def draw_figures(self, ds, history, optimiser):
fontsize = self.font_size
fontsize_title = self.font_size_title
nxmax = self.nx
nymax = self.ny
problem = history.problem
for target in problem.targets:
target.set_dataset(ds)
target_index = {}
i = 0
for target in problem.targets:
target_index[target] = i, i+target.nmisfits
i += target.nmisfits
xbest = history.get_best_model()
misfits = history.misfits[history.get_sorted_misfits_idx(chain=0), ...]
ws = problem.get_target_weights()
gcms = problem.combine_misfits(
misfits[:1, :, :],
extra_correlated_weights=optimiser.get_correlated_weights(problem),
get_contributions=True)[0, :]
w_max = num.nanmax(ws)
gcm_max = num.nanmax(gcms)
source = problem.get_source(xbest)
target_to_result = {}
all_syn_trs = []
all_syn_specs = []
results = problem.evaluate(xbest)
dtraces = []
for target, result in zip(problem.targets, results):
if not isinstance(result, WaveformMisfitResult):
dtraces.extend([None] * target.nmisfits)
continue
itarget, itarget_end = target_index[target]
assert itarget_end == itarget + 1
w = target.get_combined_weight()
if target.misfit_config.domain == 'cc_max_norm':
tref = (
result.filtered_obs.tmin + result.filtered_obs.tmax) * 0.5
for tr_filt, tr_proc, tshift in (
(result.filtered_obs,
result.processed_obs,
0.),
(result.filtered_syn,
result.processed_syn,
result.tshift)):
norm = num.sum(num.abs(tr_proc.ydata)) / tr_proc.data_len()
tr_filt.ydata /= norm
tr_proc.ydata /= norm
tr_filt.shift(tshift)
tr_proc.shift(tshift)
ctr = result.cc
ctr.shift(tref)
dtrace = ctr
else:
for tr in (
result.filtered_obs,
result.filtered_syn,
result.processed_obs,
result.processed_syn):
tr.ydata *= w
for spec in (
result.spectrum_obs,
result.spectrum_syn):
if spec is not None:
spec.ydata *= w
if result.tshift is not None and result.tshift != 0.0:
# result.filtered_syn.shift(result.tshift)
result.processed_syn.shift(result.tshift)
dtrace = make_norm_trace(
result.processed_syn, result.processed_obs,
problem.norm_exponent)
target_to_result[target] = result
dtrace.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
dtraces.append(dtrace)
result.processed_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_trs.append(result.processed_syn)
if result.spectrum_syn:
result.spectrum_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_specs.append(result.spectrum_syn)
if not all_syn_trs:
logger.warn('No traces to show!')
return
def skey(tr):
return tr.meta['normalisation_family'], tr.meta['path']
trace_minmaxs = trace.minmax(all_syn_trs, skey)
amp_spec_maxs = amp_spec_max(all_syn_specs, skey)
dminmaxs = trace.minmax([x for x in dtraces if x is not None], skey)
for tr in dtraces:
if tr:
dmin, dmax = dminmaxs[skey(tr)]
tr.ydata /= max(abs(dmin), abs(dmax))
cg_to_targets = meta.gather(
problem.waveform_targets,
lambda t: (t.path, t.codes[3]),
filter=lambda t: t in target_to_result)
cgs = sorted(cg_to_targets.keys())
for cg in cgs:
targets = cg_to_targets[cg]
frame_to_target, nx, ny, nxx, nyy = layout(
source, targets, nxmax, nymax)
figures = {}
for iy in range(ny):
for ix in range(nx):
if (iy, ix) not in frame_to_target:
continue
ixx = ix // nxmax
iyy = iy // nymax
if (iyy, ixx) not in figures:
title = '_'.join(x for x in cg if x)
item = PlotItem(
name='fig_%s_%i_%i' % (title, ixx, iyy))
item.attributes['targets'] = []
figures[iyy, ixx] = (
item, plt.figure(figsize=self.size_inch))
figures[iyy, ixx][1].subplots_adjust(
left=0.03,
right=1.0 - 0.03,
bottom=0.03,
top=1.0 - 0.06,
wspace=0.2,
hspace=0.2)
item, fig = figures[iyy, ixx]
target = frame_to_target[iy, ix]
item.attributes['targets'].append(target.string_id())
amin, amax = trace_minmaxs[
target.normalisation_family, target.path]
absmax = max(abs(amin), abs(amax))
ny_this = nymax # min(ny, nymax)
nx_this = nxmax # min(nx, nxmax)
i_this = (iy % ny_this) * nx_this + (ix % nx_this) + 1
axes2 = fig.add_subplot(ny_this, nx_this, i_this)
space = 0.5
space_factor = 1.0 + space
axes2.set_axis_off()
axes2.set_ylim(-1.05 * space_factor, 1.05)
axes = axes2.twinx()
axes.set_axis_off()
if target.misfit_config.domain == 'cc_max_norm':
axes.set_ylim(-10. * space_factor, 10.)
