Mars-Machina
A deep convolutional variational autoencoder trained on digital terrain maps of Mars from HiRise/MRO. 3D surfaces can be procedurally generated from the latent space
Dependencies
- Unity 3D
- Unity ML Agents
- https://github.com/llSourcell/Unity_ML_Agents/blob/master/docs/Using-TensorFlow-Sharp-in-Unity-(Experimental).md
- Python 3.6+
- Keras, Tensorflow==1.4, Matplotlib, Numpy, PIL, Scikit-learn
Python
to get started with your own DCVAE follow the steps below
Download a Digital terrain map from HiRise
Save the DTM image to the directory: Python/hirise/ as a png file. Make sure to trim the unnecessary regions in GIMP or photoshop before training. See the current file for an example
Train a quick model:
python autoencoder.py --lose mse --epochs 1000 --name hirise
a tensorflow graph will be saved to: Python/out/frozen_hirise.bytes
Create a game object in unity and give it some mesh components, then attach the meshGenerator.cs script
For getting started with Tensorflow in Unity see: https://github.com/pearsonkyle/Unity-Variational-Autoencoder
An example of the latent space generations after 1000 epochs of training:
