def optimize_Q(self):
batch_size = self._config['minibatch_size']
gamma = self._config['discount_factor']
num_actions = self._config['num_actions']
states, actions, next_states, rewards, dones, infos = self.memory.get_batch(
batch_size)
states = normalized(states)
next_states = normalized(next_states)
mask = np.zeros((batch_size, num_actions))
mask[np.arange(batch_size), actions] = 1
Qs_next = self.get_Q_values(next_states)
# The Q values of the terminal states is 0 by definition, so override them
Qs_next[dones] = 0
ys = rewards + gamma * (np.amax(
np.multiply(Qs_next, mask),
axis=1,
))
_, loss, loss_summary_str = self._tf_session.run(
[self._train_op, self._loss_op, self._tf_summary['loss']],
feed_dict={
self._train_op_input['states']: states,
self._train_op_input['actions']: actions,
self._train_op_input['ys']: ys,
})
return (loss, loss_summary_str)
def __call__(self, idx):
"""
computes geodesic distances to all vertices in the mesh
idx can be either an integer (single vertex index) or a list of vertex indices
or an array of bools of length n (with n the number of vertices in the mesh)
"""
u0 = np.zeros(len(self._verts))
u0[idx] = 1.0
# heat method, step 1
u = self._factored_AtLc(u0).ravel()
# heat method step 2
grad_u = 1 / (2 * self._triangle_area)[:,np.newaxis] * (
self._unit_normal_cross_e01 * u[self._tris[:,2]][:,np.newaxis]
+ self._unit_normal_cross_e12 * u[self._tris[:,0]][:,np.newaxis]
+ self._unit_normal_cross_e20 * u[self._tris[:,1]][:,np.newaxis]
)
X = - grad_u / veclen(grad_u)[:,np.newaxis]
# heat method step 3
div_Xs = np.zeros(len(self._verts))
for i1, i2, i3 in [(0, 1, 2), (1, 2, 0), (2, 0, 1)]: # for edge i2 --> i3 facing vertex i1
vi1, vi2, vi3 = self._tris[:,i1], self._tris[:,i2], self._tris[:,i3]
e1 = self._verts[vi2] - self._verts[vi1]
e2 = self._verts[vi3] - self._verts[vi1]
e_opp = self._verts[vi3] - self._verts[vi2]
cot1 = 1 / np.tan(np.arccos(
(normalized(-e2) * normalized(-e_opp)).sum(axis=1)))
cot2 = 1 / np.tan(np.arccos(
(normalized(-e1) * normalized( e_opp)).sum(axis=1)))
div_Xs += np.bincount(
vi1.astype(int),
0.5 * (cot1 * (e1 * X).sum(axis=1) + cot2 * (e2 * X).sum(axis=1)),
minlength=len(self._verts))
phi = self._factored_L(div_Xs).ravel()
phi -= phi.min()
return phi
def __init__(self, verts, tris, m=1.0):
self._verts = verts
self._tris = tris
# precompute some stuff needed later on
e01 = verts[tris[:, 1]] - verts[tris[:, 0]]
e12 = verts[tris[:, 2]] - verts[tris[:, 1]]
e20 = verts[tris[:, 0]] - verts[tris[:, 2]]
self._triangle_area = .5 * veclen(np.cross(e01, e12))
unit_normal = normalized(np.cross(normalized(e01), normalized(e12)))
self._un = unit_normal
self._unit_normal_cross_e01 = np.cross(unit_normal, -e01)
self._unit_normal_cross_e12 = np.cross(unit_normal, -e12)
self._unit_normal_cross_e20 = np.cross(unit_normal, -e20)
# parameters for heat method
h = np.mean(map(veclen, [e01, e12, e20]))
t = m * h**2
# pre-factorize poisson systems
Lc, vertex_area = compute_mesh_laplacian(verts,
tris,
area_type='lumped_mass')
A = sparse.spdiags(vertex_area, 0, len(verts), len(verts))
#self._factored_AtLc = splu((A - t * Lc).tocsc()).solve
self._factored_AtLc = cholesky((A - t * Lc).tocsc(), mode='simplicial')
#self._factored_L = splu(Lc.tocsc()).solve
self._factored_L = cholesky(Lc.tocsc(), mode='simplicial')
def train(args, model, sess, saver):
if args.fine_tuning :
saver.restore(sess, args.pre_trained_model)
print("saved model is loaded for fine-tuning!")
