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,
dataset_cfg=dataset_cfg, gloss_cfg=gloss_cfg, dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg, doptim_cfg=doptim_cfg, label_smooth=False,
sampler_cfg=sampler_cfg, dist_loss=False, device=device
)
gan_schema = GanSchema()
gan_desc = gan_schema.dump(gan_cfg)
logger.save_cfg(gan_desc)
gan = gan_cfg.get()
gan.train(use_tqdm=True)
gan.save('checkpoints', 'CelebA64')
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',
15, train_loader, test_loader, train_data_cfg, test_data_cfg, 32, 'cuda')
gan_cfg = GanConfig(
name='StackMNISTLPG', gen_cfg=gen_cfg, dis_cfg=dis_cfg,
gen_step=1, dis_step=1, gan_epoch=EPOCHS, loader_cfg=loader_cfg,
dataset_cfg=dataset_cfg, gloss_cfg=gloss_cfg, dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg, doptim_cfg=doptim_cfg, label_smooth=False,
sampler_cfg=sampler_cfg, dist_loss=False, device=device, util_cfg=util_cfg
)
gan_schema = GanSchema()
gan_desc = gan_schema.dump(gan_cfg)
logger.save_cfg(gan_desc)
gan = gan_cfg.get()
gan.train(use_tqdm=True)
gan.save('checkpoints', 'mnist')
loader_cfg = LoaderConfig('naive', batch_size=128, shuffle=True)
gen_cfg = ModelConfig('MLG',
input_size=latent_dim,
hidden_size=32,
output_size=2)
dis_cfg = ModelConfig('MLD', input_size=2, hidden_size=32, output_size=1)
gan_cfg = GanConfig(name='LearningPrior2',
gen_cfg=gen_cfg,
dis_cfg=dis_cfg,
gen_step=1,
dis_step=1,
gan_epoch=20000,
loader_cfg=loader_cfg,
dataset_cfg=dataset_cfg,
gloss_cfg=gloss_cfg,
dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg,
doptim_cfg=doptim_cfg,
label_smooth=False,
sampler_cfg=sampler_cfg,
dist_loss=False,
device=device)
gan_schema = GanSchema()
gan_desc = gan_schema.dump(gan_cfg)
logger.save_cfg(gan_desc)
gan = gan_cfg.get()
# th.nn.init.normal_(gan.gen.sample_matrix.weight)
gan.train(use_tqdm=True)
input_nc=3,
output_nc=3,
num_downs=6,
ngf=64,
use_resizeconv=True,
ex_label=True,
label_len=label_len)
dis_cfg = ModelConfig('PatchDis', input_nc=6)
gan_cfg = GanConfig(name='Pix2Pix',
gen_cfg=gen_cfg,
dis_cfg=dis_cfg,
gen_step=1,
dis_step=1,
gan_epoch=GAN_EPOCHS,
loader_cfg=loader_cfg,
dataset_cfg=dataset_cfg,
gloss_cfg=gloss_cfg,
dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg,
doptim_cfg=doptim_cfg,
device=device)
gan = gan_cfg.get()
gan.load('./p2p/' + EXP_NAME)
SAVE_DIR = './eval_result/' + EXP_NAME
os.makedirs(SAVE_DIR, exist_ok=True)
dirs = gan.eval_on_dataset(SAVE_DIR, label_len=label_len)
dirs = [['/'.join(x.split('/')[-2:]) for x in d] for d in dirs]
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)
gan_cfg = GanConfig(name='MNIST-WGAN-GP',
gen_cfg=gen_cfg,
dis_cfg=dis_cfg,
gen_step=1,
dis_step=1,
gan_epoch=EPOCHS,
loader_cfg=loader_cfg,
dataset_cfg=dataset_cfg,
gloss_cfg=gloss_cfg,
dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg,
doptim_cfg=doptim_cfg,
label_smooth=False,
sampler_cfg=sampler_cfg,
dist_loss=False,
device=device,
LAMBDA=0.1)
gan_schema = GanSchema()
gan_desc = gan_schema.dump(gan_cfg)
logger.save_cfg(gan_desc)
gan = gan_cfg.get()
gan.train(use_tqdm=True)
gan.save('checkpoints', 'mnist')
sampler_cfg = SamplerConfig(
name='Gaussian',
out_shape=latent_dim,
)
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)
# dataset_cfg = DatasetConfig('GMM', mode=mode, sig=1, num_per_mode=dataset_len // mode)
loader_cfg = LoaderConfig('naive', batch_size=128, shuffle=True)
gen_cfg = ModelConfig('MLG', input_size=latent_dim, hidden_size=32, output_size=2)
dis_cfg = ModelConfig('MLD', input_size=2, hidden_size=32, output_size=1)
gan_cfg = GanConfig(
name='Vanilla', gen_cfg=gen_cfg, dis_cfg=dis_cfg,
gen_step=1, dis_step=1, gan_epoch=15000, loader_cfg=loader_cfg,
dataset_cfg=dataset_cfg, gloss_cfg=gloss_cfg, dloss_cfg=dloss_cfg,
goptim_cfg=goptim_cfg, doptim_cfg=doptim_cfg, label_smooth=False,
sampler_cfg=sampler_cfg, device=device
)
gan_schema = GanSchema()
gan_desc = gan_schema.dump(gan_cfg)
logger.save_cfg(gan_desc)
gan = gan_cfg.get()
gan.train(use_tqdm=True)