Skip to content
Snippets Groups Projects
Commit 44d5a0cb authored by Jörg Martin's avatar Jörg Martin
Browse files

Made plots more suitable for publication

parents 5406914b 34b89949
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Experiments/configurations/eiv_naval.json
+ 1
− 1
View file @ 44d5a0cb
......@@ -15,7 +15,7 @@
"init_std_y_list": [0.5],
"gamma": 0.5,
"hidden_layers": [1024, 1024, 1024, 1024],
"fixed_std_x": 0.05,
"fixed_std_x": 0.025,
"seed_range": [0,10],
"gpu_number": 1
}
Experiments/create_tabular.py
+ 2
− 1
View file @ 44d5a0cb
......@@ -2,7 +2,8 @@ import os
import glob
import json
metrics_to_display = ['rmse','true_coverage_numerical', 'total_coverage']
metrics_to_display = ['rmse', 'coverage_numerical',
'true_coverage_numerical', 'total_coverage']
show_incomplete = True
list_of_result_files = glob.glob(os.path.join('results','*.json'))
......
Experiments/plot_prediction.py
+ 64
− 6
View file @ 44d5a0cb
......@@ -11,11 +11,20 @@ import json
import torch
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
from EIVArchitectures import Networks
from EIVTrainingRoutines import train_and_store
font = {'family' : 'DejaVu Sans',
'weight' : 'normal',
'size' : 16}
matplotlib.rc('font', **font)
linewidth = 2.0
# coverage factor
k = 1.96
......@@ -107,6 +116,7 @@ def compute_predictions_and_uncertainties(data, x_range, eiv, number_of_draws,
return_ground_truth=False,
return_normalized_func=True,
normalize=normalize)
plotting_dictionary['func'] = normalized_func
input_dim = test_data[0][0].numel()
output_dim = test_data[0][1].numel()
assert output_dim == 1
......@@ -223,8 +233,19 @@ list_x_range = [torch.linspace(-1.0,1.0, 50),
torch.linspace(-0.2,0.8, 50)]
list_color = [('red','blue')] * len(data_list)
list_number_of_draws = [((100,5), 100)] * len(data_list)
for i, (data, x_range, color, number_of_draws) in enumerate(zip(data_list,
list_x_range, list_color, list_number_of_draws)):
# create an extra zoom plot for zoom_example
zoom_example = 'linear'
# where to zoom
zoom_point = 14
# size of the zoom plot
x_zoom_radius = 0.4
y_zoom_radius = x_zoom_radius
fignum = 0
for data, x_range, color, number_of_draws in zip(data_list,
list_x_range, list_color, list_number_of_draws):
fignum += 1
eiv_plotting_dictionary = compute_predictions_and_uncertainties(
data=data,
x_range=x_range,
......@@ -237,25 +258,62 @@ for i, (data, x_range, color, number_of_draws) in enumerate(zip(data_list,
number_of_draws=number_of_draws[1])
input_dim = eiv_plotting_dictionary['input_dim']
if input_dim == 1:
plt.figure(i+1)
plt.figure(fignum)
plt.clf()
x_values, y_values = eiv_plotting_dictionary['range_points']
plt.plot(x_values.flatten(), y_values.flatten(),'-', color='k')
noisy_x_values, _ = eiv_plotting_dictionary['noisy_range_points']
plt.plot(x_values.flatten(), y_values.flatten(),'-', color='k', linewidth=linewidth)
eiv_pred = eiv_plotting_dictionary['prediction']
eiv_unc = eiv_plotting_dictionary['uncertainty']
plt.plot(x_values, eiv_pred,'-',
color=color[0])
color=color[0], linewidth=linewidth)
plt.fill_between(x_values.flatten(), eiv_pred-k * eiv_unc,
eiv_pred + k * eiv_unc,
color=color[0], alpha=0.5)
noneiv_pred = noneiv_plotting_dictionary['prediction']
noneiv_unc = noneiv_plotting_dictionary['uncertainty']
plt.plot(x_values.flatten(), noneiv_pred,'-',
color=color[1])
color=color[1], linewidth=linewidth)
plt.fill_between(x_values.flatten(), noneiv_pred-k * noneiv_unc,
noneiv_pred + k * noneiv_unc,
color=color[1], alpha=0.5)
plt.tight_layout()
plt.savefig(f'results/figures/prediction_{data}.pdf')
if data == zoom_example:
fignum += 1
func = eiv_plotting_dictionary['func']
x_point = x_values[zoom_point]
y_point = func(torch.tensor(x_point)).numpy()
noisy_x_point = noisy_x_values[zoom_point]
func_noisy_x_point = func(torch.tensor(noisy_x_point)).numpy()
plt.figure(fignum)
plt.clf()
plt.plot(x_values.flatten(), y_values.flatten(),'-', color='k', linewidth=linewidth)
plt.plot(x_values, eiv_pred,'-',
color=color[0], linewidth=linewidth)
plt.fill_between(x_values.flatten(), eiv_pred-k * eiv_unc,
eiv_pred + k * eiv_unc,
color=color[0], alpha=0.5)
plt.plot(x_values.flatten(), noneiv_pred,'-',
color=color[1], linewidth=linewidth)
plt.fill_between(x_values.flatten(), noneiv_pred-k * noneiv_unc,
noneiv_pred + k * noneiv_unc,
color=color[1], alpha=0.5)
plt.axvline(x_point, color='black', linestyle='dotted')
plt.axhline(y_point, color='black', linestyle='dotted')
plt.axvline(noisy_x_point, color='gray', linestyle='dashed')
plt.axhline(func_noisy_x_point, color='gray', linestyle='dashed')
plt.text(x_point - 0.1 * x_zoom_radius,y_point-0.9 * y_zoom_radius, r'$\zeta$', color='k')
plt.text(noisy_x_point - 0.1 * x_zoom_radius,y_point-0.9 * y_zoom_radius, r'$x$', color='gray')
plt.text(x_point - 0.92 * x_zoom_radius,y_point-0.13 * y_zoom_radius, r'$g(\zeta)$', color='k')
plt.text(x_point - 0.92 * x_zoom_radius,func_noisy_x_point-0.13 * y_zoom_radius, r'$g(x)$', color='gray')
plt.gca().set_xlim(left=x_point - x_zoom_radius,
right=x_point + x_zoom_radius)
plt.gca().set_ylim(bottom=y_point - y_zoom_radius,
top=y_point + y_zoom_radius)
plt.gca().set_aspect('equal', adjustable='box')
plt.tight_layout()
plt.savefig(f'results/figures/prediction_{data}_zoom.pdf')
else:
# multidimensional handling not included yet
pass
......