print("model path is %s"%(args.pre_trained_model))
num_imgs = len(os.listdir(args.train_Sharp_path))
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('./logs',sess.graph)
# if args.test_with_train:
# f = open("valid_logs.txt", 'w')
epoch = 1
step = num_imgs // args.batch_size
blur_imgs = util.image_loader(args.train_Blur_path, args.load_X, args.load_Y)
sharp_imgs = util.image_loader(args.train_Sharp_path, args.load_X, args.load_Y)
while epoch <= args.max_epoch:
random_index = np.random.permutation(len(blur_imgs))
for k in range(step):
s_time = time.time()
blur_batch, sharp_batch = util.batch_gen(blur_imgs, sharp_imgs, args.patch_size, args.batch_size, random_index, k, args.augmentation)
Knoise = np.random.randn(args.batch_size,64)
for t in range(args.critic_updates):
_, D_loss = sess.run([model.D_train, model.D_loss], feed_dict = {model.blur : blur_batch, model.sharp : sharp_batch,model.Knoise:Knoise, model.epoch : epoch})
if (k+1) % args.log_freq == 0:
_, G_loss,gene_K,gene_img,reg_loss,D_loss,G_loss,gp_loss = sess.run([model.G_train, model.G_loss,model.gene_K,model.gene_img,model.reg_loss,model.D_loss,model.G_loss,model.gp_loss], feed_dict = {model.blur : blur_batch, model.sharp : sharp_batch,model.Knoise:Knoise, model.epoch : epoch})
gene_K=util.normalized(gene_K)
gene_img=util.normalized(gene_img)
util.imshow(gene_K[0,:,:,0],cmap='gray')
toshow = np.hstack((sharp_batch[0]/255.0,blur_batch[0]/255.0,gene_img[0]))
util.imshow(toshow)
print("training with %d epoch %d/%d batch, D_loss: %0.2f, gp_loss: %0.2f, G_loss: %0.2f, reg_loss: %0.2f "%(epoch,k+1,step,D_loss,gp_loss,G_loss,reg_loss))
else:
_, G_loss = sess.run([model.G_train, model.G_loss], feed_dict = {model.blur : blur_batch, model.sharp : sharp_batch,model.Knoise:Knoise, model.epoch : epoch})
e_time = time.time()
# if epoch % args.log_freq == 0:
summary = sess.run(merged, feed_dict = {model.blur : blur_batch, model.sharp: sharp_batch,model.Knoise:Knoise, model.epoch : epoch})
train_writer.add_summary(summary, epoch)
# if args.test_with_train:
# test(args, model, sess, saver, f, epoch, loading = False)
print("%d training epoch completed" %(epoch))
print("D_loss : %0.4f, \t G_loss : %0.4f"%(D_loss, G_loss))
print("Elpased time : %0.4f"%(e_time - s_time))
saver.save(sess, './model/DeblurrGAN', global_step = epoch, write_meta_graph = False)
epoch += 1
saver.save(sess, './model/DeblurrGAN_last', write_meta_graph = False)
def sample_plane(plane, n, width, noise = 0): plane_repok(plane) e1, e2, _ = np.eye(3) v1 = e1 if (plane[0] == 0 and plane[1] == 0) else ut.normalized(np.array([-plane[1], plane[0], 0], 'd')) v2 = ut.normalized(np.cross(plane[:3], v1)) #print 'dot', np.dot(v1, plane[:3]), np.dot(v2, plane[:3]) #print 'sample', np.sum(np.dot(np.array([v1, v2]).T, np.random.uniform(-width/2, width/2, (2, n))).T * plane[:3], 1) center = -plane[3]*plane[:3] #print 'dot2', np.dot(center, plane[:3]) + plane[3], plane pts = np.dot(np.array([v1, v2]).T, np.random.uniform(-width/2, width/2, (2, n))).T + center #print 'ins', len(plane_inliers(plane, pts, 0.05)) pts += np.random.randn(*pts.shape)*noise return pts
def analytic_marginal_states(net, conditioned_on={}):
N = count_states(net)
S = np.zeros(N)
for i in range(N):
id_to_state(net, i)
S[i] = net.probability(conditioned_on)
return normalized(S, order=1)
def get_recent_state(self, current_observation, n=3):
# append current observation with last n observations
state = self.memory.get_recent_state(current_observation, n=3)
state = normalized(state)
# (n,h,w,c)
state = np.stack([state], axis=0)
return state
def analytic_marginal_states(net, conditioned_on={}):
N = count_states(net)
S = np.zeros(N)
for i in range(N):
id_to_state(net, i)
S[i] = net.probability(conditioned_on)
return normalized(S, order=1)
