def plot_generator(epoch, progress):
x = began.generate_x(10000, test=True)
x.unchain_backward()
x = began.to_numpy(x)
try:
plot_scatter(x,
dir=args.plot_dir,
filename="generator_scatter_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
plot_kde(x,
dir=args.plot_dir,
filename="generator_kde_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
except:
pass
def plot_samples(epoch, progress):
samples_fale = gan.generate_x(10000, from_gaussian=True)
samples_fale.unchain_backward()
samples_fale = gan.to_numpy(samples_fale)
try:
plot_scatter(samples_fale,
dir=args.plot_dir,
filename="scatter_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
plot_kde(samples_fale,
dir=args.plot_dir,
filename="kde_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
except:
pass
def plot_reconstruction(epoch, progress, x):
z = began.encode(x, test=True)
x = began.decode(z, test=True)
x.unchain_backward()
x = began.to_numpy(x)
try:
plot_scatter(
x,
dir=args.plot_dir,
filename="reconstruction_scatter_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
plot_kde(x,
dir=args.plot_dir,
filename="reconstruction_kde_epoch_{}_time_{}min".format(
epoch, progress.get_total_time()))
except:
pass
z = Variable(Tensor(np.random.normal(0, 1, (opt.batchSize, opt.nz)).astype(np.float32))).to(device)
optimizer_G.zero_grad()
# Generate a batch of images
samples_fake = generator(z)
# Adversarial loss
lossG = - torch.sum(discriminator(samples_fake / opt.scale) / opt.batchSize)
lossG.backward()
optimizer_G.step()
# ====================
# Save to tensorborad
# ====================
if(i == 0 or (i % opt.n_display_step == 0)):
board_train.add_scalar('Generater/loss_G', lossG.item(), iterations)
board_train.add_scalar('Discriminator/loss_D', lossD.item(), iterations)
# Monitor trainnig progresses
print("epoch={}, iters={}, loss_G={:.5f}, loss_C={:.5f}".format(epoch, iterations, lossG, lossD))
# ============
# Save images
# ============
generator.eval()
z_fixed = Variable(Tensor(np.random.normal(0, 1, (10000, opt.nz)).astype(np.float32))).to(device)
with torch.no_grad():
samples_fake = generator(z_fixed)
plot_scatter(samples_fake.cpu().numpy(), dir=os.path.join(opt.dir_out, opt.exper_name), filename="scatter_epoches{}".format(epoch))
plot_kde(samples_fake.cpu().numpy(), dir=os.path.join(opt.dir_out, opt.exper_name), filename="kde_epoches{}".format(epoch))
prog.add_loss_dis()
z = sampler.sample_z(config['dim_z'],
batchsize,
gaussian=config['gaussian'])
z = Variable(torch.from_numpy(z))
samples_fake = gen_net(z)
samples_fake /= config['scale']
f_fake = dis_net(samples_fake)
loss_gen = -f_fake.mean()
prog.add_loss_gen(loss_gen)
gen_optim.zero_grad()
loss_gen.backward()
gen_optim.step()
if (i + 1) % config['num_plot'] == 0:
print(i + 1)
z = sampler.sample_z(config['dim_z'],
10000,
gaussian=config['gaussian'])
z = Variable(torch.from_numpy(z))
samples_fake = gen_net(z).data.numpy()
plot.plot_scatter(samples_fake,
dir='plot',
filename='{}_scatter'.format(i + 1))
plot.plot_kde(samples_fake,
dir='plot',
filename='{}_kde'.format(i + 1))
prog.plot()
loss_critic.backward()
dis_optim.step()
prog.add_loss_dis()
z = sampler.sample_z(config['dim_z'],
batchsize,
gaussian=config['gaussian'])
z = Variable(torch.from_numpy(z).cuda())
samples_fake = gen_net(z)
samples_fake /= config['scale']
f_fake = dis_net(samples_fake)
loss_gen = -f_fake.mean()
prog.add_loss_gen(loss_gen.data.cpu().numpy()[0])
gen_optim.zero_grad()
loss_gen.backward()
gen_optim.step()
if (i + 1) % config['num_plot'] == 0:
print(i + 1)
z = sampler.sample_z(config['dim_z'],
10000,
gaussian=config['gaussian'])
z = Variable(torch.from_numpy(z).cuda())
samples_fake = gen_net(z).data.cpu().numpy()
plot.plot_scatter(samples_fake,
filename='{}_scatter'.format(i + 1),
show=True)
plot.plot_kde(samples_fake, filename='{}_kde'.format(i + 1), show=True)
prog.plot()
