tf-3dgan

Tensorflow implementation of 3D Generative Adversarial Network.

This is a tensorflow implementation of the paper "Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling"

Requirements

• tensorflow>=1.0
• visdom>=1.0.1 (for mesh visualization)
• scipy
• scikit-image
• stl (optional)

One-line installation

```
pip install scipy scikit-image stl visdom
```

Data

• Download the training data from the 3D Shapenet website
• Extract the zip and modify the path appropriately in dataIO.py

Usage

To train the model (visdom will show generated chairs after every 200 minibatches)

python 3dgan_mit_biasfree.py 0 <path_to_model_checkpoint>

To generate chairs

python 3dgan_mit_biasfree.py 1 <path_to_trained_model>

Some sample generated chairs

Source code files

File	Description
3dgan_mit_biasfree.py	3dgan as mentioned in the paper, with same hyperparams.
3dgan.py	baseline 3dgan with fully connected layer at end of discriminator.
3dgan_mit.py	3dgan as mentioned in the paper with bias in convolutional layers.
3dgan_autoencoder.py	3dgan with support for autoencoder based pre-training.
3dgan_feature_matching.py	3dgan with additional loss of feature mathcing of last layers.
dataIO.py	data input output and plotting utilities.
utils.py	tensorflow utils like leaky_relu and batch_norm layer.

Todo

• Host the trained models
• Add argparser based interface
• Add threaded dataloader
• Release the pytorch and keras versions of the GAN.
• Train for longer number of epochs to improve quality of generated chairs.