def lstsq(b, y, alpha=0.01):
"""
Batched linear least-squares for pytorch with optional L1 regularization.
Parameters
----------
b : shape(L, M, N)
y : shape(L, M)
Returns
-------
tuple of (coefficients, model, residuals)
"""
bT = b.transpose(-1, -2)
AA = torch.bmm(bT, b)
if alpha != 0:
diag = torch.diagonal(AA, dim1=1, dim2=2)
diag += alpha
RHS = torch.bmm(bT, y[:, :, None])
X, LU = torch.gesv(RHS, AA)
fit = torch.bmm(b, X)[..., 0]
res = y - fit
return X[..., 0], fit, res
def backward(ctx, grad_output):
L, = ctx.saved_variables
if ctx.upper:
L = L.t()
grad_output = grad_output.t()
# make sure not to double-count variation, since
# only half of output matrix is unique
Lbar = grad_output.tril()
P = Potrf.phi(torch.mm(L.t(), Lbar))
S = torch.gesv(P + P.t(), L.t())[0]
S = torch.gesv(S.t(), L.t())[0]
S = Potrf.phi(S)
return S, None
def lsolve(A, B):
"""
Computes the solution X to AX = B.
"""
return torch.gesv(B, A)[0]
def predict(self, pred_x, hyper=None, in_optimization=False):
if hyper is not None:
param_original = self.model.param_to_vec()
self.cholesky_update(hyper)
kernel_max = self.model.kernel.forward_on_identical().data[0]
n_pred, n_dim = pred_x.size()
pred_x_radius = torch.sqrt(torch.sum(pred_x**2, 1, keepdim=True))
assert (pred_x_radius.data > 0).all()
pred_x_sphere = pred_x / pred_x_radius
satellite = pred_x_sphere * pred_x.size(1)**0.5
one_radius = Variable(torch.ones(1, 1)).type_as(self.train_x)
K_non_ori_radius = self.model.kernel.radius_kernel(
self.train_x_nonorigin_radius, one_radius * 0)
K_non_ori_sphere = self.model.kernel.sphere_kernel(
self.train_x_nonorigin_sphere, pred_x_sphere)
K_non_ori = K_non_ori_radius.view(-1, 1) * K_non_ori_sphere
K_non_pre = self.model.kernel(self.train_x_nonorigin, pred_x)
K_non_sat = self.model.kernel(self.train_x_nonorigin, satellite)
K_ori_pre_diag = self.model.kernel.radius_kernel(
pred_x_radius, one_radius * 0)
K_ori_sat_diag = self.model.kernel.radius_kernel(
one_radius * 0, one_radius * n_dim**0.5).repeat(n_pred, 1)
K_sat_pre_diag = self.model.kernel.radius_kernel(
pred_x_radius, one_radius * n_dim**0.5)
chol_B = torch.cat([
K_non_ori, K_non_pre,
self.mean_vec.index_select(0, self.ind_nonorigin), K_non_sat
], 1)
chol_solver = torch.gesv(chol_B, self.cholesky_nonorigin)[0]
chol_solver_q = chol_solver[:, :n_pred]
chol_solver_k = chol_solver[:, n_pred:n_pred * 2]
chol_solver_y = chol_solver[:, n_pred * 2:n_pred * 2 + 1]
chol_solver_q_bar_0 = chol_solver[:, n_pred * 2 + 1:]
sol_p_sqr = kernel_max + self.model.likelihood(pred_x).view(
-1, 1) + self.jitter - (chol_solver_q**2).sum(0).view(-1, 1)
if not (sol_p_sqr.data >= 0).all():
if not in_optimization:
neg_mask = sol_p_sqr.data < 0
neg_val = sol_p_sqr.data[neg_mask]
min_neg_val = torch.min(neg_val)
max_neg_val = torch.max(neg_val)
kernel_max = self.model.kernel.forward_on_identical().data[0]
print('p')
print('negative %d/%d pred_var range %.4E(%.4E) ~ %.4E(%.4E)' %
(torch.sum(neg_mask), sol_p_sqr.numel(), min_neg_val,
min_neg_val / kernel_max, max_neg_val,
max_neg_val / kernel_max))
print('kernel max %.4E / noise variance %.4E' %
(kernel_max,
torch.exp(self.model.likelihood.log_noise_var.data)[0]))
print('jitter %.4E' % self.jitter)
print('-' * 50)
sol_p = torch.sqrt(sol_p_sqr.clamp(min=1e-12))
sol_k_bar = (
K_ori_pre_diag -
(chol_solver_q * chol_solver_k).sum(0).view(-1, 1)) / sol_p
sol_y_bar = (self.mean_vec.index_select(0, self.ind_origin) -
torch.mm(chol_solver_q.t(), chol_solver_y)) / sol_p
sol_q_bar_1 = (
K_ori_sat_diag -
(chol_solver_q * chol_solver_q_bar_0).sum(0).view(-1, 1)) / sol_p
sol_p_bar_sqr = kernel_max + self.model.likelihood(pred_x).view(
-1, 1) + self.jitter - (chol_solver_q_bar_0**2).sum(0).view(
-1, 1) - (sol_q_bar_1**2)
if not (sol_p_bar_sqr.data >= 0).all():
if not in_optimization:
neg_mask = sol_p_bar_sqr.data < 0
neg_val = sol_p_bar_sqr.data[neg_mask]
min_neg_val = torch.min(neg_val)
max_neg_val = torch.max(neg_val)
kernel_max = self.model.kernel.forward_on_identical().data[0]
print('p bar')
print('negative %d/%d pred_var range %.4E(%.4E) ~ %.4E(%.4E)' %
(torch.sum(neg_mask), sol_p_bar_sqr.numel(), min_neg_val,
min_neg_val / kernel_max, max_neg_val,
max_neg_val / kernel_max))
print('kernel max %.4E / noise variance %.4E' %
(kernel_max,
torch.exp(self.model.likelihood.log_noise_var.data)[0]))
print('jitter %.4E' % self.jitter)
print('-' * 50)
sol_p_bar = torch.sqrt(sol_p_bar_sqr.clamp(min=1e-12))
sol_k_tilde = (K_sat_pre_diag -
(chol_solver_q_bar_0 * chol_solver_k).sum(0).view(
-1, 1) - sol_k_bar * sol_q_bar_1) / sol_p_bar
pred_mean = torch.mm(
chol_solver_k.t(),
chol_solver_y) + sol_k_bar * sol_y_bar + self.model.mean(pred_x)
pred_var = self.model.kernel.forward_on_identical() - (
chol_solver_k**2).sum(0).view(-1,
1) - sol_k_bar**2 - sol_k_tilde**2
if not (pred_var.data >= 0).all():
if not in_optimization:
neg_mask = pred_var.data < 0
neg_val = pred_var.data[neg_mask]
min_neg_val = torch.min(neg_val)
max_neg_val = torch.max(neg_val)
kernel_max = self.model.kernel.forward_on_identical().data[0]
print('predictive variance')
print('negative %d/%d pred_var range %.4E(%.4E) ~ %.4E(%.4E)' %
(torch.sum(neg_mask), pred_var.numel(), min_neg_val,
min_neg_val / kernel_max, max_neg_val,
max_neg_val / kernel_max))
print('kernel max %.4E / noise variance %.4E' %
(kernel_max,
torch.exp(self.model.likelihood.log_noise_var.data)[0]))
print('jitter %.4E' % self.jitter)
print('-' * 50)
numerically_stable = (pred_var.data >= 0).all()
zero_pred_var = (pred_var.data <= 0).all()
if hyper is not None:
self.cholesky_update(param_original)
return pred_mean, pred_var.clamp(
min=1e-12), numerically_stable, zero_pred_var
def gesv_wrapper(return_dict, i, *args): return_dict[i] = torch.gesv(*args)[0]
def get_perspective_transform(src, dst):
r"""Calculates a perspective transform from four pairs of the corresponding
points.
The function calculates the matrix of a perspective transform so that:
.. math ::
\begin{bmatrix}
t_{i}x_{i}^{'} \\
t_{i}y_{i}^{'} \\
t_{i} \\
\end{bmatrix}
=
\textbf{map_matrix} \cdot
\begin{bmatrix}
x_{i} \\
y_{i} \\
1 \\
\end{bmatrix}
where
.. math ::
dst(i) = (x_{i}^{'},y_{i}^{'}), src(i) = (x_{i}, y_{i}), i = 0,1,2,3
Args:
src (Tensor): coordinates of quadrangle vertices in the source image.
dst (Tensor): coordinates of the corresponding quadrangle vertices in
the destination image.
