/
image_preference_binary_classification_based_on_ranking.py
258 lines (195 loc) · 18.5 KB
/
image_preference_binary_classification_based_on_ranking.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
import os
import json
import pickle as pkl
import numpy as np
import pandas as pd
from scipy.optimize import minimize
from collections import defaultdict
import itertools
from config import *
def main():
# Exclude participants with systematically unreliable, unstable AU4 estimation
with open(os.path.join(prep_rootdir, 'AU4_excluded_participants.json'), 'r') as f:
excluded_participants = json.load(f)
filtered_user_id_list = [p for p in user_id_list if p not in excluded_participants]
# Load the data
dfs = []
for user_id in filtered_user_id_list:
df_reaction = pd.read_csv(os.path.join(prep_rootdir, 'preprocessed_reaction_data', f"{user_id}.csv"))
df_images = pd.read_csv(os.path.join(prep_rootdir, 'preprocessed_image_data', f"{user_id}.csv"))
df = pd.merge(df_reaction, df_images, on = ['session_index', 'image_index'], how='inner')
df['user_id'] = user_id
dfs.append(df)
df_reaction_all = pd.concat(dfs)
data_dict = defaultdict(list)
data_dict['user_id'] = filtered_user_id_list
basic_metrics = ['CLIP_score', 'Aesthetic_score', 'BLIP_score', 'ImageReward_score', 'PickScore', 'HPS_V2_score']
# Ensemble of the three pre-trained scoring models
granularity = 0.1
weights = np.arange(0, 1 + granularity, granularity)
all_weight_combinations = [[w1, w2, 1 - w1 - w2] for w1 in weights for w2 in weights if (w1 + w2) <= 1]
all_weight_combinations = np.array([combo for combo in all_weight_combinations if combo[2] >= 0])
for left_out_user_id in filtered_user_id_list:
df_reaction_train = df_reaction_all[df_reaction_all['user_id'] != left_out_user_id].copy()
df_reaction_test = df_reaction_all[df_reaction_all['user_id'] == left_out_user_id].copy()
for metric in basic_metrics:
train_mean, train_std = df_reaction_train[metric].mean(), df_reaction_train[metric].std()
df_reaction_train[metric] = (df_reaction_train[metric] - train_mean) / train_std
df_reaction_test[metric] = (df_reaction_test[metric] - train_mean) / train_std
df_train_metrics_data_1, df_train_metrics_data_2 = get_data_pair(df_reaction_train, basic_metrics + ['rank_score'])
df_test_metrics_data_1, df_test_metrics_data_2 = get_data_pair(df_reaction_test, basic_metrics + ['rank_score'])
def ensemble_scores_correctness(df_metrics_data_1, df_metrics_data_2, weights):
n_combinations = weights.shape[0]
scores_pairwise_diff = ((weights[:,0:1] * np.tile(df_metrics_data_1['ImageReward_score'].to_numpy(), [n_combinations, 1])) + (weights[:,1:2] * np.tile(df_metrics_data_1['PickScore'].to_numpy(), [n_combinations, 1])) + (weights[:,2:3] * np.tile(df_metrics_data_1['HPS_V2_score'].to_numpy(), [n_combinations, 1]))) - ((weights[:,0:1] * np.tile(df_metrics_data_2['ImageReward_score'].to_numpy(), [n_combinations, 1])) + (weights[:,1:2] * np.tile(df_metrics_data_2['PickScore'].to_numpy(), [n_combinations, 1])) + (weights[:,2:3] * np.tile(df_metrics_data_2['HPS_V2_score'].to_numpy(), [n_combinations, 1])))
ranks_pairwise_diff = df_metrics_data_1['rank_score'].to_numpy() - df_metrics_data_2['rank_score'].to_numpy()
correctness = ranks_pairwise_diff * scores_pairwise_diff > 0
n_correct, n_incorrect = np.sum(correctness, axis=1) // 2, np.sum(~correctness, axis=1) // 2
return n_correct, n_incorrect
train_n_correct, train_n_incorrect = ensemble_scores_correctness(df_train_metrics_data_1, df_train_metrics_data_2, all_weight_combinations)
test_n_correct, test_n_incorrect = ensemble_scores_correctness(df_test_metrics_data_1, df_test_metrics_data_2, all_weight_combinations)
best_idx = np.argmax(train_n_correct)
best_w1, best_w2, best_w3 = all_weight_combinations[best_idx]
data_dict['ImageReward_weight'].append(best_w1)
data_dict['PickScore_weight'].append(best_w2)
data_dict['HPS_V2_weight'].append(best_w3)
data_dict['ensemble_baseline_score_n_correct'].append(test_n_correct[best_idx])
data_dict['ensemble_baseline_score_n_incorrect'].append(test_n_incorrect[best_idx])
print(f"Ensemble baseline score optimization... User {left_out_user_id} complete...")