def __init__(self, verts, tris, m=10.0):
self._verts = verts
self._tris = tris
# precompute some stuff needed later on
e01 = verts[tris[:, 1]] - verts[tris[:, 0]]
e12 = verts[tris[:, 2]] - verts[tris[:, 1]]
e20 = verts[tris[:, 0]] - verts[tris[:, 2]]
self._triangle_area = .5 * veclen(np.cross(e01, e12))
unit_normal = normalized(np.cross(normalized(e01), normalized(e12)))
self._unit_normal_cross_e01 = np.cross(unit_normal, e01)
self._unit_normal_cross_e12 = np.cross(unit_normal, e12)
self._unit_normal_cross_e20 = np.cross(unit_normal, e20)
# parameters for heat method
h = np.mean(map(veclen, [e01, e12, e20]))
t = m * h**2
# pre-factorize poisson systems
Lc, A = compute_mesh_laplacian(verts, tris)
self._factored_AtLc = splu((A - t * Lc).tocsc()).solve
self._factored_L = splu(Lc.tocsc()).solve
def __init__(self, verts, tris, m=10.0):
self._verts = verts
self._tris = tris
# precompute some stuff needed later on
e01 = verts[tris[:,1]] - verts[tris[:,0]]
e12 = verts[tris[:,2]] - verts[tris[:,1]]
e20 = verts[tris[:,0]] - verts[tris[:,2]]
self._triangle_area = .5 * veclen(np.cross(e01, e12))
unit_normal = normalized(np.cross(normalized(e01), normalized(e12)))
self._unit_normal_cross_e01 = np.cross(unit_normal, e01)
self._unit_normal_cross_e12 = np.cross(unit_normal, e12)
self._unit_normal_cross_e20 = np.cross(unit_normal, e20)
# parameters for heat method
h = np.mean(map(veclen, [e01, e12, e20]))
t = m * h ** 2
# pre-factorize poisson systems
Lc, A = compute_mesh_laplacian(verts, tris)
self._factored_AtLc = splu((A - t * Lc).tocsc()).solve
self._factored_L = splu(Lc.tocsc()).solve
def backproj(p):
ray = ut.normalized(
np.dot(mvg.pixel_ray_matrix(scan.R(root), scan.K(root)),
ut.homog(p)))
c = scan.center(root)
dist = (-plane[3] - np.dot(c, plane[:3])) / np.dot(ray, plane[:3])
assert dist >= 0
pt = c + ray * dist
print planefit.dist_to_plane(plane, pt[np.newaxis, :])
return pt
def parse(infile, lang):
assert lang in ('en', 'es')
ss, deps = _parser[lang].parse(infile)
def key((sid, t)):
return sid
by_sid = normalized(groupby(deps, key),
key=key,
filler=lambda sid: (sid, ()),
start=1)
return izip(ss, by_sid)
def __init__(self, face_idx, mesh_pts, texsize=TEXSIZE):
if texsize is None:
texsize = 256
self.face_idx = np.asarray(face_idx)
self.nfaces = len(face_idx)
self.mesh_pts = np.asarray(mesh_pts)
self.texsize = texsize
self.face_planes, self.face_center, self.face_edges, self.face_pts, self.face_juv = [], [], [], [], []
u_grid, v_grid = [
np.array(x, 'd') for x in np.mgrid[:texsize, :texsize]
]
uf = u_grid.flatten()
vf = v_grid.flatten()
for j in xrange(len(face_idx)):
p1 = mesh_pts[face_idx[j][0]]
p2 = mesh_pts[face_idx[j][1]]
p3 = mesh_pts[face_idx[j][3]]
e1 = -p1 + p3
e2 = -p1 + p2
n = ut.normalized(np.cross(e1, e2))
d = -np.dot(p1, n) #np.dot(p1, n) + d = 0
self.face_planes.append(np.concatenate([n, [d]]))
pts = p1 + (uf / (texsize - 1.))[:, na] * e1[na, :] + (
vf / (texsize - 1.))[:, na] * e2[na, :]
juv = np.zeros((len(uf), 3), 'l')
juv[:, 0] = j
juv[:, 1] = uf
juv[:, 2] = vf
self.face_juv.append(juv)
self.face_pts.append(pts)
self.face_edges.append((p1, np.array([e1, e2])))
self.face_center.append(p1 + 0.5 * e1 + 0.5 * e2)
self.face_center = np.array(self.face_center)
self.tex2juv = np.vstack(self.face_juv)
self.juv2tex = np.zeros((len(self.face_idx), texsize, texsize), 'l')
self.juv2tex[self.tex2juv[:, 0], self.tex2juv[:, 1],
self.tex2juv[:, 2]] = range(len(self.tex2juv))
self.texel_pts = np.vstack(self.face_pts)
self.face_planes = np.array(self.face_planes)
self.on_border = ut.lor((self.tex2juv[:, 1] == 0),
(self.tex2juv[:, 2] == 0),
(self.tex2juv[:, 1] == self.texsize - 1),
(self.tex2juv[:, 2] == self.texsize - 1))
self.ntexels = len(self.tex2juv)
def __call__(self, idx):
"""
computes geodesic distances to all vertices in the mesh
