def __init__(self,
config,
batchsize,
n_classes=0,
softmax=True,
comm=None,
test=False):
self.softmax = softmax
if not test:
self.l_emd = config.l_emd
self.l_re = config.l_re
self.l_patch_dis = config.l_patch_dis
self.l_gp = config.l_gp
self.l_per = config.l_per
self.comm = comm
self.config = config
self.normalize_stat = self.config.normalize_stat if hasattr(
self.config, "normalize_stat") else False
if self.normalize_stat:
self.l_re = 0
self.batchsize = batchsize
self.gen = SNGAN(n_classes=n_classes,
normalize_stat=self.normalize_stat)
self.gen.initialize_params() # initialize gamma and beta
self.dim_z = 128
params = {}
if config.initial_z == "zero":
params["z"] = L.Parameter(
np.zeros((batchsize, self.dim_z)).astype("float32"))
elif config.initial_z == "random":
params["z"] = L.Parameter(
np.random.normal(size=(batchsize,
self.dim_z)).astype("float32"))
gamma = self.gen.get_gamma()
beta = self.gen.get_beta()
for i in range(len(gamma)):
if not config.not_initial_gamma:
params[f"gamma{i + 1}"] = gamma[i]
if not config.not_initial_beta:
params[f"beta{i + 1}"] = beta[i]
# get variables of parameters
self.gamma = [g.W for g in gamma]
self.beta = [b.W for b in beta]
if config.lr_g_linear > 0:
params["g_linear"] = self.gen.l1
super(AdaSNGAN, self).__init__(**params)
if not test:
self.setup_optimizer(config.init_lr)
if config.lr_g_linear > 0:
self.g_linear.W.update_rule.hyperparam.alpha = config.lr_g_linear
class AdaSNGAN(AdaGenerator):
def __init__(self, config, batchsize, comm=None, test=False):
if not test:
self.l_emd = config.l_emd
self.l_re = config.l_re
self.l_patch_dis = config.l_patch_dis
self.l_gp = config.l_gp
self.l_per = config.l_per
self.comm = comm
self.config = config
self.normalize_stat = self.config.normalize_stat if hasattr(
self.config, "normalize_stat") else False
if self.normalize_stat:
self.l_re = 0
self.batchsize = batchsize
self.gen = SNGAN(n_classes=config.n_classes,
normalize_stat=self.normalize_stat)
self.gen.initialize_params() # initialize gamma and beta
self.dim_z = 128
params = {}
if config.initial_z == "zero":
params["z"] = L.Parameter(
np.zeros((batchsize, self.dim_z)).astype("float32"))
elif config.initial_z == "random":
params["z"] = L.Parameter(
np.random.normal(size=(batchsize,
self.dim_z)).astype("float32"))
gamma = self.gen.get_gamma()
beta = self.gen.get_beta()
for i in range(len(gamma)):
if not config.not_initial_gamma:
params[f"gamma{i + 1}"] = gamma[i]
if not config.not_initial_beta:
params[f"beta{i + 1}"] = beta[i]
# get variables of parameters
self.gamma = [g.W for g in gamma]
self.beta = [b.W for b in beta]
if config.lr_g_linear > 0:
params["g_linear"] = self.gen.l1
super(AdaSNGAN, self).__init__(**params)
if not test:
self.setup_optimizer(config.init_lr)
if config.lr_g_linear > 0:
self.g_linear.W.update_rule.hyperparam.alpha = config.lr_g_linear
def forward(self, z):
h = self.gen(z.shape[0], z=z, gamma=self.gamma, beta=self.beta)
return h
def bs_reg(self):
xp = self.xp
bs_re = 0
for g in self.gamma:
bs_re += F.mean_squared_error(g, xp.ones(g.shape, dtype="float32"))
for b in self.beta:
bs_re += F.mean_squared_error(b, xp.zeros(b.shape,
dtype="float32"))
return bs_re