def init_network(self, opt):
Base_Model.init_network(self, opt)
self.if_pool = opt.if_pool
self.multi_view = opt.multi_view
self.conditional_D = opt.conditional_D
assert len(self.multi_view) > 0
self.loss_names = ['D', 'G']
self.metrics_names = ['Mse', 'CosineSimilarity', 'PSNR']
self.visual_names = ['G_real', 'G_fake', 'G_input', 'G_Map_fake_F', 'G_Map_real_F', 'G_Map_fake_S', 'G_Map_real_S']
# identity loss
if self.opt.idt_lambda > 0:
self.loss_names += ['idt']
# feature metric loss
if self.opt.fea_m_lambda > 0:
self.loss_names += ['fea_m']
# map loss
if self.opt.map_m_lambda > 0:
self.loss_names += ['map_m']
if self.training:
self.model_names = ['G', 'D']
else: # during test time, only load Gs
self.model_names = ['G']
self.netG = factory.define_3DG(opt.noise_len, opt.input_shape, opt.output_shape,
opt.input_nc_G, opt.output_nc_G, opt.ngf, opt.which_model_netG,
opt.n_downsampling, opt.norm_G, not opt.no_dropout,
opt.init_type, self.gpu_ids, opt.n_blocks,
opt.encoder_input_shape, opt.encoder_input_nc, opt.encoder_norm,
opt.encoder_blocks, opt.skip_number, opt.activation_type, opt=opt)
if self.training:
if opt.ganloss == 'gan':
use_sigmoid = True
elif opt.ganloss == 'lsgan':
use_sigmoid = False
else:
raise ValueError()
# conditional Discriminator
if self.conditional_D:
d_input_channels = opt.input_nc_D + 1
else:
d_input_channels = opt.input_nc_D
self.netD = factory.define_D(d_input_channels, opt.ndf,
opt.which_model_netD,
opt.n_layers_D, opt.norm_D,
use_sigmoid, opt.init_type, self.gpu_ids,
opt.discriminator_feature, num_D=opt.num_D, n_out_channels=opt.n_out_ChannelsD)
if self.if_pool:
self.fake_pool = ImagePool(opt.pool_size)
def init_network(self, opt):
Base_Model.init_network(self, opt)
self.loss_names = ['D', 'G']
self.metrics_names = ['Mse', 'CosineSimilarity']
self.visual_names = [
'G_real', 'G_fake', 'G_input', 'G_map_fake', 'G_map_real'
]
# identity loss
if self.opt.idt_lambda > 0:
self.loss_names += ['idt']
# feature metric loss
if self.opt.fea_m_lambda > 0:
self.loss_names += ['fea_m']
# map loss
if self.opt.map_m_lambda > 0:
self.loss_names += ['map_m']
if self.training:
self.model_names = ['G', 'D']
else: # during test time, only load Gs
self.model_names = ['G']
self.netG = factory.define_3DG(
opt.noise_len, opt.input_shape, opt.output_shape, opt.input_nc_G,
opt.output_nc_G, opt.ngf, opt.which_model_netG, opt.n_downsampling,
opt.norm_G, not opt.no_dropout, opt.init_type, self.gpu_ids,
opt.n_blocks, opt.encoder_input_shape, opt.encoder_input_nc,
opt.encoder_norm)
if self.training:
if opt.ganloss == 'gan':
use_sigmoid = True
elif opt.ganloss == 'lsgan':
use_sigmoid = False
elif opt.ganloss == 'wgan':
self.loss_names += ['wasserstein']
use_sigmoid = False
elif opt.ganloss == 'wgan_gp':
self.loss_names += ['wasserstein', 'grad_penalty']
use_sigmoid = False
else:
raise ValueError()
self.netD = factory.define_D(opt.input_nc_D, opt.ndf,
opt.which_model_netD, opt.n_layers_D,
opt.norm_D, use_sigmoid,
opt.init_type, self.gpu_ids,
opt.discriminator_feature)
