-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgenerate_report.py
executable file
·190 lines (169 loc) · 7.93 KB
/
generate_report.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
#!/usr/bin/env python
# ------------------------------------------------------------------------------
# Comparison of regularization techniqes.
# Jan Kukacka, 11/2017
# jan.kukacka@tum.de
# ------------------------------------------------------------------------------
# Interpretation of the experiment results
# ------------------------------------------------------------------------------
import cPickle
import os
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
def load_results(filename):
assert os.path.isfile(filename), "File " + filename + " does not exist."
with open(filename) as f:
results = cPickle.load(f)
return results
def get_avg_time(histories):
repetitions = len(histories)
epochs = len(histories[0]['time'])
arr = np.empty((repetitions,epochs))
for i in xrange(repetitions):
arr[i] = histories[i]['time']
return np.mean(arr.flatten())
def get_best_test(histories, series_name):
test_score_index = ['loss', 'acc', 'top_k_categorical_accuracy'].index(series_name)
repetitions = len(histories)
arr_ep = np.empty((repetitions,))
arr_test = np.empty((repetitions,))
for i in xrange(repetitions):
arr_ep[i] = histories[i]['best_epoch']
arr_test[i] = histories[i]['test_score'][test_score_index]
ep_mean = np.mean(arr_ep)
ep_std = np.std(arr_ep)
score_mean = np.mean(arr_test)
score_std = np.std(arr_test)
return score_mean, score_std, ep_mean, ep_std
def plot_series(res, axes=None, model_index=0, dataset_size=500,
series_name='loss', timed=False, show_best_test=False,
line_params={}, dev_params={}):
histories = res[model_index]['results'][dataset_size]
repetitions = len(histories)
epochs = len(histories[0][series_name])
arr = np.empty((repetitions, epochs))
for i in xrange(repetitions):
arr[i] = histories[i][series_name]
means = np.mean(arr, axis=0)
devs = np.std(arr, axis=0)
timing = xrange(epochs)
if timed:
avg_time = get_avg_time(histories)
timing = np.array(timing)*avg_time
if axes is None:
axes = plt.gca()
if 'label' not in line_params:
line_params['label'] = res[model_index]['name'],
axes.plot(timing, means, **line_params)
axes.fill_between(timing, means+devs, means-devs, **dev_params)
axes.legend(loc='upper left', bbox_to_anchor=(1,1))
if show_best_test:
best_test_mean, best_test_std, best_ep_mean, best_ep_std = get_best_test(histories, series_name)
timing = 1
if timed:
timing = get_avg_time(histories)
axes.errorbar(best_ep_mean*timing, best_test_mean,
xerr=best_ep_std*timing, yerr=best_test_std,
marker='x', markersize=10, c=line_params['color'])
def plot_comparison(res, dataset_size, title, series_name, xlabel='epochs',
timed=False, save_png=True, save_eps=True):
plt.figure()
plt.title(title)
plt.gca().set_ylabel(series_name)
plt.gca().set_xlabel(xlabel)
colors = 'rgbc'
for i in xrange(len(res)):
color = colors[i]
line_params = {'label': res[i]['name'] + ' (training)', 'color': color}
dev_params = {'linewidth': 0, 'alpha': 0.3, 'facecolor': color}
plot_series(res, model_index=i, series_name=series_name,
dev_params=dev_params, line_params=line_params,
dataset_size=dataset_size, timed=timed, show_best_test=True)
line_params = {'label': res[i]['name'] + ' (validation)', 'color': color, 'linestyle':'--'}
plot_series(res, model_index=i, series_name='val_'+series_name,
dev_params=dev_params, line_params=line_params,
dataset_size=dataset_size, timed=timed)
if save_png:
t = 'time' if timed else 'ep'
plt.savefig('report/plot_'+t+'_'+series_name+'_{}.png'.format(dataset_size),
bbox_inches='tight')
if save_eps:
t = 'time' if timed else 'ep'
plt.savefig('report/plot_'+t+'_'+series_name+'_{}.svg'.format(dataset_size),
bbox_inches='tight')
plt.close()
def generate_report(results_filename='report/results.pkl', generate_png=True,
generate_eps=True):
res = load_results(results_filename)
# Folder for saving images
if not os.path.isdir('report'):
os.mkdir('report')
# normal charts
for key in sorted(res[0]['results']):
plot_comparison(res, key, 'Loss, {} samples'.format(key), 'loss',
save_png=generate_png, save_eps=generate_eps)
plot_comparison(res, key, 'Accuracy, {} samples'.format(key), 'acc',
save_png=generate_png, save_eps=generate_eps)
plot_comparison(res, key, 'Top-5, {} samples'.format(key),
'top_k_categorical_accuracy',save_png=generate_png,
save_eps=generate_eps)
# timed charts
for key in sorted(res[0]['results']):
plot_comparison(res, key, 'Loss, {} samples'.format(key), 'loss',
timed=True, xlabel='time (s)', save_png=generate_png,
save_eps=generate_eps)
plot_comparison(res, key, 'Accuracy, {} samples'.format(key), 'acc',
timed=True, xlabel='time (s)', save_png=generate_png,
save_eps=generate_eps)
plot_comparison(res, key, 'Top-5, {} samples'.format(key),
'top_k_categorical_accuracy', timed=True,
xlabel='time (s)',
save_png=generate_png,
save_eps=generate_eps)
if __name__ == '__main__':
generate_report()
#----------------------------------------------------------------
# Helping function for printing numeric data
def print_best_epoch(res, model_index, dataset_size):
# Convergence speed metrics
for dataset_size in sorted(res[0]['results']):
print 'Dataset size:', dataset_size
for model_index in xrange(len(res)):
histories = res[model_index]['results'][dataset_size]
repetitions = len(histories)
arr = np.empty((repetitions,))
for i in xrange(repetitions):
arr[i] = histories[i]['best_epoch']
means = np.mean(arr)
devs = np.std(arr)
print res[model_index]['name'], 'converges in {:.2f}+-{:.4f}'.format(means, devs)
def print_best_scores(res, model_index, dataset_size):
for dataset_size in sorted(res[0]['results']):
print 'Dataset size:', dataset_size
for model_index in xrange(len(res)):
histories = res[model_index]['results'][dataset_size]
repetitions = len(histories)
arr = np.empty((repetitions,3))
for i in xrange(repetitions):
arr[i] = histories[i]['test_score']
means = np.mean(arr, axis=0)
devs = np.std(arr, axis=0)
print '. ', res[model_index]['name'], 'has loss {:.2f} +- {:.4f}'.format(means[0], devs[0])
print '. ', res[model_index]['name'], 'has accuracy {:.1f}% +- {:.2f}%'.format(means[1]*100, devs[1]*100)
print '. ', res[model_index]['name'], 'has top-5 {:.1f}% +- {:.2f}%'.format(means[2]*100, devs[2]*100)
print '..'
def print_timings(res, model_index, dataset_size):
for dataset_size in sorted(res[0]['results']):
print 'Dataset size:', dataset_size
for model_index in xrange(len(res)):
histories = res[model_index]['results'][dataset_size]
repetitions = len(histories)
epochs = len(histories[0]['time'])
arr = np.empty((repetitions,epochs))
for i in xrange(repetitions):
arr[i] = histories[i]['time']
means = np.mean(arr.flatten())
devs = np.std(arr.flatten())
print '. ', res[model_index]['name'], 'needs {:.2f}s +- {:.4f}s'.format(means, devs), 'per epoch'