Returns:
Tensor: the perspective transformation.
Shape:
- Input: :math:`(B, 4, 2)` and :math:`(B, 4, 2)`
- Output: :math:`(B, 3, 3)`
"""
if not torch.is_tensor(src):
raise TypeError("Input type is not a torch.Tensor. Got {}".format(
type(src)))
if not torch.is_tensor(dst):
raise TypeError("Input type is not a torch.Tensor. Got {}".format(
type(dst)))
if not src.shape[-2:] == (4, 2):
raise ValueError("Inputs must be a Bx4x2 tensor. Got {}".format(
src.shape))
if not src.shape == dst.shape:
raise ValueError("Inputs must have the same shape. Got {}".format(
dst.shape))
if not (src.shape[0] == dst.shape[0]):
raise ValueError(
"Inputs must have same batch size dimension. Got {}".format(
src.shape, dst.shape))
def ax(p, q):
ones = torch.ones_like(p)[..., 0:1]
zeros = torch.zeros_like(p)[..., 0:1]
return torch.cat([
p[:, 0:1], p[:, 1:2], ones, zeros, zeros, zeros,
-p[:, 0:1] * q[:, 0:1], -p[:, 1:2] * q[:, 0:1]
],
dim=1)
def ay(p, q):
ones = torch.ones_like(p)[..., 0:1]
zeros = torch.zeros_like(p)[..., 0:1]
return torch.cat([
zeros, zeros, zeros, p[:, 0:1], p[:, 1:2], ones,
-p[:, 0:1] * q[:, 1:2], -p[:, 1:2] * q[:, 1:2]
],
dim=1)
# we build matrix A by using only 4 point correspondence. The linear
# system is solved with the least square method, so here
# we could even pass more correspondence
p = []
p.append(ax(src[:, 0], dst[:, 0]))
p.append(ay(src[:, 0], dst[:, 0]))
p.append(ax(src[:, 1], dst[:, 1]))
p.append(ay(src[:, 1], dst[:, 1]))
p.append(ax(src[:, 2], dst[:, 2]))
p.append(ay(src[:, 2], dst[:, 2]))
p.append(ax(src[:, 3], dst[:, 3]))
p.append(ay(src[:, 3], dst[:, 3]))
# A is Bx8x8
A = torch.stack(p, dim=1)
# b is a Bx8x1
b = torch.stack([
dst[:, 0:1, 0],
dst[:, 0:1, 1],
dst[:, 1:2, 0],
dst[:, 1:2, 1],
dst[:, 2:3, 0],
dst[:, 2:3, 1],
dst[:, 3:4, 0],
dst[:, 3:4, 1],
],
dim=1)
# solve the system Ax = b
X, LU = torch.gesv(b, A)
# create variable to return
batch_size = src.shape[0]
M = torch.ones(batch_size, 9, device=src.device, dtype=src.dtype)
M[..., :8] = torch.squeeze(X, dim=-1)
return M.view(-1, 3, 3) # Bx3x3
##############################
res1 = torch.gels(y, X)
print("Solution 1:")
print(res1[0])
# Solution 2
print(torch.matmul(torch.transpose(X, 0, 1),X))
print(torch.matmul(torch.transpose(X, 0, 1),y))
##############################
## How to compute l and r?
## Dimensions: l (2x2); r (2x1)
##############################
l = torch.matmul(torch.transpose(X, 0, 1),X)
r = torch.matmul(torch.transpose(X, 0, 1),y)
res2 = torch.gesv(r,l)
print("Solution 2:")
print(res2[0])
# Solution 3
##############################
## What is l and r?
## Dimensions: l (2x2); r (2x1)
##############################
l = torch.matmul(torch.transpose(X, 0, 1),X)
r = torch.matmul(torch.transpose(X, 0, 1),y)
res3 = torch.matmul(torch.inverse(l),r)
print("Solution 3:")
print(res3)
def solve(matrix1, matrix2):
solution, _ = torch.gesv(matrix2, matrix1)
return solution
sync()
start = time.time()
c = Kinv(x, x, b, g, alpha=alpha)
sync()
end = time.time()
print('Timing (KeOps implementation):', round(end - start, 5), 's')
###############################################################################
# Compare with a straightforward PyTorch implementation:
#
sync()
start = time.time()
K_xx = alpha * torch.eye(N) + torch.exp(-torch.sum(
(x[:, None, :] - x[None, :, :])**2, dim=2) / (2 * sigma**2))
c_py = torch.gesv(b, K_xx)[0]
sync()
end = time.time()
print('Timing (PyTorch implementation):', round(end - start, 5), 's')
print("Relative error = ", (torch.norm(c - c_py) / torch.norm(c_py)).item())
# Plot the results next to each other:
for i in range(Dv):
plt.subplot(Dv, 1, i + 1)
plt.plot(c.cpu().detach().numpy()[:40, i], '-', label='KeOps')
plt.plot(c_py.cpu().detach().numpy()[:40, i], '--', label='PyTorch')
plt.legend(loc='lower right')
plt.tight_layout()
plt.show()
###############################################################################
def compute_beta_cuda(self, X, Y):
XtX, XtY = X.permute(1, 0).mm(X), X.permute(1, 0).mm(Y)
beta_cholesky, _ = torch.gesv(XtY, XtX)
return beta_cholesky
def dynamics(self, z, u, i):
"""Dynamics model function.
Args:
z (Tensor<..., state_size>): State distribution.
u (Tensor<..., action_size>): Action vector(s).
i (Tensor<...>): Time index.
Returns:
derivatives of current state wrt to time (Tensor<..., state_size>).
"""
mc = self.mc if z.dim() == 1 else self.mc.repeat(z.shape[0])
mp1 = self.mp1 if z.dim() == 1 else self.mp1.repeat(z.shape[0])
mp2 = self.mp2 if z.dim() == 1 else self.mp2.repeat(z.shape[0])
l1 = self.l1 if z.dim() == 1 else self.l1.repeat(z.shape[0])
l2 = self.l2 if z.dim() == 1 else self.l2.repeat(z.shape[0])
mu = self.mu if z.dim() == 1 else self.mu.repeat(z.shape[0])
g = self.g if z.dim() == 1 else self.g.repeat(z.shape[0])
x_dot = z[..., 1]
theta1 = z[..., 2]
theta1_dot = z[..., 3]
theta2 = z[..., 4]
theta2_dot = z[..., 5]
dtheta = theta1 - theta2
F = u[..., 0]
angles = torch.tensor([theta1, theta2, dtheta])
sin_theta1, sin_theta2, sin_dtheta = angles.sin()
cos_theta1, cos_theta2, cos_dtheta = angles.cos()
a0 = mp2 + 2 * mc
a1 = mc * l2
a2 = l1 * theta1_dot**2
a3 = a1 * theta2_dot**2
# yapf: disable
A = torch.stack([
torch.stack([
2 * (mp1 + mp2 + mc),
-a0 * l1 * cos_theta1,
-a1 * cos_theta2
], dim=-1),
torch.stack([
-3 * a0 * cos_theta1,
(2 * a0 + 2 * mc) * l1,
3 * a1 * cos_dtheta
], dim=-1),
torch.stack([
-3 * cos_theta2,
3 * l1 * cos_dtheta,
2 * l2
], dim=-1),
], dim=-1).transpose(-2, -1)
b = torch.stack([
torch.stack([
- 2 * mu * x_dot - a0 * a2 * sin_theta1 - a3 * sin_theta2
], dim=-1),
torch.stack([
3 * a0 * g * sin_theta1 - 3 * a3 * sin_dtheta
], dim=-1),
torch.stack([
6*F/(l2*mp2) + 3 * a2 * sin_dtheta + 3 * g * sin_theta2
], dim=-1),
], dim=-1).transpose(-2, -1)
# yapf: enable
sol = torch.gesv(b, A)[0].transpose(-2, -1)
# For symplectic integration.
x_dot_dot = sol[..., 0].view(x_dot.shape)
theta1_dot_dot = sol[..., 1].view(theta1_dot.shape)
theta2_dot_dot = sol[..., 2].view(theta2_dot.shape)
return torch.stack([
x_dot,
x_dot_dot,
theta1_dot,
theta1_dot_dot,
theta2_dot,
theta2_dot_dot,
],
dim=-1)
def forward(self,
x_train,
y_train,
body_train,
x_test=None,
body_test=None,
classify=False):