else:
axes.set_ylim(
-absmax * 1.33 * space_factor, absmax * 1.33)
itarget, itarget_end = target_index[target]
assert itarget_end == itarget + 1
result = target_to_result[target]
dtrace = dtraces[itarget]
tap_color_annot = (0.35, 0.35, 0.25)
tap_color_edge = (0.85, 0.85, 0.80)
tap_color_fill = (0.95, 0.95, 0.90)
plot_taper(
axes2, result.processed_obs.get_xdata(), result.taper,
fc=tap_color_fill, ec=tap_color_edge)
obs_color = mpl_color('aluminium5')
obs_color_light = light(obs_color, 0.5)
syn_color = mpl_color('scarletred2')
syn_color_light = light(syn_color, 0.5)
misfit_color = mpl_color('scarletred2')
weight_color = mpl_color('chocolate2')
cc_color = mpl_color('aluminium5')
if target.misfit_config.domain == 'cc_max_norm':
tref = (result.filtered_obs.tmin +
result.filtered_obs.tmax) * 0.5
plot_dtrace(
axes2, dtrace, space, -1., 1.,
fc=light(cc_color, 0.5),
ec=cc_color)
plot_dtrace_vline(
axes2, tref, space, color=tap_color_annot)
elif target.misfit_config.domain == 'frequency_domain':
asmax = amp_spec_maxs[
target.normalisation_family, target.path]
fmin, fmax = \
target.misfit_config.get_full_frequency_range()
plot_spectrum(
axes2,
result.spectrum_syn,
result.spectrum_obs,
fmin, fmax,
space, 0., asmax,
syn_color=syn_color,
obs_color=obs_color,
syn_lw=1.0,
obs_lw=0.75,
color_vline=tap_color_annot,
fontsize=fontsize)
else:
plot_dtrace(
axes2, dtrace, space, 0., 1.,
fc=light(misfit_color, 0.3),
ec=misfit_color)
plot_trace(
axes, result.filtered_syn,
color=syn_color_light, lw=1.0)
plot_trace(
axes, result.filtered_obs,
color=obs_color_light, lw=0.75)
plot_trace(
axes, result.processed_syn,
color=syn_color, lw=1.0)
plot_trace(
axes, result.processed_obs,
color=obs_color, lw=0.75)
# xdata = result.filtered_obs.get_xdata()
tmarks = [
result.processed_obs.tmin,
result.processed_obs.tmax]
for tmark in tmarks:
axes2.plot(
[tmark, tmark], [-0.9, 0.1], color=tap_color_annot)
dur = tmarks[1] - tmarks[0]
for tmark, text, ha in [
(tmarks[0],
'$\\,$ ' + meta.str_duration(
tmarks[0] - source.time),
'left'),
(tmarks[1],
'$\\Delta$ ' + meta.str_duration(dur),
'right')]:
axes2.annotate(
text,
xy=(tmark, -0.9),
xycoords='data',
xytext=(
fontsize * 0.4 * [-1, 1][ha == 'left'],
fontsize * 0.2),
textcoords='offset points',
ha=ha,
va='bottom',
color=tap_color_annot,
fontsize=fontsize)
axes2.set_xlim(tmarks[0] - dur*0.1, tmarks[1] + dur*0.1)
rel_w = ws[itarget] / w_max
rel_c = gcms[itarget] / gcm_max
sw = 0.25
sh = 0.1
ph = 0.01
for (ih, rw, facecolor, edgecolor) in [
(0, rel_w, light(weight_color, 0.5),
weight_color),
(1, rel_c, light(misfit_color, 0.5),
misfit_color)]:
bar = patches.Rectangle(
(1.0 - rw * sw, 1.0 - (ih + 1) * sh + ph),
rw * sw,
sh - 2 * ph,
facecolor=facecolor, edgecolor=edgecolor,
zorder=10,
transform=axes.transAxes, clip_on=False)
axes.add_patch(bar)
scale_string = None
if target.misfit_config.domain == 'cc_max_norm':
scale_string = 'Syn/obs scales differ!'
infos = []
if scale_string:
infos.append(scale_string)
if self.nx == 1 and self.ny == 1:
infos.append(target.string_id())
else:
infos.append('.'.join(x for x in target.codes if x))
dist = source.distance_to(target)
azi = source.azibazi_to(target)[0]
infos.append(meta.str_dist(dist))
infos.append('%.0f\u00B0' % azi)
infos.append('%.3g' % ws[itarget])
infos.append('%.3g' % gcms[itarget])
axes2.annotate(
'\n'.join(infos),
xy=(0., 1.),
xycoords='axes fraction',
xytext=(2., 2.),
textcoords='offset points',
ha='left',
va='top',
fontsize=fontsize,
fontstyle='normal')
if (self.nx == 1 and self.ny == 1):
yield item, fig
del figures[iyy, ixx]
if not (self.nx == 1 and self.ny == 1):
for (iyy, ixx), (_, fig) in figures.items():
title = '.'.join(x for x in cg if x)
if len(figures) > 1:
title += ' (%i/%i, %i/%i)' % (
iyy + 1, nyy, ixx + 1, nxx)
fig.suptitle(title, fontsize=fontsize_title)
for item, fig in figures.values():
yield item, fig
def draw_figures(self, ds, history, optimiser):
color_parameter = self.color_parameter
misfit_cutoff = self.misfit_cutoff
fontsize = self.font_size
fontsize_title = self.font_size_title
nxmax = self.nx
nymax = self.ny
problem = history.problem
for target in problem.targets:
target.set_dataset(ds)
target_index = {}
i = 0
for target in problem.targets:
target_index[target] = i, i+target.nmisfits
i += target.nmisfits
gms = history.get_sorted_primary_misfits()[::-1]
models = history.get_sorted_primary_models()[::-1]
if misfit_cutoff is not None:
ibest = gms < misfit_cutoff
gms = gms[ibest]
models = models[ibest]
gms = gms[::10]
models = models[::10]
nmodels = models.shape[0]
if color_parameter == 'dist':
mx = num.mean(models, axis=0)
cov = num.cov(models.T)
mdists = core.mahalanobis_distance(models, mx, cov)
icolor = meta.ordersort(mdists)
elif color_parameter == 'misfit':
iorder = num.arange(nmodels)
icolor = iorder
elif color_parameter in problem.parameter_names:
ind = problem.name_to_index(color_parameter)
icolor = problem.extract(models, ind)
target_to_results = defaultdict(list)
all_syn_trs = []
dtraces = []
for imodel in range(nmodels):
model = models[imodel, :]
source = problem.get_source(model)
results = problem.evaluate(model)
dtraces.append([])
for target, result in zip(problem.targets, results):
w = target.get_combined_weight()
if isinstance(result, gf.SeismosizerError) or \
not isinstance(target, WaveformMisfitTarget) or \
not num.all(num.isfinite(w)):
dtraces[-1].extend([None] * target.nmisfits)
continue
itarget, itarget_end = target_index[target]
assert itarget_end == itarget + 1
if target.misfit_config.domain == 'cc_max_norm':
tref = (
result.filtered_obs.tmin + result.filtered_obs.tmax) \
* 0.5
for tr_filt, tr_proc, tshift in (
(result.filtered_obs,
result.processed_obs,
0.),
(result.filtered_syn,
result.processed_syn,
result.tshift)):
norm = num.sum(num.abs(tr_proc.ydata)) \
/ tr_proc.data_len()
tr_filt.ydata /= norm
tr_proc.ydata /= norm
tr_filt.shift(tshift)
tr_proc.shift(tshift)
ctr = result.cc
ctr.shift(tref)
dtrace = ctr
else:
for tr in (
result.filtered_obs,
result.filtered_syn,
result.processed_obs,
result.processed_syn):
tr.ydata *= w
if result.tshift is not None and result.tshift != 0.0:
# result.filtered_syn.shift(result.tshift)
result.processed_syn.shift(result.tshift)
dtrace = make_norm_trace(
result.processed_syn, result.processed_obs,
problem.norm_exponent)
target_to_results[target].append(result)
dtrace.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
dtraces[-1].append(dtrace)
result.processed_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_trs.append(result.processed_syn)
if not all_syn_trs:
logger.warn('No traces to show!')
