data_root_dir
    train
        1.jpg
        2.jpg
        ......
    val
        1.jpg
        2.jpg
        ......

CUDA_VISIBLE_DEVICES: specify which GPUs to use, for example "0,1"
epoch: number of training iterations
batch_size: sample batch size of one training step g_init_lr: initial learning rate of generator g_final_lr: final learning rate of generator
d_init_lr: initial learning rate of discriminator
d_final_lr: final learning rate of discriminator
noise_dim: dimension of noise, if use_cnn is True, shape of noise will be [noise_dim, 1, 1], if use_cnn is False, shape of noise will be [noise_dim]
d_train_times: train the generator every d_train_times times the discriminator is trained
data_root_dir: data root directory
num_workers: num_workers of pytorch data loader
img_size: size of image generated from generator, if use_cnn is True, img_size will be 2 ** len(generator_features), len(generator_features) is length of generator_features
save_img_total_step: perform image generation every save_img_total_step steps of training and save the image
img_count: the number of images generated each time
discriminator_weight_min: min value of weight of discriminator discriminator_weight_max: max value of weight of discriminator
use_cnn: if specified as True, the convolution neural network is used, otherwise the MLP is used
generator_features: channels of every layer of generator if use_cnn is True, number of neurons of every layer if use_cnn is False
img_save_dir: storage directory for generated images

python train.py