# See the autograd section for explanation of what happens here.
n = x_train.size(0)
p = x_train.size(-1)
d = torch.zeros(n, n)
nB = body_train.size(1) # number of regions/bodies
if classify: # i.e we are predicting not training
body_train = body_train > 0.5
nullB = torch.sqrt(1 - torch.sum(body_train.float().pow(2), 1))
for i in range(p):
d += 0.5 * (x_train[:, i].unsqueeze(1) - x_train[:, i].unsqueeze(0)
).pow(2) / self.lengthscale[i].pow(2)
kse = self.sigma_f.pow(2) * torch.exp(-d) + self.sigma_n.pow(
2) * torch.eye(n)
kyy = nullB.unsqueeze(0) * nullB.unsqueeze(1) * kse
for i in range(nB):
kyy += body_train[:, i].float().unsqueeze(
1) * body_train[:, i].float().unsqueeze(0) * kse
c = torch.cholesky(kyy, upper=True)
# v = torch.potrs(y_train, c, upper=True)
v, _ = torch.gesv(y_train, kyy)
if x_test is None:
out = (c, v)
if x_test is not None:
with torch.no_grad():
if classify:
body_test = body_test > 0.5 # make a distinct classifier
nullB_test = torch.sqrt(1 -
torch.sum(body_test.float().pow(2), 1))
ntest = x_test.size(0)
d = torch.zeros(ntest, n)
for i in range(p):
d += 0.5 * (x_test[:, i].unsqueeze(1) -
x_train[:, i].unsqueeze(0)
).pow(2) / self.lengthscale[i].pow(2)
kse = self.sigma_f.pow(2) * torch.exp(-d)
kfy = nullB_test.unsqueeze(1) * nullB.unsqueeze(0) * kse
for i in range(nB):
kfy += body_test[:, i].float().unsqueeze(
1) * body_train[:, i].float().unsqueeze(0) * kse
# solve
f_test = kfy.mm(v)
tmp = torch.potrs(kfy.t(), c, upper=True)
tmp = torch.sum(kfy * tmp.t(), dim=1)
cov_f = self.sigma_f.pow(2) - tmp
out = (f_test, cov_f)
return out
global_folder = os.path.join(data_folder, 'global')
for dx in dx_list:
# Set subject of observations based on the number of features
for exp_i in range(n_experiments):
# Set rho
# rho = rho_vec[exp_i]
rho = 0.001
m = n_centers[exp_i // n_sub_exp]
# Create data
X = torch.randn(n, dx).type(dtype)
W = torch.randn(dy, dx).t().type(dtype)
Y = torch.mm(X, W).type(dtype) + 2 * torch.randn(n, dy).type(dtype)
X_t_X, LU = torch.gesv(
torch.eye(dx).type(dtype), torch.mm(X.t(), X))
W_full_data = torch.mm(X_t_X, torch.mm(X.t(), Y))
# print('Plotting')
# plt.scatter(Y.numpy()[0], torch.mm(X, W).numpy()[0], alpha=0.2, color='b')
# plt.show()
global_data = {'X': X, 'W': W, 'Y': Y}
# Print info
print('\n[ INFO ] ==== Experiment %d ====' % exp_i)
print('[ INFO ] Global matrices info:')
for key, data in global_data.items():
print('\t\t- Shape of matrix %s: %s' % (key, str(data.shape)))
print('\t\t- Number of observations: ', n)
def __getitem__(self, index):
image_path = self.image_paths[index]
# albedo_path = self.albedo_paths[index]
# mask_path = self.mask_paths[index]
if self.opt.direction == 'BtoA':
input_nc = self.opt.output_nc
output_nc = self.opt.input_nc
else:
input_nc = self.opt.input_nc
output_nc = self.opt.output_nc
content = sio.loadmat(image_path)
if self.isTrain:
rgb_img = content['imag']
chrom = content['chrom']
mask = content['mask']
mask = resize(mask, [384, 512], 1)
mask[mask > 0] = 1
mask = np.mean(mask, axis=2)
#mask = skimage.morphology.binary_erosion(mask, square(1))
mask = np.expand_dims(mask, axis=2)
mask = np.repeat(mask, 3, axis=2)
l1 = content['l1']
l2 = content['l2']
rand_id = random.randint(0, 3)
light_id1 = random.randint(0, 8)
light_id2 = random.randint(0, 8)
arr_id = random.randint(0, 1)
img1 = content['im1']
img2 = content['im2']
img1 = np.nan_to_num(img1)
img2 = np.nan_to_num(img2)
img1[img1 > 1.0] = 1.0
img1[img1 < 0.0] = 0.0
img2[img2 > 1.0] = 1.0
img2[img2 < 0.0] = 0.0
[img1, img2, l1,
l2] = self.produceColor(img1, img2, arr_id, light_id1, light_id2,
l1, l2)
# img2 = self.DA(img2, rand_id)
#img1 = img1/2
#img2 = img2/2
chrom = np.nan_to_num(chrom)
#rgb_img = np.nan_to_num(rgb_img)
#l1 = np.nan_to_num(l1)
#l2 = np.nan_to_num(l2)
# img1[img1 != img1] = 0.0
# img2[img2 != img2] = 0.0
# chrom[chrom != chrom] = 0.0
# rgb_img[rgb_img != rgb_img] = 0.0
#for i in range(3):
# img1[:, :, i] = img1[:, :, i] * l1[i]
# img2[:, :, i] = img2[:, :, i] * l2[i]
rgb_img = img1 + img2
#rgb_img[rgb_img > 1.0] = 1.0
#rgb_img[rgb_img < 0.0] = 0.0
chrom[chrom > 1.0] = 1.0
chrom[chrom < 0.0] = 0.0
rgb_img = resize(rgb_img, [384, 512], 1)
img1 = resize(img1, [384, 512], 1)
img2 = resize(img2, [384, 512], 1)
chrom = resize(chrom, [384, 512], 1)
rgb_img = self.DA(rgb_img, rand_id)
chrom = self.DA(chrom, rand_id)
mask = self.DA(mask, rand_id)
img1 = self.DA(img1, rand_id)
img2 = self.DA(img2, rand_id)
# if arr_id:
# l1 = self.l1Matrix[light_id1]/255
# l2 = self.l2Matrix[light_id2]/255
# l1 = np.reshape(l1, (3, -1))
# l2 = np.reshape(l2, (3, -1))
# else:
# l1 = self.l2Matrix[light_id2]/255
# l2 = self.l1Matrix[light_id1]/255
# l1 = np.reshape(l1, (3, -1))
# l2 = np.reshape(l2, (3, -1))
# rgb_img = img1 + img2
lightColor = np.concatenate((l1, l2), axis=1)
lightColor = torch.from_numpy(lightColor).contiguous().float()
rgb_img = torch.from_numpy(np.transpose(
rgb_img, (2, 0, 1))).contiguous().float()
chrom = torch.from_numpy(np.transpose(
chrom, (2, 0, 1))).contiguous().float()
mask = torch.from_numpy(np.transpose(
mask.astype(float), (2, 0, 1))).contiguous().float()
no_albedo_nf = rgb_img / (1e-6 + chrom)
sum_albedo = torch.sum(no_albedo_nf, 0, keepdim=True)
gamma = no_albedo_nf / (sum_albedo.repeat(3, 1, 1) + 1e-6)
gamma = gamma.view(3, -1)
lightT = lightColor.t()
light = lightColor
B = torch.mm(lightT, gamma)
A = torch.mm(lightT, light)
shadings, _ = torch.gesv(B, A)
# shadings[0, :] = (shadings[0:, ] - torch.min(shadings[0, :]))/(torch.max(shadings[0:, ]) - torch.min(shadings[0, :]))
# shadings[1, :] = (shadings[1:, ] - torch.min(shadings[1, :]))/(torch.max(shadings[1:, ]) - torch.min(shadings[1, :]))
shadings[shadings != shadings] = 0.0
im1 = shadings[0, :].repeat(3, 1).view(3, rgb_img.size(1),
rgb_img.size(2))
im2 = shadings[1, :].repeat(3, 1).view(3, rgb_img.size(1),
rgb_img.size(2))
im1 = (im1 - torch.min(im1[mask > 0])) / (
torch.max(im1[mask > 0]) - torch.min(im1[mask > 0]))
im2 = (im2 - torch.min(im2[mask > 0])) / (
torch.max(im2[mask > 0]) - torch.min(im2[mask > 0]))
# remove nan values
im1[im1 != im1] = 0.0
im2[im2 != im2] = 0.0
im1[mask == 0] = 0.0
im2[mask == 0] = 0.0
im1[0, :, :] *= lightColor[0, 0]
im1[1, :, :] *= lightColor[1, 0]
im1[2, :, :] *= lightColor[2, 0]
im2[0, :, :] *= lightColor[0, 1]
im2[1, :, :] *= lightColor[1, 1]
im2[2, :, :] *= lightColor[2, 1]
# normalize the data
# im1 = torch.mul(shadings[0, :].repeat(3, 1), lightColor[:, 0].view(3, 1).repeat(1, shadings.size(1)))
# im2 = torch.mul(shadings[1, :].repeat(3, 1), lightColor[:, 1].view(3, 1).repeat(1, shadings.size(1)))
# im1 = im1.view(3, rgb_img.size(1), rgb_img.size(2))
# im2 = im2.view(3, rgb_img.size(1), rgb_img.size(2))
im1[im1 > 1] = 1.0