idx can be either an integer (single vertex index) or a list of vertex indices
or an array of bools of length n (with n the number of vertices in the mesh)
"""
u0 = np.zeros(len(self._verts))
u0[idx] = 1.0
# heat method, step 1
u = self._factored_AtLc(u0).ravel()
# heat method step 2
grad_u = 1 / (2 * self._triangle_area)[:, np.newaxis] * (
self._unit_normal_cross_e01 * u[self._tris[:, 2]][:, np.newaxis] +
self._unit_normal_cross_e12 * u[self._tris[:, 0]][:, np.newaxis] +
self._unit_normal_cross_e20 * u[self._tris[:, 1]][:, np.newaxis])
X = -grad_u / veclen(grad_u)[:, np.newaxis]
# heat method step 3
div_Xs = np.zeros(len(self._verts))
for i1, i2, i3 in [(0, 1, 2), (1, 2, 0),
(2, 0, 1)]: # for edge i2 --> i3 facing vertex i1
# 0 1 2
# 1 2 0
# 2 0 1
vi1, vi2, vi3 = self._tris[:, i1], self._tris[:,
i2], self._tris[:,
i3]
e1 = self._verts[vi2] - self._verts[vi1]
e2 = self._verts[vi3] - self._verts[vi1]
e_opp = self._verts[vi3] - self._verts[vi2]
cot1 = 1 / np.tan(
np.arccos((normalized(-e2) * normalized(-e_opp)).sum(axis=1)))
cot2 = 1 / np.tan(
np.arccos((normalized(-e1) * normalized(+e_opp)).sum(axis=1)))
div_Xs += np.bincount(vi1.astype(int),
0.5 * (cot1 * (e1 * X).sum(axis=1) + cot2 *
(e2 * X).sum(axis=1)),
minlength=len(self._verts))
phi = self._factored_L(div_Xs).ravel()
phi -= phi.min()
return phi
def __init__(self, verts, tris, m=1.0):
self._verts = verts
self._tris = tris
# precompute some stuff needed later on
e01 = verts[tris[:,1]] - verts[tris[:,0]]
e12 = verts[tris[:,2]] - verts[tris[:,1]]
e20 = verts[tris[:,0]] - verts[tris[:,2]]
self._triangle_area = .5 * veclen(np.cross(e01, e12))
unit_normal = normalized(np.cross(normalized(e01), normalized(e12)))
self._un = unit_normal
self._unit_normal_cross_e01 = np.cross(unit_normal, -e01)
self._unit_normal_cross_e12 = np.cross(unit_normal, -e12)
self._unit_normal_cross_e20 = np.cross(unit_normal, -e20)
# parameters for heat method
h = np.mean(map(veclen, [e01, e12, e20]))
t = m * h ** 2
# pre-factorize poisson systems
Lc, vertex_area = compute_mesh_laplacian(verts, tris, area_type='lumped_mass')
A = sparse.spdiags(vertex_area, 0, len(verts), len(verts))
#self._factored_AtLc = splu((A - t * Lc).tocsc()).solve
self._factored_AtLc = cholesky((A - t * Lc).tocsc(), mode='simplicial')
#self._factored_L = splu(Lc.tocsc()).solve
self._factored_L = cholesky(Lc.tocsc(), mode='simplicial')
def m_deep_with_shortcut(m,
p=None,
marg=None,
fro=None,
to=None,
cpt='marginal'):
"""constructs an m_deep_bistable model with a single additional connection from node 'fro' to node 'to'
(fro must be greater than to to prevent cycles)
"""
net = m_deep_bistable(m, p, marg)
if type(fro) is int:
fro = 'X%d' % fro
if type(to) is int:
to = 'X%d' % to
fro = net.get_node_by_name(fro)
to = net.get_node_by_name(to)
if cpt is 'random':
cpt = np.random.random((2, 2))
sums = cpt.sum(axis=0)
cpt[:, 0] /= sums[0]
cpt[:, 1] /= sums[1]
elif cpt is 'marginal':
if p is None: p = net.get_node_by_name('X1').get_table()[0, 0]
dists = {}
def store_dists(n, d):
dists[n] = d
net.bfs_traverse([fro], store_dists)
marg = compute_marginal_for_given_p(dists[to], p)
cpt = np.array([[marg, 1 - marg], [1 - marg, marg]])
prev_table = to.get_table()
prev_parents = net.parents(to)
# making room for a new parent
shape = tuple([fro.size()] + list(prev_table.shape))
table = np.zeros(shape)
# populate new table such that P(to|parents,fro) = P(to|fro)P(to|parents)
for i in range(fro.size()):
for j in range(to.size()):
table[i, ..., j] = cpt[i, j] * prev_table[..., j]
# renormalize such that P(to|a particular configuration of parents) = 1
table = normalized(table, axis=-1, order=1)
net.cpt([fro] + prev_parents + [to], table)
return net
def __init__(self, face_idx, mesh_pts, texsize = TEXSIZE):
