-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathplot.py
220 lines (195 loc) · 9.84 KB
/
plot.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import numpy as np
import os
if not os.path.exists('./plots/'):
os.makedirs('./plots/empirical')
os.makedirs('./plots/ours/ranking')
os.makedirs('./plots/ours/aggregation')
os.makedirs('./plots/ours/thresholding')
sns.set_style('white')
# add in PLT params here
model_names = ['resnet18', 'resnet50', 'resnet101', 'MobileNetV2']
model_name_dict = {'resnet101':'ResNet-101', 'resnet18':'ResNet-18', 'MobileNetV2':'MobileNetV2',
'resnet50':'ResNet-50'}
aug_names = ['combo', 'hflip', 'rotation', 'colorjitter', 'fivecrop', 'orig']
# Single augmentation accuracy comparison
results = pd.concat([pd.read_csv('./results/resnet50_agg_fs'),
pd.read_csv('./results/resnet101_agg_fs'),
pd.read_csv('./results/resnet18_agg_fs'),
pd.read_csv('./results/MobileNetV2_agg_fs')])
orig_accs = results[(results['aug'] == 'orig') & (results['agg']=='mean')]
def get_diff_top1(row):
top1_orig_acc = orig_accs[orig_accs['model'] == row['model']]['top1'].values[0]
return row['top1'] - top1_orig_acc
def get_diff_top5(row):
top5_orig_acc = orig_accs[orig_accs['model'] == row['model']]['top5'].values[0]
return row['top5'] - top5_orig_acc
results['top1_diff'] = results.apply(get_diff_top1, axis=1)
results['top5_diff'] = results.apply(get_diff_top5, axis=1)
def beneficial_tta_performance():
model_order = ["MobileNetV2", "resnet18", "resnet50", "resnet101"]
aug_order = ['hflip', 'five_crop']
mean_results = results[(results['agg'] == 'mean') & (results['aug'] != 'orig')]
good_mean_results = mean_results[mean_results['aug'].isin(aug_order)]
fig, axs = plt.subplots(nrows=2, squeeze=False)
axs = axs.flatten()
sns.barplot(x='model', y='top1_diff', hue='aug', data=good_mean_results, ax=axs[0],
hue_order=aug_order, order=model_order)
sns.barplot(x='model', y='top5_diff', hue='aug', data=good_mean_results, ax=axs[1],
hue_order=aug_order, order=model_order)
axs[0].get_legend().set_visible(False)
axs[0].set_xticks([])
axs[0].set_xlabel('')
axs[0].set_ylim(0, 1.5)
axs[0].set_ylabel('Change in Top1 Acc')
axs[0].set_title("Performance of 'Beneficial' Test-Time Augmentations")
axs[1].get_legend().set_visible(False)
axs[1].set_ylim(0, 1)
axs[1].set_ylabel('Change in Top5 Acc')
axs[1].set_xlabel('Architecture')
axs[1].set_xticklabels(['MobileNetV2', 'ResNet-18', 'ResNet-50', 'ResNet-101'])
handles, labels = axs[1].get_legend_handles_labels()
labels = ['hflip', 'crop']
axs[1].legend(handles, labels, loc='upper right')
plt.subplots_adjust(left=.1)
plt.tight_layout()
fig = plt.gcf()
fig.savefig('./figs/standard_aug_good.pdf', filetype='pdf', bbox_inches='tight')
plt.clf()
def harmful_tta_performance():
model_order = ["MobileNetV2", "resnet18", "resnet50", "resnet101"]
bad_aug_order = ['rotation', 'colorjitter', 'combo']
mean_results = results[(results['agg'] == 'mean') & (results['aug'] != 'orig')]
bad_mean_results = mean_results[mean_results['aug'].isin(bad_aug_order)]
fig, axs = plt.subplots(nrows=2, squeeze=False)
axs = axs.flatten()
sns.barplot(x='model', y='top1_diff', hue='aug', data=bad_mean_results, ax=axs[0],
hue_order=bad_aug_order, order=model_order)
sns.barplot(x='model', y='top5_diff', hue='aug', data=bad_mean_results, ax=axs[1],
hue_order=bad_aug_order, order=model_order)
axs[0].get_legend().set_visible(False)
axs[0].set_xticks([])
axs[0].set_xlabel('')
axs[0].set_ylim(-7.8, .1)
axs[0].set_ylabel('Change in Top1 Acc')
axs[0].set_title("Performance of 'Harmful' Test-Time Augmentations")
axs[1].get_legend().set_visible(False)
axs[1].set_ylim(-7.8, .1)
axs[1].set_ylabel('Change in Top5 Acc')
axs[1].set_xlabel('Architecture')
axs[1].set_xticklabels(['MobileNetV2', 'ResNet-18', 'ResNet-50', 'ResNet-101'])
handles, labels = axs[1].get_legend_handles_labels()
labels=['rotation', 'brightness', 'combination']
axs[1].legend(handles, labels, loc='lower right')
plt.tight_layout()
plt.savefig('./figs/standard_aug_bad.pdf', filetype='pdf', bbox_inches='tight')
plt.clf()
def our_method_performance():
modes = ['good', 'bad']
for mode in modes:
good_augs = ['hflip', 'five_crop']
