ regularization	$R_{\text{LC}}=\mathbb{E}_{x\sim p_{d}}[\lvert\lvert D(x) - \alpha_{F}\rvert\rvert^{2}] + \mathbb{E}_{\hat{x}\sim p_{g}}[\lvert\lvert D(G(\hat{x})) - \alpha_{R}\rvert\rvert^{2}]$

Dependencies

Dirac-GAN is written in PyTorch 1.8.1. No GPU is required! All additional dependencies can be seen in the requirements.txt file. To install all dependencies simply run:

pip install -r requirements.txt

Older version of PyTorch may also allows running the code without issues.

Usage

The implementation provides a simple GUI to run all Dirac-GAN experiments with different settings. Simply run:

python main.py

Set the desired parameters in the GUI and click on "Run training" to perform training. This could take a few seconds. If the training is finished all results are plotted and shown.

References

[1] @inproceedings{Mescheder2018,
    title={Which training methods for GANs do actually converge?},
    author={Mescheder, Lars and Geiger, Andreas and Nowozin, Sebastian},
    booktitle={International conference on machine learning},
    pages={3481--3490},
    year={2018},
    organization={PMLR}
}

R1 regularization

Thanks for creating this. Your code helped me to understand and implement R1 regularization.

Here is a patch to allow the main window to be resized with the widgets being laid out automatically:

--- a/main.py
+++ b/main.py
@@ -2,7 +2,7 @@ import sys
 import matplotlib
 
 matplotlib.use('Qt5Agg')
-from PyQt5.QtWidgets import QApplication, QMainWindow, QPushButton, QComboBox, QCheckBox, QSizePolicy
+from PyQt5.QtWidgets import QApplication, QMainWindow, QPushButton, QComboBox, QCheckBox, QSizePolicy, QWidget, QVBoxLayout, QHBoxLayout, QSpacerItem
 from matplotlib.backends.backend_qt5agg import FigureCanvasQTAgg as FigureCanvas
 from matplotlib.figure import Figure
 import torch
@@ -103,7 +103,6 @@ class App(QMainWindow):
         # Make list box to choose regularization
         self.instance_noise_box = QCheckBox("Instance noise", self)
         self.instance_noise_box.resize(300, 50)
-        self.instance_noise_box.move(self.width - 300, self.height - 200)
         # Make list box to choose regularization
         self.regularization_box = QComboBox(self)
         self.regularization_box.addItem("None")
@@ -111,7 +110,6 @@ class App(QMainWindow):
         self.regularization_box.addItem("RLC")
         self.regularization_box.addItem("R2 gradient penalty")
         self.regularization_box.resize(300, 50)
-        self.regularization_box.move(self.width - 300, self.height - 150)
         # Make list box to choose GAN loss
         self.gan_list_box = QComboBox(self)
         self.gan_list_box.addItem("Standard GAN")
@@ -122,13 +120,26 @@ class App(QMainWindow):
         self.gan_list_box.addItem("DRAGAN")
         self.gan_list_box.addItem("Hinge GAN")
         self.gan_list_box.resize(300, 50)
-        self.gan_list_box.move(self.width - 300, self.height - 100)
         # Make run button
         run_button = QPushButton("button", self)
         run_button.setText("Run training")
         run_button.resize(300, 50)
-        run_button.move(self.width - 300, self.height - 50)
         run_button.clicked.connect(self.__train)
+
+        vlayout = QVBoxLayout()
+        vlayout.addItem(QSpacerItem(0, 0, QSizePolicy.Fixed, QSizePolicy.Expanding));
+        vlayout.addWidget(self.instance_noise_box)
+        vlayout.addWidget(self.regularization_box)
+        vlayout.addWidget(self.gan_list_box)
+        vlayout.addWidget(run_button)
+
+        hlayout = QHBoxLayout()
+        hlayout.addWidget(self.dynamic_plot)
+        hlayout.addLayout(vlayout)
+
+        widget = QWidget()
+        widget.setLayout(hlayout)
+        self.setCentralWidget(widget)
         # Show everything
         self.show()

christophreich1996 / dirac-gan Goto Github PK

dirac-gan's Introduction

Dirac-GAN

Dependencies

Usage

References

dirac-gan's People

Contributors

Stargazers

Watchers

Forkers

dirac-gan's Issues

R1 regularization

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent