Hi,

In our previous post, Get Started: DCGAN for Fashion-MNIST, you learned how to train a DCGAN to generate grayscale Fashion-MNIST images. Check out our new tutorial by Margaret: GAN Training Challenges: DCGAN for Color Images, to learn how to train a DCGAN with fashion images in color. We will also discuss the GAN training challenges and evaluation measures.
 

Image

We take the code of DCGAN for grayscale images from our previous post and make adjustments to train on color images:

  1. Data: a color image dataset Clothing & Models from Kaggle, clothing images scraped from Zalando.com.
  2. Generator: adjust the upsampling in the model architecture to generate a color image. 
  3. Discriminator: adjust the input image shape from 28×28×1 to 64×64×3. 

We will discuss the GAN training challenges of non-convergence and mode collapse. You will learn not only the theory but hands-on experiments to simulate the training failures, which helps you understand how to improve. For example, experiment with changing the model architecture, learning rates, vector noise dimension, etc. 

GAN models are difficult to evaluate. We will share how to evaluate them qualitatively and quantitatively.

This lesson is part of PyImageSearch University under GANs 101, and the direct link is here.

Click here to join PyImageSearch University

Have fun learning! 
 

Adrian Rosebrock
Chief PyImageSearcher 

P.S. PyImageSearch University helps you master computer vision, deep learning, and OpenCV with content from experts. It has 28 courses (updated regularly), 39+ hours of on-demand video, and access to centralized code repos for all 500+ tutorials on the PyImageSearch blog. In addition, you can follow each lesson easily with code in Colab notebooks in a web browser on any OS.