# Image preference classification based on AUcomb valence score on only the selected data
exp_decay_coef_list = np.arange(0+0.1, 2.1+0.1, 0.1)
AU4_weight_list = np.arange(0, 1 + 0.1, 0.1)
diff_threshold_list = np.arange(0.0, 1.0+0.02, 0.02)
AUcomb_sel_classification_coefs_list = np.array(list(itertools.product(exp_decay_coef_list, AU4_weight_list, diff_threshold_list)))
for user_id_idx, left_out_user_id in enumerate(filtered_user_id_list):
df_reaction_train = df_reaction_all[df_reaction_all['user_id'] != left_out_user_id].copy()
df_reaction_test = df_reaction_all[df_reaction_all['user_id'] == left_out_user_id].copy()
for metric in basic_metrics:
train_mean, train_std = df_reaction_train[metric].mean(), df_reaction_train[metric].std()
df_reaction_train[metric] = (df_reaction_train[metric] - train_mean) / train_std
df_reaction_test[metric] = (df_reaction_test[metric] - train_mean) / train_std
df_reaction_train['ensemble_baseline_score'] = np.vectorize(lambda u: data_dict['ImageReward_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['ImageReward_score'].to_numpy() + np.vectorize(lambda u: data_dict['PickScore_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['PickScore'].to_numpy() + np.vectorize(lambda u: data_dict['HPS_V2_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['HPS_V2_score'].to_numpy()
df_reaction_test['ensemble_baseline_score'] = np.vectorize(lambda u: data_dict['ImageReward_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['ImageReward_score'].to_numpy() + np.vectorize(lambda u: data_dict['PickScore_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['PickScore'].to_numpy() + np.vectorize(lambda u: data_dict['HPS_V2_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['HPS_V2_score'].to_numpy()
df_train_metrics_data_1, df_train_metrics_data_2 = get_data_pair(df_reaction_train, basic_metrics + ['rank_score', 'ensemble_baseline_score', 'AU4_activation_value', 'AU12_activation_value'])
df_test_metrics_data_1, df_test_metrics_data_2 = get_data_pair(df_reaction_test, basic_metrics + ['rank_score', 'ensemble_baseline_score', 'AU4_activation_value', 'AU12_activation_value'])
def AUcomb_valence_scores_correctness(df_metrics_data_1, df_metrics_data_2, coefs):
n_combinations = coefs.shape[0]
exp_decay_coefs = coefs[:,0:1]
AU4_coefs = coefs[:,1:2]
AU12_coefs = 1 - AU4_coefs
diff_thresholds = coefs[:,2:3]
AU4_scores = - (1 - np.exp(-exp_decay_coefs * np.tile(df_metrics_data_1['AU4_activation_value'].to_numpy(), [n_combinations, 1])))
AU12_scores = (1 - np.exp(-exp_decay_coefs * np.tile(df_metrics_data_1['AU12_activation_value'].to_numpy(), [n_combinations, 1])))
data1_scores = AU4_coefs * AU4_scores + AU12_coefs * AU12_scores