if texsize is None:
texsize = 256
self.face_idx = np.asarray(face_idx)
self.nfaces = len(face_idx)
self.mesh_pts = np.asarray(mesh_pts)
self.texsize = texsize
self.face_planes, self.face_center, self.face_edges, self.face_pts, self.face_juv = [], [], [], [], []
u_grid, v_grid = [np.array(x, 'd') for x in np.mgrid[:texsize, :texsize]]
uf = u_grid.flatten()
vf = v_grid.flatten()
for j in xrange(len(face_idx)):
p1 = mesh_pts[face_idx[j][0]]
p2 = mesh_pts[face_idx[j][1]]
p3 = mesh_pts[face_idx[j][3]]
e1 = -p1 + p3
e2 = -p1 + p2
n = ut.normalized(np.cross(e1, e2))
d = -np.dot(p1, n) #np.dot(p1, n) + d = 0
self.face_planes.append(np.concatenate([n, [d]]))
pts = p1 + (uf/(texsize - 1.))[:, na]*e1[na, :] + (vf/(texsize - 1.))[:, na]*e2[na, :]
juv = np.zeros((len(uf), 3), 'l')
juv[:, 0] = j
juv[:, 1] = uf
juv[:, 2] = vf
self.face_juv.append(juv)
self.face_pts.append(pts)
self.face_edges.append((p1, np.array([e1, e2])))
self.face_center.append(p1 + 0.5*e1 + 0.5*e2)
self.face_center = np.array(self.face_center)
self.tex2juv = np.vstack(self.face_juv)
self.juv2tex = np.zeros((len(self.face_idx), texsize, texsize), 'l')
self.juv2tex[self.tex2juv[:, 0], self.tex2juv[:, 1], self.tex2juv[:, 2]] = range(len(self.tex2juv))
self.texel_pts = np.vstack(self.face_pts)
self.face_planes = np.array(self.face_planes)
self.on_border = ut.lor((self.tex2juv[:, 1] == 0),
(self.tex2juv[:, 2] == 0),
(self.tex2juv[:, 1] == self.texsize-1),
(self.tex2juv[:, 2] == self.texsize-1))
self.ntexels = len(self.tex2juv)
def greedy_project(scan, mesh, stable_angle=np.radians(70)):
filled = np.zeros(mesh.ntexels, 'bool')
texel_colors = np.zeros((mesh.ntexels, 3))
visible_by_view = np.zeros((scan.length, mesh.ntexels), 'bool')
stable_by_view = np.zeros((scan.length, mesh.ntexels), 'bool')
for frame in xrange(scan.length):
visible = mesh.texel_visible(scan, frame)
visible_by_view[frame] = visible
for j in xrange(mesh.nfaces):
d = np.dot(
ut.normalized(-mesh.face_center[j] + scan.center(frame)),
mesh.face_planes[j][:3])
assert -1 <= d <= 1
angle = np.arccos(abs(d))
print frame, j, np.rad2deg(angle), (angle <= stable_angle)
on_face = (mesh.tex2juv[:, 0] == j)
stable_by_view[frame,
on_face] = visible_by_view[frame, on_face] * (
angle <= stable_angle)
print 'stable texels', np.sum(stable_by_view[frame])
projectable = stable_by_view.copy()
unstable = np.all(-stable_by_view, axis=0)
print 'Unstable faces:', np.unique(mesh.tex2juv[unstable, 0])
projectable[:, unstable] = visible_by_view[:, unstable]
unused = range(scan.length)
while np.sum(projectable) > 0:
# it's ok to project onto a face if it's (1) unfilled (2) it's either stable, or it's unstable in every view
frame = max(ut.shuffled(unused),
key=lambda f: np.sum(projectable[f, :]))
projs = scan.project(frame, mesh.texel_pts)
ok = projectable[frame]
print 'chose', frame, np.sum(
projectable[frame, :]), np.sum(ok), 'projecting to', np.unique(
mesh.tex2juv[projectable[frame, :], 0])
texel_colors[ok] = ig.lookup_bilinear(scan.im(frame), projs[ok, 0],
projs[ok, 1])
projectable[:, ok] = 0
unused.remove(frame)
print 'projected', frame, np.sum(filled)
return [texel_colors]
def sample_marginal_states(net, evidence, samples, when=None): """Computes S[i] = vector of marginal probabilities that net is in state id i. If given, when(net) is evaluated to decide whether each sample is included """ n_states = count_states(net) # estimate starting distribution over states by sampling S = np.zeros(n_states) def do_count_state(i, net): if when is None or when(net): S[state_to_id(net, net.state_vector())] += 1 gibbs_sample(net, evidence, do_count_state, samples, 1) return normalized(S, order=1)
def sample_marginal_states(net, evidence, samples, when=None):
"""Computes S[i] = vector of marginal probabilities that net is in state id i.
