def main(fmt='png'):
nsamples = 8
npixels = 256
tol = 1e-6 # for example reconstructions, run GD until ||x-x*|| <= tol
objfun = get_dataset(nsamples, npixels, lower_level_solver='gd')
dirname = 'examples'
outfolder = os.path.join(MAIN_OUTFOLDER, dirname)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
# Plot the training data (true & noisy images only)
plot_training_data(objfun, 6, npixels, filename=os.path.join(outfolder, 'gd_demo_data'),
ymin=-0.3, ymax=1.3, legend_loc='center left', fmt=fmt, font_size='x-small',
nrows=2, axis_font_size='small', figsize=cm2inch(1.3*8,1.3*3.5))
param_combinations = [] # alpha, eps_l2, eps_tv, xlabel, ylabel
# first row (vary alpha)
param_combinations.append((1.0, 1e-3, 1e-3, r'$\alpha=1$', 'TV\nweight'))
param_combinations.append((1e-1, 1e-3, 1e-3, r'$\alpha=0.1$', None))
param_combinations.append((1e-2, 1e-3, 1e-3, r'$\alpha=0.01$', None))
# second row (vary eps_tv)
param_combinations.append((1.0, 1e-3, 1e0, r'$\nu=1$', 'TV\nsmoothing'))
param_combinations.append((1.0, 1e-3, 1e-1, r'$\nu=0.1$', None))
param_combinations.append((1.0, 1e-3, 1e-2, r'$\nu=0.01$', None))
# third row (vary eps_l2)
param_combinations.append((1.0, 1e0, 1e-3, r'$\xi=1$', 'L2\npenalty'))
param_combinations.append((1.0, 1e-1, 1e-3, r'$\xi=0.1$', None))
param_combinations.append((1.0, 1e-2, 1e-3, r'$\xi=0.01$', None))
plot_example_recons(objfun, 0, tol, npixels, param_combinations, 3, 3,
filename=os.path.join(outfolder, 'gd_demo_recons'),
figsize=cm2inch(1.3 * 8, 1.3 * 4.3), ymin=-0.3, ymax=1.3,
legend_loc='center left', fmt=fmt, font_size='x-small', axis_font_size='small')
return
def main(fmt='png'):
# Generate main plots for 2D denoising
setting_filestem = '3param2d_demo'
infolder = os.path.join(RESULTS_FOLDER, setting_filestem)
outfolder = os.path.join(MAIN_OUTFOLDER, setting_filestem)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
results_list = get_results(setting_filestem, infolder=infolder, fista_dynamic_only=False)
# Basic objective reduction
make_obj_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_obj_redn'),
use_logscale=False, legend_loc='upper right', fmt=fmt, ymin=None, ymax=None,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small', legend_ncol=2)
# niters plot
make_niters_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_niters'),
use_logscale=True, legend_loc=None, fmt=fmt, cumulative=True, ymin=None, ymax=None,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small', xtick_freq=25)
# Parameter plots
for param_idx, param_name in [(0, 'alpha'), (1, 'eps_l2'), (2, 'eps_tv')]:
make_param_plot(results_list, param_idx, filename=os.path.join(outfolder, setting_filestem + '_param_%s' % param_name),
use_logscale=True, legend_loc=None, fmt=fmt, param_is_log=True, ymin=None, ymax=None,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small')
# Build final reconstructions (this is slow, so save results before plotting)
build_final_reconstructions2d(results_list, outfolder, fista_dynamic_only=True,
lower_level_solver='fista', lower_level_niters=2000)
# and plot final reconstructions
img_idx = [0, 3, 7, 12, 20]
plot_final_reconstructions2d_vertical(results_list, img_idx, outfolder,
filename=os.path.join(outfolder, setting_filestem + '_recons'),
fmt=fmt, fista_dynamic_only=True, axis_font_size='small',
figsize=cm2inch(1.3*6,1.3*(3*len(img_idx))))
# sigma v alpha comparison
setting_files = []
setting_files.append('3param2d_demo')
setting_files.append('3param2d_demo_noise2')
setting_files.append('3param2d_demo_noise3')
setting_files.append('3param2d_demo_noise4')
setting_files.append('3param2d_demo_noise5')
setting_files.append('3param2d_demo_noise6')
setting_files.append('3param2d_demo_noise7')
setting_files.append('3param2d_demo_noise8')
plot_alpha_sigma_relationship(setting_files, filename=os.path.join(outfolder, setting_filestem + '_alpha_sigma'),
ymin=None, ymax=None, legend_loc='lower right', lw=1.5,
fmt=fmt, font_size='x-small', axis_font_size='small', markersize=6,
solver_name='fista_dynamic', figsize=cm2inch(1.3*7, 1.3*2.5))
return
def main_3param(fmt='png'):
# Generate main plots for 3-param denoising
setting_filestem = '3param_reg6'