AU4_scores = - (1 - np.exp(-exp_decay_coefs * np.tile(df_metrics_data_2['AU4_activation_value'].to_numpy(), [n_combinations, 1])))
AU12_scores = (1 - np.exp(-exp_decay_coefs * np.tile(df_metrics_data_2['AU12_activation_value'].to_numpy(), [n_combinations, 1])))
data2_scores = AU4_coefs * AU4_scores + AU12_coefs * AU12_scores
scores_pairwise_diff = data1_scores - data2_scores
ranks_pairwise_diff = df_metrics_data_1['rank_score'].to_numpy() - df_metrics_data_2['rank_score'].to_numpy()
compare_idx = np.abs(scores_pairwise_diff) >= diff_thresholds
all_metrics_n_correct = {}
all_metrics_n_incorrect = {}
for metric in (basic_metrics + ['ensemble_baseline_score', 'AUcomb_valence_score']):
if metric == 'AUcomb_valence_score':
metric_scores_pairwise_diff = scores_pairwise_diff
correctness = ranks_pairwise_diff * metric_scores_pairwise_diff > 0
n_correct, n_incorrect = np.sum(np.logical_and(correctness, compare_idx), axis=1) // 2, np.sum(np.logical_and(~correctness, compare_idx), axis=1) // 2
else:
metric_scores_pairwise_diff = (df_metrics_data_1[metric] - df_metrics_data_2[metric]).to_numpy()
correctness = ranks_pairwise_diff * metric_scores_pairwise_diff > 0
n_correct, n_incorrect = np.sum(np.logical_and(np.tile(correctness, [n_combinations, 1]), compare_idx), axis=1) // 2, np.sum(np.logical_and(np.tile(~correctness, [n_combinations, 1]), compare_idx), axis=1) // 2
all_metrics_n_correct[metric] = n_correct
all_metrics_n_incorrect[metric] = n_incorrect
return all_metrics_n_correct, all_metrics_n_incorrect
all_metrics_train_n_correct, all_metrics_train_n_incorrect = AUcomb_valence_scores_correctness(df_train_metrics_data_1, df_train_metrics_data_2, AUcomb_sel_classification_coefs_list)
all_metrics_test_n_correct, all_metrics_test_n_incorrect = AUcomb_valence_scores_correctness(df_test_metrics_data_1, df_test_metrics_data_2, AUcomb_sel_classification_coefs_list)
best_idx = np.argmax(all_metrics_train_n_correct['AUcomb_valence_score'] - 2 * all_metrics_train_n_incorrect['AUcomb_valence_score'])
best_exp_decay, best_AU4_coef, best_diff_threshold = AUcomb_sel_classification_coefs_list[best_idx]
best_AU12_coef = 1 - best_AU4_coef
data_dict['exp_decay_coef'].append(best_exp_decay)
data_dict['AU4_coef'].append(best_AU4_coef)
data_dict['AU12_coef'].append(best_AU12_coef)
data_dict['diff_threshold'].append(best_diff_threshold)
for metric in (basic_metrics + ['ensemble_baseline_score', 'AUcomb_valence_score']):
data_dict[f"AUcomb_selected_{metric}_n_correct"].append(all_metrics_test_n_correct[metric][best_idx])
data_dict[f"AUcomb_selected_{metric}_n_incorrect"].append(all_metrics_test_n_incorrect[metric][best_idx])
print(f"AUcomb valence score optimization... User {left_out_user_id} complete...")
