def forward(self):
self.z, self.conv0, self.conv2, self.conv3, self.conv4 = self.encode(
self.real_A)
self.z_b, self.conv0_b, self.conv2_b, self.conv3_b, self.conv4_b = self.encode(
self.real_B)
self.z_tf = self.transform(self.z, self.real_RT)
self.depth_tf = self.depthdecode(self.z_tf)
self.depth_a = self.depthdecode(self.z)
self.depth_b = self.depthdecode(self.z_b)
self.warp(self.real_A, self.depth_tf, self.real_RT)
_, self.conv0_w, self.conv2_w, self.conv3_w, self.conv4_w = self.encode(
self.warp_fake_B)
self.fake_A = self.decode(self.z, self.conv0, self.conv2, self.conv3,
self.conv4)
self.conv0_tf, _, _ = inverse_warp(self.conv0, self.depth_tf,
self.real_RT, self.intrinsics)
self.conv2_tf, _, _ = inverse_warp(
self.conv2,
torch.nn.functional.upsample(self.depth_tf, scale_factor=0.25),
self.real_RT, self.get_K(self.intrinsics, 0.25))
self.conv3_tf, _, _ = inverse_warp(
self.conv3,
torch.nn.functional.upsample(self.depth_tf, scale_factor=0.125),
self.real_RT, self.get_K(self.intrinsics, 0.125))
self.conv4_tf, _, _ = inverse_warp(
self.conv4,
torch.nn.functional.upsample(self.depth_tf, scale_factor=0.0625),
self.real_RT, self.get_K(self.intrinsics, 0.0625))
self.fake_B, self.fake_B3, self.fake_B2, self.fake_B1 = self.decode(
self.z_tf, self.conv0_tf, self.conv2_tf, self.conv3_tf,
self.conv4_tf)
def get_high_res(self, image, pose, z=None):
image_small = cv2.resize(image, (256, 256))
image_small = torch.from_numpy(image_small / 128. - 1).permute(
(2, 0, 1)).contiguous().unsqueeze(0)
image_small = Variable(image_small).to(self.device).float()
RT = self.get_RT(pose)
z = self.enc(image_small) if z is None else z
z_tf = self.transform(z, RT)
depth = self.decode(z_tf)
depth = F.upsample(depth, scale_factor=4, mode='bilinear')
intrinsics = self.intrinsics[:1, :, :] * 4
intrinsics[0, 2, 2] = 1
image = image / 128. - 1
image = torch.from_numpy(image).permute(
(2, 0, 1)).contiguous().unsqueeze(0).to(self.device).float()
image, _, _ = inverse_warp(image, depth, RT, intrinsics)
image = tensor2im(image.data.detach())
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
return image
def warp(self, image, depth, RT):
self.warp_fake_B, self.flow, self.mask = inverse_warp(
image, depth, RT, self.intrinsics)