def mouse_press_optimize(self, event):
self.event = event
p = self._pt_xyz(event)
if p is not None:
print "Optimizing particle near", p
n = self.state.obj_closest_particle(p)
old_err = self.state.error
_ = opt.do_levmarq_particles(self.state, np.array([n]), max_iter=2)
new_err = self.state.error
print '{}->{}'.format(old_err, new_err)
self.state.set_tile(self.state.oshape)
self.set_field()
self.draw()
def check_add_particles(st,
guess,
rad='calc',
do_opt=True,
im_change_frac=0.2,
min_derr='3sig',
**kwargs):
"""
Checks whether to add particles at a given position by seeing if adding
the particle improves the fit of the state.
Parameters
----------
st : :class:`peri.states.State`
The state to check adding particles to.
guess : [N,3] list-like
The positions of particles to check to add.
rad : {Float, ``'calc'``}, optional.
The radius of the newly-added particles. Default is ``'calc'``,
which uses the states current radii's median.
do_opt : Bool, optional
Whether to optimize the particle position before checking if it
should be kept. Default is True (optimizes position).
im_change_frac : Float
How good the change in error needs to be relative to the change in
the difference image. Default is 0.2; i.e. if the error does not
decrease by 20% of the change in the difference image, do not add
the particle.
min_derr : Float or '3sig'
The minimal improvement in error to add a particle. Default
is ``'3sig' = 3*st.sigma``.
Returns
-------
accepts : Int
The number of added particles
new_poses : [N,3] list
List of the positions of the added particles. If ``do_opt==True``,
then these positions will differ from the input 'guess'.
"""
# FIXME does not use the **kwargs, but needs b/c called with wrong kwargs
if min_derr == '3sig':
min_derr = 3 * st.sigma
accepts = 0
new_poses = []
if rad == 'calc':
rad = guess_add_radii(st)
message = ('-' * 30 + 'ADDING' + '-' * 30 +
'\n Z\t Y\t X\t R\t|\t ERR0\t\t ERR1')
with log.noformat():
CLOG.info(message)
for a in range(guess.shape[0]):
p0 = guess[a]
absent_err = st.error
absent_d = st.residuals.copy()
ind = st.obj_add_particle(p0, rad)
if do_opt:
# the slowest part of this
opt.do_levmarq_particles(st,
ind,
damping=1.0,
max_iter=1,
run_length=3,
eig_update=False,
include_rad=False)
present_err = st.error
present_d = st.residuals.copy()
dont_kill = should_particle_exist(absent_err,
present_err,
absent_d,
present_d,
im_change_frac=im_change_frac,
min_derr=min_derr)
if dont_kill:
accepts += 1
p = tuple(st.obj_get_positions()[ind].ravel())
r = tuple(st.obj_get_radii()[ind].ravel())
new_poses.append(p)
part_msg = '%2.2f\t%3.2f\t%3.2f\t%3.2f\t|\t%4.3f \t%4.3f' % (
p + r + (absent_err, st.error))
with log.noformat():
CLOG.info(part_msg)
else:
st.obj_remove_particle(ind)
if np.abs(absent_err - st.error) > 1e-4:
raise RuntimeError('updates not exact?')
return accepts, new_poses
def add_subtract_misfeatured_tile(st,
tile,
rad='calc',
max_iter=3,
invert='guess',
max_allowed_remove=20,
minmass=None,
use_tp=False,
**kwargs):
"""
Automatically adds and subtracts missing & extra particles in a region
of poor fit.
Parameters
----------
st: :class:`peri.states.State`
The state to add and subtract particles to.
tile : :class:`peri.util.Tile`
The poorly-fit region to examine.
rad : Float or 'calc', optional
The initial radius for added particles; added particles radii are
not fit until the end of add_subtract. Default is ``'calc'``, which
uses the median radii of active particles.
max_iter : Int, optional
The maximum number of loops for attempted adds at one tile location.
Default is 3.
invert : {'guess', True, False}, optional
Whether to invert the image for feature_guess -- True for dark
particles on a bright background, False for bright particles. The
default is to guess from the state's current particles.
max_allowed_remove : Int, optional
The maximum number of particles to remove. If the misfeatured tile
contains more than this many particles, raises an error. If it
contains more than half as many particles, logs a warning. If more
than this many particles are added, they are optimized in blocks of
``max_allowed_remove``. Default is 20.
Other Parameters
----------------
im_change_frac : Float on [0, 1], optional.
