-
Notifications
You must be signed in to change notification settings - Fork 47
/
generate_waterbirds.py
185 lines (156 loc) · 6.91 KB
/
generate_waterbirds.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
import os
import numpy as np
import random
import pandas as pd
from PIL import Image
from tqdm import tqdm
from dataset_utils import crop_and_resize, combine_and_mask
################ Paths and other configs - Set these #################################
cub_dir = '/u/scr/nlp/CUB_200_2011'
places_dir = '/u/scr/nlp/places365'
output_dir = '/u/scr/nlp/dro/cub/data'
dataset_name = 'waterbird_complete95_forest2water2'
target_places = [
['bamboo_forest', 'forest/broadleaf'], # Land backgrounds
['ocean', 'lake/natural']] # Water backgrounds
val_frac = 0.2 # What fraction of the training data to use as validation
confounder_strength = 0.95 # Determines relative size of majority vs. minority groups
######################################################################################
images_path = os.path.join(cub_dir, 'images.txt')
df = pd.read_csv(
images_path,
sep=" ",
header=None,
names=['img_id', 'img_filename'],
index_col='img_id')
### Set up labels of waterbirds vs. landbirds
# We consider water birds = seabirds and waterfowl.
species = np.unique([img_filename.split('/')[0].split('.')[1].lower() for img_filename in df['img_filename']])
water_birds_list = [
'Albatross', # Seabirds
'Auklet',
'Cormorant',
'Frigatebird',
'Fulmar',
'Gull',
'Jaeger',
'Kittiwake',
'Pelican',
'Puffin',
'Tern',
'Gadwall', # Waterfowl
'Grebe',
'Mallard',
'Merganser',
'Guillemot',
'Pacific_Loon'
]
water_birds = {}
for species_name in species:
water_birds[species_name] = 0
for water_bird in water_birds_list:
if water_bird.lower() in species_name:
water_birds[species_name] = 1
species_list = [img_filename.split('/')[0].split('.')[1].lower() for img_filename in df['img_filename']]
df['y'] = [water_birds[species] for species in species_list]
### Assign train/tesst/valid splits
# In the original CUB dataset split, split = 0 is test and split = 1 is train
# We want to change it to
# split = 0 is train,
# split = 1 is val,
# split = 2 is test
train_test_df = pd.read_csv(
os.path.join(cub_dir, 'train_test_split.txt'),
sep=" ",
header=None,
names=['img_id', 'split'],
index_col='img_id')
df = df.join(train_test_df, on='img_id')
test_ids = df.loc[df['split'] == 0].index
train_ids = np.array(df.loc[df['split'] == 1].index)
val_ids = np.random.choice(
train_ids,
size=int(np.round(val_frac * len(train_ids))),
replace=False)
df.loc[train_ids, 'split'] = 0
df.loc[val_ids, 'split'] = 1
df.loc[test_ids, 'split'] = 2
### Assign confounders (place categories)
# Confounders are set up as the following:
# Y = 0, C = 0: confounder_strength
# Y = 0, C = 1: 1 - confounder_strength
# Y = 1, C = 0: 1 - confounder_strength
# Y = 1, C = 1: confounder_strength
df['place'] = 0
train_ids = np.array(df.loc[df['split'] == 0].index)
val_ids = np.array(df.loc[df['split'] == 1].index)
test_ids = np.array(df.loc[df['split'] == 2].index)
for split_idx, ids in enumerate([train_ids, val_ids, test_ids]):
for y in (0, 1):
if split_idx == 0: # train
if y == 0:
pos_fraction = 1 - confounder_strength
else:
pos_fraction = confounder_strength
else:
pos_fraction = 0.5
subset_df = df.loc[ids, :]
y_ids = np.array((subset_df.loc[subset_df['y'] == y]).index)
pos_place_ids = np.random.choice(
y_ids,
size=int(np.round(pos_fraction * len(y_ids))),
replace=False)
df.loc[pos_place_ids, 'place'] = 1
for split, split_label in [(0, 'train'), (1, 'val'), (2, 'test')]:
print(f"{split_label}:")
split_df = df.loc[df['split'] == split, :]
print(f"waterbirds are {np.mean(split_df['y']):.3f} of the examples")
print(f"y = 0, c = 0: {np.mean(split_df.loc[split_df['y'] == 0, 'place'] == 0):.3f}, n = {np.sum((split_df['y'] == 0) & (split_df['place'] == 0))}")
print(f"y = 0, c = 1: {np.mean(split_df.loc[split_df['y'] == 0, 'place'] == 1):.3f}, n = {np.sum((split_df['y'] == 0) & (split_df['place'] == 1))}")
print(f"y = 1, c = 0: {np.mean(split_df.loc[split_df['y'] == 1, 'place'] == 0):.3f}, n = {np.sum((split_df['y'] == 1) & (split_df['place'] == 0))}")
print(f"y = 1, c = 1: {np.mean(split_df.loc[split_df['y'] == 1, 'place'] == 1):.3f}, n = {np.sum((split_df['y'] == 1) & (split_df['place'] == 1))}")
### Assign places to train, val, and test set
place_ids_df = pd.read_csv(
os.path.join(places_dir, 'categories_places365.txt'),
sep=" ",
header=None,
names=['place_name', 'place_id'],
index_col='place_id')
target_place_ids = []
for idx, target_places in enumerate(target_places):
place_filenames = []
for target_place in target_places:
target_place_full = f'/{target_place[0]}/{target_place}'
assert (np.sum(place_ids_df['place_name'] == target_place_full) == 1)
target_place_ids.append(place_ids_df.index[place_ids_df['place_name'] == target_place_full][0])
print(f'train category {idx} {target_place_full} has id {target_place_ids[idx]}')
# Read place filenames associated with target_place
place_filenames += [
f'/{target_place[0]}/{target_place}/{filename}' for filename in os.listdir(
os.path.join(places_dir, 'data_large', target_place[0], target_place))
if filename.endswith('.jpg')]
random.shuffle(place_filenames)
# Assign each filename to an image
indices = (df.loc[:, 'place'] == idx)
assert len(place_filenames) >= np.sum(indices),\
f"Not enough places ({len(place_filenames)}) to fit the dataset ({np.sum(df.loc[:, 'place'] == idx)})"
df.loc[indices, 'place_filename'] = place_filenames[:np.sum(indices)]
### Write dataset to disk
output_subfolder = os.path.join(output_dir, dataset_name)
os.makedirs(output_subfolder, exist_ok=True)
df.to_csv(os.path.join(output_subfolder, 'metadata.csv'))
for i in tqdm(df.index):
# Load bird image and segmentation
img_path = os.path.join(cub_dir, 'images', df.loc[i, 'img_filename'])
seg_path = os.path.join(cub_dir, 'segmentations', df.loc[i, 'img_filename'].replace('.jpg','.png'))
img_np = np.asarray(Image.open(img_path).convert('RGB'))
seg_np = np.asarray(Image.open(seg_path).convert('RGB')) / 255
# Load place background
# Skip front /
place_path = os.path.join(places_dir, 'data_large', df.loc[i, 'place_filename'][1:])
place = Image.open(place_path).convert('RGB')
img_black = Image.fromarray(np.around(img_np * seg_np).astype(np.uint8))
combined_img = combine_and_mask(place, seg_np, img_black)
output_path = os.path.join(output_subfolder, df.loc[i, 'img_filename'])
os.makedirs('/'.join(output_path.split('/')[:-1]), exist_ok=True)
combined_img.save(output_path)