for batch_idx, (data, target) in enumerate(train_loader, 0):
netD.zero_grad()
mb_size = data.size(0)
X.data.resize_(data.size()).copy_(data)
X.data.resize_(mb_size, x_dim)
z.data.resize_(mb_size, z_dim).normal_(0, 1)
label.data.resize_(mb_size)
G_share_sample = netG_share(z)
G_indep_sample = create_netG_indeps_sample(netG_indeps, G_share_sample)
############################
# (1) Update D network:
###########################
D_real = netD(X)
D_fake = netD_fake(G_indep_sample, netD)
D_loss = compute_dloss(D_real, D_fake, label)
D_exp.add_scalar_value('D_loss',
D_loss.data[0],
step=batch_idx + it * train_size)
D_loss.backward(retain_variables=True)
D_solver.step()
############################
# (2) Update G network:
###########################
D_fake = netD_fake(G_indep_sample, netD)
G_losses, index = compute_gloss(D_fake, label)
mutil_backward(G_losses, netG_share, netG_indeps, index)
mutil_steps(G_losses, G_share_solver, G_indep_solver, index)
add2experiments(G_losses, G_exps, step=batch_idx + it * train_size)
mb_size = data.size(0)
X.data.resize_(data.size()).copy_(data)
X.data.resize_(mb_size, x_dim)
z.data.resize_(mb_size, z_dim).normal_(0, 1)
label.data.resize_(mb_size)
cnt = batch_idx + it * train_size
G_indep_sample = create_netG_indeps_sample(netG_indeps, z)
G_share_sample = create_netG_share_sample(netG_share, G_indep_sample)
############################
# (1) Update D network: vaillgan loss
###########################
if not netD_continue_trian:
D_real = netD(X)
D_fake = netD_fake(G_share_sample, netD)
D_loss = compute_dloss(D_real, D_fake, label)
D_exp.add_scalar_value('D_loss', D_loss.data[0], step=cnt)
D_loss.backward(retain_variables=True)
D_solver.step()
else:
D_real = netD(X).data.numpy()[0][0]
D_fake = netD_fake(G_share_sample, netD)
D_fake = compute_mean(D_fake, 10.0).data.numpy()[0]
print 'D_real:{}----/D_fake:{}'.format(D_real, D_fake)
############################
# (2) Update G network:
###########################
D_fake = netD_fake(G_share_sample, netD)
G_losses, index = compute_gloss(D_fake, label)