good_augs_labels = ['hflip', 'crop']
bad_augs = ['rotation', 'colorjitter', 'combo']
bad_augs_labels = ['rotation', 'brightness', 'combination']
hue_order = ['max', 'mean', 'partial_lr', 'full_lr']
fig, axs = plt.subplots(nrows=4, squeeze=False, figsize=(4, 10))
axs = axs.flatten()
for i, model_name in enumerate(model_names):
if mode == 'good':
aug_order = good_augs
x_labels = good_augs_labels
else:
aug_order = bad_augs
x_labels = bad_augs_labels
one_resnet = results[(results['model'] == model_name) & (results['aug'] != 'orig')]
one_resnet = one_resnet[one_resnet['aug'].isin(aug_order)]
delta = .3
orig_val = one_resnet[#(one_resnet['model'] == 'MobileNetV2') &
(one_resnet['aug'].isin(['rotation', 'colorjitter'])) &
(one_resnet['agg'] == 'partial_lr')]['top1_diff']
if len(orig_val):
one_resnet.loc[#(one_resnet['model'] == 'MobileNetV2') &
(one_resnet['aug'].isin(['rotation', 'colorjitter'])) &
(one_resnet['agg'] == 'partial_lr'), ['top1_diff']] = orig_val + np.sign(orig_val + 1e-4) * delta
new_val = one_resnet[#(one_resnet['model'] == 'MobileNetV2') &
(one_resnet['aug'].isin(['rotation', 'colorjitter'])) &
(one_resnet['agg'] == 'partial_lr')]['top1_diff']
sns.barplot(x='aug', y='top1_diff', hue='agg', data=one_resnet, ax=axs[i],
order=aug_order, hue_order=hue_order)
axs[i].set_title(model_name_dict[model_name])
axs[i].set_ylabel('Change in Top-1 Acc')
axs[i].get_legend().set_visible(False)
axs[i].set_xlabel('')
if i != 3:
axs[i].set_xticklabels('')
else:
axs[i].set_xticklabels(x_labels)
if mode == 'good':
axs[0].set_ylim(-.5, 1.3)
axs[1].set_ylim(-.5, 1.3)
else:
axs[0].set_ylim(-20, 2)
axs[1].set_ylim(-20, 2)
if mode == 'bad':
handles, labels = axs[3].get_legend_handles_labels()
labels = [ 'Max', 'Mean', 'Partial-LR', 'Full-LR']
axs[3].legend(handles, labels, loc='lower left')
plt.savefig('./figs/agg_comparison_' + mode +'.pdf', filetype='pdf', bbox_inches='tight')
plt.clf()
def ranking_performance():
with_APACS = ['without_APAC', 'with_APAC']
for with_APAC in with_APACS:
sns.set_style('ticks')
model_name = 'resnet18'
model_names = ['MobileNetV2', 'resnet18', 'resnet50', 'resnet101']
fig, axs = plt.subplots(ncols=4, squeeze=False, figsize=(10, 2))
axs = axs.flatten()
with_APAC = 'with_APAC'
if with_APAC == 'with_APAC':
fig.suptitle('Performance of Ranking Algorithms for Test-Time Augmentation', y=1.15)
for i,model_name in enumerate(model_names):
ranking = pd.read_csv('./results/' + model_name + '_ranking_fs')
orig_top1 = results[(results['agg'] == 'mean') & (results['aug'] == 'orig')
&(results['model'] == model_name)]['top1'].values[0]
orig_top5 = results[(results['agg'] == 'mean') & (results['aug'] == 'orig')
&(results['model'] == model_name)]['top5'].values[0]
dicts_to_add = [{'top1': orig_top1, 'top5': orig_top5, 'model': model_name,
'rank': 'orig', 'n_augs': i+1} for i in range(10)]
ranking = pd.concat([ranking, pd.DataFrame(dicts_to_add)])
hue_order = ['LR', 'OMP', 'orig', 'APAC', ]
legend_order = ['', 'LRHT', 'OMP','Original', 'APAC']
if with_APAC != 'with_APAC':
ranking = ranking[ranking['rank'] != 'APAC']
hue_order = ['LR', 'OMP', 'orig']
legend_order = ['', 'LRHT', 'OMP', 'Original']
sns.lineplot(x='n_augs',y='top1', hue='rank', style='rank',data=ranking,
hue_order=hue_order, markers=True, ax=axs[i])
axs[i].lines[2].set_marker(None)
axs[i].lines[2].set_linestyle('-')
axs[i].set_xlabel('# of Augmentations')
axs[i].set_title(model_name_dict[model_name])
axs[i].get_legend().set_visible(False)
if with_APAC == 'with_APAC':
axs[i].set_ylim(50, 80)
axs[i].set_xticklabels([])
axs[i].set_xlabel('')
if i != 0:
axs[i].set_ylabel('')
else:
axs[i].set_ylabel('Top-1 Accuracy')
if i == 3 and with_APAC == 'with_APAC':
handles, labels = axs[i].get_legend_handles_labels()
labels = legend_order
axs[i].legend(handles[1:], labels[1:], loc='lower right', prop={'size': 8})
lines = axs[i].get_lines()
for i, line in enumerate(lines):
line.set_zorder(10 - i)
plt.subplots_adjust(left=.00, right=.99)
plt.savefig('./figs/ranking_' + with_APAC + '.pdf', filetype='pdf', bbox_inches='tight')
plt.clf()
beneficial_tta_performance()
harmful_tta_performance()
our_method_performance()
ranking_performance()