This is an abstract implementation of a DCGAN architecture proposed in Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. The model is developed in pytorch and trained on CelebA face dataset.
The architecture has been implemented using the following:
- Python 3.5
- Scipy
- Torchvision
- Tensorflow 1.7.0
- Tensorboard
Tensorflow and Tensorboard are used for visualization and monitoring purposes, thus they are not mandatory.
To start training the dcgan model use:
python main.py --dataPath /path/to/celebA
The logger.py file is used to create and update the model's instance for Tensorboard. To monitor the training process use:
tensorboard --logdir='./logs' --port 6006
and use your browser to access the localhost at the specified port.
Some indicative examples of face images generated by the dcgan generator are:
This work is based on the PyTorch examples. The Tensorboard support is provided from yunjey.