def set_input(self, input):
"""
the structure of the input.
{
'image':Bx3x96x96 FloatTensor
'pose':Bx1 FloatTensor
'identity':Bx1 FloatTensor
}
test_pose (B): used for the test='initial learning rate
"""
self.load_input(input)
self.image = torch.squeeze(torch.stack(self.image, dim=0))
self.batchsize = len(self.pose)
self.pose = torch.LongTensor(self.pose)
self.frontal_pose = torch.LongTensor(
np.random.randint(self.N_p, size=self.batchsize))
# if self.is_Train:
self.input_pose = one_hot(self.frontal_pose, self.N_p)
# else:
# self.input_pose = one_hot(test_pose.long(), self.N_p)
self.identity = torch.LongTensor(self.identity)
self.fake_identity = torch.zeros(
self.batchsize).long() # 0 indicates fake
self.noise = torch.FloatTensor(
np.random.normal(loc=0.0,
scale=0.3,
size=(self.batchsize, self.N_z)))
#cuda
if self.opt.gpu_ids:
self.image = self.image.cuda()
self.pose = self.pose.cuda()
self.frontal_pose = self.frontal_pose.cuda()
self.input_pose = self.input_pose.cuda()
self.identity = self.identity.cuda()
self.fake_identity = self.fake_identity.cuda()
self.noise = self.noise.cuda()
self.image = Variable(self.image)
self.pose = Variable(self.pose)
self.frontal_pose = Variable(self.frontal_pose)
self.input_pose = Variable(self.input_pose)
self.identity = Variable(self.identity)
self.fake_identity = Variable(self.fake_identity)
self.noise = Variable(self.noise)
def set_input(self, input):
"""
the structure of the input.
{
'image':Bx3x96x96 FloatTensor
'pose':Bx1 FloatTensor
'identity':Bx1 FloatTensor
}
test_pose (B): used for the test='initial learning rate
"""
self.load_input(input)
self.batchsize = []
self.rand_pose = []
self.frontal_pose = []
self.input_pose = []
self.fake_identity = []
self.noise = []
self.frontal_image = [[] for r in range(len(self.pose))]
self.profile_image = [[] for r in range(len(self.pose))]
# self.real_frontal_pose = [[] for r in range(len(self.pose))]
# self.real_profile_pose = [[] for r in range(len(self.pose))]
# self.frontal_id = [[] for r in range(len(self.pose))]
self.profile_id = [[] for r in range(len(self.pose))]
self.frontal_name = [[] for r in range(len(self.pose))]
self.profile_name = [[] for r in range(len(self.pose))]
for j in range(len(self.pose)):
for e in range(len(self.pose[j])):
if self.pose[j][e]:
self.frontal_image[j].append(self.image[j][e])
# self.real_frontal_pose[j].append(self.pose[j][e])
# self.frontal_id[j].append(self.identity[j][e])
self.frontal_name[j].append(self.name[j][e])
else:
self.profile_image[j].append(self.image[j][e])
# self.real_profile_pose[j].append(self.pose[j][e])
self.profile_id[j].append(self.identity[j][e])
self.profile_name[j].append(self.name[j][e])
self.frontal_image[j] = torch.squeeze(
torch.stack(self.frontal_image[j], dim=0))
self.profile_image[j] = torch.squeeze(
torch.stack(self.profile_image[j], dim=0))
# self.real_frontal_pose[j] = torch.LongTensor(self.real_frontal_pose[j])
# self.real_profile_pose[j] = torch.LongTensor(self.real_profile_pose[j])
# self.frontal_id[j] = torch.LongTensor(self.frontal_id[j])
self.profile_id[j] = torch.LongTensor(self.profile_id[j])
x = torch.split(self.identity[j], 1)
self.identity[j] = torch.cat((self.identity[j], x[0]), 0)
if self.is_Train:
self.batchsize.append((self.pose[j]).size(0))
else:
self.batchsize.append((self.profile_id[j]).size(0))
self.rand_pose.append(
torch.LongTensor(
np.random.randint(self.N_p, size=self.batchsize[j] + 1)))
self.frontal_pose.append(torch.ones(self.batchsize[j] + 1).long())
x = torch.split(self.profile_id[j], 1)
self.profile_id[j] = torch.cat((self.profile_id[j], x[0]), 0)
if self.is_Train:
self.input_pose.append(one_hot(self.rand_pose[j], self.N_p))
else:
self.input_pose.append(one_hot(self.frontal_pose[j], self.N_p))
self.fake_identity.append(
torch.zeros(self.batchsize[j] + 1).long()) # 0 indicates fake
self.noise.append(
torch.FloatTensor(
np.random.normal(loc=0.0,
scale=0.3,
size=(self.batchsize[j] + 1, self.N_z))))
#cuda
if self.opt.gpu_ids:
self.image[j] = self.image[j].cuda()
self.pose[j] = self.pose[j].cuda()
self.rand_pose[j] = self.rand_pose[j].cuda()
self.frontal_pose[j] = self.frontal_pose[j].cuda()
self.input_pose[j] = self.input_pose[j].cuda()
self.identity[j] = self.identity[j].cuda()
self.fake_identity[j] = self.fake_identity[j].cuda()
self.noise[j] = self.noise[j].cuda()
self.frontal_image[j] = self.frontal_image[j].cuda()
self.profile_image[j] = self.profile_image[j].cuda()
# self.real_frontal_pose[j] = self.real_frontal_pose[j].cuda()
# self.real_profile_pose[j] = self.real_profile_pose[j].cuda()
# self.frontal_id[j] = self.frontal_id[j].cuda()
self.profile_id[j] = self.profile_id[j].cuda()
self.image[j] = Variable(self.image[j])
self.pose[j] = Variable(self.pose[j])
self.rand_pose[j] = Variable(self.rand_pose[j])
self.frontal_pose[j] = Variable(self.frontal_pose[j])
self.input_pose[j] = Variable(self.input_pose[j])
self.identity[j] = Variable(self.identity[j])
self.fake_identity[j] = Variable(self.fake_identity[j])
self.noise[j] = Variable(self.noise[j])
self.frontal_image[j] = Variable(self.frontal_image[j])
self.profile_image[j] = Variable(self.profile_image[j])
# self.real_frontal_pose[j] = Variable(self.real_frontal_pose[j])
# self.real_profile_pose[j] = Variable(self.real_profile_pose[j])
# self.frontal_id[j] = Variable(self.frontal_id[j])
self.profile_id[j] = Variable(self.profile_id[j])