def get_timing_vs_ndata_at_const_nfreq(nvals,
nharmonics,
max_freq,
filename=None,
overwrite=True,
time_slow=True):
#if template is None:
template = get_default_template(nharmonics=nharmonics)
template.precompute()
# load saved results
results = {}
if filename is None:
filename = './saved_results/timing_results_nh%d_maxfrq%.1f.pkl' % (
nharmonics, max_freq)
if not filename is None and os.path.exists(filename):
old_results = pickle.load(open(filename, 'rb'))
results.update(old_results)
for n in nvals:
if n in results:
continue
x, y, dy = generate_random_signal(n)
x[0] = 0
x[-1] = 1
# time FTP
print("timing: n = %d, h = %d, ftp" % (n, nharmonics))
model = FastTemplatePeriodogram(template=template)
model.fit(x, y, dy)
t0 = time()
frq, p = model.autopower(maximum_frequency=max_freq)
tftp = time() - t0
print(" done in %.4f seconds" % (tftp))
if time_slow:
print("timing: n = %d, h = %d, slow" % (n, nharmonics))
model = SlowTemplatePeriodogram(template=template)
model.fit(x, y, dy)
t0 = time()
p = model.power(frq)
tslow = time() - t0
print(" done in %.4f seconds" % (tslow))
else:
tslow = -1
results[n] = (tftp, tslow)
# save
if not filename is None and overwrite:
pickle.dump(results, open(filename, 'wb'))
return zip(*[results[n] for n in nvals])
def get_modeler(ndata, nh, precompute=True):
t, y, yerr = get_data(n=ndata)
template = get_template(nh=nh)
if precompute:
template.precompute()
modeler = FastTemplatePeriodogram(template=template)
modeler.fit(t, y, yerr)
return modeler
def get_timing_vs_nharmonics(x,
y,
yerr,
hvals,
filename=None,
overwrite=True,
only_use_saved_data=False):
# load old results
results = {}
if not filename is None and os.path.exists(filename):
old_results = pickle.load(open(filename, 'rb'))
results.update(old_results)
# return if nothing to do
if all([h in results for h in hvals]):
return hvals, [results[h] for h in hvals]
if only_use_saved_data:
return zip(*[(h, results[h]) for h in hvals if h in results])
for h in hvals:
if h in results:
continue
print(" H = ", h)
template = get_default_template(nharmonics=h)
#template.precompute()
model = FastTemplatePeriodogram(template=template)
model.fit(x, y, yerr)
t0 = time()
model.autopower()
results[h] = time() - t0
print(" %.4f seconds" % (results[h]))
if not filename is None and overwrite:
pickle.dump(results, open(filename, 'wb'))
return hvals, [results[h] for h in hvals]
def plot_templates_and_periodograms(x,
y,
err,
y_temp,
freq_val=None,
hfac=None,
ofac=None,
nharms=None,
settings=default_settings):
p_ftps = []
phi_data = None if freq_val is None else (x * freq_val -
settings['phi0']) % 1.0
for i, nharm in enumerate(nharms):
#model.templates.values()[0].nharmonics = nharm
#model.templates.values()[0].precompute()
template = Template.from_sampled(y_temp, nharmonics=nharm)
template.precompute()
model = FastTemplatePeriodogram(template=template)
# Run FTP
model.fit(x, y, err)
frq, p = model.autopower(samples_per_peak=ofac, nyquist_factor=hfac)
p_ftps.append(p)
# get axes boundaries
bounds_ax_tmp, bounds_ax_pdg = \
get_boundaries_for_axtmp_and_axpdg(max(frq), settings)
# initialize figure
f = plt.figure(figsize=(2 * settings['axsize'], settings['axsize']))
ax_pdg = f.add_axes(bounds_ax_pdg)
ax_tmp = f.add_axes(bounds_ax_tmp)
settings = translate_color(ax_pdg, settings)
# get full template
phi0 = np.linspace(0, 1, len(y_temp))
y0 = y_temp
ymin, ymax = min(-y0), max(-y0)
# x position for text (H = ...); has to be in ax_tmp data coordinates
x0 = settings['annotate_x0'] / bounds_ax_tmp[2]
# functions for normalizing templates
tmpnorm = settings['tmp_height_frac'] / (ymax - ymin)
yoffset = 0.5 * (1 - settings['tmp_height_frac'])
tmp_transform = lambda yt, et: (
(-yt - ymin) * tmpnorm + yoffset, et * tmpnorm
if not et is None else None)
# normalize data and template
ydata, edata = tmp_transform(y, err)
y0, _ = tmp_transform(y0, None)
#colorfunc = lambda i : "%.5f"%(settings['colorfunc_a'] * (float(len(nharms) - i - 1) / float(len(nharms) - 1)) \
# + settings['colorfunc_b']) if i < len(nharms) - 1 \
# else settings['answer_color']
colorfunc = lambda i: settings['answer_color']
for i, (p, h) in enumerate(zip(p_ftps, nharms)):
offset = len(p_ftps) - i - 1
ytext = offset + 0.5
lw = 1 if i < len(p_ftps) - 1 else 1
# plot periodogram
ax_pdg.plot(frq, p + offset, color=colorfunc(i), lw=lw, zorder=20)
ax_pdg.plot(frq,
p + offset,
color=ax_pdg.get_facecolor(),
lw=3,
zorder=19)
if (i == len(p_ftps) - 1) and settings['plot_bls']:
frq_bls, p_bls = get_bls_periodogram(x, y, err, hfac=3, ofac=10)
ax_pdg.plot(frq_bls,
p_bls + offset,
color=settings['color_bls'],
alpha=settings['alpha_bls'],
zorder=18)
ax_pdg.plot(frq_bls,
p_bls + offset,
color=ax_pdg.get_facecolor(),
zorder=17,
lw=3)
ax_pdg.text(0.02,
0.92,
"Box Least Squares",
color=settings['color_bls'],
ha='left',
va='top',
bbox=settings['bbox'],
fontsize=settings['annotation_fontsize'],
transform=ax_pdg.transAxes)
if (i > 0 and settings['plot_multiharmonic_periodogram']):
# plot multiharmonic periodogram
frq_mh, p_mh = get_multiharmonic_periodogram(x, y, err, h)
color = settings['color_multiharmonic_periodogram']
alpha = settings['alpha_multiharmonic_periodogram']
ax_pdg.plot(frq_mh,
p_mh + offset,
color=color,
alpha=alpha,
zorder=16)
ax_pdg.plot(frq_mh,
p_mh + offset,
color=ax_pdg.get_facecolor(),
zorder=15,
lw=3)
if i == 1:
df0 = (2 * freq_val - min(frq_mh)) / (max(frq_mh) -
min(frq_mh))
ind = int(df0 * len(frq_mh))
dx = max(ax_pdg.get_xlim()) - min(ax_pdg.get_xlim())
#ax_pdg.annotate('Multiharmonic Lomb Scargle',
# xy = (frq_mh[ind], p_mh[ind] + offset + 0.03),
# xycoords = 'data',
# xytext = (frq_mh[ind] - 0.015 * dx, 1.45 + offset),
# textcoords = 'data',
# horizontalalignment = 'right',
# verticalalignment = 'bottom',
# color = settings['font_color'],
# arrowprops = dict(ec=ax_pdg.get_facecolor(), fc=settings['font_color'],
# lw=1.5, arrowstyle='simple'),
# fontsize = settings['annotation_fontsize'],
# bbox = settings['bbox'])
ax_pdg.text(0.02,
0.98,
"Multi-harmonic Lomb Scargle",
fontsize=settings['annotation_fontsize'],
bbox=settings['bbox'],
color=color,
ha='left',
va='top',
transform=ax_pdg.transAxes)
# get truncated template
#template.nharmonics = h
template = Template.from_sampled(y_temp, nharmonics=h)
template.precompute()
phi = np.linspace(0, 1, 100)
ytmp_trunc = template(phi)
# normalize truncated template
ytmp_trunc, _ = tmp_transform(ytmp_trunc, None)
# plot truncated template
ax_tmp.plot(phi, ytmp_trunc + offset, color=colorfunc(i), lw=lw)
# plot data
ax_tmp.errorbar(phi_data,
ydata + offset,
yerr=edata,
**settings['data_params'])
# write H = ...
ax_tmp.text(x0,
ytext,
"H = %d" % (h),
va='center',
ha='left',
color=settings['font_color'],
fontsize=settings['label_fontsize'])
# Write axis labels
ax_tmp.set_xlabel('Phase')
ax_pdg.set_xlabel('Frequency $[d^{-1}]$')
# Draw line for correct frequency
if freq_val is not None:
ax_pdg.axvline(freq_val, ls=':', color='k')
# Write titles
ytitle = settings['top'] + settings['title_pad']
xtitle_pdg = settings['left'] + bounds_ax_tmp[2] + settings['wspace'] \
+ 0.5 * bounds_ax_pdg[2]
xtitle_tmp = settings['left'] + 0.5 * bounds_ax_tmp[2]
f.text(xtitle_pdg,
ytitle,
"Template periodogram",
va='bottom',
ha='center',
color=settings['font_color'],
fontsize=settings['title_fontsize'])
f.text(xtitle_tmp,
ytitle,
"Template fits",
va='bottom',
ha='center',
color=settings['font_color'],
fontsize=settings['title_fontsize'])
# Set other properties
ax_pdg.set_xlim(0, int(max(frq) / settings['locator_axpdg'])\
* settings['locator_axpdg'])
ax_pdg.set_ylim(0, len(nharms))
ax_pdg.xaxis.set_major_locator(MultipleLocator(settings['locator_axpdg']))
ax_tmp.set_xlim(0, 1)
ax_tmp.set_ylim(*ax_pdg.get_ylim())
ax_tmp.set_yticks(ax_pdg.get_yticks())
ax_tmp.xaxis.set_major_locator(MultipleLocator(settings['locator_axtmp']))
# for both axes...
for ax in [ax_pdg, ax_tmp]:
# turn of ytick labels
[label.set_visible(False) for label in ax.get_yticklabels()]
ax.yaxis.set_major_locator(MultipleLocator(1))
clean_up_axis(ax, settings)
clean_up_figure(f, settings)
def plot_accuracy(x,
y,
yerr,
y_temp,
nharmonics,
compare_with=10,
settings=default_settings):
template = Template.from_sampled(y_temp, nharmonics=10)
# if comparing to large nharmonics, set template now
if isinstance(compare_with, float) or isinstance(compare_with, int):
template = Template.from_sampled(y_temp, nharmonics=int(compare_with))
#template.nharmonics = int(compare_with)
template.precompute()
# Set the reference model
true_model = SlowTemplatePeriodogram(template=template) \
if compare_with == 'slow_version' \
else FastTemplatePeriodogram(template=template)
# fit data
true_model.fit(x, y, yerr)
results, frq, p_ans = {}, None, None
label_formatter = lambda kind, h=None : \
"$P_{\\rm %s}(\\omega%s)$"\
%(kind, "" if h is None else "|H=%d"%(h))
corrlabel = lambda R: "$R = %.3f$" % (R)
# store results from the reference model
# (if the reference model is FastTemplatePeriodogram)
if isinstance(true_model, FastTemplatePeriodogram):
frq, p_ans = true_model.autopower()
results = {'ans': (frq, p_ans)}
# add results from all desired harmonics
for h in nharmonics:
# set template harmonics
#template.nharmonics = h
template = Template.from_sampled(y_temp, nharmonics=h)
template.precompute()
# create & fit modeler
model = FastTemplatePeriodogram(template=template)
model.fit(x, y, yerr)
# compute periodogram
results[h] = model.autopower()
# compute results for reference model
# (if the reference model is gatspy)
if not 'ans' in results and isinstance(true_model,
SlowTemplatePeriodogram):
frq = results[h][0]
p_ans = true_model.power(frq)
results['ans'] = (frq, p_ans)
# Set up plot geometry
nplots = len(nharmonics)
if not settings['nrows'] is None:
nrows = settings['nrows']
else:
nrows = max([int(sqrt(nplots)), 1])
