Implementation of paper 'Unsupervised representation learning with Deep Convolutional Generative Adversarial Networks', Alec Radford, Luke Metz and Soumith Chintala.
This implementation is in python3 using Keras framework with Tensorflow as backend.
The model is trained on the LUSN dataset.
The original dataset is in a somewhat awkward format (lmdb) and the widely-used bedroom category is very large (43GB), and it requires a python2-only script to download it.
Therefore there is a repackaged version as a simple folder of jpgs, containing a random sample. The partial dataset with images in JPG format can be found at LSUN bedroom scene 20% sample on Kaggle and is prepared by Jeremy Howard.
When reading the images, folder arrangement of this dataset should be carefully taken care of.
dcgan_LSUN.py: This is the code for python implementation of Unsupervised representation learning with Deep Convolutional Generative Adversarial Networksimages: DCGAN generated bedrooms at every sample interval (5) when training the model for 100 epochs
The model hasn't been trained on all the epochs due to hardware constraints. So, the images folder will only contain 1 image from first epoch (trained for 1 epoch just for the sake of testing whether the code is working or not).
I will be training the model and will update the repo soon.
Clone the repository, change your present working directory to the cloned directory, Now create a now folder in this directory named images to save the generated images after every sampled interval and now train the model. Below commands accomplishes these steps.
$ git clone https://github.com/manideep2510/DCGAN_LSUN.git
$ cd DCGAN_LSUN
$ mkdir images
$ python dcgan_LSUN.py
Download the dataset from this link.
In the dcgan_LSUN.py code when we are reading the images into a numpy array, take care of the images path carefully. You will be required to change that part of code as your paths for the images would be diffirent than mine.
Architecture guidelines for stable Deep Convolutional GANs,
- Replace any pooling layers with strided convolutions (discriminator) and fractional-strided convolutions (generator).
- Use batchnorm in both the generator and the discriminator.
- Remove fully connected hidden layers for deeper architectures.
- Use ReLU activation in generator for all layers except for the output, which uses Tanh.
- Use LeakyReLU activation in the discriminator for all layers.
[1] Unsupervised representation learning with Deep Convolutional Generative Adversarial Networks', Alec Radford, Luke Metz and Soumith Chintala.
[2] Large-Scale Scene Understanding
[3] LSUN bedroom scene 20% sample
[4] Keras