return
def skey(tr):
return tr.meta['normalisation_family'], tr.meta['path']
trace_minmaxs = trace.minmax(all_syn_trs, skey)
dtraces_all = []
for dtraces_group in dtraces:
dtraces_all.extend(dtraces_group)
dminmaxs = trace.minmax([
dtrace_ for dtrace_ in dtraces_all if dtrace_ is not None], skey)
for tr in dtraces_all:
if tr:
dmin, dmax = dminmaxs[skey(tr)]
tr.ydata /= max(abs(dmin), abs(dmax))
cg_to_targets = meta.gather(
problem.waveform_targets,
lambda t: (t.path, t.codes[3]),
filter=lambda t: t in target_to_results)
cgs = sorted(cg_to_targets.keys())
from matplotlib import colors
cmap = cm.ScalarMappable(
norm=colors.Normalize(vmin=num.min(icolor), vmax=num.max(icolor)),
cmap=plt.get_cmap('coolwarm'))
imodel_to_color = []
for imodel in range(nmodels):
imodel_to_color.append(cmap.to_rgba(icolor[imodel]))
for cg in cgs:
targets = cg_to_targets[cg]
frame_to_target, nx, ny, nxx, nyy = layout(
source, targets, nxmax, nymax)
figures = {}
for iy in range(ny):
for ix in range(nx):
if (iy, ix) not in frame_to_target:
continue
ixx = ix // nxmax
iyy = iy // nymax
if (iyy, ixx) not in figures:
title = '_'.join(x for x in cg if x)
item = PlotItem(
name='fig_%s_%i_%i' % (title, ixx, iyy))
item.attributes['targets'] = []
figures[iyy, ixx] = (
item, plt.figure(figsize=self.size_inch))
figures[iyy, ixx][1].subplots_adjust(
left=0.03,
right=1.0 - 0.03,
bottom=0.03,
top=1.0 - 0.06,
wspace=0.2,
hspace=0.2)
item, fig = figures[iyy, ixx]
target = frame_to_target[iy, ix]
item.attributes['targets'].append(target.string_id())
amin, amax = trace_minmaxs[
target.normalisation_family, target.path]
absmax = max(abs(amin), abs(amax))
ny_this = nymax # min(ny, nymax)
nx_this = nxmax # min(nx, nxmax)
i_this = (iy % ny_this) * nx_this + (ix % nx_this) + 1
axes2 = fig.add_subplot(ny_this, nx_this, i_this)
space = 0.5
space_factor = 1.0 + space
axes2.set_axis_off()
axes2.set_ylim(-1.05 * space_factor, 1.05)
axes = axes2.twinx()
axes.set_axis_off()
if target.misfit_config.domain == 'cc_max_norm':
axes.set_ylim(-10. * space_factor, 10.)
else:
axes.set_ylim(-absmax*1.33 * space_factor, absmax*1.33)
itarget, itarget_end = target_index[target]
assert itarget_end == itarget + 1
for imodel, result in enumerate(target_to_results[target]):
syn_color = imodel_to_color[imodel]
dtrace = dtraces[imodel][itarget]
tap_color_annot = (0.35, 0.35, 0.25)
tap_color_edge = (0.85, 0.85, 0.80)
tap_color_fill = (0.95, 0.95, 0.90)
plot_taper(
axes2,
result.processed_obs.get_xdata(),
result.taper,
fc=tap_color_fill, ec=tap_color_edge, alpha=0.2)
obs_color = mpl_color('aluminium5')
obs_color_light = light(obs_color, 0.5)
plot_dtrace(
axes2, dtrace, space, 0., 1.,
fc='none',
ec=syn_color)
# plot_trace(
# axes, result.filtered_syn,
# color=syn_color_light, lw=1.0)
if imodel == 0:
plot_trace(
axes, result.filtered_obs,
color=obs_color_light, lw=0.75)
plot_trace(
axes, result.processed_syn,
color=syn_color, lw=1.0, alpha=0.3)
plot_trace(
axes, result.processed_obs,
color=obs_color, lw=0.75, alpha=0.3)
if imodel != 0:
continue
xdata = result.filtered_obs.get_xdata()
axes.set_xlim(xdata[0], xdata[-1])
tmarks = [
result.processed_obs.tmin,
result.processed_obs.tmax]
for tmark in tmarks:
axes2.plot(
[tmark, tmark], [-0.9, 0.1],
color=tap_color_annot)
dur = tmarks[1] - tmarks[0]
for tmark, text, ha in [
(tmarks[0],
'$\\,$ ' + meta.str_duration(
tmarks[0] - source.time),
'left'),
(tmarks[1],
'$\\Delta$ ' + meta.str_duration(
dur),
'right')]:
axes2.annotate(
text,
xy=(tmark, -0.9),
xycoords='data',
xytext=(
fontsize*0.4 * [-1, 1][ha == 'left'],
fontsize*0.2),
textcoords='offset points',
ha=ha,
va='bottom',
color=tap_color_annot,
fontsize=fontsize)
axes2.set_xlim(
tmarks[0] - dur*0.1, tmarks[1] + dur*0.1)
scale_string = None
if target.misfit_config.domain == 'cc_max_norm':
scale_string = 'Syn/obs scales differ!'
infos = []
if scale_string:
infos.append(scale_string)
if self.nx == 1 and self.ny == 1:
infos.append(target.string_id())
else:
infos.append('.'.join(x for x in target.codes if x))
dist = source.distance_to(target)
azi = source.azibazi_to(target)[0]
infos.append(meta.str_dist(dist))
infos.append(u'%.0f\u00B0' % azi)
axes2.annotate(
'\n'.join(infos),
xy=(0., 1.),
xycoords='axes fraction',
xytext=(2., 2.),
textcoords='offset points',
ha='left',
va='top',
fontsize=fontsize,
fontstyle='normal')
if (self.nx == 1 and self.ny == 1):
yield item, fig
del figures[iyy, ixx]
if not (self.nx == 1 and self.ny == 1):
for (iyy, ixx), (_, fig) in figures.items():
title = '.'.join(x for x in cg if x)
if len(figures) > 1:
title += ' (%i/%i, %i/%i)' % (iyy+1, nyy, ixx+1, nxx)
fig.suptitle(title, fontsize=fontsize_title)
for item, fig in figures.values():
yield item, fig
def draw_figures(self, problem, dataset, history):
target_index = {}
i = 0
for target in problem.targets:
target_index[target] = i, i+target.nmisfits
i += target.nmisfits
ws = problem.get_target_weights()
if history:
misfits = history.misfits[history.get_sorted_misfits_idx(), ...]
gcms = problem.combine_misfits(
misfits[:1, :, :], get_contributions=True)[0, :]
event = problem.base_source
cg_to_targets = meta.gather(
problem.waveform_targets,
lambda t: (t.path, t.codes[3]))
cgs = sorted(cg_to_targets.keys())
for cg in cgs:
cg_str = '.'.join(cg)
targets = cg_to_targets[cg]
if len(targets) == 0:
continue
assert all(target_index[target][0] == target_index[target][1] - 1
for target in targets)
itargets = num.array(
[target_index[target][0] for target in targets])
labels = ['.'.join(x for x in t.codes[:3] if x) for t in targets]
azimuths = num.array([event.azibazi_to(t)[0] for t in targets])
distances = num.array([t.distance_to(event) for t in targets])
item = PlotItem(
name='seismic_stations_weights_%s' % cg_str,
title=u'Station weights (%s)' % cg_str,
description=u'\n\nMarkers are scaled according to the '
u'weighting factor of the corresponding target\'s '
u'contribution in the misfit function.')
fig, ax, legend = self.plot_station_distribution(
azimuths, distances, ws[itargets], labels)
legend.set_title(
'Weight',
prop=dict(size=self.font_size))
yield (item, fig)
if history:
item = PlotItem(
name='seismic_stations_contributions_%s' % cg_str,
title=u'Station misfit contributions (%s)' % cg_str,
description=u'\n\nMarkers are scaled according to their '
u'misfit contribution for the globally best '
u'source model.')
fig, ax, legend = self.plot_station_distribution(
azimuths, distances, gcms[itargets], labels)
legend.set_title(
'Contribution',
prop=dict(size=self.font_size))
yield (item, fig)
def draw_figures(self, problem, dataset, history):
target_index = {}
i = 0
for target in problem.targets:
target_index[target] = i, i + target.nmisfits
i += target.nmisfits
misfits = history.misfits[history.get_sorted_misfits_idx(), ...]
gcms = problem.combine_misfits(misfits[:1, :, :],
get_contributions=True)[0, :]
models = history.get_sorted_primary_models()[::-1]
origin = problem.base_source
targets = [
t for t in problem.targets if isinstance(t, PhasePickTarget)
]
ntargets = len(targets)
nmodels = history.nmodels
tts = num.zeros((nmodels, ntargets, 2))
sdata = []
for imodel in range(nmodels):
model = models[imodel, :]
source = problem.get_source(model)
sdata.append(
(origin.distance_to(source), origin.azibazi_to(source)[0]))
results = problem.evaluate(model, targets=targets)
for itarget, result in enumerate(results):
result = results[itarget]
tts[imodel, itarget, :] = result.tobs, result.tsyn
ok = num.all(num.isfinite(tts), axis=2)
ok = num.all(ok, axis=0)
targets_ok = [
target for (itarget, target) in enumerate(targets) if ok[itarget]
]
tts = tts[:, ok, :]
residuals = tts[:, :, 0] - tts[:, :, 1]
mean_residuals = num.mean(residuals, axis=0)
rlo, rhi = num.percentile(mean_residuals, [10., 90.])
residual_amax = max(abs(rlo), abs(rhi))
target_to_residual = dict(
(target, residual)
for (target, residual) in zip(targets_ok, mean_residuals))
cg_to_targets = meta.gather(targets, lambda t: (t.path, ))
cgs = sorted(cg_to_targets.keys())
for cg in cgs:
cg_str = '.'.join(cg)
targets = cg_to_targets[cg]
if len(targets) == 0:
continue
assert all(target_index[target][0] == target_index[target][1] - 1
for target in targets)
itargets = num.array(
[target_index[target][0] for target in targets])
labels = ['.'.join(x for x in t.codes[:3] if x) for t in targets]
azimuths = num.array([origin.azibazi_to(t)[0] for t in targets])
distances = num.array([t.distance_to(origin) for t in targets])
residuals = num.array(
[target_to_residual.get(t, num.nan) for t in targets])
item = PlotItem(
name='picks_contributions_%s' % cg_str,
title=u'Pick residuals and contributions (%s)' % cg_str,
description=u'\n\nMarkers are scaled according to their '
u'misfit contribution for the globally best '
u'source model.')
fig, axes, legend = self.plot_station_distribution(
azimuths,
distances,
gcms[itargets],
labels,
colors=residuals,
cnorm=(-residual_amax, residual_amax),
cmap='RdYlBu',
clabel='$T_{obs} - T_{syn}$ [s]',
scatter_kwargs=dict(alpha=1.0),
legend_title='Contribution')
sources = [problem.get_source(x) for x in models[::10]]
azimuths = num.array([origin.azibazi_to(s)[0] for s in sources])
distances = num.array([s.distance_to(origin) for s in sources])
axes.plot(azimuths * d2r,
distances,
'o',
ms=2.0,
color='black',
alpha=0.3)
yield (item, fig)
def draw_figures(self, ds, history):
fontsize = self.font_size
fontsize_title = self.font_size_title
problem = history.problem
for target in problem.targets:
target.set_dataset(ds)
target_index = dict(
(target, i) for (i, target) in enumerate(problem.targets))
gms = problem.combine_misfits(history.misfits)
isort = num.argsort(gms)
gms = gms[isort]
models = history.models[isort, :]
misfits = history.misfits[isort, :]
xbest = models[0, :]
ws = problem.get_target_weights()
gcms = problem.combine_misfits(misfits[:1, :, :],
get_contributions=True)[0, :]
w_max = num.nanmax(ws)
gcm_max = num.nanmax(gcms)
source = problem.get_source(xbest)
target_to_result = {}
all_syn_trs = []
all_syn_specs = []
results = problem.evaluate(xbest)
dtraces = []
for target, result in zip(problem.targets, results):
if not isinstance(result, WaveformMisfitResult):
dtraces.append(None)
continue
itarget = target_index[target]
w = target.get_combined_weight()
if target.misfit_config.domain == 'cc_max_norm':
tref = (result.filtered_obs.tmin +
result.filtered_obs.tmax) * 0.5
for tr_filt, tr_proc, tshift in ((result.filtered_obs,
result.processed_obs, 0.),
(result.filtered_syn,
result.processed_syn,
result.tshift)):
norm = num.sum(num.abs(tr_proc.ydata)) / tr_proc.data_len()
tr_filt.ydata /= norm
tr_proc.ydata /= norm
tr_filt.shift(tshift)
tr_proc.shift(tshift)
ctr = result.cc
ctr.shift(tref)
dtrace = ctr
else:
for tr in (result.filtered_obs, result.filtered_syn,
result.processed_obs, result.processed_syn):
tr.ydata *= w
for spec in (result.spectrum_obs, result.spectrum_syn):
if spec is not None:
spec.ydata *= w
if result.tshift is not None and result.tshift != 0.0:
# result.filtered_syn.shift(result.tshift)
result.processed_syn.shift(result.tshift)
dtrace = make_norm_trace(result.processed_syn,
result.processed_obs,
problem.norm_exponent)
target_to_result[target] = result
dtrace.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
dtraces.append(dtrace)
result.processed_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_trs.append(result.processed_syn)
if result.spectrum_syn:
result.spectrum_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_specs.append(result.spectrum_syn)
if not all_syn_trs:
logger.warn('no traces to show')
return []
def skey(tr):
return tr.meta['normalisation_family'], tr.meta['path']
trace_minmaxs = trace.minmax(all_syn_trs, skey)
amp_spec_maxs = amp_spec_max(all_syn_specs, skey)
dminmaxs = trace.minmax([x for x in dtraces if x is not None], skey)
for tr in dtraces:
if tr:
dmin, dmax = dminmaxs[skey(tr)]
tr.ydata /= max(abs(dmin), abs(dmax))
cg_to_targets = meta.gather(problem.waveform_targets,
lambda t: (t.path, t.codes[3]),
filter=lambda t: t in target_to_result)
cgs = sorted(cg_to_targets.keys())
figs = []
for cg in cgs:
targets = cg_to_targets[cg]
nframes = len(targets)
nx = int(math.ceil(math.sqrt(nframes)))
ny = (nframes - 1) // nx + 1
nxmax = 4
nymax = 4
nxx = (nx - 1) // nxmax + 1
nyy = (ny - 1) // nymax + 1
# nz = nxx * nyy
xs = num.arange(nx) // ((max(2, nx) - 1.0) / 2.)
ys = num.arange(ny) // ((max(2, ny) - 1.0) / 2.)
xs -= num.mean(xs)
ys -= num.mean(ys)
fxs = num.tile(xs, ny)
fys = num.repeat(ys, nx)
data = []
for target in targets:
azi = source.azibazi_to(target)[0]
dist = source.distance_to(target)
x = dist * num.sin(num.deg2rad(azi))
y = dist * num.cos(num.deg2rad(azi))
data.append((x, y, dist))
gxs, gys, dists = num.array(data, dtype=num.float).T
iorder = num.argsort(dists)
gxs = gxs[iorder]
gys = gys[iorder]
targets_sorted = [targets[ii] for ii in iorder]
gxs -= num.mean(gxs)
gys -= num.mean(gys)
gmax = max(num.max(num.abs(gys)), num.max(num.abs(gxs)))
if gmax == 0.:
gmax = 1.
gxs /= gmax
gys /= gmax
dists = num.sqrt((fxs[num.newaxis, :] - gxs[:, num.newaxis])**2 +
(fys[num.newaxis, :] - gys[:, num.newaxis])**2)
distmax = num.max(dists)
availmask = num.ones(dists.shape[1], dtype=num.bool)
frame_to_target = {}
for itarget, target in enumerate(targets_sorted):
iframe = num.argmin(
num.where(availmask, dists[itarget], distmax + 1.))
availmask[iframe] = False
iy, ix = num.unravel_index(iframe, (ny, nx))
frame_to_target[iy, ix] = target
figures = {}
for iy in range(ny):
for ix in range(nx):
if (iy, ix) not in frame_to_target:
continue
ixx = ix // nxmax
iyy = iy // nymax
if (iyy, ixx) not in figures:
title = '_'.join(x for x in cg if x)
item = PlotItem(name='fig_%s_%i_%i' %
(title, ixx, iyy))
item.attributes['targets'] = []
figures[iyy,
ixx] = (item,
plt.figure(figsize=self.size_inch))
figures[iyy, ixx][1].subplots_adjust(left=0.03,
right=1.0 - 0.03,
bottom=0.03,
top=1.0 - 0.06,
wspace=0.2,
hspace=0.2)
figs.append(figures[iyy, ixx])
item, fig = figures[iyy, ixx]
target = frame_to_target[iy, ix]
item.attributes['targets'].append(target.string_id())
amin, amax = trace_minmaxs[target.normalisation_family,
target.path]
absmax = max(abs(amin), abs(amax))
ny_this = nymax # min(ny, nymax)
nx_this = nxmax # min(nx, nxmax)
i_this = (iy % ny_this) * nx_this + (ix % nx_this) + 1
axes2 = fig.add_subplot(ny_this, nx_this, i_this)
space = 0.5
space_factor = 1.0 + space
axes2.set_axis_off()
axes2.set_ylim(-1.05 * space_factor, 1.05)
axes = axes2.twinx()
axes.set_axis_off()
if target.misfit_config.domain == 'cc_max_norm':
axes.set_ylim(-10. * space_factor, 10.)
else:
axes.set_ylim(-absmax * 1.33 * space_factor,
absmax * 1.33)
itarget = target_index[target]
result = target_to_result[target]
dtrace = dtraces[itarget]
tap_color_annot = (0.35, 0.35, 0.25)
tap_color_edge = (0.85, 0.85, 0.80)
tap_color_fill = (0.95, 0.95, 0.90)
plot_taper(axes2,
result.processed_obs.get_xdata(),
result.taper,
fc=tap_color_fill,
ec=tap_color_edge)
obs_color = mpl_color('aluminium5')
obs_color_light = light(obs_color, 0.5)
syn_color = mpl_color('scarletred2')
syn_color_light = light(syn_color, 0.5)
misfit_color = mpl_color('scarletred2')
weight_color = mpl_color('chocolate2')
cc_color = mpl_color('aluminium5')
if target.misfit_config.domain == 'cc_max_norm':
tref = (result.filtered_obs.tmin +
result.filtered_obs.tmax) * 0.5
plot_dtrace(axes2,
dtrace,
space,
-1.,
1.,
fc=light(cc_color, 0.5),
ec=cc_color)
plot_dtrace_vline(axes2,
tref,
space,
color=tap_color_annot)
elif target.misfit_config.domain == 'frequency_domain':
asmax = amp_spec_maxs[target.normalisation_family,
target.path]
fmin, fmax = \
target.misfit_config.get_full_frequency_range()
plot_spectrum(axes2,
result.spectrum_syn,
result.spectrum_obs,
fmin,
fmax,
space,
0.,
asmax,
syn_color=syn_color,
obs_color=obs_color,
syn_lw=1.0,
obs_lw=0.75,
color_vline=tap_color_annot,
fontsize=fontsize)
else:
plot_dtrace(axes2,
dtrace,
space,
0.,
1.,
fc=light(misfit_color, 0.3),
ec=misfit_color)
plot_trace(axes,
result.filtered_syn,
color=syn_color_light,
lw=1.0)
plot_trace(axes,
result.filtered_obs,
color=obs_color_light,
lw=0.75)
plot_trace(axes,
result.processed_syn,
color=syn_color,
lw=1.0)
plot_trace(axes,
result.processed_obs,
color=obs_color,
lw=0.75)
xdata = result.filtered_obs.get_xdata()
axes.set_xlim(xdata[0], xdata[-1])
tmarks = [
result.processed_obs.tmin, result.processed_obs.tmax
]
for tmark in tmarks:
axes2.plot([tmark, tmark], [-0.9, 0.1],
color=tap_color_annot)
for tmark, text, ha in [
(tmarks[0],
'$\\,$ ' + meta.str_duration(tmarks[0] - source.time),
'right'),
(tmarks[1], '$\\Delta$ ' +
meta.str_duration(tmarks[1] - tmarks[0]), 'left')
]:
axes2.annotate(
text,
xy=(tmark, -0.9),
xycoords='data',
xytext=(fontsize * 0.4 * [-1, 1][ha == 'left'],
fontsize * 0.2),
textcoords='offset points',
ha=ha,
va='bottom',
color=tap_color_annot,
fontsize=fontsize)
rel_w = ws[itarget] / w_max
rel_c = gcms[itarget] / gcm_max
sw = 0.25
sh = 0.1
ph = 0.01
for (ih, rw, facecolor, edgecolor) in [
(0, rel_w, light(weight_color, 0.5), weight_color),
(1, rel_c, light(misfit_color, 0.5), misfit_color)
]:
bar = patches.Rectangle(
(1.0 - rw * sw, 1.0 - (ih + 1) * sh + ph),
rw * sw,
sh - 2 * ph,
facecolor=facecolor,
edgecolor=edgecolor,
zorder=10,
transform=axes.transAxes,
clip_on=False)
axes.add_patch(bar)
scale_string = None
if target.misfit_config.domain == 'cc_max_norm':
scale_string = 'Syn/obs scales differ!'
infos = []
if scale_string:
infos.append(scale_string)
infos.append('.'.join(x for x in target.codes if x))
dist = source.distance_to(target)
azi = source.azibazi_to(target)[0]
infos.append(meta.str_dist(dist))
infos.append('%.0f\u00B0' % azi)
infos.append('%.3g' % ws[itarget])
infos.append('%.3g' % gcms[itarget])
axes2.annotate('\n'.join(infos),
xy=(0., 1.),
xycoords='axes fraction',
xytext=(2., 2.),
textcoords='offset points',
ha='left',
va='top',
fontsize=fontsize,
fontstyle='normal')
for (iyy, ixx), (_, fig) in figures.items():
title = '.'.join(x for x in cg if x)
if len(figures) > 1:
title += ' (%i/%i, %i/%i)' % (iyy + 1, nyy, ixx + 1, nxx)
fig.suptitle(title, fontsize=fontsize_title)
return figs
def draw_figures(self, ds, history):
color_parameter = self.color_parameter
misfit_cutoff = self.misfit_cutoff
fontsize = self.font_size
fontsize_title = self.font_size_title
problem = history.problem
for target in problem.targets:
target.set_dataset(ds)
target_index = dict(
(target, i) for (i, target) in enumerate(problem.targets))
gms = problem.combine_misfits(history.misfits)
isort = num.argsort(gms)[::-1]
gms = gms[isort]
models = history.models[isort, :]
if misfit_cutoff is not None:
ibest = gms < misfit_cutoff
gms = gms[ibest]
models = models[ibest]
gms = gms[::10]
models = models[::10]
nmodels = models.shape[0]
if color_parameter == 'dist':
mx = num.mean(models, axis=0)
cov = num.cov(models.T)
mdists = core.mahalanobis_distance(models, mx, cov)
icolor = meta.ordersort(mdists)
elif color_parameter == 'misfit':
iorder = num.arange(nmodels)
icolor = iorder
elif color_parameter in problem.parameter_names:
ind = problem.name_to_index(color_parameter)
icolor = problem.extract(models, ind)
target_to_results = defaultdict(list)
all_syn_trs = []
dtraces = []
for imodel in range(nmodels):
model = models[imodel, :]
source = problem.get_source(model)
results = problem.evaluate(model)
dtraces.append([])
for target, result in zip(problem.targets, results):
if isinstance(result, gf.SeismosizerError):
dtraces[-1].append(None)
continue
if not isinstance(target, WaveformMisfitTarget):
dtraces[-1].append(None)
continue
itarget = target_index[target]
w = target.get_combined_weight()
if target.misfit_config.domain == 'cc_max_norm':
tref = (
result.filtered_obs.tmin + result.filtered_obs.tmax) \
* 0.5
for tr_filt, tr_proc, tshift in ((result.filtered_obs,
result.processed_obs,
0.),
(result.filtered_syn,
result.processed_syn,
result.tshift)):
norm = num.sum(num.abs(tr_proc.ydata)) \
/ tr_proc.data_len()
tr_filt.ydata /= norm
tr_proc.ydata /= norm
tr_filt.shift(tshift)
tr_proc.shift(tshift)
ctr = result.cc
ctr.shift(tref)
dtrace = ctr
else:
for tr in (result.filtered_obs, result.filtered_syn,
result.processed_obs, result.processed_syn):
tr.ydata *= w
if result.tshift is not None and result.tshift != 0.0:
# result.filtered_syn.shift(result.tshift)
result.processed_syn.shift(result.tshift)
dtrace = make_norm_trace(result.processed_syn,
result.processed_obs,
problem.norm_exponent)
target_to_results[target].append(result)
dtrace.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
dtraces[-1].append(dtrace)
result.processed_syn.meta = dict(
normalisation_family=target.normalisation_family,
path=target.path)
all_syn_trs.append(result.processed_syn)
if not all_syn_trs:
logger.warn('no traces to show')
return []
def skey(tr):
return tr.meta['normalisation_family'], tr.meta['path']
trace_minmaxs = trace.minmax(all_syn_trs, skey)
dtraces_all = []
for dtraces_group in dtraces:
dtraces_all.extend(dtraces_group)
dminmaxs = trace.minmax(
[dtrace_ for dtrace_ in dtraces_all if dtrace_ is not None], skey)
for tr in dtraces_all:
if tr:
dmin, dmax = dminmaxs[skey(tr)]
tr.ydata /= max(abs(dmin), abs(dmax))
cg_to_targets = meta.gather(problem.waveform_targets,
lambda t: (t.path, t.codes[3]),
filter=lambda t: t in target_to_results)
cgs = sorted(cg_to_targets.keys())
from matplotlib import colors
cmap = cm.ScalarMappable(norm=colors.Normalize(vmin=num.min(icolor),
vmax=num.max(icolor)),
cmap=plt.get_cmap('coolwarm'))
imodel_to_color = []
for imodel in range(nmodels):
imodel_to_color.append(cmap.to_rgba(icolor[imodel]))
figs = []
for cg in cgs:
targets = cg_to_targets[cg]
nframes = len(targets)
nx = int(math.ceil(math.sqrt(nframes)))
ny = (nframes - 1) // nx + 1
nxmax = 4
nymax = 4
nxx = (nx - 1) // nxmax + 1
nyy = (ny - 1) // nymax + 1
# nz = nxx * nyy
xs = num.arange(nx) / ((max(2, nx) - 1.0) / 2.)
ys = num.arange(ny) / ((max(2, ny) - 1.0) / 2.)
