In this project, we investigate how to provide artistic insipiration by generating a vast variety of logo designs. While there exists a small amount of literature on this highly unstable and large problem relating to logo generation, prior work hasn't been able to demonstrate stable generative models without the use intricate unsupervised clustering of the images. We show the creation of multiple stable generative adversarial networks (GANs) that are able to produce logos by optimizing the loss function used. We experiment with vanilla DCGAN's, WGAN, WGAN-GP, and LSGAN this provides better gradients to the generator to learn during the early epochs. Our results show that WGAN-GP was the most robust model in all criteria metrics used to evaluate logo generation (see paper for full details).
-
Clone the repository
-
Download the dataset titled "LLD-icon-sharp HDF5" from https://data.vision.ee.ethz.ch/sagea/lld/. Store this HDF5 file in the "datasets" folder in your cloned repository.
-
Run
python models/train.pyto run the model.
Customize model by changing parameters for models/train.py . This includes the following:
| Parameter | Bash Script | Options |
|---|---|---|
| Epochs | '--epochs' | Any |
| Dataset | '--dataset' | 'logo', 'cifar10', 'fashion_mnist', 'mnist', 'celeba' |
| Learning Rate | '--lr' | Any |
| Loss Mode (i.e. Wasserstein, Least Squares, minimax, etc) | '--adversarial_loss_mode' | 'gan', 'lsgan', 'wgan' |
| Gradient Penalty Mode | '--gradient_penalty_mode' | 'none', '1-gp', '0-gp', 'lp' |
| Gradient Penalty Weight | '--gradient_penalty_weight' | Any |
| Experiment Name | '--experiment_name' | Any |
To train DCGAN for 50 epochs with learning rate 0.0001, run
python models/train.py --epoch=50 --experiment_name=dcgan --lr=0.0001
To train LSGAN for 50 epochs with learning rate 0.0001, run
python models/train.py --epoch=50 --adversarial_loss_mode=lsgan --experiment_name=lsgan --lr=0.0001
To train WGAN-GP with gp-mode 1 for 50 epochs with learning rate 0.0001, run
python models/train.py --epoch=50 --adversarial_loss_mode=wgan --gradient_penalty_mode=1-gp --experiment_name=wgan --lr=0.0001
Create an "outputs" folder inside of your models folder. Within the output folder, a "summaries" folder will be created that stores summaries of the models that are automatically saved (using TensorboardX) during training. Also the images generated at each time step will be stored in "samples_training" (also within outputs) and the folder is created by train.py.
In order to find inception score, train models using CIFAR-10 dataset (use '----dataset="cifar10' argument in command line) and then using the following repository: https://github.com/sbarratt/inception-score-pytorch. The models were too large to store in this repository without being corrupted.
- train.py: source code to train generator and discriminatorr and store samples of the generated images
- module.py: generator and discriminator classes
- data.py: preprocessing datasets for train.py
- imlib/pylib/torchlib/torchprob: various util files/folders
See requirements.txt
- Adit Gupta
- Max Yuan
- Chae Young Lee
- Darren Liang
This project was completed as part of the CPSC452 (Deep Learning Theory) final project with Prof. Krishnaswamy.