markdtw / least-squares-gan

GAN that adopts the least squares loss function for the discriminator in tensorflow

Least Squares Generative Adversarial Networks

Tensorflow implementation of Least Squares Generative Adversarial Networks by Mao et al (LSGAN).

Prerequisites

• Python 2.7+
• NumPy
• SciPy
• tqdm
• Tensorflow r1.0+
• lmdb (for processing LSUN dataset only)

Data

• LSUN Scene Classification
• CelebA

Preparation

1. Clone this repo, create ckpt/ folder:

   git clone https://github.com/markdtw/least-squares-gan.git
   cd least-squares-gan
   mkdir ckpt

2. To train on LSUN, use the provided tools to download and extract. For example:

   python download.py -c conference_room
   unzip conference_room_train_lmdb.zip
   python data.py export conference_room_train_lmdb --out_dir conference_room_train_images --flat

   I replaced .webp from this line to .jpg

3. To train on CelebA, I use this file to download. Shout out to carpedm20.

4. Now you are good to go, first time training on LSUN will center-crop all the images to 224x224 and store them in a new folder.

Train

Train on LSUN conference room with default settings:

python main.py --train

Train on CelebA with default settings:

python main.py --train --dataset=CelebA

Train from a previous checkpoint at epoch X:

python main.py --train --modelpath=ckpt/lsgan-LSUN<CelebA>-X

Check out tunable hyper-parameters:

python main.py

Some results

Epoch 10:

Epoch 25:

Epoch 45:

Results from epoch 45 is already nice and crispy.

Generator loss:

Discriminator loss:

Notes

• The model will save 40 generated pictures in log/ folder every epoch.
• Initialization is important! Default initialization with tf.xavier_initializer will lead to either D or G's gradient vanishing problem, instead I use tf.truncated_normal_initializer which is identical to DCGAN original implementation to solve the problem.
• Issues are more than welcome!

Resources

• The paper
• Highly based on This repo