def train_GE(self, x, l_recons=None, l_norm=None):
"""
:param x:
:param l_recons:
:param l_norm:
:return:
"""
bs = x.size(0)
real_labels = cuda(Variable(torch.ones(bs)))
fake_labels = cuda(Variable(torch.zeros(bs)))
z_e = self.E(x)
real_logit = self.D(x, z_e)
z = self.sample_z(bs)
x_fake = self.G(z)
fake_logit = self.D(x_fake, z)
z_norm = torch.sqrt(torch.sum(z_e ** 2, dim=-1))
g_loss = self.compute_ge_loss(real_logit, fake_logit, fake_labels, real_labels)
g_loss_out = g_loss.data.cpu().numpy()
if self.bigan_type != cte.BiGAN:
x_recons = self.G(z_e)
x_recons_loss = self.get_x_recons_loss(x, x_recons)
g_loss += l_recons * x_recons_loss.mean()
if self.bigan_type in [cte.PMDGAN, cte.MLPMDGAN, cte.EPMDGAN]:
z_e_loss = self.get_ze_loss(z_e)
g_loss += l_norm * z_e_loss.mean()
return g_loss_out, z_norm, g_loss
def train_D(self, x):
bs = x.size(0)
real_labels = cuda(Variable(torch.ones(bs)))
fake_labels = cuda(Variable(torch.zeros(bs)))
z_e = self.E(x)
real_logit = self.D(x, z_e)
z = self.sample_z(bs)
x_fake = self.G(z)
fake_logit = self.D(x_fake, z)
d_loss = self.compute_d_loss(real_logit, real_labels, fake_logit, fake_labels)
return d_loss
def sample3(self, z):
z = torch.tensor(z, dtype=torch.float)
z = Variable(z).detach()
z = cuda(z)
x = self.G(z)
x = x.data.cpu().numpy()
return x
def sample2(self, z, tensor=False):
z = torch.tensor(z)
z = Variable(z).detach()
z = cuda(z)
x = self.G(z)
if not tensor:
return x.data.cpu().numpy()
else:
return x
def psnr(self, loader_sorted):
psnr_list = list()
for i, x in enumerate(loader_sorted):
x = cuda(x)
z = self.E(x)
x_recons = self.G(z)
psnr = loss.PSNR_torch(x, x_recons, axis=(-3, -2, -1))
psnr = psnr.data.cpu().numpy()
psnr_list.extend(psnr)
return np.array(psnr_list)
def get_z(self, x_input):
n_imgs = x_input.shape[0]
z_list = list()
n_batches = int(np.ceil(n_imgs / 128))
for i in range(n_batches):
x = torch.tensor(x_input[i * 128:(i + 1) * 128]).float()
x = cuda(x)
z = self.E(x)
z_list.extend(z.data.cpu().numpy())
z_list = np.array(z_list)
return z_list
def get_prediction(self, x_input, hard=True):
n_imgs = x_input.shape[0]
out_list = list()
n_batches = int(np.ceil(n_imgs / 128))
for i in range(n_batches):
x = torch.tensor(x_input[i * 128:(i + 1) * 128]).float()
x = lib.cuda(x)
out = self.forward(x)
out = torch.argmax(out, dim=1) if hard else F.softmax(out, dim=1)
out_list.extend(out.data.cpu().numpy())
return np.array(out_list)
def sample4(self, bs):
# It’s more efficient than any other autograd setting - it will use
# the absolute minimal amount of memory to evaluate the model. volatile
# also determines that requires_grad is False.
n_bs = bs // 64 + 1
x_list = list()
z_list = list()
for i in range(n_bs):
z = Variable(torch.randn(64, self.z_dim)).detach()
z = cuda(z)
x = self.G(z)
x_list.extend(x.cpu().data.numpy())
z_list.extend(z.cpu().data.numpy())
return np.array(x_list[:bs]), np.array(z_list[:bs])
def get_data_batch(self, data_curr):
if not self.d_batch_on:
return data_curr
if self.bigan_type in [cte.EPMDGAN] and self.data_batch is None:
self.data_batch = cuda(data_curr)
return data_curr
data = torch.cat((cuda(data_curr), self.data_batch), 0)
self.model.eval()
data_recons = self.model.G(self.model.E(data))
psnr = loss_fn.PSNR_torch(data, data_recons, axis=(-3, -2, -1))
idx_sorted = psnr.argsort(descending=False) # From smaller to larger
self.model.train()
half_batch_size = self.batch_size // 3
if np.random.uniform() > self.l_perc:
self.data_batch = cuda(data_curr)
else:
self.data_batch = data[idx_sorted[:half_batch_size]]
self.model.train()
return data[
idx_sorted[:self.
batch_size]] # Keep the images with the worst reconstruction
def sample_z(self, bs):
z = torch.randn((bs, self.z_dim))
z = cuda(z)
z = Variable(z)
return z
def preprocess_batch(self, x):
data = Variable(x).float()
data = cuda(data)
return data
