def __init__(self, use_cuda=False, threshold=0.001, patch_size=19):
super(AffineShapeEstimator, self).__init__()
self.threshold = threshold
self.use_cuda = use_cuda
self.PS = patch_size
self.gx = nn.Conv2d(1,
1,
kernel_size=(1, 3),
padding=(0, 1),
bias=False)
self.gx.weight.data = torch.from_numpy(
np.array([[[[0.5, 0, -0.5]]]], dtype=np.float32))
self.gy = nn.Conv2d(1,
1,
kernel_size=(3, 1),
padding=(1, 0),
bias=False)
self.gy.weight.data = torch.from_numpy(
np.array([[[[0.5], [0], [-0.5]]]], dtype=np.float32))
self.gk = torch.from_numpy(
CircularGaussKernel(kernlen=patch_size,
circ_zeros=True).astype(np.float32))
self.gk = Variable(self.gk, requires_grad=False)
if use_cuda:
self.gk = self.gk.cuda()
return
def extract_patches(self, scale_pyramid, LAFs, pyr_idxs, level_idxs, PS = 19, gauss_mask = False, use_cuda = False):
patches_list = []
if gauss_mask:
mask = torch.from_numpy(CircularGaussKernel(kernlen = PS, circ_zeros = False).astype(np.float32))
mask = Variable(mask)
if use_cuda:
mask = mask.cuda()
for i in range(len(scale_pyramid)):
cur_idxs = pyr_idxs == i #torch.nonzero((pyr_idxs == i).data)
for j in range(1, len(level_idxs) - 1):
cur_lvl_idxs = torch.nonzero(((level_idxs == j) * cur_idxs).data)
if len(cur_lvl_idxs.size()) == 0:
continue
curr_aff = LAFs[cur_lvl_idxs.view(-1), :,:]
grid = torch.nn.functional.affine_grid(curr_aff, torch.Size((cur_lvl_idxs.size(0),
1,
PS,
PS)))
patches_list.append(torch.nn.functional.grid_sample(scale_pyramid[i][j].expand(curr_aff.size(0),
scale_pyramid[i][0].size(1),
scale_pyramid[i][0].size(2),
scale_pyramid[i][0].size(3)), grid))
patches = torch.cat(patches_list, dim = 0)
if gauss_mask:
patches = patches * mask.unsqueeze(0).unsqueeze(0).expand(patches.size(0),1,PS,PS)
return patches
def __init__(self, mrSize=3.0, patch_size=None):
super(OrientationDetector, self).__init__()
if patch_size is None:
patch_size = 32
self.PS = patch_size
self.bin_weight_kernel_size, self.bin_weight_stride = self.get_bin_weight_kernel_size_and_stride(
self.PS, 1)
self.mrSize = mrSize
self.num_ang_bins = 36
self.gx = nn.Conv2d(1, 1, kernel_size=(1, 3), bias=False)
self.gx.weight.data = torch.from_numpy(
np.array([[[[0.5, 0, -0.5]]]], dtype=np.float32))
self.gy = nn.Conv2d(1, 1, kernel_size=(3, 1), bias=False)
self.gy.weight.data = torch.from_numpy(
np.array([[[[0.5], [0], [-0.5]]]], dtype=np.float32))
self.angular_smooth = nn.Conv1d(1,
1,
kernel_size=3,
padding=1,
bias=False)
self.angular_smooth.weight.data = torch.from_numpy(
np.array([[[0.33, 0.34, 0.33]]], dtype=np.float32))
self.gk = 10. * torch.from_numpy(
CircularGaussKernel(kernlen=self.PS).astype(np.float32))
self.gk = Variable(self.gk, requires_grad=False)
return
def __init__(self, threshold = 0.001, patch_size = 19):
super(AffineShapeEstimator, self).__init__()
self.threshold = threshold;
self.PS = patch_size
self.gx = nn.Conv2d(1, 1, kernel_size=(1,3), bias = False)
self.gx.weight.data = torch.from_numpy(np.array([[[[-1, 0, 1]]]], dtype=np.float32))
self.gy = nn.Conv2d(1, 1, kernel_size=(3,1), bias = False)
self.gy.weight.data = torch.from_numpy(np.array([[[[-1], [0], [1]]]], dtype=np.float32))
self.gk = torch.from_numpy(CircularGaussKernel(kernlen = self.PS, sigma = (self.PS / 2) /3.0).astype(np.float32))
self.gk = Variable(self.gk, requires_grad=False)
return
def __init__(self, PS=16):
super(BaumResNetEll, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(1, 16, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(16, affine=False), nn.ReLU(),
nn.Conv2d(16, 32, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(32, affine=False), nn.ReLU(),
nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(32, affine=False), nn.ReLU(),
nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Dropout(0.1),
nn.Conv2d(64, 3, kernel_size=4, bias=True), nn.Tanh())
self.PS = PS
self.features.apply(weights_init)
self.gx = nn.Conv2d(1, 1, kernel_size=(1, 3), bias=False)
self.gx.weight.data = torch.from_numpy(
np.array([[[[0.5, 0, -0.5]]]], dtype=np.float32))
self.gy = nn.Conv2d(1, 1, kernel_size=(3, 1), bias=False)
self.gy.weight.data = torch.from_numpy(
np.array([[[[0.5], [0], [-0.5]]]], dtype=np.float32))
self.gxx = nn.Conv2d(1, 1, kernel_size=(1, 3), bias=False)
self.gxx.weight.data = torch.from_numpy(
np.array([[[[1.0, -2.0, 1.0]]]], dtype=np.float32))
self.gyy = nn.Conv2d(1, 1, kernel_size=(3, 1), bias=False)
self.gyy.weight.data = torch.from_numpy(
np.array([[[[1.0], [-2.0], [1.0]]]], dtype=np.float32))
self.gk = torch.from_numpy(
CircularGaussKernel(kernlen=PS,
circ_zeros=False).astype(np.float32))
self.gk = Variable(self.gk, requires_grad=False)
return
def calculate_weights(self, sigma):
kernel = CircularGaussKernel(sigma=sigma, circ_zeros=False)
h, w = kernel.shape
halfSize = float(h) / 2.
self.pad = int(np.floor(halfSize))
return torch.from_numpy(kernel.astype(np.float32)).view(1, 1, h, w)