If given, when(net) is evaluated to decide whether each sample is included
"""
n_states = count_states(net)
# estimate starting distribution over states by sampling
S = np.zeros(n_states)
def do_count_state(i, net):
if when is None or when(net):
S[state_to_id(net, net.state_vector())] += 1
gibbs_sample(net, evidence, do_count_state, samples, 1)
return normalized(S, order=1)
def m_deep_with_shortcut(m, p=None, marg=None, fro=None, to=None, cpt='marginal'):
"""constructs an m_deep_bistable model with a single additional connection from node 'fro' to node 'to'
(fro must be greater than to to prevent cycles)
"""
net = m_deep_bistable(m, p, marg)
if type(fro) is int:
fro = 'X%d' % fro
if type(to) is int:
to = 'X%d' % to
fro = net.get_node_by_name(fro)
to = net.get_node_by_name(to)
if cpt is 'random':
cpt = np.random.random((2,2))
sums = cpt.sum(axis=0)
cpt[:,0] /= sums[0]
cpt[:,1] /= sums[1]
elif cpt is 'marginal':
if p is None: p = net.get_node_by_name('X1').get_table()[0,0]
dists = {}
def store_dists(n,d): dists[n] = d
net.bfs_traverse([fro], store_dists)
marg = compute_marginal_for_given_p(dists[to], p)
cpt = np.array([[marg, 1-marg], [1-marg, marg]])
prev_table = to.get_table()
prev_parents = net.parents(to)
# making room for a new parent
shape = tuple([fro.size()] + list(prev_table.shape))
table = np.zeros(shape)
# populate new table such that P(to|parents,fro) = P(to|fro)P(to|parents)
for i in range(fro.size()):
for j in range(to.size()):
table[i,...,j] = cpt[i,j] * prev_table[...,j];
# renormalize such that P(to|a particular configuration of parents) = 1
table = normalized(table, axis=-1, order=1)
net.cpt([fro] + prev_parents + [to], table)
return net
def greedy_project(scan, mesh, stable_angle = np.radians(70)):
filled = np.zeros(mesh.ntexels, 'bool')
texel_colors = np.zeros((mesh.ntexels, 3))
visible_by_view = np.zeros((scan.length, mesh.ntexels), 'bool')
stable_by_view = np.zeros((scan.length, mesh.ntexels), 'bool')
for frame in xrange(scan.length):
visible = mesh.texel_visible(scan, frame)
visible_by_view[frame] = visible
for j in xrange(mesh.nfaces):
d = np.dot(ut.normalized(-mesh.face_center[j] + scan.center(frame)), mesh.face_planes[j][:3])
assert -1 <= d <= 1
angle = np.arccos(abs(d))
print frame, j, np.rad2deg(angle), (angle <= stable_angle)
on_face = (mesh.tex2juv[:, 0] == j)
stable_by_view[frame, on_face] = visible_by_view[frame, on_face]*(angle <= stable_angle)
print 'stable texels', np.sum(stable_by_view[frame])
projectable = stable_by_view.copy()
unstable = np.all(-stable_by_view, axis = 0)
print 'Unstable faces:', np.unique(mesh.tex2juv[unstable, 0])
projectable[:, unstable] = visible_by_view[:, unstable]
unused = range(scan.length)
while np.sum(projectable) > 0:
# it's ok to project onto a face if it's (1) unfilled (2) it's either stable, or it's unstable in every view
frame = max(ut.shuffled(unused), key = lambda f : np.sum(projectable[f, :]))
projs = scan.project(frame, mesh.texel_pts)
ok = projectable[frame]
print 'chose', frame, np.sum(projectable[frame, :]), np.sum(ok), 'projecting to', np.unique(mesh.tex2juv[projectable[frame, :], 0])
texel_colors[ok] = ig.lookup_bilinear(scan.im(frame), projs[ok, 0], projs[ok, 1])
projectable[:, ok] = 0
unused.remove(frame)
print 'projected', frame, np.sum(filled)
return [texel_colors]
def _fit_1(self, sequence, V, f0, m0, eps, max_iter, relax):
"""Run the Expectation-Maximization algorithm.