im2[im2 > 1] = 1.0
im1[im1 < 0] = 0.0
im2[im2 < 0] = 0.0
# rgb_img = 2*rgb_img - 1.0
img1 = torch.from_numpy(np.transpose(
img1, (2, 0, 1))).contiguous().float()
img2 = torch.from_numpy(np.transpose(
img2, (2, 0, 1))).contiguous().float()
return {
'rgb_img': rgb_img,
'chrom': chrom,
'im1': im1,
'im2': im2,
'A_paths': image_path,
'mask': mask,
'img1': img1,
'img2': img2
}
else:
rgb_img = content['imag']
#chrom = content['chrom']
#mask = content['chrom']
#im1 = content['im1']
#im2 = content['im2']
rgb_img[rgb_img > 1] = 1
rgb_img[rgb_img < 0] = 0
# rgb_img = 2*rgb_img - 1.0
#img1 = content['im1']
#img2 = content['im2']
rgb_img = torch.from_numpy(np.transpose(
rgb_img, (2, 0, 1))).contiguous().float()
#chrom = torch.from_numpy(np.transpose(chrom, (2, 0, 1))).contiguous().float()
#mask = torch.from_numpy(np.transpose(mask.astype(float), (2, 0, 1))).contiguous().float()
#im1 = torch.from_numpy(np.transpose(im1, (2, 0, 1))).contiguous().float()
#im2 = torch.from_numpy(np.transpose(im2, (2, 0, 1))).contiguous().float()
return {'rgb_img': rgb_img, 'A_paths': image_path}
def solve(A, b):
return torch.gesv(b, A)[0].contiguous()
def manitest(input_image,
net,
mode,
maxIter=50000,
lim=None,
hs=None,
cuda_on=True,
stop_when_found=None,
verbose=True):
def list_index(a_list, inds):
return [a_list[i] for i in inds]
def group_chars(mode):
if mode == 'rotation':
hs = torch.Tensor([pi / 20])
elif mode == 'translation':
hs = torch.Tensor([0.25, 0.25])
elif mode == 'rotation+scaling':
hs = torch.Tensor([pi / 20, 0.1])
elif mode == 'rotation+translation':
hs = torch.Tensor([pi / 20, 0.25, 0.25])
elif mode == 'scaling+translation':
hs = torch.Tensor([0.1, 0.25, 0.25])
elif mode == 'similarity':
hs = torch.Tensor([pi / 20, 0.5, 0.5, 0.1])
else:
raise NameError('Wrong mode name entered')
if cuda_on:
hs.cuda()
return hs
def gen_simplices(cur_vec, cur_dim):
nonlocal num_simpl
nonlocal simpls
if cur_dim == dimension_group + 1:
if not cur_vec:
return
simpls.append(cur_vec)
num_simpl = num_simpl + 1
return
if (n_vec[2 * cur_dim - 2] == i or n_vec[2 * cur_dim - 1] == i):
cur_vec = cur_vec + [i]
gen_simplices(cur_vec, cur_dim + 1)
else:
gen_simplices(cur_vec, cur_dim + 1)
if (n_vec[2 * cur_dim - 2] != -1):
cur_vec_l = cur_vec + [n_vec[2 * cur_dim - 2]]
gen_simplices(cur_vec_l, cur_dim + 1)
if (n_vec[2 * cur_dim - 1] != -1):
cur_vec_r = cur_vec + [n_vec[2 * cur_dim - 1]]
gen_simplices(cur_vec_r, cur_dim + 1)
def check_oob(coord):
inside = 1
for u in range(len(coord)):
if (coord[u] > lim[u, 1] + 1e-8 or coord[u] < lim[u, 0] - 1e-8):
inside = 0
break
return inside
def get_and_create_neighbours(cur_node):
nonlocal id_max, coords, dist, visited, neighbours, W, ims
#1)Generate coordinates of neighbouring nodes
for l in range(dimension_group):
#Generate coordinates
coordsNeighbour1 = coords[cur_node].clone()
coordsNeighbour1[l] += hs[l]
coordsNeighbour2 = coords[cur_node].clone()
coordsNeighbour2[l] -= hs[l]
if check_oob(coordsNeighbour1):
#Can we find a similar coordinate?
dists = (torch.stack(coords, 0) -
coordsNeighbour1.repeat(len(coords), 1)).abs().sum(
dim=1)
II1 = (dists < 1e-6).nonzero()
if not II1.size():
id_max += 1
#create node: i) coords, ii)visited, iii)distance
coords.append(coordsNeighbour1)
dist.append(np.inf)
visited.append(0)
#Assing the NodeID to IDNeighbours
neighbours.append([-1] * 2 * dimension_group)
neighbours[cur_node][2 * l] = id_max
#Do the reverse
neighbours[id_max][2 * l + 1] = cur_node
W.append(None)
ims.append([])
else:
#Node already exists
neighbours[cur_node][2 * l] = II1[0, 0]
#Do the reverse
neighbours[II1[0, 0]][2 * l + 1] = cur_node
if check_oob(coordsNeighbour2):
#Can we find a similar coordinate?
dists = (torch.stack(coords, 0) -
coordsNeighbour2.repeat(len(coords), 1)).abs().sum(
dim=1)
II2 = (dists < 1e-6).nonzero()
if not II2.size():
id_max += 1
#create node: i) coords, ii)visited, iii)distance
coords.append(coordsNeighbour2)
dist.append(np.inf)
visited.append(0)
#Assing the NodeID to IDNeighbours
neighbours.append([-1] * 2 * dimension_group)
neighbours[cur_node][2 * l + 1] = id_max
#Do the reverse
neighbours[id_max][2 * l] = cur_node
W.append(None)
ims.append([])
else:
#Node already exists
neighbours[cur_node][2 * l + 1] = II2[0, 0]
#Do the reverse
neighbours[II2[0, 0]][2 * l] = cur_node
def generate_metric(cur_node):
nonlocal ims, W
tau = coords[cur_node]
tfm = g.para2tfm(tau, mode, 1)
I = tfm(input_image)
ims[cur_node] = I
J = g.jacobian(input_image, I, tfm, mode, 1)
J_n = J.resize_(J.size()[0], n)
curW = J_n.mm(J_n.transpose(0, 1))
W[cur_node] = curW
def evaluate_classifier(cur_node):
nonlocal manitest_score, manitest_image, fooling_tfm, out_label
x = Variable(ims[cur_node].unsqueeze(0))
output = net(x)
_, k_I = torch.max(output.data, 1)
pred_label = k_I[0]
if pred_label != input_label:
manitest_score = dist[cur_node] / input_image.norm()
manitest_image = ims[cur_node]
fooling_tfm = g.para2tfm(coords[cur_node], mode, 1)
out_label = pred_label
return True
return False
###
e = g.init_param(mode)
if cuda_on:
net.cuda()
input_image = input_image.cuda()
e = e.cuda()
dimension_group = e.size()[0]
n = functools.reduce(operator.mul, input_image.size(), 1)
stop_flag = False
point_dists = None
if stop_when_found is not None:
stop_flag = True
num_stopping_points = stop_when_found.size()[0]
point_dists = torch.Tensor(num_stopping_points)
remaining_points = num_stopping_points
if hs is None:
hs = group_chars(mode)
if lim is None:
lim = np.zeros((dimension_group, 2))
lim[:, 0] = -np.inf
lim[:, 1] = np.inf
dist = [0.0]
visited = [0]
coords = [e]
ims = [input_image]
W = [None]
id_max = 0
neighbours = [[-1] * 2 * dimension_group]
#Generate input label
x = Variable(input_image.unsqueeze(0))
output = net(x)
_, k_I = torch.max(output.data, 1)
input_label = k_I[0]
#Output Variables
manitest_score = np.inf
manitest_image = input_image.clone()
fooling_tfm = e
out_label = input_label
for k in range(maxIter):
if k % 100 == 0 and verbose:
print('>> k = {}'.format(k))
tmp_vec = np.array(dist[0:id_max + 1]) #copy the list
tmp_vec[np.asarray(visited) == 1] = np.inf
i = np.argmin(tmp_vec)
visited[i] = 1
#evaluate the classifier and check if it is fooled
if stop_flag:
dists = torch.norm(
coords[i].repeat(num_stopping_points, 1) - stop_when_found, 2,
1)
if dists.min() < 1e-6:
_, ind = torch.min(dists, 0)
point_dists[ind[0, 0]] = dist[i]
remaining_points -= 1
if remaining_points == 0:
break
elif evaluate_classifier(i):
break
get_and_create_neighbours(i)
for j in neighbours[i]:
if j == -1:
continue
#Consider unknown neighbours only
if visited[j]:
continue
#Look at the neighbours of j (vector of size 2*dimension_group)
n_vec = neighbours[j]
num_simpl = 1
simpls = []
