def pre_train(self):
embedding_optim_cfg = OptimConfig('Adam', lr=1e-2)
self.embedding_optim = embedding_optim_cfg.get(
self.gen.embedding.parameters())
embedding_loss_cfg = LossConfig('MSE')
self.embedding_loss = embedding_loss_cfg.get()
self.gen_optim = self.goptim_cfg.get(
filter(lambda x: type(x) != SampleMatrix, self.gen.parameters()))
def pre_train(self):
embedding_optim_cfg = OptimConfig('Adam', lr=1e-3)
self.embedding_optim = embedding_optim_cfg.get(
self.gen.embedding.parameters())
embedding_loss_cfg = LossConfig('MSE')
self.embedding_loss = embedding_loss_cfg.get()
self.gen_optim = goptim_cfg.get(
filter(lambda x: type(x) != SampleMatrix, self.gen.parameters()))
logger.create_scalar('js')
logger.create_scalar('coverage')
logger.create_scalar('coverage_HQ')
logger.create_scalar('confidence')
logger.create_scalar('ratio')
LATENT_DIM = 100
IMAGE_SIZE = 64
DATASET_LEN = 20000
PRELOAD_LEN = 20000
EPOCHS = 50000
device = 'cuda'
gloss_cfg = LossConfig('BCE')
dloss_cfg = LossConfig('BCE')
sampler_cfg = SamplerConfig(
name='Onehot',
out_shape=PRELOAD_LEN,
latent_dim=LATENT_DIM,
random_sampling=False
)
goptim_cfg = OptimConfig('Adam', lr=0.0002)
doptim_cfg = OptimConfig('Adam', lr=0.0002)
dataset_cfg = DatasetConfig('CelebA', dataset_len=DATASET_LEN, image_size=IMAGE_SIZE, preload_len=PRELOAD_LEN)
loader_cfg = LoaderConfig('naive', batch_size=128, shuffle=True)
gen_cfg = ModelConfig(
'EDCG', input_size=LATENT_DIM,
hidden_size=128, output_size=3,
data_num=PRELOAD_LEN, out_64=IMAGE_SIZE == 64
)
dis_cfg = ModelConfig('DCD', input_size=3, hidden_size=64, output_size=1, out_64=IMAGE_SIZE == 64)
gan_cfg = GanConfig(
name='CelebAGAN', gen_cfg=gen_cfg, dis_cfg=dis_cfg,
gen_step=1, dis_step=1, gan_epoch=EPOCHS, loader_cfg=loader_cfg,
if __name__ == '__main__':
MODE = 10
LATENT_DIM = 100
IMAGE_SIZE = 32
DATASET_LEN = 20000
EPOCHS = 30000
device = 'cuda'
gloss_cfg = LossConfig('BCE')
dloss_cfg = LossConfig('BCE')
sampler_cfg = SamplerConfig(
name='Onehot',
out_shape=DATASET_LEN,
latent_dim=LATENT_DIM
)
goptim_cfg = OptimConfig('Adam', lr=1e-3)
doptim_cfg = OptimConfig('Adam', lr=1e-3)
dataset_cfg = DatasetConfig('MNIST', train=True, stack=False, along_width=False, size=32)
loader_cfg = LoaderConfig('naive', batch_size=128, shuffle=True)
gen_cfg = ModelConfig('EDCG', input_size=LATENT_DIM, hidden_size=128, output_size=1, data_num=DATASET_LEN)
dis_cfg = ModelConfig('DCD', input_size=1, hidden_size=128, output_size=1)
train_data_cfg = DatasetConfig('MNIST', stack=False, train=True, along_width=False)
test_data_cfg = DatasetConfig('MNIST', stack=False, train=False, along_width=False)
train_loader = LoaderConfig('naive', batch_size=128, shuffle=True)
test_loader = LoaderConfig('naive', batch_size=128, shuffle=True)
util_cfg = UtilityModelConfig('NaiveClassifier', False, '/home/bourgan/gan_dev/checkpoints/mnist_naive.pth.tar',
GAN_EPOCHS = 50000
DECREASE_LR_EPOCHS = 25000
IMG_SIZE = 256
BATCH_SIZE = 1
label_len = int(args.label_len)
EXP_NAME = 'refocus_final_' + str(label_len)
# logger = SummaryWriter('./log/' + EXP_NAME)
LR = 2e-4
gloss_cfg = LossConfig('MSE')
dloss_cfg = LossConfig('MSE')
goptim_cfg = OptimConfig('Adam', lr=LR, beta=(0.5, 0.999))
doptim_cfg = OptimConfig('Adam', lr=LR, beta=(0.5, 0.999))
dataset_cfg = DatasetConfig('FlowerFull',
size=IMG_SIZE,
num=label_len,
train=False)
loader_cfg = LoaderConfig('naive', batch_size=1, shuffle=False)
# gen_cfg = ModelConfig('ResNetGen', input_nc=3, output_nc=3, ngf=64, n_blocks=6)
gen_cfg = ModelConfig('UNetGen',
input_nc=3,
output_nc=3,
num_downs=6,
ngf=64,
use_resizeconv=True,
os.path.join(logger.img_dir,
'sampler_modes_{0:05d}.jpg'.format(postfix)))
plt.close()
if __name__ == '__main__':
dataset_len = 2000
mode = 10
latent_dim = 10
device = 'cuda'
gloss_cfg = LossConfig('BCE')
dloss_cfg = LossConfig('BCE')
sampler_loss_cfg = LossConfig('MSE')
sampler_optim_cfg = OptimConfig('Adam', lr=0.001)
sampler_cfg = SamplerConfig(
name='Learning',
out_shape=latent_dim,
data_num=dataset_len,
batch_size=2000,
epoch=50,
alpha=0,
loss_cfg=sampler_loss_cfg,
optim_cfg=sampler_optim_cfg,
)
goptim_cfg = OptimConfig('Adam', lr=1e-3)
doptim_cfg = OptimConfig('Adam', lr=1e-3)
# dataset_cfg = DatasetConfig('Spiral', mode=mode, sig=1, num_per_mode=dataset_len // mode)