df_reaction_all['AU4_valence_score'] = - (1 - np.exp(-np.vectorize(lambda u: data_dict['exp_decay_coef'][data_dict['user_id'].index(u)])(df_reaction_all['user_id']) * df_reaction_all['AU4_activation_value'].to_numpy()))
df_reaction_all['AU12_valence_score'] = (1 - np.exp(-np.vectorize(lambda u: data_dict['exp_decay_coef'][data_dict['user_id'].index(u)])(df_reaction_all['user_id']) * df_reaction_all['AU12_activation_value'].to_numpy()))
df_reaction_all['AUcomb_valence_score'] = np.vectorize(lambda u: data_dict['AU4_coef'][data_dict['user_id'].index(u)])(df_reaction_all['user_id']) * df_reaction_all['AU12_valence_score'] * df_reaction_all['AU4_valence_score'] + np.vectorize(lambda u: data_dict['AU12_coef'][data_dict['user_id'].index(u)])(df_reaction_all['user_id']) * df_reaction_all['AU12_valence_score']
# Compute the performance of all scores
for user_id in filtered_user_id_list:
df_participant = df_reaction_all[df_reaction_all['user_id'] == user_id].copy()
participant_scores_pairwise_diff = defaultdict(list)
for session in np.unique(df_participant['session_index']):
df_session = df_participant[df_participant['session_index'] == session].copy()
scores_pairwise_diff = {}
for metric in (basic_metrics + ['rank_score', 'AUcomb_valence_score']):
scores = df_session[metric].to_numpy()
scores_pairwise_diff[metric] = (scores[None,:] - scores[:,None]).flatten()
compare_idx = np.where(scores_pairwise_diff['rank_score'] != 0)[0]
for metric in (basic_metrics + ['rank_score', 'AUcomb_valence_score']):
scores_pairwise_diff[metric] = scores_pairwise_diff[metric][compare_idx]
participant_scores_pairwise_diff[metric].extend(scores_pairwise_diff[metric])
for metric in basic_metrics + ['AUcomb_valence_score']:
correctness = np.array(participant_scores_pairwise_diff['rank_score']) * np.array(participant_scores_pairwise_diff[metric]) > 0
n_correct, n_incorrect = np.sum(correctness) // 2, np.sum(~correctness) // 2
data_dict[f"{metric}_n_correct"].append(n_correct)
data_dict[f"{metric}_n_incorrect"].append(n_incorrect)
# Integrate AUcomb valence score with the pre-trained scoring models
AUcomb_coef_list = np.arange(0, 10.0+0.1, 0.1)
for (baseline_score, ensemble_score) in [('ImageReward_score', 'ensemble_ImageReward_AUcomb_score'), ('PickScore', 'ensemble_PickScore_AUcomb_score'), ('HPS_V2_score', 'ensemble_HPS_V2_AUcomb_score'), ('ensemble_baseline_score', 'ensemble_all_score')]:
for user_id_idx, left_out_user_id in enumerate(filtered_user_id_list):
df_reaction_train = df_reaction_all[df_reaction_all['user_id'] != left_out_user_id].copy()
df_reaction_test = df_reaction_all[df_reaction_all['user_id'] == left_out_user_id].copy()
for metric in basic_metrics:
train_mean, train_std = df_reaction_train[metric].mean(), df_reaction_train[metric].std()
df_reaction_train[metric] = (df_reaction_train[metric] - train_mean) / train_std
df_reaction_test[metric] = (df_reaction_test[metric] - train_mean) / train_std
df_reaction_train['ensemble_baseline_score'] = np.vectorize(lambda u: data_dict['ImageReward_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['ImageReward_score'].to_numpy() + np.vectorize(lambda u: data_dict['PickScore_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['PickScore'].to_numpy() + np.vectorize(lambda u: data_dict['HPS_V2_weight'][data_dict['user_id'].index(u)])(df_reaction_train['user_id']) * df_reaction_train['HPS_V2_score'].to_numpy()
