def summarize_smooth(self, noise=0.1):
''' Find best estimate for diffeomorphism
looking at each singularly. '''
maximum_likelihood_index = np.zeros(self.shape, dtype='int32')
variance = np.zeros(self.shape, dtype='float32')
epsilon = None
for c in coords_iterate(self.shape):
k = self.flattening.cell2index[c]
sim = self.neighbor_similarity_flat[k]
if epsilon is None:
epsilon = np.random.randn(*sim.shape)
sim_min = sim.min()
sim_max = sim.max()
if sim_max == sim_min:
best_index = 0
variance[c] = 0
else:
std = noise * (sim_max - sim_min)
best = np.argmin(sim + epsilon * std)
best_index = self.neighbor_indices_flat[k][best]
variance[c] = sim[best]
maximum_likelihood_index[c] = best_index
d = self.flattening.flat2coords(maximum_likelihood_index)
variance = variance - variance.min()
vmax = variance.max()
if vmax > 0:
variance *= (1 / vmax)
return Diffeomorphism2D(d, variance)
def consistency_based_uncertainty(action, normalize_distance):
'''
Update the uncertainties for the action by the improved uncertainty
classification based on comparing the diffeomorphism with its inverse.
'''
# print('Using length score function %s' % length_score)
d1 = action.diffeo.d
d2 = action.diffeo_inv.d
v1, v2 = consistency_based_uncertainty2(d1, d2, normalize_distance)
D1 = Diffeomorphism2D(d1, v1)
D2 = Diffeomorphism2D(d2, v2)
a = DiffeoAction(label=action.label,
diffeo=D1,
diffeo_inv=D2,
original_cmd=action.original_cmd)
return a
def DDSFromSymbolic(resolution, symdiffeosystem): # @UnusedVariable
"""
Creates a DiffeoSystem from synthetic diffeomorphisms.
"""
diffeo2s_config = get_diffeo2s_config()
diffeo2dds_config = get_diffeo2dds_config()
_, symdds = diffeo2dds_config.symdds.instance_smarter(symdiffeosystem)
logger.info('Creating symbolic diffeomorphism (resolution = %d)' %
resolution)
diffeoactions = []
for _, action in enumerate(symdds.actions):
id_diffeo, diffeo = parse_diffeo_spec(diffeo2s_config,
action['diffeo'])
label = action.get('label', id_diffeo)
original_cmd = np.array(action['original_cmd'])
logger.info('Getting symbolic diffeomorphism %r' % id_diffeo)
shape = (resolution, resolution)
viewport = SquareDomain([[-1, +1], [-1, +1]])
manifold = diffeo.get_topology()
D, Dinfo = diffeo_from_function_viewport(diffeo, manifold, viewport,
shape)
D2d = Diffeomorphism2D(D, Dinfo)
diffeo_inv = diffeo.get_inverse()
D_inv, Dinfo_inv = \
diffeo_from_function_viewport(diffeo_inv, manifold, viewport, shape)
D2d_inv = Diffeomorphism2D(D_inv, Dinfo_inv)
action = DiffeoAction(label=label,
diffeo=D2d,
diffeo_inv=D2d_inv,
original_cmd=original_cmd)
diffeoactions.append(action)
dds = DiffeoSystem('unnamed', actions=diffeoactions)
return dds
def summarize_averaged(self, n=10, noise=0.1):
d = []
for _ in range(n):
diff = self.summarize_smooth(noise)
d.append(diff.d)
# print('.')
ds = np.array(d, 'float')
avg = ds.mean(axis=0)
# var = diff.variance
var = ds[:, :, :, 0].var(axis=0) + ds[:, :, :, 1].var(axis=0)
# print var.shape
assert avg.shape == diff.d.shape
return Diffeomorphism2D(avg, var)
def get_value(self):
'''
Find maximum likelihood estimate for diffeomorphism looking
at each pixel singularly.
Returns a Diffeomorphism2D.
'''
if not self.initialized():
msg = 'Cannot summarize() because not initialized yet.'