ncols = 1
while ncols * nrows < nplots:
ncols += 1
figsize = (settings['axsize'] * ncols, settings['axsize'] * nrows)
# create figure
f, axes = plt.subplots(nrows, ncols, figsize=figsize)
# ensure we have a list of axes
if not hasattr(axes, '__iter__'):
axes = [axes]
# some plotting definitions
ans_label = label_formatter('slow') if compare_with == 'slow_version'\
else label_formatter('FTP', int(compare_with))
ax_label_params = dict(fontsize=settings['label_fontsize'],
color=settings['font_color'])
ax_annotation_params = dict(fontsize=settings['annotation_fontsize'],
color=settings['font_color'],
bbox=settings['bbox'])
scatter_params = settings['scatter_params']
#print scatter_params
for i, h in enumerate(nharmonics):
# select the axes instance
r, c = i / ncols, i % ncols
ax = axes[c] if ncols >= 1 and nrows == 1 else axes[r][c]
settings = translate_color(ax, settings)
p = results[h][1]
# scatterplot
ax.plot([0, 1], [0, 1], ls=':', color=settings['font_color'], zorder=2)
ax.scatter(p, p_ans, **scatter_params)
# write the pearson R correlation
ax.text(0.05,
0.9,
corrlabel(pearsonr(p_ans, p)[0]),
transform=ax.transAxes,
ha='left',
va='bottom',
zorder=10,
**ax_annotation_params)
# many of the gatspy periodogram values are 0;
# write pearson R using only non-zero P_gatspy values
#if compare_with == 'slow_version':
# nonzero_p_ans, nonzero_p = zip(*filter(lambda (Pans, P) : Pans > 0, zip(p_ans, p)))
# Rnonzero = pearsonr(nonzero_p_ans, nonzero_p)[0]
# ax.text(0.05, 0.9 - 0.03, "%s; $P_{\\rm non-lin. opt.} > 0$"\
# %(corrlabel(Rnonzero)),zorder=10,
# transform=ax.transAxes, ha='left', va='top', **ax_annotation_params)
# set plot properties
ax.set_xlabel(label_formatter('FTP', h), **ax_label_params)
if c == 0:
ax.set_ylabel(ans_label, **ax_label_params)
else:
[label.set_visible(False) for label in ax.get_yticklabels()]
ax.set_xlim(settings['x_and_y_min'], 1)
ax.set_ylim(settings['x_and_y_min'], 1)
name_list = ["0", "0.25", "0.5", "0.75", "1"]
pos_list = [0, 0.25, 0.5, 0.75, 1]
ax.xaxis.set_major_locator(FixedLocator((pos_list)))
ax.xaxis.set_major_formatter(FixedFormatter((name_list)))
ax.yaxis.set_major_locator(FixedLocator((pos_list)))
ax.yaxis.set_major_formatter(FixedFormatter((name_list)))
clean_up_axis(ax, settings)
adjust_figure(f, settings)
clean_up_figure(f, settings)
def plot_timing_vs_ndata_const_freq(settings=default_settings):
f, ax = plt.subplots(figsize=(settings['axsize'], settings['axsize']))
settings = translate_color(ax, settings)
#fname = os.path.join(settings['results_dir'], settings['timing_filename'])
nharms = settings['nharmonics']
if not hasattr(nharms, '__iter__'):
nharms = [nharms]
x, y, dy = generate_random_signal(10)
x[0] = 0
x[-1] = 1
xoffset = 0.08 * (settings['xlim'][1] - settings['xlim'][0])
nfrq = len(FastTemplatePeriodogram().fit(
x, y, dy).autofrequency(maximum_frequency=settings['max_freq']))
for i, h in enumerate(nharms):
time_slow = (h == 3)
tftp, tslow = get_timing_vs_ndata_at_const_nfreq(settings['ndata'],
h,
settings['max_freq'],
time_slow=time_slow)
label = None if h < max(nharms) else 'Fast template periodogram'
color = settings['scatter_params_ftp']['c']
lw = settings['linewidth']
spars = {}
spars.update(settings['scatter_params_ftp'])
fudge = 0.7
ls = '-' if h == max(nharms) else ':'
#ls = '-'
q = 1.
if len(nharms) > 1:
q = float(h - min(nharms)) / float(max(nharms) - min(nharms))
spars['alpha'] = fudge * q + (1 - fudge)
ax.scatter(settings['ndata'], tftp, label=label, **spars)
ax.plot(settings['ndata'],
tftp,
color=color,
lw=lw,
alpha=spars['alpha'],
ls=ls)
# now label this
ax.text(settings['ndata'][-1] + xoffset,
tftp[-1],
"$H = %d$" % (h),
ha='left',
va='center',
color=settings['font_color'],
fontsize=settings['annotation_fontsize'],
bbox=settings['bbox'])
if time_slow:
ax.scatter(settings['ndata'],
tslow,
label='Non-linear optimization',
**settings['scatter_params_nonlin'])
ax.plot(settings['ndata'],
tslow,
color=settings['scatter_params_nonlin']['c'],
lw=settings['linewidth'])
ax.text(0.05,
0.7,
"$N_f=%d$" % (nfrq),
ha='left',
va='top',
transform=ax.transAxes)
ax.set_xlabel('Number of datapoints')
ax.set_ylabel("Execution time [s]")
ax.set_title('Constant baseline')
ax.set_xscale('log')
ax.set_yscale('log')
ax.set_xlim(*settings['xlim'])
ax.set_ylim(*settings['ylim'])
ax.legend(loc=settings['legend_loc'])
adjust_figure(f, settings)
clean_up_axis(ax, settings)
clean_up_figure(f, settings)
def get_timing_vs_ndata(nvals,
nharmonics,
filename=None,
overwrite=True,
time_gatspy=True,
only_use_saved_data=False,
time_lomb_scargle=False):
#if template is None:
template = get_default_template(nharmonics=nharmonics)
#template.precompute()
# load saved results
results = {}
if not filename is None and os.path.exists(filename):
old_results = pickle.load(open(filename, 'rb'))
results.update(old_results)
else:
results = {name: [] for name in ['nfreqs', 'ndata', 'tftp', 'tgats']}
if only_use_saved_data:
return select_from_dict(results, nvals)
# return if nothing to do
if all([n in results for n in nvals]):
return [results[n] for n in nvals]
for n in nvals:
if n in results['ndata']:
continue
results['ndata'].append(n)
x, y, dy = generate_random_signal(n)
# time FTP
print("timing: n = %d, h = %d, ftp" % (n, nharmonics))
model = FastTemplatePeriodogram(template=template)
model.fit(x, y, dy)
t0 = time()
frq, p = model.autopower()
results['tftp'].append(time() - t0)
print(" done in %.4f seconds" % (results['tftp'][-1]))
results['nfreqs'].append(len(frq))
if time_gatspy:
# time GATSPY
print("timing: n = %d, gatspy" % (n))
model = SlowTemplatePeriodogram(template=template)
model.fit(x, y, dy)
t0 = time()
p = model.power(frq)
results['tgats'].append(time() - t0)
print(" done in %.4f seconds" % (results['tgats'][-1]))
else:
results['tgats'].append(-1)
# SORT results
results = sort_timing_dict(results)
# save
if not filename is None and overwrite:
pickle.dump(results, open(filename, 'wb'))
return select_from_dict(results, nvals)
else:
Ttemp, Ytemp = pickle.load(
open_results('template_filename', tconf, 'rb'))
template = Template.from_sampled(Ytemp,
nharmonics=tconf['nharm_answer'])
x, y, err = generate_random_signal(ndata,
sigma,
freq=freq,
template=template)
template.precompute()
# build model
model = FastTemplatePeriodogram(template=template)
y_temp = template(np.linspace(0, 1, 100))
print("plotting timing vs ndata at constant freq")
plot_timing_vs_ndata_const_freq(
settings=conf['timing_vs_ndata_const_freq'])
print("plotting timing vs nharmonics")
plot_timing_vs_nharmonics(conf['timing_vs_nharmonics'])
print("plotting timing vs ndata")
plot_timing_vs_ndata(conf['timing_vs_ndata'])
#print("plotting nobs dt for surveys")
#plot_nobs_dt_for_surveys(settings=conf['nobs_dt_for_surveys'])