df_reaction_test['ensemble_baseline_score'] = np.vectorize(lambda u: data_dict['ImageReward_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['ImageReward_score'].to_numpy() + np.vectorize(lambda u: data_dict['PickScore_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['PickScore'].to_numpy() + np.vectorize(lambda u: data_dict['HPS_V2_weight'][data_dict['user_id'].index(u)])(df_reaction_test['user_id']) * df_reaction_test['HPS_V2_score'].to_numpy()
df_train_metrics_data_1, df_train_metrics_data_2 = get_data_pair(df_reaction_train, basic_metrics + ['rank_score', 'ensemble_baseline_score', 'AUcomb_valence_score'])
df_test_metrics_data_1, df_test_metrics_data_2 = get_data_pair(df_reaction_test, basic_metrics + ['rank_score', 'ensemble_baseline_score', 'AUcomb_valence_score'])
def ensemble_scores_correctness(df_metrics_data_1, df_metrics_data_2, weights):
n_combinations = weights.shape[0]
scores_pairwise_diff = (np.tile(df_metrics_data_1[baseline_score].to_numpy(), [n_combinations, 1]) + np.expand_dims(weights, 1) * np.tile(df_metrics_data_1['AUcomb_valence_score'].to_numpy(), [n_combinations, 1])) - (np.tile(df_metrics_data_2[baseline_score].to_numpy(), [n_combinations, 1]) + np.expand_dims(weights, 1) * np.tile(df_metrics_data_2['AUcomb_valence_score'].to_numpy(), [n_combinations, 1]))
ranks_pairwise_diff = df_metrics_data_1['rank_score'].to_numpy() - df_metrics_data_2['rank_score'].to_numpy()
correctness = ranks_pairwise_diff * scores_pairwise_diff > 0
n_correct, n_incorrect = np.sum(correctness, axis=1) // 2, np.sum(~correctness, axis=1) // 2
return n_correct, n_incorrect
train_n_correct, train_n_incorrect = ensemble_scores_correctness(df_train_metrics_data_1, df_train_metrics_data_2, AUcomb_coef_list)
test_n_correct, test_n_incorrect = ensemble_scores_correctness(df_test_metrics_data_1, df_test_metrics_data_2, AUcomb_coef_list)
best_idx = np.argmax(train_n_correct)
best_AUcomb_coef = AUcomb_coef_list[best_idx]
data_dict[f'AUcomb_coef_for_{ensemble_score}'].append(best_AUcomb_coef)
data_dict[f'{ensemble_score}_n_correct'].append(test_n_correct[best_idx])
data_dict[f'{ensemble_score}_n_incorrect'].append(test_n_incorrect[best_idx])
print(f"{ensemble_score} optimization... User {left_out_user_id} complete...")
df_results = pd.DataFrame(data_dict)
os.makedirs(results_rootdir, exist_ok=True)
df_results.to_csv(os.path.join(results_rootdir, 'image_preference_binary_classification_based_on_ranking.csv'), index=False)
def get_data_pair(df_reaction, metrics):
all_metrics_data_list1 = defaultdict(list)
all_metrics_data_list2 = defaultdict(list)
for user_id in np.unique(df_reaction['user_id']):
df_participant = df_reaction[df_reaction['user_id'] == user_id].copy()
participant_scores_pairwise_diff = []
participant_ranks_pairwise_diff = []
for session in np.unique(df_participant['session_index']):
df_session = df_participant[df_participant['session_index'] == session]
metric_data_list1 = {}
metric_data_list2 = {}
for metric in metrics:
scores = df_session[metric].to_numpy()
scores_list1 = np.tile(scores[None,:], [scores.size, 1]).flatten()
scores_list2 = np.tile(scores[:,None], [1, scores.size]).flatten()
metric_data_list1[metric] = scores_list1
metric_data_list2[metric] = scores_list2
compare_idx = np.where(metric_data_list1['rank_score'] != metric_data_list2['rank_score'])[0]
for metric in metrics:
all_metrics_data_list1[metric].extend(metric_data_list1[metric][compare_idx])
all_metrics_data_list2[metric].extend(metric_data_list2[metric][compare_idx])
df_metrics_data_1 = pd.DataFrame(dict(all_metrics_data_list1))
df_metrics_data_2 = pd.DataFrame(dict(all_metrics_data_list2))
return df_metrics_data_1, df_metrics_data_2
if __name__ == '__main__':
main()