raise Diffeo2dEstimatorInterface.NotReady(msg)
certainty = np.zeros(self.shape, dtype='float32')
certainty.fill(np.nan)
dd = diffeo_identity(self.shape)
dd[:] = -1
for i in range(self.nsensels):
if self.inference_method == DiffeomorphismEstimatorFaster.Order:
eord_score = self.neig_eord_score[i, :]
best = np.argmin(eord_score)
if self.inference_method == DiffeomorphismEstimatorFaster.Similarity:
esim_score = self.neig_esim_score[i, :]
best = np.argmin(esim_score)
jc = self.flat_structure.neighbor_cell(i, best)
ic = self.flat_structure.flattening.index2cell[i]
if self.inference_method == DiffeomorphismEstimatorFaster.Order:
certain = -np.min(eord_score) / np.mean(eord_score)
if self.inference_method == DiffeomorphismEstimatorFaster.Similarity:
first = np.sort(esim_score)[:10]
certain = -(first[0] - np.mean(first[1:]))
# certain = -np.min(esim_score) / np.mean(esim_score)
# certain = np.min(esim_score) / self.num_samples
# certain = -np.mean(esim_score) / np.min(esim_score)
dd[ic[0], ic[1], 0] = jc[0]
dd[ic[0], ic[1], 1] = jc[1]
certainty[ic[0], ic[1]] = certain
certainty = certainty - certainty.min()
vmax = certainty.max()
if vmax > 0:
certainty *= (1.0 / vmax)
return Diffeomorphism2D(dd, certainty)
def commutator(a1, a2):
# TODO: use uncertainty
d1 = a1.d
d1_inv = a1.d_inv
d2 = a2.d
d2_inv = a2.d_inv
C = diffeo_compose
z = C(C(C(d1, d2), d1_inv), d2_inv)
diffeo = Diffeomorphism2D(z)
label = "[%s,%s]" % (a1.label, a2.label)
primitive = False
invertible = True
original_cmd = None
return Action(diffeo, label, primitive, invertible, original_cmd)
def summarize_continuous(self, quivername):
center = np.zeros(list(self.shape) + [2], dtype='float32')
spread = np.zeros(self.shape, dtype='float32')
maxerror = np.zeros(self.shape, dtype='float32')
minerror = np.zeros(self.shape, dtype='float32')
# from PIL import Image # @UnresolvedImport
# sim_image = Image.new('L', np.flipud(self.shape) * np.flipud(self.lengths))
# zer_image = Image.new('L', np.flipud(self.shape) * np.flipud(self.lengths))
for c in coords_iterate(self.shape):
# find index in flat array
k = self.flattening.cell2index[c]
# Look at the average similarities of the neihgbors
sim = self.neighbor_similarity_flat[k]
sim_square = sim.reshape(
self.lengths).astype('float32') / self.num_samples
from diffeo2d_learn.library.simple.display import sim_square_modify
sim_square, minerror[c], maxerror[c] = sim_square_modify(
sim_square,
np.min(self.neighbor_similarity_flat) / self.num_samples,
np.max(self.neighbor_similarity_flat) / self.num_samples)
# avg_square = (np.max(sim_square) + np.min(sim_square)) / 2
sim_zeroed = np.zeros(sim_square.shape)
sim_zeroed[sim_square > 0.85] = sim_square[sim_square > 0.85]
from diffeo2d_learn.library.simple.display import get_cm
center[c], spread[c] = get_cm(sim_square)
# p0 = tuple(np.flipud(np.array(c) * self.lengths))
# sim_image.paste(Image.fromarray((sim_square * 255).astype('uint8')), p0 + tuple(p0 + self.lengths))
# zer_image.paste(Image.fromarray((sim_zeroed * 255).astype('uint8')), p0 + tuple(p0 + self.lengths))
# sim_image.save(quivername + 'simimage.png')
# zer_image.save(quivername + 'simzeroed.png')
from diffeo2d_learn.library.simple.display import display_disp_quiver
display_disp_quiver(center, quivername)
from diffeo2d_learn.library.simple.display import display_continuous_stats
display_continuous_stats(center, spread, minerror, maxerror,
quivername)
dcont = displacement_to_coord(center)
diff = dcont.astype('int')
diff = get_valid_diffeomorphism(diff)
diffeo2d = Diffeomorphism2D(diff)
return diffeo2d
def make_hard(dd):
assert isinstance(dd, Diffeomorphism2D)
if use_isomorphism_heuristics:
stats = diffeo_stats(dd.d)
per = np.percentile(stats.norm, norm_percentile)
limit = per * factor
# / 3.0
# print('norm mean/mean: %g %g' % (np.mean(stats.norm), np.median(stats.norm)))
# for i in range(0, 100, 5):
# print(' %3d%% = %g' % (i, np.percentile(stats.norm, i)))
# limit = np.percentile(stats.norm, info_percentile)
# if limit <= 1:
# print('limit was %g' % limit)
# limit = 4
variance = (stats.norm > limit).astype('float')
logger.info('---hard choices---')
logger.info(' per: %g pixels * %g =' % (per, factor))
logger.info('limit: %g pixels' % limit)
logger.info(' vis: %.1f%% ' % (100 * np.mean(variance)))
else:
variance = (dd.variance > info_threshold).astype('float')
return Diffeomorphism2D(dd.d, variance)
def get_value(self):
'''
Find maximum likelihood estimate for diffeomorphism looking
at each pixel singularly.
Returns a Diffeomorphism2D.
'''
if not self.initialized():
msg = 'No data seen yet'
raise Diffeo2dEstimatorInterface.NotReady(msg)
maximum_likelihood_index = np.zeros(self.shape, dtype='int32')
variance = np.zeros(self.shape, dtype='float32')
E2 = np.zeros(self.shape, dtype='float32')
E3 = np.zeros(self.shape, dtype='float32')
E4 = np.zeros(self.shape, dtype='float32')
num_problems = 0
i = 0
# for each coordinate
for c in coords_iterate(self.shape):
# find index in flat array
k = self.flattening.cell2index[c]
# Look at the average similarities of the neihgbors
sim = self.neighbor_similarity_flat[k]
sim_min = sim.min()
sim_max = sim.max()
if sim_max == sim_min:
# if all the neighbors have the same similarity
best_index = 0
variance[c] = 0 # minimum information
maximum_likelihood_index[c] = best_index
else:
best = np.argmin(sim)
best_index = self.neighbor_indices_flat[k][best]
# uncertainty ~= similarity of the best pixel
variance[c] = sim[best]
maximum_likelihood_index[c] = best_index
E2[c] = self.neighbor_similarity_best[k] / self.num_samples
# Best match error
E3[c] = np.min(
self.neighbor_num_bestmatch_flat[k]) / self.num_samples
E4[c] = np.min(self.neighbor_argsort_flat[k]) / self.num_samples
i += 1
d = self.flattening.flat2coords(maximum_likelihood_index)
if num_problems > 0:
logger.info('Warning, %d were not informative.' % num_problems)
pass
# normalization for this variance measure
vmin = variance.min()
variance = variance - vmin
vmax = variance.max()
if vmax > 0:
variance *= (1 / vmax)
# return maximum likelihood plus uncertainty measure
return Diffeomorphism2D(d, variance) # , E2, E3, E4)