Tensorflow implementation of DTLC-GAN (CVPR 2018): Generative Adversarial Image Synthesis with Decision Tree Latent Controller.
-
Prerequisites
- Tensorflow 1.9
- Python 3.6
-
Training
- Important Arguments (See the others in train.py)
att: attribute to learn (default:'')ks: # of outputs of each node of each layer (default:[2, 3, 3])lambdas: loss weights of each layer (default:[1.0, 1.0, 1.0])--n_d: # of d steps in each iteration (default:1)--n_g: # of g steps in each iteration (default:1)--loss_mode: gan loss (choices:[gan, lsgan, wgan, hinge], default:gan)--gp_mode: type of gradient penalty (choices:[none, dragan, wgan-gp], default:none)--norm: normalization (choices:[batch_norm, instance_norm, layer_norm, none], default:batch_norm)--experiment_name: name for current experiment (default:default)
- Example
CUDA_VISIBLE_DEVICES=0 \ python train.py \ --att Eyeglasses \ --ks 2 3 3 \ --lambdas 1 1 1 \ --n_d 1 \ --n_g 1 \ --loss_mode hinge \ --gp_mode dragan \ --norm layer_norm \ --experiment_name att{Eyeglasses}_ks{2-3-3}_lambdas{1-1-1}_continuous_last{False}_loss{hinge}_gp{dragan}_norm{layer_norm}
- Important Arguments (See the others in train.py)
- Celeba dataset
- Images should be placed in ./data/img_align_celeba/*.jpg
- Attribute labels should be placed in ./data/list_attr_celeba.txt
- the above links might be inaccessible, the alternatives are
- img_align_celeba.zip
- list_attr_celeba.txt
- Eyeglasses, 3 layers
- Bangs, 3 layers