Parameters:
- sequence: sequence to fit
- V : erasure parameters
- f : initial estimate of position frequency matrix
- m : initial estimate of a priori model probability
- eps : stop when relative change in expected log-likelihood
is less than this
- relax : parameter to take probabilities away from zero
"""
W = f0.shape[0] - 1 # size of window
n = len(V) # sequence length
it = 0 # current iteration
f1 = f0 # current estimation of f
m1 = m0 # current estimation of m
E1 = None # current expected log-likelihood
while True:
pM = np.ones((n,))
pB = np.ones((n,))
for i, X in self._roll(sequence, W):
pM[i], pB[i] = MEME._score(X, f1)
pM *= m1
pB *= 1.0 - m1
Z = pM/(pM + pB)
E2 = np.sum(Z*np.log(pM) + (1 - Z)*np.log(pB))
# Correct for overlapping, enforce Z[i:(i+W)] <= 1 for i=1...n
# MEME hack to account for dependence between sub-sequences.
for i in xrange(len(Z)):
s = Z[i:i+W].sum()
if s > 1.0:
Z[i:i+W] /= s
# Update m and f
qB = 0.0
qM = 0.0
c = np.zeros(f0.shape, dtype=float)
for i, X in self._roll(sequence, W):
ZB = 1.0 - Z[i]
ZM = Z[i]*V[i]
for j, k in enumerate(X):
c[0 , k] += ZB
c[j + 1, k] += ZM
qB += ZB
qM += ZM
m2 = qM/(qB + qM)
f2 = (1.0 - relax)*util.normalized(c) + relax
# Check convergence
if E1:
err = (E1 - E2)/E1
if err < eps:
break
E1 = E2
it += 1
if it > max_iter:
if self.logger:
self.logger.warning(
'MEME.fit: max iterations reached without convergence (err={0})'.format(err))
break
f1 = f2
m1 = m2
return f2, m2, E2, Z
def eig_steadystate(A): # steady state distribution of A_ff transition matrix is largest eigenvector (eigenvalue=1) w,v = np.linalg.eig(A) inds = np.argsort(w) S_steady_state = np.abs(v[:,inds[-1]]) return normalized(S_steady_state, order=1)
def label_box(seq,
root=0,
side_len1=None,
side_len2=None,
side_len3=None,
y_flip=True,
mode='normal'):
print seq
if type(seq) == type(''):
scan = dset.Scan(seq, None)
else:
scan = seq
seq = scan.path
if mode == 'normal':
_, _, tracks = dset.read_bundler(scan.bundle_file, scan.full_shape)
pts = np.array([t[0] for t in tracks])
proj = scan.project(root, pts)
w = 1
pylab.clf()
im_with_pts = ig.draw_pts(scan.im(root), proj, width=w)
pylab.imshow(im_with_pts)
rect = ut.bbox2d(pylab.ginput(2, timeout=-1))
#rect = (1782.005828476269, 1431.7364696086595, 529.75936719400488, 354.40549542048279)
print rect
ok = ut.land(rect[0] <= proj[:, 0], proj[:, 0] <= rect[0] + rect[2],
rect[1] <= proj[:, 1], proj[:, 1] <= rect[1] + rect[3])
pts_in_box = pts[ok]
thresh = pylab.dist(scan.center(root), scan.center(root + 1)) / 50.
