def face_sample(real=False):
import gatedpixelcnn
z = gatedpixelcnn.facefeatures_sample(real)
solver = Solver(_cfg(real), try_load_best=True)
x = solver.sample(z)
if real:
fn = "results/vqvaer_face_sample_1.png"
else:
fn = "results/vqvae_face_sample_1.png"
torchvision.utils.save_image(x, fn, range=(-1, 1), normalize=True)
utils.show_pics(x, normalize=True, range=(-1, 1))
def mnist_sample():
cfg = Config()
solver = Solver(cfg, try_load_best=True)
x = solver.sample((32, 28, 28)).float() / cfg.input_dim
x = x[:, None, :, :]
utils.show_pics(x)
def mk_sample():
sample.fill_(0)
net.eval()
for i in range(28):
for j in range(28):
out = net(sample)
probs = F.softmax(out[:, :, i, j], dim=1)
sample[:, :, i, j] = torch.multinomial(probs, 1).float() / 255.
is_infer = False
if is_infer:
mk_sample()
utils.show_pics(sample)
else:
tr, te = dataloader.load_mnist(select=8)
epochs = 30
for epoch in range(last_epoch + 1, epochs + last_epoch + 1):
# train
err_tr = []
time_tr = time.time()
net.train()
for i, (input, _) in enumerate(tr):
input = input.cuda()
target = (input[:, 0] * 255).long()
out = net(input)
loss = F.cross_entropy(out, target)
data_cpu = d[0]
data = data_cpu.to(self.device)
recon_batch, mu, logvar = self.model(data)
test_loss += loss_function(recon_batch, data, mu,
logvar).item()
if i == 0:
choose_orig = data_cpu
choose_recon = recon_batch.view(-1, 3, 64, 64).cpu()
choose_z = mu.cpu()
test_loss /= len(test_loader)
return test_loss, choose_orig, choose_recon, choose_z
def sample(self):
self.model.eval()
with torch.no_grad():
sample_z = torch.randn(64, 100)
z = sample_z.to(self.device)
sample_data = self.model.decode(z).cpu().view(64, 3, 64, 64)
return sample_data, sample_z
solver = Solver()
is_infer = False
if is_infer:
sample_data, _ = solver.sample()
utils.show_pics(sample_data)
else:
solver.train(30)
is_infer = False
checkpoint_path = "results/dcgan.pt"
if is_read_checkpoint and os.path.isfile(checkpoint_path):
checkpoint = torch.load(checkpoint_path)
netG.load_state_dict(checkpoint['netG'])
netD.load_state_dict(checkpoint['netD'])
optimizerG.load_state_dict(checkpoint['optG'])
optimizerD.load_state_dict(checkpoint['optD'])
last_epoch = checkpoint['epoch']
print('Load checkpoint! Last Epoch: %d' % last_epoch)
if is_infer:
fake = netG(fixed_noise)
fake = fake.detach().cpu()
utils.show_pics(fake, normalize=True, range=(-1, 1))
else:
normalizer = torchvision.transforms.Normalize((0.5, 0.5, 0.5),
(0.5, 0.5, 0.5))
train_loader, test_loader = dataloader.load_faces(normalizer)
# graph,现在tensorboard似乎好像不能支持2个graph
noise = torch.randn(64, nz, 1, 1, device=device)
writer.add_graph(netG, noise)
fixed_images = utils.first_batch(test_loader)
writer.add_graph(netD, fixed_images.to(device))
# original img
img_grid = torchvision.utils.make_grid(fixed_images,
range=(-1, 1),
normalize=True)
def face_reconstruct(real=False):
solver = Solver(_cfg(real), try_load_best=True)
tr, te = _load(real)
x = utils.random_choices(te.dataset, k=32)
x_tilde = solver.reconstruct(x)
utils.show_pics(torch.cat([x, x_tilde]), normalize=True, range=(-1, 1))