This project generates pictures of trees using Generative Adversarial Networks (GANS). The neural networks used were built using PyTorch. The dataset of tree images used to train the discriminator network was the CIFAR100 Dataset, which was pulled into the project using TorchVision.
The pictures generated, as well as the pictures used for training, are 3-color-channel 32x32 pictures.
This project requires PyTorch and TorchVision in order to run.
pip3 install torch torchvision
Neural network parameters can be set at the top of main.py.
Below the NN parameters, the include_list can be set. This is a list
of classes that the data loader will pull from the CIFAR100 dataset.
Currently, it is only set to pine trees. But I have also included the
commented out list of all trees.
The GAN can be run by executing main.py in a Python 3 interpreter.
Image output is at /data/images/VIS_GAN/.
CIFAR100 dataset must be obtained from CIFAR and is not distributed with this repo.
At epoch 0, the Generator is merely generating random noise:
Very quickly (epoch 3), the adjust to the general structure:
Overtime (epoch 100) colors become less noisy:
Eventually (epoch 500), we start to see some variety come out of the generator.
I am impressed by how quickly the generator is able to form the general structure of the tree. Over more epochs (around 10,000), though, the results start to regress and overfit to the classifiers model of a tree. Resulting in images that are less tree-like than even earlier epochs. I suspect that this is because of the limited dataset used.