self.fake_pool = ImagePool(opt.pool_size)
def init_network(self, opt):
Base_Model.init_network(self, opt)
self.if_pool = opt.if_pool
self.multi_view = opt.multi_view
assert len(self.multi_view) > 0
self.metrics_names = ['Mse', 'CosineSimilarity', 'PSNR']
self.visual_names = [
'G_real', 'G_fake', 'G_input1', 'G_input2', 'G_Map_fake_F',
'G_Map_real_F', 'G_Map_fake_S', 'G_Map_real_S'
]
self.netG = factory.define_3DG(opt.noise_len,
opt.input_shape,
opt.output_shape,
opt.input_nc_G,
opt.output_nc_G,
opt.ngf,
opt.which_model_netG,
opt.n_downsampling,
opt.norm_G,
not opt.no_dropout,
opt.init_type,
self.gpu_ids,
opt.n_blocks,
opt.encoder_input_shape,
opt.encoder_input_nc,
opt.encoder_norm,
opt.encoder_blocks,
opt.skip_number,
opt.activation_type,
opt=opt)
self.loss_names = ['idt']
self.model_names = ['G']
# map loss
if self.opt.map_projection_lambda > 0:
self.loss_names += ['map_m']
def init_network(self, opt):
Base_Model.init_network(self, opt)
self.if_pool = opt.if_pool
self.multi_view = opt.multi_view
self.conditional_D = opt.conditional_D
self.order_map_list = [(0, 1, 2, 3, 4), (0, 1, 3, 2, 4),
(0, 1, 4, 2, 3)]
assert len(self.multi_view) >= 0
self.loss_names = ['D', 'G']
self.metrics_names = ['Mse', 'CosineSimilarity', 'PSNR']
self.visual_names = [
'G_real', 'G_fake', 'G_input', 'G_Map_fake_F', 'G_Map_real_F',
'G_Map_fake_S', 'G_Map_real_S'
]
self.model_names = ['G']
if self.training:
self.model_names += ['D']
# identity loss
if self.opt.idt_lambda > 0:
self.loss_names += ['idt']
# feature metric loss
if self.opt.feature_D_lambda > 0:
self.loss_names += ['fea_m']
self.loss_names += ['D_Map', 'G_Map']
# feature metric loss
if self.opt.feature_D_map_lambda > 0:
self.loss_names += ['fea_m_Map']
# map loss
if self.opt.map_projection_lambda > 0:
self.loss_names += ['idt_Map']
# models
self.model_names += ['D_Map']
self.netG = factory.define_3DG(opt.noise_len,
opt.input_shape,
opt.output_shape,
opt.input_nc_G,
opt.output_nc_G,
opt.ngf,
opt.which_model_netG,
opt.n_downsampling,
opt.norm_G,
not opt.no_dropout,
opt.init_type,
self.gpu_ids,
opt.n_blocks,
opt.encoder_input_shape,
opt.encoder_input_nc,
opt.encoder_norm,
opt.encoder_blocks,
opt.skip_number,
opt.activation_type,
opt=opt)
if self.training:
# out of discriminator is not probability when
# GAN loss is LSGAN
use_sigmoid = False
# conditional Discriminator
if self.conditional_D:
d_input_channels = opt.input_nc_D + 1
else:
d_input_channels = opt.input_nc_D
self.netD = factory.define_D(d_input_channels,
opt.ndf,
opt.which_model_netD,
opt.n_layers_D,
opt.norm_D,
use_sigmoid,
opt.init_type,
self.gpu_ids,
opt.discriminator_feature,
num_D=opt.num_D,
n_out_channels=opt.n_out_ChannelsD)
self.netD_Map = factory.define_D(
len(self.multi_view),
opt.map_ndf,
opt.map_which_model_netD,
opt.map_n_layers_D,
opt.map_norm_D,
use_sigmoid,
opt.init_type,
self.gpu_ids,
opt.discriminator_feature,
opt.map_num_D,
n_out_channels=opt.map_n_out_ChannelsD)
if self.if_pool:
self.fake_pool = ImagePool(opt.pool_size)
self.fake_pool_Map = ImagePool(opt.map_pool_size)