print(info_optimizer, file=f)
print('----', file=f)
if gpu_mode:
d_real_flag, d_fake_flag = d_real_flag.cuda(), d_fake_flag.cuda()
G.cuda()
D.cuda()
BCE_loss = nn.BCELoss().cuda()
CE_loss = nn.CrossEntropyLoss().cuda()
MSE_loss = nn.MSELoss().cuda()
else:
BCE_loss = nn.BCELoss()
CE_loss = nn.CrossEntropyLoss()
MSE_loss = nn.MSELoss()
d_loss_fn, g_loss_fn = loss_norm_gp.get_losses_fn('wgan')
train_hist = {}
train_hist['D_loss'] = []
train_hist['G_loss'] = []
train_hist['info_loss'] = []
train_hist['per_epoch_time'] = []
train_hist['total_time'] = []
#------------------train----------------------
D.train()
print('training start!!')
start_time = time.time()
for i in tqdm.trange(epoch):
G.train()
epoch_start_time = time.time()
pin_memory=use_gpu,
drop_last=True)
# model
D = model.Discriminator_v1_1(x_dim=1, dim=dim).to(device)
G = model.Generator_v1_1(x_dim=z_dim, dim=dim).to(device)
#save model in txt
with open('./output/%s/setting.txt' % experiment_name, 'a') as f:
print('----', file=f)
print(G, file=f)
print('----', file=f)
print(D, file=f)
# gan loss function
d_loss_fn, g_loss_fn = loss_norm_gp.get_losses_fn(
'gan') #'gan', 'lsgan', 'wgan', 'hinge_v1', 'hinge_v2'
# optimizer
d_optimizer = torch.optim.Adam(D.parameters(),
lr=d_learning_rate,
betas=(0.5, 0.999))
g_optimizer = torch.optim.Adam(G.parameters(),
lr=g_learning_rate,
betas=(0.5, 0.999))
info_optimizer = optim.Adam(itertools.chain(G.parameters(), D.parameters()),
lr=0.0001,
betas=(0.6, 0.95),
amsgrad=True) #G,D都更新
# = train =