def extract_features(self, image, settings={}, must_preserve_size=False):
sett = self.bedges_settings().copy()
sett['radius'] = 0
sett['preserve_size'] = False or must_preserve_size
if 1:
unspread_edges = ag.features.bedges(image, **sett)
else:
# LEAVE-BEHIND: From multi-channel images
unspread_edges = ag.features.bedges_from_image(image, **sett)
# Now do spreading
edges = ag.features.bspread(unspread_edges,
spread=self.bedges_settings()['spread'],
radius=self.bedges_settings()['radius'])
feats = self.extract_parts(edges, unspread_edges, settings=settings)
sett = self.settings
sett.update(settings)
psize = sett.get('subsample_size', (1, 1))
feats = gv.sub.subsample(feats, psize)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i]
for i in range(2))
lower = (buf[0] // 2, buf[1] // 2)
upper = tuple(image.shape[i] - (buf[i] - lower[i]) for i in range(2))
return gv.ndfeature(feats, lower=lower, upper=upper)
def _response_map(self, feats, mixcomp, strides=(1, 1)):
sh = self.svms[mixcomp]['weights'].shape
pmult = self.settings.get('padding_multiple_of_object', 0.5)
padding = (int(sh[0] * pmult), int(sh[1] * pmult), 0)
if min(feats.shape[:2]) < 2:
return np.array([]), None, padding
bigger = gv.ndfeature.zeropad(feats, padding)
from .fast import multifeature_real_correlate2d
#index = 26
#res = multifeature_correlate2d(bigger[...,index:index+1], weights[...,index:index+1].astype(np.float64))
weights = self.svms[mixcomp]['weights']
if bigger.shape[0] < weights.shape[0] or bigger.shape[
1] < weights.shape[1]:
return np.zeros((0, 0, 0)), None, padding
res = self.svms[mixcomp]['intercept'] + \
multifeature_real_correlate2d(bigger.astype(np.float64), weights, strides=strides)
lower, upper = gv.ndfeature.inner_frame(bigger, (sh[0] / 2, sh[1] / 2))
res = gv.ndfeature(res, lower=lower, upper=upper)
return res, bigger, padding
def _response_map(self, feats, mixcomp, strides=(1, 1)):
sh = self.svms[mixcomp]['weights'].shape
pmult = self.settings.get('padding_multiple_of_object', 0.5)
padding = (int(sh[0]*pmult), int(sh[1]*pmult), 0)
if min(feats.shape[:2]) < 2:
return np.array([]), None, padding
bigger = gv.ndfeature.zeropad(feats, padding)
from .fast import multifeature_real_correlate2d
#index = 26
#res = multifeature_correlate2d(bigger[...,index:index+1], weights[...,index:index+1].astype(np.float64))
weights = self.svms[mixcomp]['weights']
if bigger.shape[0] < weights.shape[0] or bigger.shape[1] < weights.shape[1]:
return np.zeros((0, 0, 0)), None, padding
res = self.svms[mixcomp]['intercept'] + \
multifeature_real_correlate2d(bigger.astype(np.float64), weights, strides=strides)
lower, upper = gv.ndfeature.inner_frame(bigger, (sh[0]/2, sh[1]/2))
res = gv.ndfeature(res, lower=lower, upper=upper)
return res, bigger, padding
def extract_features(self, image, settings={}, must_preserve_size=False):
sett = self.bedges_settings().copy()
sett['radius'] = 0
sett['preserve_size'] = False or must_preserve_size
if 1:
unspread_edges = ag.features.bedges(image, **sett)
else:
# LEAVE-BEHIND: From multi-channel images
unspread_edges = ag.features.bedges_from_image(image, **sett)
# Now do spreading
edges = ag.features.bspread(unspread_edges, spread=self.bedges_settings()['spread'], radius=self.bedges_settings()['radius'])
feats = self.extract_parts(edges, unspread_edges, settings=settings)
sett = self.settings
sett.update(settings)
psize = sett.get('subsample_size', (1, 1))
feats = gv.sub.subsample(feats, psize)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
return gv.ndfeature(feats, lower=lower, upper=upper)
def extract_features(self, image, settings={}, must_preserve_size=False):
sett = self.settings
sett.update(settings)
radii = sett.get('spread_radii', (0, 0))
psize = sett.get('subsample_size', (1, 1))
spread_shape = (radii[0] * 2 + 1, radii[1] * 2 + 1)
self._net.layers[-1]._shape = spread_shape
self._net.layers[-1]._strides = psize
feats = self._net.extract(image[np.newaxis])[0]
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
return gv.ndfeature(feats, lower=lower, upper=upper)
def extract_features(self, image, settings={}, must_preserve_size=False):
sett = self.settings
sett.update(settings)
radii = sett.get('spread_radii', (0, 0))
psize = sett.get('subsample_size', (1, 1))
spread_shape = (radii[0] * 2 + 1, radii[1] * 2 + 1)
self._net.layers[-1]._shape = spread_shape
self._net.layers[-1]._strides = psize
feats = self._net.extract(image[np.newaxis])[0]
buf = tuple(image.shape[i] - feats.shape[i] * psize[i]
for i in range(2))
lower = (buf[0] // 2, buf[1] // 2)
upper = tuple(image.shape[i] - (buf[i] - lower[i]) for i in range(2))
return gv.ndfeature(feats, lower=lower, upper=upper)
def extract_features(self,
image,
settings={},
must_preserve_size=False,
dropout=None):
sett = self.bedges_settings().copy()
sett['radius'] = 0
if 1:
#unspread_edges = ag.features.bedges(image, **sett)
#unspread_edges = gv.gradients.extract(image, orientations=8)
unspread_edges = _extract_edges(
self.bedges_settings(),
self.settings,
image,
must_preserve_size=must_preserve_size)
else:
# LEAVE-BEHIND: From multi-channel images
unspread_edges = ag.features.bedges_from_image(image, **sett)
# Now do spreading
edges = ag.features.bspread(unspread_edges,
spread=self.bedges_settings()['spread'],
radius=self.bedges_settings()['radius'])
# TODO Temporary
#sett['preserve_size'] = True
#unspread_edges = ag.features.bedges(image, **sett)
th = self.threshold_in_counts(self.settings['threshold'],
edges.shape[-1],
sett['contrast_insensitive'])
# TODO : Since we're using a hard-coded tau
if self.settings.get('tau', 0) == 0:
sett = self.settings.copy()
sett.update(settings)
psize = sett.get('subsample_size', (1, 1))
ORI = self.settings.get('orientations', 1)
POL = self.settings.get('polarities', 1)
P = ORI * POL
H = P // 2
if POL == 2 and not self.settings.get('no_collapse', False):
part_to_feature = np.zeros(self.parts.shape[0], dtype=np.int64)
for f in range(part_to_feature.shape[0]):
thepart = f // P
ori = f % H
v = thepart * H + ori
part_to_feature[f] = v
else:
part_to_feature = np.arange(self.parts.shape[0],
dtype=np.int64)
# Rotation spreading?
rotspread = sett.get('rotation_spreading_radius', 0)
if rotspread == 0:
between_feature_spreading = None
else:
between_feature_spreading = np.zeros(
(self.num_parts, rotspread * 2 + 1), dtype=np.int32)
for f in range(self.num_parts):
thepart = f // ORI
ori = f % ORI
for i in range(rotspread * 2 + 1):
between_feature_spreading[
f, i] = thepart * ORI + (ori - rotspread + i) % ORI
feats = ag.features.extract_parts(
edges,
unspread_edges,
self._log_parts,
self._log_invparts,
th,
self.settings['patch_frame'],
spread_radii=sett.get('spread_radii', (0, 0)),
subsample_size=psize,
part_to_feature=part_to_feature,
stride=self.settings.get('part_coding_stride', 1),
between_feature_spreading=between_feature_spreading)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i]
for i in range(2))
lower = (buf[0] // 2, buf[1] // 2)
upper = tuple(image.shape[i] - (buf[i] - lower[i])
for i in range(2))
else:
sett = self.settings.copy()
sett.update(settings)
feats = self.extract_parts(edges,
unspread_edges,
settings=sett,
dropout=dropout)
psize = sett.get('subsample_size', (1, 1))
feats = gv.sub.subsample(feats, psize)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i]
for i in range(2))
lower = (buf[0] // 2, buf[1] // 2)
upper = tuple(image.shape[i] - (buf[i] - lower[i])
for i in range(2))
# Now collapse the polarities
#feats = feats.reshape((int(feats.shape[0]//2), 2) + feats.shape[1:])
feats = feats.reshape(feats.shape[:2] + (feats.shape[2] // 2, 2))
feats = feats.max(axis=-1)
# TODO: Experiment
if 0:
Q = np.load('Q.npy')
new_feats_shape = feats.shape
new_feats = np.empty(new_feats_shape, dtype=np.uint8)
for i, j in itr.product(range(feats.shape[0]),
range(feats.shape[1])):
# Transform the basis 0.572
#new_feats[i,j] = np.dot(Q[:,-ARTS:].T, feats[i,j])
new_feats[i, j] = (np.fabs(
np.dot(Q.T, feats[i, j].astype(float))) > 15)
feats = new_feats
return gv.ndfeature(feats, lower=lower, upper=upper)
def extract_features(self, image, settings={}, dropout=None):
sett = self.bedges_settings().copy()
sett['radius'] = 0
if 1:
#unspread_edges = ag.features.bedges(image, **sett)
#unspread_edges = gv.gradients.extract(image, orientations=8)
unspread_edges = self._extract_edges(image)
else:
# LEAVE-BEHIND: From multi-channel images
unspread_edges = ag.features.bedges_from_image(image, **sett)
# Now do spreading
edges = ag.features.bspread(unspread_edges, spread=self.bedges_settings()['spread'], radius=self.bedges_settings()['radius'])
# TODO Temporary
#sett['preserve_size'] = True
#unspread_edges = ag.features.bedges(image, **sett)
th = self.threshold_in_counts(self.settings['threshold'], edges.shape[-1])
# TODO : Since we're using a hard-coded tau
if self.settings.get('tau', 0) == 0:
sett = self.settings.copy()
sett.update(settings)
psize = sett.get('subsample_size', (1, 1))
if 0:
feats = ag.features.extract_parts_adaptive_EXPERIMENTAL(edges, unspread_edges,
self._log_parts,
self._log_invparts,
th,
self.settings['patch_frame'],
spread_radii=sett.get('spread_radii', (0, 0)),
subsample_size=psize,
collapse=2,
accept_threshold=15)
elif 1:
# These are experiments with the new edge type
#import gv.fast
#th = self.settings['amp_threshold']
#feats = gv.fast.extract_parts(edges, unspread_edges, amps,
feats = ag.features.extract_parts(edges, unspread_edges,
self._log_parts,
self._log_invparts,
th,
self.settings['patch_frame'],
spread_radii=sett.get('spread_radii', (0, 0)),
subsample_size=psize,
part_to_feature=np.arange(self.num_features*2)//2,
stride=self.settings.get('part_coding_stride', 1))
else:
all_feats = []
for i in range(10):
feats = ag.features.extract_parts(edges, unspread_edges,
self._log_parts,
self._log_invparts,
th,
self.settings['patch_frame'],
spread_radii=sett.get('spread_radii', (0, 0)),
subsample_size=psize,
collapse=2,
accept_threshold=-100000)
all_feats.append(feats)
all_feats = np.asarray(all_feats)
feats = all_feats.max(axis=0)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
else:
sett = self.settings.copy()
sett.update(settings)
feats = self.extract_parts(edges, unspread_edges, settings=sett, dropout=dropout)
psize = sett.get('subsample_size', (1, 1))
feats = gv.sub.subsample(feats, psize)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
# Now collapse the polarities
#feats = feats.reshape((int(feats.shape[0]//2), 2) + feats.shape[1:])
feats = feats.reshape(feats.shape[:2] + (feats.shape[2]//2, 2))
feats = feats.max(axis=-1)
# TODO: Experiment
if 0:
Q = np.load('Q.npy')
new_feats_shape = feats.shape
new_feats = np.empty(new_feats_shape, dtype=np.uint8)
#import pdb; pdb.set_trace()
for i, j in itr.product(range(feats.shape[0]), range(feats.shape[1])):
# Transform the basis 0.572
#new_feats[i,j] = np.dot(Q[:,-ARTS:].T, feats[i,j])
new_feats[i,j] = (np.fabs(np.dot(Q.T, feats[i,j].astype(float))) > 15)
feats = new_feats
return gv.ndfeature(feats, lower=lower, upper=upper)
def extract_features(self, image, settings={}, must_preserve_size=False, dropout=None):
sett = self.bedges_settings().copy()
sett['radius'] = 0
if 1:
#unspread_edges = ag.features.bedges(image, **sett)
#unspread_edges = gv.gradients.extract(image, orientations=8)
unspread_edges = _extract_edges(self.bedges_settings(), self.settings, image, must_preserve_size=must_preserve_size)
else:
# LEAVE-BEHIND: From multi-channel images
unspread_edges = ag.features.bedges_from_image(image, **sett)
# Now do spreading
edges = ag.features.bspread(unspread_edges, spread=self.bedges_settings()['spread'], radius=self.bedges_settings()['radius'])
# TODO Temporary
#sett['preserve_size'] = True
#unspread_edges = ag.features.bedges(image, **sett)
th = self.threshold_in_counts(self.settings['threshold'], edges.shape[-1], sett['contrast_insensitive'])
# TODO : Since we're using a hard-coded tau
if self.settings.get('tau', 0) == 0:
sett = self.settings.copy()
sett.update(settings)
psize = sett.get('subsample_size', (1, 1))
ORI = self.settings.get('orientations', 1)
POL = self.settings.get('polarities', 1)
P = ORI * POL
H = P // 2
if POL == 2 and not self.settings.get('no_collapse', False):
part_to_feature = np.zeros(self.parts.shape[0], dtype=np.int64)
for f in range(part_to_feature.shape[0]):
thepart = f // P
ori = f % H
v = thepart * H + ori
part_to_feature[f] = v
else:
part_to_feature = np.arange(self.parts.shape[0], dtype=np.int64)
# Rotation spreading?
rotspread = sett.get('rotation_spreading_radius', 0)
if rotspread == 0:
between_feature_spreading = None
else:
between_feature_spreading = np.zeros((self.num_parts, rotspread*2 + 1), dtype=np.int32)
for f in range(self.num_parts):
thepart = f // ORI
ori = f % ORI
for i in range(rotspread*2 + 1):
between_feature_spreading[f,i] = thepart * ORI + (ori - rotspread + i) % ORI
feats = ag.features.extract_parts(edges, unspread_edges,
self._log_parts,
self._log_invparts,
th,
self.settings['patch_frame'],
spread_radii=sett.get('spread_radii', (0, 0)),
subsample_size=psize,
part_to_feature=part_to_feature,
stride=self.settings.get('part_coding_stride', 1),
between_feature_spreading=between_feature_spreading)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
else:
sett = self.settings.copy()
sett.update(settings)
feats = self.extract_parts(edges, unspread_edges, settings=sett, dropout=dropout)
psize = sett.get('subsample_size', (1, 1))
feats = gv.sub.subsample(feats, psize)
buf = tuple(image.shape[i] - feats.shape[i] * psize[i] for i in range(2))
lower = (buf[0]//2, buf[1]//2)
upper = tuple(image.shape[i] - (buf[i]-lower[i]) for i in range(2))
# Now collapse the polarities
#feats = feats.reshape((int(feats.shape[0]//2), 2) + feats.shape[1:])
feats = feats.reshape(feats.shape[:2] + (feats.shape[2]//2, 2))
feats = feats.max(axis=-1)
# TODO: Experiment
if 0:
Q = np.load('Q.npy')
new_feats_shape = feats.shape
new_feats = np.empty(new_feats_shape, dtype=np.uint8)
for i, j in itr.product(range(feats.shape[0]), range(feats.shape[1])):
# Transform the basis 0.572
#new_feats[i,j] = np.dot(Q[:,-ARTS:].T, feats[i,j])
new_feats[i,j] = (np.fabs(np.dot(Q.T, feats[i,j].astype(float))) > 15)
feats = new_feats
return gv.ndfeature(feats, lower=lower, upper=upper)