gen_simplices([], 1)
if W[j] is None:
generate_metric(j)
for j_ in range(num_simpl - 1):
X = torch.stack(list_index(coords,
simpls[j_])) - coords[j].repeat(
len(simpls[j_]), 1)
if cuda_on:
v = torch.cuda.FloatTensor(list_index(
dist, simpls[j_])).unsqueeze(1)
one_vector = torch.ones(v.size()).cuda()
else:
v = torch.FloatTensor(list_index(dist,
simpls[j_])).unsqueeze(1)
one_vector = torch.ones(v.size())
M_prime = (X.mm(W[j]).mm(X.transpose(0, 1)))
try:
invM_prime_v, _ = torch.gesv(v, M_prime)
except:
invM_prime_v = v * np.inf
try:
invM_prime_1, _ = torch.gesv(one_vector, M_prime)
except:
invM_prime_1 = one_vector * np.inf
invM_prime_v.transpose_(0, 1)
# one_vector.squeeze_()
# v.squeeze_()
#Solve second order equation
# dz^2 * one_vector' * invM_prime * one_vector
# - 2 * dz * one_vector' * invM_prime * v + v' * invM_prime * v - 1
Delta = (invM_prime_v.sum()
)**2 - invM_prime_1.sum() * (invM_prime_v.mm(v) - 1)
Delta = Delta[0, 0]
if Delta >= 0:
#Compute solution
x_c = (invM_prime_v.sum() +
np.sqrt(Delta)) / invM_prime_1.sum()
#Test that it is not on the border of the simplex
te, _ = torch.gesv(x_c - v, M_prime)
if te.min() > 0:
dist[j] = min(dist[j], x_c)
return manitest_score, manitest_image, fooling_tfm, dist, coords, input_label, out_label, point_dists, k
def conditional(Xnew,
X,
kern,
f,
full_cov=False,
q_sqrt=None,
whiten=False,
jitter_level=1e-6):
"""
Given F, representing the GP at the points X, produce the mean and
(co-)variance of the GP at the points Xnew.
Additionally, there may be Gaussian uncertainty about F as represented by
q_sqrt. In this case `f` represents the mean of the distribution and
q_sqrt the square-root of the covariance.
Additionally, the GP may have been centered (whitened) so that
p(v) = N( 0, I)
f = L v
thus
p(f) = N(0, LL^T) = N(0, K).
In this case 'f' represents the values taken by v.
The method can either return the diagonals of the covariance matrix for
each output of the full covariance matrix (full_cov).
We assume K independent GPs, represented by the columns of f (and the
last dimension of q_sqrt).
:param Xnew: data matrix, size N x D.
:param X: data points, size M x D.
:param kern: GPflow kernel.
:param f: data matrix, M x K, representing the function values at X,
for K functions.
:param q_sqrt: matrix of standard-deviations or Cholesky matrices,
size M x K or M x M x K.
:param whiten: boolean of whether to whiten the representation as
described above.
:return: two element tuple with conditional mean and variance.
"""
# compute kernel stuff
num_data = X.size(0) # M
num_func = f.size(1) # K
Kmn = kern.K(X, Xnew)
Kmm = kern.K(X) + Variable(torch.eye(num_data,
out=X.data.new())) * jitter_level
Lm = torch.potrf(Kmm, upper=False)
# Compute the projection matrix A
A, _ = torch.gesv(Kmn, Lm)
# compute the covariance due to the conditioning
if full_cov:
fvar = kern.K(Xnew) - torch.matmul(A.t(), A)
fvar = fvar.unsqueeze(0).expand(num_func, -1, -1) # K x N x N
else:
fvar = kern.Kdiag(Xnew) - (A**2).sum(0)
fvar = fvar.unsqueeze(0).expand(num_func, -1) # K x N
# fvar is K x N x N or K x N
# another backsubstitution in the unwhitened case (complete the inverse of the cholesky decomposition)
if not whiten:
A, _ = torch.gesv(A, Lm.t())
# construct the conditional mean
fmean = torch.matmul(A.t(), f)
if q_sqrt is not None:
if q_sqrt.dim() == 2:
LTA = A * q_sqrt.t().unsqueeze(2) # K x M x N
elif q_sqrt.dim() == 3:
L = batch_tril(q_sqrt.permute(2, 0, 1)) # K x M x M
# A_tiled = tf.tile(tf.expand_dims(A, 0), tf.stack([num_func, 1, 1])) # I don't think I need this
LTA = torch.matmul(L.transpose(-2, -1), A) # K x M x N
else: # pragma: no cover
raise ValueError("Bad dimension for q_sqrt :{}".format(
q_sqrt.dim()))
if full_cov:
fvar = fvar + torch.matmul(LTA.t(), LTA) # K x N x N
else:
fvar = fvar + (LTA**2).sum(1) # K x N
fvar = fvar.permute(*range(fvar.dim() - 1, -1, -1)) # N x K or N x N x K
return fmean, fvar
def Linear_Time_Iteration(A, B, C, F_initial, mu, epsilon):
"""
This function will find the linear time iteration solution to the system of equations in the form of
AX(-1) + BX + CE[X(+1)] + epsilon = 0
with a recursive solution in the form of X = FX(-1) + Q*epsilon
Parameters
----------
A : torch, array_like, dtype=float
The matrix of coefficients next to endogenous variables entering with a lag
B : torch, array_like, dtype=float
The matrix of coefficients next to endogenous, contemporanous variables
C : torch, array_like, dtype=float
The matrix of coefficients next to endogenous variables entering with a lead
F : torch, array_like, dtype=float
The initial guess for F
mu : number, dtype=float
Small positive real number to be multiplied by a conformable identity matrix
epsilon : number, dtype=float
Threshold value, should be set to a small value like 1e-16
Returns
-------
F : torch, array_like, dtype=float
The matrix of coefficients next to the endogenous variable in the solution
Q : torch, array_like, dtype=float
The matrix of coefficients next to the disturbance term in the solution
Notes
-----
"""
F = F_initial
S = zeros(*A.shape)
# F.requires_grad_()
# S.requires_grad_()
Id = eye(*A.shape) * mu
Ch = C
Bh = (B + 2 * mm(C, Id))
Ah = (mm(C, matrix_power(Id, 2)) + mm(B, Id) + A)
metric = 1
iter = 1
while metric > epsilon:
if iter % 10000 == 0:
print(iter)
F = -gesv(Ah, (Bh + mm(Ch, F)))[0]
S = -gesv(Ch, (Bh + mm(Ah, S)))[0]
metric1 = max(abs(Ah + mm(Bh, F) + mm(Ch, (mm(F, F)))))
metric2 = max(abs(mm(Ah, mm(S, S)) + mm(Bh, S) + Ch))
metric = max(metric1, metric2)
iter += 1
if iter > 1000000:
break
# eig_F = max(abs(eig(F)[0]))
# eig_S = max(abs(eig(S)[0]))
# if (eig_F > 1) or (eig_S > 1) or (mu > 1-eig_S):
# print('Conditions of Proposition 3 violated')
F = F + Id
Q = -inverse(B + mm(C, F))
return F, Q
def predict(self, pred_x, hyper=None):
if hyper is not None:
param_original = self.model.param_to_vec()
self.model.vec_to_param(hyper)
kernel_on_input_map = deepcopy(self.model.kernel)
kernel_on_input_map.input_map = id_transform
train_x_input_map = self.model.kernel.input_map(self.train_x)
pred_x_input_map = self.model.kernel.input_map(pred_x)
train_origin_point_mask = train_x_input_map.data[:, 0] == 0
n_train_origin = torch.sum(train_origin_point_mask)
n_train_other = self.train_x.size(0) - n_train_origin
train_point_ind = torch.sort(train_origin_point_mask,
0,
descending=True)[1]
train_origin_ind = train_point_ind[:n_train_origin]
train_other_ind = train_point_ind[n_train_origin:]
train_x_other = self.train_x[train_other_ind]
train_x_origin = self.train_x[train_origin_ind]
train_x_other_input_map = train_x_input_map[train_other_ind]
train_y_other = self.train_y[train_other_ind]
train_y_origin = self.train_y[train_origin_ind]
