Least Squares Generative Adversarial Networks

Implementation of LSGANs in Tensorflow. Official repo for the paper can be found here.

Requirements

• Python 2.7
• Tensorflow v1.0

Datasets

• LSUN

Results

The results aren't as good as the paper shows, and I'm still investigating why.

Churches from LSUN

Bedrooms from LSUN

Training

Data

I used the LSUN church and bedroom dataset. The images are resized to 112x112 (same size as the generator produces).

How to

Train

python train.py --DATA_DIR=[/path/to/images/] --DATASET=[dataset] --BATCH_SIZE=[batch_size]

For example, if you have the LSUN dataset

pytohn train.py --DATA_DIR=/mnt/lsun/church/images/ --DATASET=church

View Results

To see a fancy picture such as the one on this page, simply run

python createPhotos.py checkpoints/church/

or wherever your model is saved.

If you see the following as your "results", then you did not provide the complete path to your checkpoint, and this is from the model's initialized weights.