def run_spherical_relp_stills_pred_param(self, verbose=True):
if verbose:
print 'Testing derivatives for SphericalRelpStillsPredictionParameterisation'
print '====================================================================='
# Build a prediction parameterisation for the stills experiment
pred_param = SphericalRelpStillsPredictionParameterisation(
self.stills_experiments,
detector_parameterisations = [self.det_param],
beam_parameterisations = [self.s0_param],
xl_orientation_parameterisations = [self.xlo_param],
xl_unit_cell_parameterisations = [self.xluc_param])
# Predict the reflections in place. Must do this ahead of calculating
# the analytical gradients so quantities like s1 are correct
from dials.algorithms.refinement.prediction import ExperimentsPredictor
ref_predictor = ExperimentsPredictor(self.stills_experiments,
spherical_relp=True)
ref_predictor.update()
ref_predictor.predict(self.reflections)
# get analytical gradients
an_grads = pred_param.get_gradients(self.reflections)
fd_grads = self.get_fd_gradients(pred_param, ref_predictor)
# compare FD with analytical calculations
for i, (an_grad, fd_grad) in enumerate(zip(an_grads, fd_grads)):
if verbose: print "\nParameter {0}: {1}". format(i, fd_grad['name'])
for idx, name in enumerate(["dX_dp", "dY_dp", "dDeltaPsi_dp"]):
if verbose: print name
for a, b in zip(an_grad[name], fd_grad[name]):
if name == 'dDeltaPsi_dp':
# DeltaPsi errors are much worse than X, Y errors!
# FIXME, look into this further
assert approx_equal(a,b, eps=5e-3)
else:
assert approx_equal(a,b, eps=5e-6)
if verbose: print "OK"
if verbose: print
pred_param.compose(reflections) an_grads = pred_param.get_gradients(reflections) # get finite difference gradients p_vals = pred_param.get_param_vals() deltas = [1.e-7] * len(p_vals) for i in range(len(deltas)): val = p_vals[i] p_vals[i] -= deltas[i] / 2. pred_param.set_param_vals(p_vals) pred_param.compose(reflections) ref_predictor.update() ref_predictor.predict(reflections) rev_state = reflections['xyzcal.mm'].deep_copy() p_vals[i] += deltas[i] pred_param.set_param_vals(p_vals) pred_param.compose(reflections) ref_predictor.update() ref_predictor.predict(reflections) fwd_state = reflections['xyzcal.mm'].deep_copy() p_vals[i] = val fd = (fwd_state - rev_state)
def run_stills_pred_param(self, verbose = False):
if verbose:
print 'Testing derivatives for StillsPredictionParameterisation'
print '========================================================'
# Build a prediction parameterisation for the stills experiment
pred_param = StillsPredictionParameterisation(self.stills_experiments,
detector_parameterisations = [self.det_param],
beam_parameterisations = [self.s0_param],
xl_orientation_parameterisations = [self.xlo_param],
xl_unit_cell_parameterisations = [self.xluc_param])
# Predict the reflections in place. Must do this ahead of calculating
# the analytical gradients so quantities like s1 are correct
from dials.algorithms.refinement.prediction import ExperimentsPredictor
ref_predictor = ExperimentsPredictor(self.stills_experiments)
ref_predictor.update()
ref_predictor.predict(self.reflections)
# get analytical gradients
an_grads = pred_param.get_gradients(self.reflections)
fd_grads = self.get_fd_gradients(pred_param, ref_predictor)
for i, (an_grad, fd_grad) in enumerate(zip(an_grads, fd_grads)):
# compare FD with analytical calculations
if verbose: print "\nParameter {0}: {1}". format(i, fd_grad['name'])
for idx, name in enumerate(["dX_dp", "dY_dp", "dDeltaPsi_dp"]):
if verbose: print name
a = fd_grad[name]
b = an_grad[name]
abs_error = a - b
denom = a + b
fns = five_number_summary(abs_error)
