CPEG 589 – Advanced Deep Learning Lecture 3

CPEG 589 ? Advanced Deep Learning Lecture 4

1

Outline

WGAN Review PyTorch Fundamentals GAN Implementation in PyTorch ? DCGAN, WGAN

2

GAN ? Optimal Value of D, V(G,D), JS Divergence

D =

+

=> D*(x) =

When pg = pdata D(x) = ?

V(G*,D*) is optimal when JS divergence between pdata(x) and pg(x) are equal => V(G,D)* = 2(||) - log 4

It accomplishes this by iteratively minimizing (implicitly) the JS divergence between pdata(x) and pg(x) during the training process

During the training, the generator may collapse to a setting where it rotates through a small set of similar outputs. This is a common failure case for GANs, commonly referred to as Mode Collapse (it is missing some of the modes of the multi-modal data)

There is no good evaluation metric to indicate how is the training progressing and whether it is progressing in the right direction.

3

The above led to the motivation for Wasserstein GAN

Python Tensorflow and PyTorch Installation

Install Anaconda (Individual Edition) for Python 3.7 for your OS, e.g., for Windows 10, the URL is:

When you install, make sure you check the PATH environment variable for Anaconda

4

Tensorflow 2.x Installation

If your computer has a GPU, install the CUDA SDK 10.1 from Nvidia site. Extract it to a folder. Then set a path to the bin folder where cudnn_xx resides

Launch command prompt and issue the following instructions from the command prompt, one at a time: conda create -n tensorflow2x pip python=3.7 activate tensorflow2x pip install --upgrade tensorflow

Other packages as needed, can be installed from the Visual Studio or from the command prompt (once you are in the activated tensorflow2x environment)

For example to install opencv, issue: pip install opencv-python

5

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download