-
Notifications
You must be signed in to change notification settings - Fork 2
/
data.py
178 lines (144 loc) · 6.39 KB
/
data.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
import numpy as np
import pandas as pd
import scipy.special
from tqdm import tqdm
from config import *
def get_dists(dist_type: str='facial'):
'''
dist_type
'''
dists_fnames = {
'facial': oj(DIR_PROCESSED, 'dists_pairwise_facial.npy'),
'vgg': oj(DIR_PROCESSED, 'dists_pairwise_vgg.npy'),
'gan': oj(DIR_PROCESSED, 'dists_pairwise_gan.npy'),
'facenet_vgg2': oj(DIR_PROCESSED, 'dists_pairwise_facial_facenet_vgg2.npy'),
'facenet_casia': oj(DIR_PROCESSED, 'dists_pairwise_facial_facenet_casia.npy'),
}
return np.load(open(dists_fnames[dist_type], 'rb'))
def load_all_labs(cached_file=oj(DIR_PROCESSED, 'df.pkl'),
dir_ims=oj(DIR_CELEBA, 'ims'),
celeba_id_fname=oj(DIR_CELEBA, 'Anno', 'identity_CelebA.txt'),
celeba_attr_fname=oj(DIR_CELEBA, 'Anno', 'list_attr_celeba.txt'),
mapping_file=oj(DIR_CELEBA, 'mapping.txt'),
race_preds_file=oj(DIR_PROCESSED, 'race.pkl'),
quality_scores_file=oj(DIR_PROCESSED, 'quality_scores.pkl'),
background_stats_file=oj(DIR_PROCESSED, 'background_stats.pkl'),
pose_file=oj(DIR_PROCESSED, 'pose.pkl')):
'''Load all labels for each image in celeba-hq. If not using the cached df,
have to precompute each of individual .pkl files using scripts in the
matching_banchmarking folder
'''
if os.path.exists(cached_file):
return pd.read_pickle(cached_file)
print('loading + processing labels...')
df = load_ids(dir_ims, celeba_id_fname, mapping_file)
labs, labs_full = load_labs(celeba_attr_fname, mapping_file)
# load in auxilary properties
race_pred_labs = pd.read_pickle(race_preds_file)
quality = pd.read_pickle(quality_scores_file)
background = pd.read_pickle(background_stats_file)
pose = pd.read_pickle(pose_file)
for k in labs.keys():
df[k] = labs[k].values
for k in labs_full.keys():
df[k] = labs_full[k].values
for k in race_pred_labs.keys():
df[k + '_pred'] = race_pred_labs[k].values
for k in background.keys():
df['background_' + k] = background[k].values
df['black_or_white'] = 2 * np.ones(df.shape[0]).astype(int)
df['black_or_white'].loc[df['race4_pred'] == 'Black'] = 0
df['black_or_white'].loc[df['race4_pred'] == 'White'] = 1
# process and add head pose angles
angles = np.array([ang for ang in range(66)])
for k_new, k_orig in zip(['yaw', 'pitch', 'roll'], ['yaws', 'pitches', 'rolls']):
vals = []
for i in tqdm(range(df.shape[0])):
arr = pose.iloc[i][k_orig].flatten()
preds = scipy.special.softmax(arr)
vals.append(np.sum(preds * angles) * 3 - 99)
df[k_new] = vals
df['quality'] = [s[0, 0] for s in quality['scores']]
df['fname_id'] = df['fname_final'].str.slice(stop=-4)
for i, race in enumerate(['White', 'Black', 'Asian', 'Indian']):
df[race + '_prob'] = [x[i] for x in df['race_scores_fair_4_pred'].values]
# clean up some labels
# remove id errors (can eventually move this into data.py)
# fix wrongly split ids
IDS_TO_MERGE = {
6329: 491 # same image, has different ids
}
for i in IDS_TO_MERGE:
df.loc[df.id == i, 'id'] = IDS_TO_MERGE[i]
# fix erroneous label
IDS_TO_FLIP_GENDER = [
8469
]
for i in IDS_TO_FLIP_GENDER:
df.loc[df.id == i, 'gender'] = 1
# replace value for some attributes by the mode over all images with this id
attrs = ['gender', 'race_pred', 'race4_pred']
for attr in attrs:
for i in tqdm(df.id.unique()):
idxs = df.id == i
mode = df[idxs][attr].mode().values[0] # get mode if there is disagreement
df.loc[idxs, attr] = mode
# remove some unnecesary keys
df.columns = df.columns.str.lower()
df = df[[k for k in df.keys()
if not 'md5' in k and not 'scores' in k
and not 'idx' in k and not 'orig_file' in k
and not 'img_names_pred' in k
and not 'face_name_align_pred' in k]]
df.keys()
# cache the dataframe
df.to_pickle(cached_file)
df.to_csv(cached_file[:-4] + '.csv')
return df
def load_labs(celeba_attr_fname, mapping_file, N_IMS=30000):
'''Load labels for celeba-hq
'''
remap = pd.read_csv(mapping_file, delim_whitespace=True)
labs_full = pd.read_csv(celeba_attr_fname, delim_whitespace=True, skiprows=1)
labs_full = labs_full.loc[[remap.iloc[i]['orig_file'] for i in range(N_IMS)]] # for i in range(labs_full.shape[0])]
labs_full = labs_full == 1
# print(labs_full.head())
labs = pd.DataFrame()
# print(labs.keys())
# print(labs_full.keys())
# large is more male
labs['gender'] = labs_full['Male']
# larger is longer
labs['hair-length'] = ~(labs_full['Bald'] | labs_full['Receding_Hairline']) # Bangs, Receding_Hairline
# larger is more
labs['facial-hair'] = ~(labs_full['No_Beard']) | labs_full['Mustache'] | labs_full['Goatee']
# labs_full['Mustache'] # Goatee, Mustache, No_Beard, 5_o_Clock_Shadow
# higher is more
labs['makeup'] = labs_full['Heavy_Makeup'] # | labs_full['Wearing_Lipstick'] # Wearing_Lipstick
# higher is darker
labs['skin-color'] = labs_full['Pale_Skin']
# older is more positive
labs['age'] = ~labs_full['Young']
# make into int
labs = labs.astype(int)
labs_full = labs_full.astype(int)
return labs, labs_full
def load_ids(dir_ims, celeba_id_fname, mapping_file):
'''Load IDs for celeba-hq
'''
ids_orig = pd.read_csv(celeba_id_fname, delim_whitespace=True, header=None)
ids_orig = ids_orig.rename(columns={0: 'orig_file', 1: 'id'})
remap = pd.read_csv(mapping_file, delim_whitespace=True)
# labels for celeb-a (not hq)
# (vals, counts) = np.unique(ids_orig.id.values, return_counts=True)
# plt.hist(counts)
# plt.xlabel('number of images with same id\nin celeb-a (hq has less)')
# plt.ylabel('count')
# plt.show()
fnames = sorted([f for f in os.listdir(oj(dir_ims)) if '.jpg' in f])
ids = remap.merge(ids_orig, on='orig_file', how='left')
ids['fname_final'] = fnames
(vals, counts) = np.unique(ids.id.values, return_counts=True)
id_to_count = {v: c for (v, c) in zip(vals, counts)}
ids['count_with_this_id'] = ids['id'].map(id_to_count)
return ids