if verbose: print (" summary of absolute errors: %9.6f %9.6f %9.6f " + \
"%9.6f %9.6f") % fns
assert flex.max(flex.abs(abs_error)) < 0.0003
# largest absolute error found to be about 0.00025 for dY/dp of
# Crystal0g_param_3. Reject outlying absolute errors and test again.
iqr = fns[3] - fns[1]
# skip further stats on errors with an iqr of near zero, e.g. dDeltaPsi_dp
# for detector parameters, which are all equal to zero
if iqr < 1.e-10:
continue
sel1 = abs_error < fns[3] + 1.5 * iqr
sel2 = abs_error > fns[1] - 1.5 * iqr
sel = sel1 & sel2
tst = flex.max_index(flex.abs(abs_error.select(sel)))
tst_val = abs_error.select(sel)[tst]
n_outliers = sel.count(False)
if verbose: print (" {0} outliers rejected, leaving greatest " + \
"absolute error: {1:9.6f}").format(n_outliers, tst_val)
# largest absolute error now 0.000086 for dX/dp of Beam0Mu2
assert abs(tst_val) < 0.00009
# Completely skip parameters with FD gradients all zero (e.g. gradients of
# DeltaPsi for detector parameters)
sel1 = flex.abs(a) < 1.e-10
if sel1.all_eq(True):
continue
# otherwise calculate normalised errors, by dividing absolute errors by
# the IQR (more stable than relative error calculation)
norm_error = abs_error / iqr
fns = five_number_summary(norm_error)
if verbose: print (" summary of normalised errors: %9.6f %9.6f %9.6f " + \
"%9.6f %9.6f") % fns
# largest normalised error found to be about 25.7 for dY/dp of
# Crystal0g_param_3.
try:
assert flex.max(flex.abs(norm_error)) < 30
except AssertionError as e:
e.args += ("extreme normalised error value: {0}".format(
flex.max(flex.abs(norm_error))),)
raise e
# Reject outlying normalised errors and test again
iqr = fns[3] - fns[1]
if iqr > 0.:
sel1 = norm_error < fns[3] + 1.5 * iqr
sel2 = norm_error > fns[1] - 1.5 * iqr
sel = sel1 & sel2
tst = flex.max_index(flex.abs(norm_error.select(sel)))
tst_val = norm_error.select(sel)[tst]
n_outliers = sel.count(False)
# most outliers found for for dY/dp of Crystal0g_param_3 (which had
# largest errors, so no surprise there).
try:
assert n_outliers < 250
except AssertionError as e:
e.args += ("too many outliers rejected: {0}".format(n_outliers),)
raise e
if verbose: print (" {0} outliers rejected, leaving greatest " + \
"normalised error: {1:9.6f}").format(n_outliers, tst_val)
# largest normalied error now about -4. for dX/dp of Detector0Tau1
assert abs(tst_val) < 4.5
if verbose: print
return
# Add in flags and ID columns by copying into standard reflection table tmp = flex.reflection_table.empty_standard(len(obs_refs)) tmp.update(obs_refs) obs_refs = tmp # Invent some variances for the centroid positions of the simulated data im_width = 0.1 * pi / 180. px_size = mydetector[0].get_pixel_size() var_x = flex.double(len(obs_refs), (px_size[0] / 2.)**2) var_y = flex.double(len(obs_refs), (px_size[1] / 2.)**2) var_phi = flex.double(len(obs_refs), (im_width / 2.)**2) obs_refs['xyzobs.mm.variance'] = flex.vec3_double(var_x, var_y, var_phi) # Re-predict using the stills reflection predictor stills_ref_predictor = ExperimentsPredictor(stills_experiments) stills_ref_predictor.update() obs_refs_stills = stills_ref_predictor.predict(obs_refs) # Set 'observed' centroids from the predicted ones obs_refs_stills['xyzobs.mm.value'] = obs_refs_stills['xyzcal.mm'] ############################### # Undo known parameter shifts # ############################### xlo_param.set_param_vals(xlo_p_vals[0]) xluc_param.set_param_vals(xluc_p_vals[0]) # make a refiner from dials.algorithms.refinement.refiner import phil_scope from libtbx.phil import parse