If adding or removing a particle decreases the error less than
``im_change_frac``*the change in the image, the particle is deleted.
Default is 0.2.
min_derr : {Float, ``'3sig'``}, optional
The minimum change in the state's error to keep a particle in the
image. Default is ``'3sig'`` which uses ``3*st.sigma``.
do_opt : Bool, optional
Set to False to avoid optimizing particle positions after adding
them. Default is True.
minmass : Float, optional
The minimum mass for a particle to be identified as a feature, as
used by trackpy. Defaults to a decent guess.
use_tp : Bool, optional
Set to True to use trackpy to find missing particles inside the
image. Not recommended since trackpy deliberately cuts out particles
at the edge of the image. Default is False.
Outputs
-------
n_added : Int
The change in the number of particles, i.e. ``n_added-n_subtracted``
ainds: List of ints
The indices of the added particles.
Notes
--------
The added/removed positions returned are whether or not the
position has been added or removed ever. It's possible/probably that
a position is added, then removed during a later iteration.
Algorithm is:
1. Remove all particles within the tile.
2. Feature and add particles to the tile.
3. Optimize the added particles positions only.
4. Run 2-3 until no particles have been added.
5. Optimize added particle radii
Because all the particles are removed within a tile, it is important
to set max_allowed_remove to a reasonable value. Otherwise, if the
tile is the size of the image it can take a long time to remove all
the particles and re-add them.
"""
if rad == 'calc':
rad = guess_add_radii(st)
if invert == 'guess':
invert = guess_invert(st)
# 1. Remove all possibly bad particles within the tile.
initial_error = np.copy(st.error)
rinds = np.nonzero(tile.contains(st.obj_get_positions()))[0]
if rinds.size >= max_allowed_remove:
CLOG.fatal('Misfeatured region too large!')
raise RuntimeError
elif rinds.size >= max_allowed_remove / 2:
CLOG.warn('Large misfeatured regions.')
elif rinds.size > 0:
rpos, rrad = st.obj_remove_particle(rinds)
# 2-4. Feature & add particles to the tile, optimize, run until none added
n_added = -rinds.size
added_poses = []
for _ in range(max_iter):
if invert:
im = 1 - st.residuals[tile.slicer]
else:
im = st.residuals[tile.slicer]
guess, _ = _feature_guess(im, rad, minmass=minmass, use_tp=use_tp)
accepts, poses = check_add_particles(st,
guess + tile.l,
rad=rad,
do_opt=True,
**kwargs)
added_poses.extend(poses)
n_added += accepts
if accepts == 0:
break
else: # for-break-else
CLOG.warn('Runaway adds or insufficient max_iter')
# 5. Optimize added pos + rad:
ainds = []
for p in added_poses:
ainds.append(st.obj_closest_particle(p))
if len(ainds) > max_allowed_remove:
for i in range(0, len(ainds), max_allowed_remove):
opt.do_levmarq_particles(st,
np.array(ainds[i:i + max_allowed_remove]),
include_rad=True,
max_iter=3)
elif len(ainds) > 0:
opt.do_levmarq_particles(st, ainds, include_rad=True, max_iter=3)
# 6. Ensure that current error after add-subtracting is lower than initial
did_something = (rinds.size > 0) or (len(ainds) > 0)
if did_something & (st.error > initial_error):
CLOG.info('Failed addsub, Tile {} -> {}'.format(
tile.l.tolist(), tile.r.tolist()))
if len(ainds) > 0:
_ = st.obj_remove_particle(ainds)
if rinds.size > 0:
for p, r in zip(rpos.reshape(-1, 3), rrad.reshape(-1)):
_ = st.obj_add_particle(p, r)
n_added = 0
ainds = []
return n_added, ainds
def add_subtract(st,
max_iter=7,
max_npart='calc',
max_mem=2e8,
always_check_remove=False,
**kwargs):
"""
Automatically adds and subtracts missing & extra particles.
Operates by removing bad particles then adding missing particles on
repeat, until either no particles are added/removed or after `max_iter`
attempts.
Parameters
----------
st: :class:`peri.states.State`
The state to add and subtract particles to.
max_iter : Int, optional
The maximum number of add-subtract loops to use. Default is 7.
