Moving to Larger Datasets

Generating digits is fun. We can have way more fun generating other stuff, such as human faces and bedroom photos. To generate good complex images like these, we need more training samples than the 60,000 samples that MNIST offers. In this section, we will download two much larger datasets (CelebA and LSUN) and train the DCGAN on them to get more complex generated samples.

Generating human faces from the CelebA dataset

The CelebFaces Attributes (CelebA) dataset is a large-scale face attributes dataset with more than 200,000 celebrity images, each with 40 attribute annotations. We need to download the cropped and aligned images. We won’t need any attribute annotation here, so we only need to download the file named img_align_celeba.zip, which is no more than 2 GB in size.