This code add residual block and simplified the source code. If you want to visualize the training process, you need to type python -m visdom.server to opeh visdom.
Since it have residual block, you best to augment data by data_augment.py, which can get better results.
Paper: Unsupervised Representation Learning with Deep Convolutional Generative Adversial Networks
data: 5 w anime images dataset [80k/13GB]baidu pandownload password: g5qa
pip install -r requirements.txt
Training model
python main.py train --gpu=True --vis=True
Test examples:
python main.py generate --vis=False
complete parameters:
data = './data/' # the road to store data
ndf = 128 # the channel of the first convolutional layer of the discriminator net
ngf = 128 # the channel of the last convolutional layer of the generator net
batch_size = 256
img_size = 64
max_epoch = 50 # numbers of iterations
lr = 1e-4 # learning rate
beta = 0.5 # Adam optimizer beat_1
nz = 100 # the channel of noise 100 x 1 x 1
gpu = True # Use GPU
d_every = 1 # Train discriminator every 1 batch
g_every = 5 # Train generator every 5 batch
vis = True # Use visdom
plot_every = 20 # Visdom plot every 20 batch
net_path = './checkpoints'
gen_num = 512 # generate 512 images
gen_select_num = 64 # select best 64 images
gen_img = 'result.png'
gen_mean = 0
gen_std = 1