Experiments/plot_summary.py
+ 91
− 17
View file @ 44d5a0cb
......@@ -10,10 +10,17 @@ import glob
import json
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
font = {'family' : 'DejaVu Sans',
'weight' : 'normal',
'size' : 16}
matplotlib.rc('font', **font)
k = 2
# load in all available result files
list_of_result_files = glob.glob(os.path.join('results','*.json'))
results = {}
......@@ -46,6 +53,7 @@ metric = 'rmse'
data_list = results.keys()
colors = ['red', 'blue']
ymax = 0.8
minimal_bar_size = ymax * 1.5e-3
# read out EiV and non-EiV results for all datasets
metric_results = [
(save_readout(results[data]['eiv'], metric),
......@@ -55,7 +63,7 @@ metric_results = [
# create figure
plt.figure(1)
plt.clf()
plt.title('RMSE')
plt.gcf().canvas.manager.set_window_title('RMSE')
# plot bars
for i, ([(eiv_metric_mean, eiv_metric_std),
......@@ -66,24 +74,28 @@ for i, ([(eiv_metric_mean, eiv_metric_std),
assert noneiv_metric_mean is not None
if eiv_metric_std is not None:
assert noneiv_metric_std is not None
eiv_bar_size = max(eiv_metric_std, minimal_bar_size)
noneiv_bar_size = max(noneiv_metric_std, minimal_bar_size)
plt.plot(i+1, eiv_metric_mean, '^', color=colors[0])
plt.bar(i+1,
height = 2*eiv_metric_std,
width = 0.1,
bottom = eiv_metric_mean - eiv_metric_std,
height = 2*eiv_bar_size,
width = 0.3,
bottom = eiv_metric_mean - eiv_bar_size,
color=colors[0],
alpha=0.5)
plt.plot(i+1, noneiv_metric_mean, '^', color=colors[1])
plt.bar(i+1,
height = 2 * k *noneiv_metric_std,
width = 0.1,
bottom = noneiv_metric_mean - k* noneiv_metric_std,
height = 2 * k *noneiv_bar_size,
width = 0.3,
bottom = noneiv_metric_mean - k* noneiv_bar_size,
color=colors[1],
alpha=0.5)
plt.ylim(bottom=0, top=ymax)
ax = plt.gca()
ax.set_xticks(np.arange(1,len(data_list)+1))
ax.set_xticklabels(data_list, rotation='vertical')
plt.ylabel('RMSE')
plt.tight_layout()
plt.savefig('results/figures/RMSE_bar_plot.pdf')
## coverage plot
......@@ -92,6 +104,7 @@ metric = 'true_coverage_numerical'
data_list = ['linear','quadratic','cubic','sine']
colors = ['red', 'blue']
ymax = 1.0
minimal_bar_size = ymax * 1.5e-3
# read out EiV and non-EiV results for all datasets
metric_results = [
(save_readout(results[data]['eiv'], metric),
......@@ -101,7 +114,7 @@ metric_results = [
# create figure
plt.figure(2)
plt.clf()
plt.title('coverage (ground truth)')
plt.gcf().canvas.manager.set_window_title('coverage (ground truth)')
# plot bars
for i, ([(eiv_metric_mean, eiv_metric_std),
......@@ -112,18 +125,20 @@ for i, ([(eiv_metric_mean, eiv_metric_std),
assert noneiv_metric_mean is not None
if eiv_metric_std is not None:
assert noneiv_metric_std is not None
eiv_bar_size = max(eiv_metric_std, minimal_bar_size)
noneiv_bar_size = max(noneiv_metric_std, minimal_bar_size)
plt.plot(i+1, eiv_metric_mean, '^', color=colors[0])
plt.bar(i+1,
height = 2*eiv_metric_std,
height = 2*eiv_bar_size,
width = 0.1,
bottom = eiv_metric_mean - eiv_metric_std,
bottom = eiv_metric_mean - eiv_bar_size,
color=colors[0],
alpha=0.5)
plt.plot(i+1, noneiv_metric_mean, '^', color=colors[1])
plt.bar(i+1,
height = 2 * k *noneiv_metric_std,
height = 2 * k *noneiv_bar_size,
width = 0.1,
bottom = noneiv_metric_mean - k* noneiv_metric_std,
bottom = noneiv_metric_mean - k* noneiv_bar_size,
color=colors[1],
alpha=0.5)
plt.axhline(0.95,0.0,1.0,color='k', linestyle='dashed')
......@@ -131,6 +146,8 @@ plt.ylim(bottom=0, top=ymax)
ax = plt.gca()