Terminates after either max_iter loops or when nothing has changed.
max_npart : Int or 'calc', optional
The maximum number of particles to add before optimizing the non-psf
globals. Default is ``'calc'``, which uses 5% of the initial number
of particles.
max_mem : Int, optional
The maximum memory to use for optimization after adding max_npart
particles. Default is 2e8.
always_check_remove : Bool, optional
Set to True to always check whether to remove particles. If ``False``,
only checks for removal while particles were removed on the previous
attempt. Default is False.
Other Parameters
----------------
invert : Bool, optional
``True`` if the particles are dark on a bright background, ``False``
if they are bright on a dark background. Default is ``True``.
min_rad : Float, optional
Particles with radius below ``min_rad`` are automatically deleted.
Default is ``'calc'`` = median rad - 25* radius std.
max_rad : Float, optional
Particles with radius above ``max_rad`` are automatically deleted.
Default is ``'calc'`` = median rad + 15* radius std, but you should
change this for your particle sizes.
min_edge_dist : Float, optional
Particles closer to the edge of the padded image than this are
automatically deleted. Default is 2.0.
check_rad_cutoff : 2-element float list.
Particles with ``radii < check_rad_cutoff[0]`` or ``> check...[1]``
are checked if they should be deleted (not automatic). Default is
``[3.5, 15]``.
check_outside_im : Bool, optional
Set to True to check whether to delete particles whose positions are
outside the un-padded image.
rad : Float, optional
The initial radius for added particles; added particles radii are
not fit until the end of ``add_subtract``. Default is ``'calc'``,
which uses the median radii of active particles.
tries : Int, optional
The number of particles to attempt to remove or add, per iteration.
Default is 50.
im_change_frac : Float, optional
How good the change in error needs to be relative to the change in
the difference image. Default is 0.2; i.e. if the error does not
decrease by 20% of the change in the difference image, do not add
the particle.
min_derr : Float, optional
The minimum change in the state's error to keep a particle in the
image. Default is ``'3sig'`` which uses ``3*st.sigma``.
do_opt : Bool, optional
Set to False to avoid optimizing particle positions after adding.
minmass : Float, optional
The minimum mass for a particle to be identified as a feature,
as used by trackpy. Defaults to a decent guess.
use_tp : Bool, optional
Set to True to use trackpy to find missing particles inside the
image. Not recommended since trackpy deliberately cuts out particles
at the edge of the image. Default is ``False``.
Returns
-------
total_changed : Int
The total number of adds and subtracts done on the data. Not the
same as ``changed_inds.size`` since the same particle or particle
index can be added/subtracted multiple times.
added_positions : [N_added,3] numpy.ndarray
The positions of particles that have been added at any point in the
add-subtract cycle.
removed_positions : [N_added,3] numpy.ndarray
The positions of particles that have been removed at any point in
the add-subtract cycle.
Notes
------
Occasionally after the intial featuring a cluster of particles is
featured as 1 big particle. To fix these mistakes, it helps to set
max_rad to a physical value. This removes the big particle and allows
it to be re-featured by (several passes of) the adds.
The added/removed positions returned are whether or not the position
has been added or removed ever. It's possible that a position is
added, then removed during a later iteration.
"""
if max_npart == 'calc':
max_npart = 0.05 * st.obj_get_positions().shape[0]
total_changed = 0
_change_since_opt = 0
removed_poses = []
added_poses0 = []
added_poses = []
nr = 1 # Check removal on the first loop
for _ in range(max_iter):
if (nr != 0) or (always_check_remove):
nr, rposes = remove_bad_particles(st, **kwargs)
na, aposes = add_missing_particles(st, **kwargs)
current_changed = na + nr
removed_poses.extend(rposes)
added_poses0.extend(aposes)
total_changed += current_changed
_change_since_opt += current_changed
if current_changed == 0:
break
elif _change_since_opt > max_npart:
_change_since_opt *= 0
CLOG.info('Start add_subtract optimization.')
opt.do_levmarq(st,
opt.name_globals(
st, remove_params=st.get('psf').params),
max_iter=1,
run_length=4,
num_eig_dirs=3,
max_mem=max_mem,
eig_update_frequency=2,
rz_order=0,
use_accel=True)
CLOG.info('After optimization:\t{:.6}'.format(st.error))
# Optimize the added particles' radii:
for p in added_poses0:
i = st.obj_closest_particle(p)
opt.do_levmarq_particles(st, np.array([i]), max_iter=2, damping=0.3)
added_poses.append(st.obj_get_positions()[i])
return total_changed, np.array(removed_poses), np.array(added_poses)