plane, _ = planefit.fit_plane_ransac(pts_in_box, thresh)
if plane[1] < 0 and y_flip:
plane *= -1
ins = planefit.plane_inliers(plane, pts, thresh)
pylab.clf()
colors = np.zeros_like(pts)
colors[:, 0] = 255
colors[ins] = (0, 255, 0)
im_ins = ig.draw_pts(scan.im(root),
map(ut.itup, proj),
map(ut.itup, colors),
width=w)
pylab.clf()
pylab.imshow(im_ins)
if not input('ok? '):
return
print 'click 2 points (used to recalibrate the plane)'
rect = ut.bbox2d(pylab.ginput(2, timeout=-1))
ok = ut.land(rect[0] <= proj[:, 0], proj[:, 0] <= rect[0] + rect[2],
rect[1] <= proj[:, 1], proj[:, 1] <= rect[1] + rect[3])
pts_in_box = pts[ok]
print 'plane before', plane
plane[3] = -np.median(np.dot(pts_in_box, plane[:3]))
print 'plane after', plane[3]
if 1:
print 'hack'
im_ins = scan.im(root)
pylab.clf()
pylab.imshow(im_ins)
print 'click 3 base points'
px = pylab.ginput(3, timeout=-1)
#px = [(2270.2989175686921, 1482.9937552039967), (2297.2764363030801, 1555.8330557868442), (2405.1865112406322, 1550.4375520399667)]
def backproj(p):
ray = ut.normalized(
np.dot(mvg.pixel_ray_matrix(scan.R(root), scan.K(root)),
ut.homog(p)))
c = scan.center(root)
dist = (-plane[3] - np.dot(c, plane[:3])) / np.dot(ray, plane[:3])
assert dist >= 0
pt = c + ray * dist
print planefit.dist_to_plane(plane, pt[np.newaxis, :])
return pt
sc = 1.
while 1:
cb = np.array(map(backproj, px))
v1 = cb[0] - cb[1]
v2 = cb[2] - cb[1]
if side_len1 is None:
side_len1 = 0.5 * (np.linalg.norm(v1) + np.linalg.norm(v2))
if side_len2 is None:
side_len2 = side_len1
if side_len3 is None:
side_len3 = side_len1
a1 = sc * side_len1
a2 = sc * side_len2
a3 = sc * side_len3
print 'side1', a1, 'side2', a2, 'side3', a3, 'thresh =', thresh, \
'v1 =', np.linalg.norm(v1), 'v2 = ', np.linalg.norm(v2)
R = np.zeros((3, 3))
cr = ut.normalized(np.cross(v1, plane[:3]))
cr *= np.sign(np.dot(cr, v2))
R[0] = a1 * ut.normalized(v1)
R[1] = a2 * ut.normalized(cr)
R[2] = a3 * ut.normalized(plane[:3])
print ut.normax(R, 1)
mesh_pts = []
for zi in xrange(2):
for yi in xrange(2):
for xi in xrange(2):
mesh_pts.append(cb[1] + R[0] * xi + R[1] * yi +
R[2] * zi)
face_idx = -1 + np.array(
[[1, 2, 4, 3],
np.array([1, 2, 4, 3]) + 4, [1, 2, 2 + 4, 1 + 4],
[2, 4, 4 + 4, 2 + 4], [4, 3, 3 + 4, 4 + 4],
[3, 1, 1 + 4, 3 + 4]])
mesh = box.Mesh(face_idx, mesh_pts, texsize=128)
# show a preview
scan_ = dset.Scan(seq)
ig.show(
[[1 + i,
box.draw_faces(mesh, scan_, i, hires=0),
scan_.im(i)] for i in [root, root + 1]])
if input('ok? '):
box.save_mesh(ut.pjoin(seq, 'cube.mat'), mesh)
break
else:
sc = float(input('scale? '))
time.sleep(2)
else:
mesh = box.load_from_mat(ut.pjoin(seq, 'cube.mat'))
scan = dset.Scan(seq, use_cams_file=False)
print 'Already marked as bad:'
good_cams_file = os.path.join(scan.path, 'good_cams.txt')
if os.path.exists(good_cams_file):
inds = map(int, open(good_cams_file, 'r').read().split())
file_ids = map(scan.file_index, scan.im_files)
bad = sorted(set(file_ids) - set(inds))
print '\n'.join(map(str, bad))
if 1:
ig.show([[
scan.file_index(scan.im_files[frame]),
box.draw_faces(mesh, scan, frame, hires=0)
] for frame in xrange(scan.length)])
inp = input('Bad cameras (as string): ')
if inp != 'skip':
bad_cams = map(int, inp.split())
all_idx = map(scan.file_index, scan.im_files)
good_cams = sorted(set(all_idx) - set(bad_cams))
ut.write_lines(ut.pjoin(seq, 'good_cams.txt'), map(str, good_cams))
def eig_steadystate(A):
# steady state distribution of A_ff transition matrix is largest eigenvector (eigenvalue=1)
w, v = np.linalg.eig(A)
inds = np.argsort(w)
S_steady_state = np.abs(v[:, inds[-1]])
return normalized(S_steady_state, order=1)
def plane_from_3(pts): assert pts.shape == (3, 3) w = ut.normalized(np.cross(pts[1] - pts[0], pts[2] - pts[0])) return np.array(list(w) + [-np.dot(w, pts[0])])