eye_mat = Variable(torch.eye(n_train_other)).type_as(self.train_x)
K_other_noise = kernel_on_input_map(
train_x_other_input_map) + torch.diag(
self.model.likelihood(train_x_other))
K_other_noise_inv, _ = torch.gesv(eye_mat, K_other_noise)
mean_vec_other = train_y_other - self.model.mean(train_x_other)
mean_vec_origin = train_y_origin - self.model.mean(train_x_origin)
k_pred_other = kernel_on_input_map(pred_x_input_map,
train_x_other_input_map)
shared_part = k_pred_other.mm(K_other_noise_inv)
pred_mean = torch.mm(shared_part,
mean_vec_other) + self.model.mean(pred_x)
pred_var = self.model.kernel.forward_on_identical() - (
shared_part * k_pred_other).sum(1, keepdim=True)
slide_origin = pred_x_input_map.clone()
slide_origin[:, 0] = 0
K_other_origin = kernel_on_input_map(train_x_other_input_map,
slide_origin)
Ainv_B = K_other_noise_inv.mm(K_other_origin)
identity_part = self.model.likelihood(self.train_x[:1] * 0)
onemat_part = self.model.kernel.forward_on_identical() - (
Ainv_B * K_other_origin).sum(0).view(-1, 1)
identity_const = 1.0 / identity_part
onemat_const = -onemat_part / identity_part / (
identity_part + onemat_part * n_train_origin)
k_pred_origin = torch.cat([
kernel_on_input_map(pred_x_input_map[i:i + 1],
slide_origin[i:i + 1])
for i in range(pred_x.size(0))
], 0)
k_vec = (k_pred_other * Ainv_B.t()).sum(1).view(-1, 1) - k_pred_origin
y_vec = torch.mm(mean_vec_other.t(), Ainv_B).view(
-1, 1) - torch.mean(mean_vec_origin)
slide_origin_mean_adjustment = (
identity_const +
onemat_const * n_train_origin) * k_vec * y_vec * n_train_origin
slide_origin_quad_adjustment = (
identity_const +
onemat_const * n_train_origin) * k_vec**2 * n_train_origin
if hyper is not None:
self.model.vec_to_param(param_original)
return pred_mean + slide_origin_mean_adjustment, pred_var - slide_origin_quad_adjustment
def b_inv(A): eye = A.new_ones(A.size(-1)).diag().expand_as(A) b_inv, _ = torch.gesv(eye, A) return b_inv
def bgesv(B, A):
return torch.stack([torch.gesv(b, a)[0] for b, a in zip(B, A)])
Xt = torch.FloatTensor(X).cuda()
yt = torch.FloatTensor(y).cuda()
# cur_time = 0
# for i in range(10):
# #log_reg = LogisticRegression(solver=opt, random_state=123)
# start = time.time()
# thetas_np = train_logistic(X, y)
# #thetas = train_logistic_torch(Xt, yt)
#
# #log_reg.fit(X, y)
# cur_time += time.time()-start
# print("Time is {}".format(cur_time/10))
cur_time = 0
pre = torch.gesv(Xt.t(), Xt.t() @ Xt)[0]
cov = Xt.t() @ Xt
for i in range(10):
#log_reg = LogisticRegression(solver=opt, random_state=123)
start = time.time()
# thetas_np = train_logistic(X, y)
theta = torch.gesv((Xt.t() @ yt).view(-1, 1), cov)
#log_reg.fit(X, y)
cur_time += time.time() - start
print("Time is {}".format(cur_time / 10))
cur_time = 0
for i in range(10):
#log_reg = LogisticRegression(solver=opt, random_state=123)
start = time.time()
# thetas_np = train_logistic(X, y)
thetas = train_logistic_torch(Xt, yt)
def _depth_warping(depth_maps_1, depth_maps_2, img_masks, translation_vectors,
rotation_matrices, intrinsic_matrices, epsilon):
# Generate a meshgrid for each depth map to calculate value
# BxHxWxC
depth_maps_1 = torch.mul(depth_maps_1, img_masks)
depth_maps_2 = torch.mul(depth_maps_2, img_masks)
depth_maps_1 = depth_maps_1.permute(0, 2, 3, 1)
depth_maps_2 = depth_maps_2.permute(0, 2, 3, 1)
img_masks = img_masks.permute(0, 2, 3, 1)
num_batch, height, width, channels = depth_maps_1.shape
y_grid, x_grid = torch.meshgrid([
torch.arange(start=0, end=height, dtype=torch.float32).cuda(),
torch.arange(start=0, end=width, dtype=torch.float32).cuda()
])
x_grid = x_grid.view(1, height, width, 1)
y_grid = y_grid.view(1, height, width, 1)
ones_grid = torch.ones((1, height, width, 1), dtype=torch.float32).cuda()
# intrinsic_matrix_inverse = intrinsic_matrix.inverse()
eye = torch.eye(3).float().cuda().view(1, 3, 3).expand(
intrinsic_matrices.shape[0], -1, -1)
intrinsic_matrices_inverse, _ = torch.gesv(eye, intrinsic_matrices)
rotation_matrices_inverse = rotation_matrices.transpose(1, 2)
# The following is when we have different intrinsic matrices for samples within a batch
temp_mat = torch.bmm(intrinsic_matrices, rotation_matrices_inverse)
W = torch.bmm(temp_mat, -translation_vectors)
M = torch.bmm(temp_mat, intrinsic_matrices_inverse)
mesh_grid = torch.cat((x_grid, y_grid, ones_grid),
dim=-1).view(height, width, 3, 1)
intermediate_result = torch.matmul(M.view(-1, 1, 1, 3, 3),
mesh_grid).view(-1, height, width, 3)
depth_maps_2_calculate = W.view(-1, 3).narrow(
dim=-1, start=2, length=1).view(-1, 1, 1, 1) + torch.mul(
depth_maps_1,
intermediate_result.narrow(dim=-1, start=2, length=1).view(
-1, height, width, 1))
# expand operation doesn't allocate new memory (repeat does)
depth_maps_2_calculate = torch.where(img_masks > 0.5,
depth_maps_2_calculate, epsilon)
depth_maps_2_calculate = torch.where(depth_maps_2_calculate > 0.0,
depth_maps_2_calculate, epsilon)
# This is the source coordinate in coordinate system 2 but ordered in coordinate system 1 in order to warp image 2 to coordinate system 1
u_2 = (W.view(-1, 3).narrow(dim=-1, start=0, length=1).view(-1, 1, 1, 1) +
torch.mul(
depth_maps_1,
intermediate_result.narrow(dim=-1, start=0, length=1).view(
-1, height, width, 1))) / (depth_maps_2_calculate)
v_2 = (W.view(-1, 3).narrow(dim=-1, start=1, length=1).view(-1, 1, 1, 1) +
torch.mul(
depth_maps_1,
intermediate_result.narrow(dim=-1, start=1, length=1).view(
-1, height, width, 1))) / (depth_maps_2_calculate)
W_2 = torch.bmm(intrinsic_matrices, translation_vectors)
M_2 = torch.bmm(torch.bmm(intrinsic_matrices, rotation_matrices),
intrinsic_matrices_inverse)
temp = torch.matmul(M_2.view(-1, 1, 1, 3, 3), mesh_grid).view(
-1, height, width, 3).narrow(dim=-1, start=2,
length=1).view(-1, height, width, 1)
depth_maps_1_calculate = W_2.view(-1, 3).narrow(
dim=-1, start=2, length=1).view(-1, 1, 1, 1) + torch.mul(
depth_maps_2, temp)
depth_maps_1_calculate = torch.mul(img_masks, depth_maps_1_calculate)
u_2_flat = u_2.view(-1)
v_2_flat = v_2.view(-1)
warped_depth_maps_2 = _bilinear_interpolate(depth_maps_1_calculate,
u_2_flat, v_2_flat).view(
num_batch, 1, height,
width)
# binarize
intersect_masks = torch.where(
_bilinear_interpolate(img_masks, u_2_flat, v_2_flat) * img_masks >=
0.9,
torch.tensor(1.0).float().cuda(),
torch.tensor(0.0).float().cuda()).view(num_batch, 1, height, width)
return [warped_depth_maps_2, intersect_masks]
def chol_solve(B, A):
c = torch.potrf(A)
s1 = torch.gesv(B, c.transpose(0, 1))[0]
s2 = torch.gesv(s1, c)[0]
return s2
def get_fantasy_strategy(self, inputs, targets, full_inputs, full_targets, full_output):
"""
Returns a new PredictionStrategy that incorporates the specified inputs and targets as new training data.