xs -= num.mean(xs)
ys -= num.mean(ys)
fxs = num.tile(xs, ny)
fys = num.repeat(ys, nx)
data = []
for target in targets:
azi = source.azibazi_to(target)[0]
dist = source.distance_to(target)
x = dist * num.sin(num.deg2rad(azi))
y = dist * num.cos(num.deg2rad(azi))
data.append((x, y, dist))
gxs, gys, dists = num.array(data, dtype=num.float).T
iorder = num.argsort(dists)
gxs = gxs[iorder]
gys = gys[iorder]
targets_sorted = [targets[ii] for ii in iorder]
gxs -= num.mean(gxs)
gys -= num.mean(gys)
gmax = max(num.max(num.abs(gys)), num.max(num.abs(gxs)))
if gmax == 0.:
gmax = 1.
gxs /= gmax
gys /= gmax
dists = num.sqrt((fxs[num.newaxis, :] - gxs[:, num.newaxis])**2 +
(fys[num.newaxis, :] - gys[:, num.newaxis])**2)
distmax = num.max(dists)
availmask = num.ones(dists.shape[1], dtype=num.bool)
frame_to_target = {}
for itarget, target in enumerate(targets_sorted):
iframe = num.argmin(
num.where(availmask, dists[itarget], distmax + 1.))
availmask[iframe] = False
iy, ix = num.unravel_index(iframe, (ny, nx))
frame_to_target[iy, ix] = target
figures = {}
for iy in range(ny):
for ix in range(nx):
if (iy, ix) not in frame_to_target:
continue
ixx = ix // nxmax
iyy = iy // nymax
if (iyy, ixx) not in figures:
title = '_'.join(x for x in cg if x)
item = PlotItem(name='fig_%s_%i_%i' %
(title, ixx, iyy))
item.attributes['targets'] = []
figures[iyy,
ixx] = (item,
plt.figure(figsize=self.size_inch))
figures[iyy, ixx][1].subplots_adjust(left=0.03,
right=1.0 - 0.03,
bottom=0.03,
top=1.0 - 0.06,
wspace=0.2,
hspace=0.2)
figs.append(figures[iyy, ixx])
item, fig = figures[iyy, ixx]
target = frame_to_target[iy, ix]
item.attributes['targets'].append(target.string_id())
amin, amax = trace_minmaxs[target.normalisation_family,
target.path]
absmax = max(abs(amin), abs(amax))
ny_this = nymax # min(ny, nymax)
nx_this = nxmax # min(nx, nxmax)
i_this = (iy % ny_this) * nx_this + (ix % nx_this) + 1
axes2 = fig.add_subplot(ny_this, nx_this, i_this)
space = 0.5
space_factor = 1.0 + space
axes2.set_axis_off()
axes2.set_ylim(-1.05 * space_factor, 1.05)
axes = axes2.twinx()
axes.set_axis_off()
if target.misfit_config.domain == 'cc_max_norm':
axes.set_ylim(-10. * space_factor, 10.)
else:
axes.set_ylim(-absmax * 1.33 * space_factor,
absmax * 1.33)
itarget = target_index[target]
for imodel, result in enumerate(target_to_results[target]):
syn_color = imodel_to_color[imodel]
dtrace = dtraces[imodel][itarget]
tap_color_annot = (0.35, 0.35, 0.25)
tap_color_edge = (0.85, 0.85, 0.80)
tap_color_fill = (0.95, 0.95, 0.90)
plot_taper(axes2,
result.processed_obs.get_xdata(),
result.taper,
fc=tap_color_fill,
ec=tap_color_edge,
alpha=0.2)
obs_color = mpl_color('aluminium5')
obs_color_light = light(obs_color, 0.5)
plot_dtrace(axes2,
dtrace,
space,
0.,
1.,
fc='none',
ec=syn_color)
# plot_trace(
# axes, result.filtered_syn,
# color=syn_color_light, lw=1.0)
if imodel == 0:
plot_trace(axes,
result.filtered_obs,
color=obs_color_light,
lw=0.75)
plot_trace(axes,
result.processed_syn,
color=syn_color,
lw=1.0,
alpha=0.3)
plot_trace(axes,
result.processed_obs,
color=obs_color,
lw=0.75,
alpha=0.3)
if imodel != 0:
continue
xdata = result.filtered_obs.get_xdata()
axes.set_xlim(xdata[0], xdata[-1])
tmarks = [
result.processed_obs.tmin,
result.processed_obs.tmax
]
for tmark in tmarks:
axes2.plot([tmark, tmark], [-0.9, 0.1],
color=tap_color_annot)
for tmark, text, ha in [
(tmarks[0], '$\\,$ ' +
meta.str_duration(tmarks[0] - source.time),
'right'),
(tmarks[1], '$\\Delta$ ' +
meta.str_duration(tmarks[1] - tmarks[0]), 'left')
]:
axes2.annotate(
text,
xy=(tmark, -0.9),
xycoords='data',
xytext=(fontsize * 0.4 * [-1, 1][ha == 'left'],
fontsize * 0.2),
textcoords='offset points',
ha=ha,
va='bottom',
color=tap_color_annot,
fontsize=fontsize)
scale_string = None
if target.misfit_config.domain == 'cc_max_norm':
scale_string = 'Syn/obs scales differ!'
infos = []
if scale_string:
infos.append(scale_string)
infos.append('.'.join(x for x in target.codes if x))
dist = source.distance_to(target)
azi = source.azibazi_to(target)[0]
infos.append(meta.str_dist(dist))
infos.append(u'%.0f\u00B0' % azi)
axes2.annotate('\n'.join(infos),
xy=(0., 1.),
xycoords='axes fraction',
xytext=(2., 2.),
textcoords='offset points',
ha='left',
va='top',
fontsize=fontsize,
fontstyle='normal')
for (iyy, ixx), (_, fig) in figures.items():
title = '.'.join(x for x in cg if x)
if len(figures) > 1:
title += ' (%i/%i, %i/%i)' % (iyy + 1, nyy, ixx + 1, nxx)
fig.suptitle(title, fontsize=fontsize_title)
return figs