infolder = os.path.join(RESULTS_FOLDER, setting_filestem)
outfolder = os.path.join(MAIN_OUTFOLDER, setting_filestem)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
results_list = get_results(setting_filestem, infolder=infolder, fista_dynamic_only=False)
# Basic objective reduction
make_obj_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_obj_redn'),
use_logscale=True, legend_loc='upper right', fmt=fmt, ymin=2e0, ymax=1e2,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small', legend_ncol=2)
# niters plot
make_niters_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_niters'),
use_logscale=True, legend_loc=None, fmt=fmt, cumulative=True, ymin=None, ymax=None,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small', xtick_freq=25)
# reconstructions
plot_final_reconstructions(results_list, nimgs=6, filename=os.path.join(outfolder, setting_filestem + '_recons'),
legend_loc='center left', fmt=fmt, ymin=None, ymax=None,
lower_level_solver='fista', lower_level_niters=1000, fista_dynamic_only=True,
figsize=cm2inch(1.3 * 8, 1.3 * 3.5), font_size='x-small', nrows=2, axis_font_size='small')
# reconstructions with beta=1e-4
alt_setting_filestem = '3param_reg4'
alt_infolder = os.path.join(RESULTS_FOLDER, alt_setting_filestem)
alt_results_list = get_results(alt_setting_filestem, infolder=alt_infolder, fista_dynamic_only=False)
plot_final_reconstructions(alt_results_list, nimgs=6,
filename=os.path.join(outfolder, alt_setting_filestem + '_recons'),
legend_loc='center left', fmt=fmt, ymin=None, ymax=None,
lower_level_solver='fista', lower_level_niters=1000, fista_dynamic_only=True,
figsize=cm2inch(1.3 * 8, 1.3 * 3.5), font_size='x-small', nrows=2, axis_font_size='small')
# reconstructions after different numbers of iterations
eval_idx = [1, 10, 20, 100]
make_recons_evolution_plot(setting_filestem, eval_idx, img_idx=0, nrows=2,
filename=os.path.join(outfolder, setting_filestem + '_recons_during_solve'),
solver_name='fista_dynamic', lower_level_solver='fista', lower_level_niters=1000,
legend_loc='center left', legend_ncol=1, axis_font_size='small', font_size='x-small',
figsize=cm2inch(1.3 *8, 1.3 * 3.5), ymin=-0.3, ymax=1.3, fmt=fmt)
return
def main_1param(fmt='png'):
# Generate main plots for 1-param denoising
setting_filestem = '1param_start0_budget20'
infolder = os.path.join(RESULTS_FOLDER, setting_filestem)
outfolder = os.path.join(MAIN_OUTFOLDER, setting_filestem)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
results_list = get_results(setting_filestem, infolder=infolder, fista_dynamic_only=False)
# Basic objective reduction
make_obj_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_obj_redn'),
use_logscale=True, legend_loc='upper right', fmt=fmt, ymin=8e-1, ymax=2e3,
figsize=cm2inch(8, 4), font_size='x-small', axis_font_size='small', legend_ncol=2)
# Zoomed in objective reduction
make_obj_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_obj_redn_zoom'),
use_logscale=False, legend_loc=None, fmt=fmt, ymin=1.3, ymax=3.0,
figsize=cm2inch(8,4), font_size='x-small', axis_font_size='small')
# Alpha plot
make_param_plot(results_list, 0, filename=os.path.join(outfolder, setting_filestem + '_param_alpha'),
use_logscale=True, legend_loc=None, fmt=fmt, param_is_log=True, ymin=None, ymax=None,
figsize=cm2inch(8,4), font_size='x-small', axis_font_size='small')
# niters plot
make_niters_plot(results_list, filename=os.path.join(outfolder, setting_filestem + '_niters'),
use_logscale=True, legend_loc=None, fmt=fmt, cumulative=True, ymin=None, ymax=None,
figsize=cm2inch(8,4), font_size='x-small', axis_font_size='small', xtick_freq=5)
plot_final_reconstructions(results_list, nimgs=6, filename=os.path.join(outfolder, setting_filestem + '_recons'),
legend_loc='center left', fmt=fmt, ymin=-0.3, ymax=1.3,
lower_level_solver='fista', lower_level_niters=1000, fista_dynamic_only=True,
figsize=cm2inch(1.3 * 8, 1.3 * 3.5), font_size='x-small', nrows=2, axis_font_size='small')
# alpha trajectories for different starting points
alt_starting_points = ['1param_start1_budget20', '1param_start2_budget20', '1param_startm1_budget20', '1param_startm2_budget20']
for alt_setting_filestem in alt_starting_points:
if 'start1' in alt_setting_filestem: # has a legend, so a bit special
legend_loc = 'lower right'
legend_ncol = 2
ymin = 1e-2
else:
legend_loc = None
legend_ncol = 1
ymin = None
alt_infolder = os.path.join(RESULTS_FOLDER, alt_setting_filestem)
alt_results_list = get_results(alt_setting_filestem, infolder=alt_infolder, fista_dynamic_only=False)
make_param_plot(alt_results_list, 0, filename=os.path.join(outfolder, alt_setting_filestem + '_param_alpha'),
use_logscale=True, legend_loc=legend_loc, fmt=fmt, param_is_log=True, ymin=ymin, ymax=None,
figsize=cm2inch(1.5 * 6, 1.5 * 3), font_size='x-small', legend_ncol=legend_ncol,
axis_font_size='small')
# Robustness to starting point [width=9cm, height=2.5cm]
filestems = ['1param_start%s_budget20' % s for s in ['m2', 'm1', '0', '1', '2']]
plot_starting_point_robust(filestems, filename=os.path.join(outfolder, '1param_start_robust'),
font_size='x-small', figsize=cm2inch(1.3*8,1.3*2.5), fmt=fmt, legend_loc='upper left',
legend_ncol=1, axis_font_size='small')
return
def main(fmt='png'):
dirname = 'examples'
outfolder = os.path.join(MAIN_OUTFOLDER, dirname)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
# compare error bounds [width=3.6cm, height=2.5cm, but scaled]
print("============")
print("Nesterov quadratic with GD")
print("============")
compare_bounds_single_problem('nesterov',
'gd',
200,
n=10,
filename=os.path.join(
outfolder, 'nesterov10_gd200'),
legend_loc='lower left',
fmt=fmt,
font_size='x-small',
axis_font_size='small',
ymin=None,
ymax=None,
figsize=cm2inch(1.5 * 3.6, 1.5 * 2.5),
lw=1.5)
print("============")
print("Nesterov quadratic with FISTA")
print("============")
compare_bounds_single_problem('nesterov',
'fista',
200,
n=10,
filename=os.path.join(
outfolder, 'nesterov10_fista200'),
legend_loc=None,
fmt=fmt,
font_size='x-small',
axis_font_size='small',
ymin=None,
ymax=None,
figsize=cm2inch(1.5 * 3.6, 1.5 * 2.5),
lw=1.5)
print("============")
print("Denoising with GD")
print("============")
compare_bounds_single_problem('denoising',
'gd',
2000,
filename=os.path.join(
outfolder, 'denoising_gd2000'),
legend_loc=None,
fmt=fmt,
font_size='x-small',
axis_font_size='small',
ymin=None,
ymax=None,
figsize=cm2inch(1.5 * 3.6, 1.5 * 2.5),
lw=1.5)
print("============")
print("Denoising with FISTA")
print("============")
compare_bounds_single_problem('denoising',
'fista',
1000,
filename=os.path.join(
outfolder, 'denoising_fista1000'),
legend_loc=None,
fmt=fmt,
font_size='x-small',
axis_font_size='small',
ymin=None,
ymax=None,
figsize=cm2inch(1.5 * 3.6, 1.5 * 2.5),
lw=1.5)
return
def main(fmt='png'):
# Generate main plots for MRI sampling
setting_filestem = 'inpainting_mri_demo2_several_noise'
infolder = os.path.join(RESULTS_FOLDER, setting_filestem)
outfolder = os.path.join(MAIN_OUTFOLDER, setting_filestem)
if not os.path.isdir(outfolder):
os.makedirs(outfolder, exist_ok=True)
results_list = get_results(setting_filestem,
infolder=infolder,
fista_dynamic_only=False)
# Basic objective reduction
make_obj_plot(results_list,
filename=os.path.join(outfolder,
setting_filestem + '_obj_redn'),
use_logscale=False,
legend_loc='upper right',
fmt=fmt,
ymin=0.3,
ymax=5.0,
figsize=cm2inch(8, 4),
font_size='x-small',
axis_font_size='small',
legend_ncol=2)
# niters plot
make_niters_plot(results_list,
filename=os.path.join(outfolder,
setting_filestem + '_niters'),
use_logscale=True,
legend_loc=None,
fmt=fmt,
cumulative=True,
ymin=1e4,
ymax=None,
figsize=cm2inch(8, 4),
font_size='x-small',
axis_font_size='small',
xtick_freq=500)
# final sampling patterns
make_sampling_pattern_plot(results_list,
filename=os.path.join(
outfolder,
setting_filestem + '_sampling_patterns'),
fmt=fmt,
axis_font_size='small',
apply_fftshift=True,
cmap='gray',
sampling_thresh=0.001,
figsize=cm2inch(1.3 * 8, 1.3 * 4))
# reconstructions
make_mri_reconstruction_plot(results_list,
filename=os.path.join(
outfolder, setting_filestem + '_recons'),
nimgs=6,
ymin=None,
ymax=None,
legend_loc='center left',
fmt=fmt,
font_size='x-small',
fista_dynamic_only=True,
lower_level_solver='fista',
lower_level_niters=2000,
nrows=2,
axis_font_size='small',
sampling_thresh=0.001,
figsize=cm2inch(1.3 * 8, 1.3 * 3.5))
return