This method is primary responsible for updating the mean and covariance caches. To add fantasy data to a
GP model, use the :meth:`~gpytorch.models.ExactGP.get_fantasy_model` method.
Args:
- :attr:`inputs` (Tensor `m x d` or `b x m x d`): Locations of fantasy observations.
- :attr:`targets` (Tensor `m` or `b x m`): Labels of fantasy observations.
- :attr:`full_inputs` (Tensor `n+m x d` or `b x n+m x d`): Training data concatenated with fantasy inputs
- :attr:`full_targets` (Tensor `n+m` or `b x n+m`): Training labels concatenated with fantasy labels.
- :attr:`full_output` (:class:`gpytorch.distributions.MultivariateNormal`): Prior called on full_inputs
Returns:
- :class:`DefaultPredictionStrategy`
A `DefaultPredictionStrategy` model with `n + m` training examples, where the `m` fantasy examples have
been added and all test-time caches have been updated.
"""
full_mean, full_covar = full_output.mean, full_output.lazy_covariance_matrix
batch_shape = full_inputs[0].shape[:-2]
full_mean = full_mean.view(*batch_shape, -1)
num_train = self.num_train
# Evaluate fant x train and fant x fant covariance matrices, leave train x train unevaluated.
fant_fant_covar = full_covar[..., num_train:, num_train:]
fant_mean = full_mean[..., num_train:]
mvn = self.likelihood(MultivariateNormal(fant_mean, fant_fant_covar), inputs)
fant_fant_covar = mvn.covariance_matrix
fant_train_covar = delazify(full_covar[..., num_train:, :num_train])
self.fantasy_inputs = inputs
self.fantasy_targets = targets
"""
Compute a new mean cache given the old mean cache.
We have \\alpha = K^{-1}y, and we want to solve [K U; U' S][a; b] = [y; y_f], where U' is fant_train_covar,
S is fant_fant_covar, and y_f is (targets - fant_mean)
To do this, we solve the bordered linear system of equations for [a; b]:
AQ = U # Q = fant_solve
[S - U'Q]b = y_f - U'\\alpha ==> b = [S - U'Q]^{-1}(y_f - U'\\alpha)
a = \\alpha - Qb
"""
# Get cached K inverse decomp. (or compute if we somehow don't already have the covariance cache)
K_inverse = self.lik_train_train_covar.root_inv_decomposition()
fant_solve = K_inverse.matmul(fant_train_covar.transpose(-2, -1))
# Solve for "b", the lower portion of the *new* \\alpha corresponding to the fantasy points.
schur_complement = fant_fant_covar - fant_train_covar.matmul(fant_solve)
small_system_rhs = targets - fant_mean - fant_train_covar.matmul(self.mean_cache)
# Schur complement of a spd matrix is guaranteed to be positive definite
if small_system_rhs.requires_grad or schur_complement.requires_grad:
# TODO: Delete this part of the if statement when PyTorch implements cholesky_solve derivative.
fant_cache_lower = torch.gesv(small_system_rhs.unsqueeze(-1), schur_complement)[0]
else:
fant_cache_lower = cholesky_solve(small_system_rhs, torch.cholesky(schur_complement))
# Get "a", the new upper portion of the cache corresponding to the old training points.
fant_cache_upper = self.mean_cache.unsqueeze(-1) - fant_solve.matmul(fant_cache_lower)
fant_cache_upper = fant_cache_upper.squeeze(-1)
fant_cache_lower = fant_cache_lower.squeeze(-1)
# New mean cache.
fant_mean_cache = torch.cat((fant_cache_upper, fant_cache_lower), dim=-1)
"""
Compute a new covariance cache given the old covariance cache.
We have access to K \\approx LL' and K^{-1} \\approx R^{-1}R^{-T}, where L and R are low rank matrices
resulting from Lanczos (see the LOVE paper).
To update R^{-1}, we first update L:
[K U; U' S] = [L 0; A B][L' A'; 0 B']
Solving this matrix equation, we get:
K = LL' ==> L = L
U = LA' ==> A = UR^{-1}
S = AA' + BB' ==> B = cholesky(S - AA')
Once we've computed Z = [L 0; A B], we have that the new kernel matrix [K U; U' S] \approx ZZ'. Therefore,
we can form a pseudo-inverse of Z directly to approximate [K U; U' S]^{-1/2}.
"""
# [K U; U' S] = [L 0; lower_left schur_root]
batch_shape = fant_train_covar.shape[:-2]
L_inverse = self.covar_cache
L = delazify(self.lik_train_train_covar.root_decomposition().root)
m, n = L.shape[-2:]
lower_left = fant_train_covar.matmul(L_inverse)
schur_root = torch.cholesky(fant_fant_covar - lower_left.matmul(lower_left.transpose(-2, -1)))
upper_right = torch.zeros(m, schur_root.size(-1), device=L.device, dtype=L.dtype)
# Form new root Z = [L 0; lower_left schur_root]
num_fant = schur_root.size(-2)
m, n = L.shape[-2:]
new_root = torch.zeros(*batch_shape, m + num_fant, n + num_fant, device=L.device, dtype=L.dtype)
new_root[..., :m, :n] = L
new_root[..., :m, n:] = upper_right
new_root[..., m:, :n] = lower_left
new_root[..., m:, n:] = schur_root
# Use pseudo-inverse of Z as new inv root
# TODO: Replace pseudo-inverse calculation with something more stable than normal equations once
# one of torch.svd, torch.qr, or torch.pinverse works in batch mode.
cap_mat = new_root.transpose(-2, -1).matmul(new_root)
if cap_mat.requires_grad or new_root.requires_grad:
# TODO: Delete this part of the if statement when PyTorch implements cholesky_solve derivative.
new_covar_cache = torch.gesv(new_root.transpose(-2, -1), cap_mat)[0].transpose(-2, -1)
else:
new_covar_cache = cholesky_solve(new_root.transpose(-2, -1), torch.cholesky(cap_mat))
new_covar_cache = new_covar_cache.transpose(-2, -1)
# Create new DefaultPredictionStrategy object
new_num_train = full_inputs[0].size(len(batch_shape))
fant_strat = self.__class__(
num_train=new_num_train,
train_inputs=full_inputs,
train_mean=full_mean,
train_train_covar=full_covar,
train_labels=full_targets,
likelihood=self.likelihood,
non_batch_train=(len(batch_shape) == 0),
)
setattr(fant_strat, "_memoize_cache", {"mean_cache": fant_mean_cache, "covar_cache": new_covar_cache})
return fant_strat
def b_inv(b_mat):
eye = torch.rand(b_mat.size(0), b_mat.size(1), b_mat.size(2)).cuda()
b_inv, _ = torch.gesv(eye, b_mat)
return b_inv
def forward(self, im, pts_before, pts_after):
'''
Deforms image according to movement of pts_before and pts_after
Args)
im torch.Tensor object of size NxCxHxW
pts_before torch.Tensor object of size NxTx2 (T is # control pts)
pts_after torch.Tensor object of size NxTx2 (T is # control pts)
'''
# check input requirements
assert (4 == im.dim())
assert (3 == pts_after.dim())
assert (3 == pts_before.dim())
N = im.size()[0]
assert (N == pts_after.size()[0] and N == pts_before.size()[0])
assert (2 == pts_after.size()[2] and 2 == pts_before.size()[2])
T = pts_after.size()[1]
assert (T == pts_before.size()[1])
H = im.size()[2]
W = im.size()[3]
if self.normalize:
pts_after = pts_after.clone()
pts_after[:, :, 0] /= 0.5 * W
pts_after[:, :, 1] /= 0.5 * H
pts_after -= 1
pts_before = pts_before.clone()
pts_before[:, :, 0] /= 0.5 * W
pts_before[:, :, 1] /= 0.5 * H
pts_before -= 1
def construct_P():
'''
Consturcts matrix P of size NxTx3 where
P[n,i,0] := 1
P[n,i,1:] := pts_after[n]
'''
# Create matrix P with same configuration as 'pts_after'
P = pts_after.new_zeros((N, T, 3))
P[:, :, 0] = 1
P[:, :, 1:] = pts_after
return P
def calc_U(pt1, pt2):
'''
Calculate distance U between pt1 and pt2
U(r) := r**2 * log(r)
where
r := |pt1 - pt2|_2
Args)
pt1 torch.Tensor object, last dim is always 2
pt2 torch.Tensor object, last dim is always 2
'''
assert (2 == pt1.size()[-1])
assert (2 == pt2.size()[-1])
diff = pt1 - pt2
sq_diff = diff**2
sq_diff_sum = sq_diff.sum(-1)
r = sq_diff_sum.sqrt()
# Adds 1e-6 for numerical stability
return (r**2) * torch.log(r + 1e-6)
def construct_K():
'''
Consturcts matrix K of size NxTxT where
K[n,i,j] := U(|pts_after[n,i] - pts_after[n,j]|_2)
'''
# Assuming the number of control points are small enough,
# We just use for-loop for easy-to-read code
# Create matrix K with same configuration as 'pts_after'
K = pts_after.new_zeros((N, T, T))
for i in range(T):
for j in range(T):
K[:, i, j] = calc_U(pts_after[:, i, :], pts_after[:, j, :])
return K
def construct_L():
'''
Consturcts matrix L of size Nx(T+3)x(T+3) where
L[n] = [[ K[n] P[n] ]]
[[ P[n]^T 0 ]]
'''
P = construct_P()
K = construct_K()
# Create matrix L with same configuration as 'K'
L = K.new_zeros((N, T + 3, T + 3))
# Fill L matrix
L[:, :T, :T] = K
L[:, :T, T:(T + 3)] = P
L[:, T:(T + 3), :T] = P.transpose(1, 2)
return L
def construct_uv_grid():
'''
Returns H x W x 2 tensor uv with UV coordinate as its elements
uv[:,:,0] is H x W grid of x values
uv[:,:,1] is H x W grid of y values
'''
u_range = torch.arange(start=-1.0,
end=1.0,
step=2.0 / W,
device=im.device)
assert (W == u_range.size()[0])
u = u_range.new_zeros((H, W))
u[:] = u_range
v_range = torch.arange(start=-1.0,
end=1.0,
step=2.0 / H,
device=im.device)
assert (H == v_range.size()[0])
vt = v_range.new_zeros((W, H))
vt[:] = v_range
v = vt.transpose(0, 1)
return torch.stack([u, v], dim=2)
L = construct_L()
VT = pts_before.new_zeros((N, T + 3, 2))
# Use delta x and delta y as known heights of the surface
VT[:, :T, :] = pts_before - pts_after
# Solve Lx = VT
# x is of shape (N, T+3, 2)
# x[:,:,0] represents surface parameters for dx surface
# (dx values as surface height (z))
# x[:,:,1] represents surface parameters for dy surface
# (dy values as surface height (z))
x, _ = torch.gesv(VT, L)
uv = construct_uv_grid()
uv_batch = uv.repeat((N, 1, 1, 1))
def calc_dxdy():
'''
Calculate surface height for each uv coordinate
Returns NxHxWx2 tensor
'''
# control points of size NxTxHxWx2
cp = uv.new_zeros((H, W, N, T, 2))
cp[:, :, :] = pts_after
cp = cp.permute([2, 3, 0, 1, 4])
U = calc_U(uv, cp) # U value matrix of size NxTxHxW
w, a = x[:, :
T, :], x[:,
T:, :] # w is of size NxTx2, a is of size Nx3x2
w_x, w_y = w[:, :, 0], w[:, :, 1] # NxT each
a_x, a_y = a[:, :, 0], a[:, :, 1] # Nx3 each
dx = (a_x[:, 0].repeat((H, W, 1)).permute(2, 0, 1) +
torch.einsum('nhwd,nd->nhw', uv_batch, a_x[:, 1:]) +
torch.einsum('nthw,nt->nhw', U, w_x)) # dx values of NxHxW
dy = (a_y[:, 0].repeat((H, W, 1)).permute(2, 0, 1) +
torch.einsum('nhwd,nd->nhw', uv_batch, a_y[:, 1:]) +
torch.einsum('nthw,nt->nhw', U, w_y)) # dy values of NxHxW
return torch.stack([dx, dy], dim=3)
dxdy = calc_dxdy()
flow_field = uv + dxdy
return F.grid_sample(im, flow_field)
def b_inv(b_mat):
eye = b_mat.new_ones(b_mat.size(-1)).diag().expand_as(b_mat)
b_inv, _ = torch.gesv(eye, b_mat)
return b_inv
def matrix_solve(self, matrix, rhs, adjoint=None):
import ipdb
ipdb.set_trace()
return torch.gesv(rhs, matrix)[0]
def compute_pseudo_gradient(parameters, lr):
theta = torch.cat([x.grad.data.flatten() for x in parameters]).cpu()
H = torch.cat([(i * theta).unsqueeze(dim=-1) for i in theta], dim=1)
U = torch.eye(H.size(0)) + lr * H
pseudo_grad, _ = torch.gesv(theta, U)
return pseudo_grad.cuda()
def forward(self, y_train, phi, sq_lambda, L, m_test):
# extract hyperparameters
sigma_f = torch.exp(self.log_sigma_f)
lengthscale = torch.exp(self.log_lengthscale)
sigma_n = torch.exp(self.log_sigma_n)
# number of basis functions
m = sq_lambda.size(0)
# input dimension
dim = sq_lambda.size(1)
if self.covfunc.type == 'matern':
lengthscale = lengthscale.repeat(1, dim).view(dim)
# See the autograd section for explanation of what happens here.
n = y_train.size(0)
lprod = torch.ones(1)
omega_sum = torch.zeros(m, 1)
for q in range(dim):
lprod = lprod.mul(lengthscale[q].pow(2))
omega_sum = omega_sum.add(lengthscale[q].pow(2) *
sq_lambda[:, q].view(m, 1).pow(2))
if self.covfunc.type == 'matern':
inv_lambda_diag = \
(
math.pow( 2.0, dim ) * math.pow( math.pi, dim/2.0 )
*math.gamma( self.covfunc.nu + dim/2.0 )
*math.pow( 2.0*self.covfunc.nu, self.covfunc.nu )
*( (2.0*self.covfunc.nu + omega_sum).mul(lprod.pow(-0.5)) ).pow(-self.covfunc.nu-dim/2.0)
.div( math.gamma(self.covfunc.nu)*lprod.pow(self.covfunc.nu) )
).pow(-1.0) .view(m).mul(sigma_f.pow(-2.0))
elif self.covfunc.type == 'se':
inv_lambda_diag = (sigma_f.pow(-2).mul(lprod.pow(-0.5)).mul(
torch.exp(0.5 * omega_sum))).mul(
math.pow(2.0 * math.pi, -dim / 2)).view(m)
Z = phi.t().mm(phi) + torch.diag(inv_lambda_diag).mul(sigma_n.pow(2))
phi_lam = torch.cat((phi, inv_lambda_diag.sqrt().diag().mul(sigma_n)),
0) # [Phi; sign*sqrt(Lambda^-1)]
_, r = torch.qr(phi_lam)
v, _ = torch.gesv(
phi.t().mm(y_train.view(n, 1)), Z
) # X,LU = torch.gesv(B, A); AX=B => v=(Phi'*Phi+sign^2I)\(Phi'*y)
# compute phi_star
nt = m_test.size(0)
phi_star = torch.ones(m, nt)
for q in range(dim):
phi_star = phi_star.mul(
torch.sin(sq_lambda[:, q].view(m, 1) *
(m_test[:, q].view(1, nt) + L[q])).div(
math.sqrt(L[q])))
# predict
f_test = phi_star.t().mm(v)
tmp, _ = torch.trtrs(phi_star, r.t(),
upper=False) # solves r^T*u=phi_star, u=r*x
tmp, _ = torch.trtrs(tmp, r) # solves r*x=u
cov_f = torch.sum(phi_star.t().mul(tmp.t()), dim=1).mul(sigma_n.pow(2))
out = (f_test, cov_f)
return out
def forward(ctx, b, a):
# TODO see if one can backprop through LU
X, LU = torch.gesv(b, a)
ctx.save_for_backward(X, a)
ctx.mark_non_differentiable(LU)
return X, LU
def backward(ctx, grad_output, grad_LU=None):
X, a = ctx.saved_variables
grad_b, _ = torch.gesv(grad_output, a.t())
grad_a = -torch.mm(grad_b, X.t())
return grad_b, grad_a
def b_inv(b_mat, device):
eye = torch.rand(b_mat.size(0), b_mat.size(1), b_mat.size(2)).to(device)
b_inv, _ = torch.gesv(eye, b_mat)
return b_inv