ax.set_xticks(np.arange(1,len(data_list)+1))
ax.set_xticklabels(data_list, rotation='vertical')
plt.ylabel('coverage ground truth')
plt.tight_layout()
plt.savefig('results/figures/true_coverage_bar_plot.pdf')
## noisy coverage plot
......@@ -139,6 +156,7 @@ metric = 'coverage_numerical'
data_list = results.keys()
colors = ['red', 'blue']
ymax = 1.0
minimal_bar_size = ymax * 1.5e-3
# read out EiV and non-EiV results for all datasets
metric_results = [
(save_readout(results[data]['eiv'], metric),
......@@ -148,7 +166,7 @@ metric_results = [
# create figure
plt.figure(3)
plt.clf()
plt.title('coverage (noisy labels)')
plt.gcf().canvas.manager.set_window_title('coverage (noisy labels)')
# plot bars
for i, ([(eiv_metric_mean, eiv_metric_std),
......@@ -159,22 +177,78 @@ for i, ([(eiv_metric_mean, eiv_metric_std),
assert noneiv_metric_mean is not None
if eiv_metric_std is not None:
assert noneiv_metric_std is not None
eiv_bar_size = max(eiv_metric_std, minimal_bar_size)
noneiv_bar_size = max(noneiv_metric_std, minimal_bar_size)
plt.plot(i+1, eiv_metric_mean, '^', color=colors[0])
plt.bar(i+1,
height = 2*eiv_metric_std,
height = 2*eiv_bar_size,
width = 0.1,
bottom = eiv_metric_mean - eiv_metric_std,
bottom = eiv_metric_mean - eiv_bar_size,
color=colors[0],
alpha=0.5)
plt.plot(i+1, noneiv_metric_mean, '^', color=colors[1])
plt.bar(i+1,
height = 2 * k *noneiv_metric_std,
height = 2 * k *noneiv_bar_size,
width = 0.1,
bottom = noneiv_metric_mean - k* noneiv_metric_std,
bottom = noneiv_metric_mean - k* noneiv_bar_size,
color=colors[1],
alpha=0.5)
plt.ylim(bottom=0, top=ymax)
ax = plt.gca()
ax.set_xticks(np.arange(1,len(data_list)+1))
ax.set_xticklabels(data_list, rotation='vertical')
plt.ylabel('coverage (epis. unc. / ' + r'$u(x)$' + ')' )
plt.tight_layout()
plt.savefig('results/figures/noisy_coverage_bar_plot.pdf')
## coverage by total uncertainty
metric = 'total_coverage'
data_list = results.keys()
colors = ['red', 'blue']
ymax = 1.0
minimal_bar_size = ymax * 1.5e-3
# read out EiV and non-EiV results for all datasets
metric_results = [
(save_readout(results[data]['eiv'], metric),
save_readout(results[data]['noneiv'], metric))
for data in data_list]
# create figure
plt.figure(4)
plt.clf()
plt.gcf().canvas.manager.set_window_title('total_coverage')
# plot bars
for i, ([(eiv_metric_mean, eiv_metric_std),
(noneiv_metric_mean, noneiv_metric_std)],\
data) in\
enumerate(zip(metric_results, data_list)):
if eiv_metric_mean is not None:
assert noneiv_metric_mean is not None
if eiv_metric_std is not None:
assert noneiv_metric_std is not None
eiv_bar_size = max(eiv_metric_std, minimal_bar_size)
noneiv_bar_size = max(noneiv_metric_std, minimal_bar_size)
plt.plot(i+1, eiv_metric_mean, '^', color=colors[0])
plt.bar(i+1,
height = 2*eiv_bar_size,
width = 0.3,
bottom = eiv_metric_mean - eiv_bar_size,
color=colors[0],
alpha=0.5)
plt.plot(i+1, noneiv_metric_mean, '^', color=colors[1])
plt.bar(i+1,
height = 2 * k *noneiv_bar_size,
width = 0.3,
bottom = noneiv_metric_mean - k* noneiv_bar_size,
color=colors[1],
alpha=0.5)
plt.ylim(bottom=0, top=ymax)
ax = plt.gca()
ax.set_xticks(np.arange(1,len(data_list)+1))
ax.set_xticklabels(data_list, rotation='vertical')
plt.axhline(0.95,0.0,1.0,color='k', linestyle='dashed')
plt.ylabel('coverage (total unc.)')
plt.tight_layout()
plt.savefig('results/figures/total_coverage_bar_plot.pdf')
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment