def create_nets(config, input_dim, num):
'''create muti netG, and append into nets.
- Params:
@config: config for netG
@input_dim: the dim of input
@num: num of netG
- Returns
mutils nets
'''
nets = []
for i in range(num):
netG = build_netG(config[3], input_dim)
nets.append(netG)
return nets
def __init__(self, opt):
super(_testGan, self).__init__(opt)
self.netG = build_netG(opt.g_model, opt.z_dim)
self.netD = build_netD(opt.d_model, opt.x_dim)
netD.apply(weight_init)
netG.apply(weight_init)
x = torch.FloatTensor(opt.mb_size, opt.x_dim, opt.img_size, opt.img_size)
z = torch.FloatTensor(opt.mb_size, opt.z_dim, 1, 1)
label = torch.FloatTensor(opt.mb_size)
if self.cuda:
netD.cuda()
netG.cuda()
x, z = x.cuda(), z.cuda()
label = label.cuda()
def create_nets(config, input_dim, num, type='G'):
'''create muti netG, and append into nets.
- Params:
@config: config for netG
@input_dim: the dim of input
@num: num of netG
@type: G nets or D nets
- Returns
mutils nets
'''
nets = []
for i in range(num):
if type == 'G':
netG = build_netG(config, input_dim)
nets.append(netG)
elif type == 'D':
netD = build_netD(config, input_dim)
nets.append(netD)
return nets
h_dim = 128 x_dim_w, x_dim_h = train_loader.dataset.train_data.size()[1:3] resize_w, resize_h = 64, 64 x_dim = 3 train_size = train_loader.dataset.train_data.size()[0] y_dim = 10 lr = 1e-3 cnt = 0 nets_num = 10 cuda = False niter = 24 # build gans netD = build_netD(config['D'][5], x_dim) netG = build_netG(config['G'][6], z_dim) print netD # init gans netD.apply(weight_init) netG.apply(weight_init) # build gans's solver G_solver = optim.Adam(netG.parameters(), lr=lr, betas=(0.5, 0.999)) D_solver = optim.Adam(netD.parameters(), lr=lr, betas=(0.5, 0.999)) # build exps of netG and netD # G_exp = create_sigle_experiment(cc, 'G_loss') # D_exp = create_sigle_experiment(cc, 'D_loss')
x_dim_w, x_dim_h = train_loader.dataset.train_data.size()[1:3]
x_dim = x_dim_w * x_dim_h
train_size = train_loader.dataset.train_data.size()[0]
y_dim = 10
lr = 1e-3
cnt = 0
nets_num = 10
cuda = False
cc = CrayonClient(hostname="localhost")
cc.remove_all_experiments()
D_exp = create_sigle_experiment(cc, 'D_loss')
G_exps = create_experiments(cc, 10)
netG_share = build_netG(config['G'][2], z_dim)
netG_indeps = create_nets(config['G'], h_dim, nets_num)
netD = build_netD(config['D'][2], x_dim)
init_network(netG_share)
init_network(netG_indeps)
init_network(netD)
D_solver = optim.Adam(netD.parameters(), lr=lr)
G_share_solver = optim.Adam(netG_share.parameters(), lr=lr)
G_indep_solver = create_optims(netG_indeps, [
lr,
])
G_solvers = create_couple2one_optims(netG_share, netG_indeps, [
lr,
])
def __init__(self, opt):
super(_competitionGan, self).__init__(opt)
self.opt = opt
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.condition_D = opt.condition_D
self.mb_size = opt.mb_size
self.Lambda = opt.Lambda
self.continue_train = opt.continue_train
self.train = opt.train
self.test = True if self.continue_train and self.train else False
self.savepath = '{}{}/'.format(opt.savepath, opt.gans_type)
self.cnt = 0
self.netGs = []
for index in range(self.nums):
netG = build_netG(opt.g_model, opt.z_dim)
self.netGs.append(netG)
self.netD = build_netD(opt.d_model, opt.x_dim, opt.condition_D)
X = torch.FloatTensor(opt.mb_size, opt.x_dim, opt.img_size,
opt.img_size)
Z = torch.FloatTensor(opt.mb_size, opt.z_dim, 1, 1)
real_like_sample = torch.FloatTensor(opt.mb_size, opt.x_dim,
opt.img_size, opt.img_size)
fake_like_sample = torch.FloatTensor(opt.mb_size, opt.x_dim,
opt.img_size, opt.img_size)
label = torch.FloatTensor(opt.mb_size)
self.criterionGAN = torch.nn.BCELoss()
self.criterionL1 = torch.nn.L1Loss()
if self.cuda:
self.netD.cuda()
for index in range(self.nums):
self.netGs[index].cuda()
self.criterionGAN.cuda()
self.criterionL1.cuda()
X, Z = X.cuda(), Z.cuda()
real_like_sample, fake_like_sample = real_like_sample.cuda(
), fake_like_sample.cuda()
label = label.cuda()
self.X = Variable(X)
self.Z = Variable(Z)
self.real_like_sample = Variable(real_like_sample)
self.fake_like_sample = Variable(fake_like_sample)
self.label = Variable(label)
info.log("Train: {} Continue: {} Test: {}".format(
self.train, self.continue_train, self.test))
if self.opt.cc:
self.create_tensorboard()
self.index_exp = create_sigle_experiment(self.cc, 'index')
self.D_solver = torch.optim.Adam(self.netD.parameters(),
lr=2e-4,
betas=(0.5, 0.999))
self.G_solvers = create_optims(self.netGs, [2e-4, (0.5, 0.999)])
if opt.train == False:
self.load_networkG(self.opt.g_network_path)
self.load_networkD(self.opt.d_network_path)
else:
init_network(self.netD)
init_network(self.netGs)
display_cnt = 100 iter = 2 nets_num = 10 cuda = False netD_continue_trian = True cc = CrayonClient(hostname="localhost") cc.remove_all_experiments() D_exp = create_sigle_experiment(cc, 'D_loss') D_preb_real = create_sigle_experiment(cc, 'preb_real') D_preb_fake = create_sigle_experiment(cc, 'preb_fake') G_exps = create_experiments(cc, 10) netG_indeps = create_nets(config['G'][2], z_dim, nets_num) netG_share = build_netG(config['G'][3], h_dim) netD = build_netD(config['D'][2], x_dim) print netG_indeps print netG_share init_network(netG_share) init_network(netG_indeps) init_network(netD) D_solver = optim.Adam(netD.parameters(), lr=lr) G_share_solver = optim.Adam(netG_share.parameters(), lr=lr) G_indep_solver = create_optims(netG_indeps, [lr, (0.5, 0.999)]) X = torch.FloatTensor(mb_size, x_dim) z = torch.FloatTensor(mb_size, z_dim)
