forked from JunhongXu/planet-kaggle-pytorch
/
baseline_ensembles.py
199 lines (161 loc) · 6.1 KB
/
baseline_ensembles.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
import cv2
import numpy as np
import torch.nn as nn
import torch
from torch.utils.data import DataLoader
from torch.autograd import Variable
from data.kgdataset import KgForestDataset, toTensor
from torchvision.transforms import Normalize, Compose, Lambda
import glob
from planet_models.resnet_planet import resnet18_planet, resnet34_planet, resnet50_planet, resnet152_planet
from planet_models.densenet_planet import densenet161, densenet121, densenet169
from util import predict, f2_score, pred_csv
def default(imgs):
return imgs
def rotate90(imgs):
for index, img in enumerate(imgs):
imgs[index] = cv2.transpose(img, (1, 0, 2))
return imgs
def rotate180(imgs):
for index, img in enumerate(imgs):
imgs[index] = cv2.flip(img, -1)
return imgs
def rotate270(imgs):
for index, img in enumerate(imgs):
img = cv2.transpose(img, (1, 0, 2))
imgs[index] = cv2.flip(img, -1)
return imgs
def horizontalFlip(imgs):
for index, img in enumerate(imgs):
img = cv2.flip(img, 1)
imgs[index] = img
return imgs
def verticalFlip(imgs):
for index, img in enumerate(imgs):
img = cv2.flip(img, 0)
imgs[index] = img
return imgs
mean = [0.31151703, 0.34061992, 0.29885209]
std = [0.16730586, 0.14391145, 0.13747531]
threshold = [0.23166666666666666, 0.19599999999999998, 0.18533333333333335,
0.08033333333333334, 0.20199999999999999, 0.16866666666666666,
0.20533333333333334, 0.27366666666666667, 0.2193333333333333,
0.21299999999999999, 0.15666666666666665, 0.096666666666666679,
0.21933333333333335, 0.058666666666666673, 0.19033333333333333,
0.25866666666666666, 0.057999999999999996]
transforms = [default, rotate90, rotate180, rotate270, verticalFlip, horizontalFlip]
models = [# resnet18_planet, resnet34_planet, resnet50_planet, densenet121, densenet161, densenet169
resnet152_planet
]
# save probabilities to files for debug
def probs(dataloader):
"""
returns a numpy array of probabilities (n_transforms, n_models, n_imgs, 17)
use transforms to find the best threshold
use models to do ensemble method
"""
n_transforms = len(transforms)
n_models = len(models)
n_imgs = dataloader.dataset.num
imgs = dataloader.dataset.images.copy()
probabilities = np.empty((n_transforms, n_models, n_imgs, 17))
for t_idx, transform in enumerate(transforms):
t_name = str(transform).split()[1]
dataloader.dataset.images = transform(imgs)
for m_idx, model in enumerate(models):
name = str(model).split()[1]
net = model().cuda()
net = nn.DataParallel(net)
net.load_state_dict(torch.load('models/{}.pth'.format(name)))
net.eval()
# predict
m_predictions = predict(net, dataloader)
# save
np.savetxt(X=m_predictions, fname='probs/{}_{}.txt'.format(t_name, name))
probabilities[t_idx, m_idx] = m_predictions
return probabilities
def find_best_threshold(labels, probabilities):
threshold = np.zeros(17)
# iterate over transformations
for t_idx in range(len(transforms)):
# iterate over class labels
t = np.ones(17) * 0.15
selected_preds = probabilities[t_idx]
selected_preds = np.mean(selected_preds, axis=0)
best_thresh = 0.0
best_score = 0.0
for i in range(17):
for r in range(500):
r /= 500
t[i] = r
preds = (selected_preds > t).astype(int)
score = f2_score(labels, preds)
if score > best_score:
best_thresh = r
best_score = score
t[i] = best_thresh
print('Transform index {}, score {}, threshold {}, label {}'.format(t_idx, best_score, best_thresh, i))
print('Transform index {}, threshold {}, score {}'.format(t_idx, t, best_score))
threshold = threshold + t
threshold = threshold / len(transforms)
return threshold
def get_validation_loader():
validation = KgForestDataset(
split='validation-3000',
transform=Compose(
[
Lambda(lambda x: toTensor(x)),
Normalize(mean=mean, std=std)
]
),
height=256,
width=256
)
valid_dataloader = DataLoader(validation, batch_size=256, shuffle=False)
return valid_dataloader
def get_test_dataloader():
test_dataset = KgForestDataset(
split='test-61191',
transform=Compose(
[
Lambda(lambda x: toTensor(x)),
Normalize(mean=mean, std=std)
]
),
label_csv=None
)
test_dataloader = DataLoader(test_dataset, batch_size=512)
return test_dataloader
def do_thresholding(names, labels):
preds = np.empty((len(transforms), len(models), 3000, 17))
for t_idx in range(len(transforms)):
for m_idx in range(len(models)):
preds[t_idx, m_idx] = np.loadtxt(names[t_idx + m_idx])
print(names)
t = find_best_threshold(labels=labels, probabilities=preds)
return t
def get_files(exclude='resnet18'):
file_names = glob.glob('probs/*.txt')
file_names = [name for name in file_names if exclude not in name]
return file_names
def predict_test(t):
preds = np.zeros((61191, 17))
for index, model in enumerate(models):
name = str(model).split()[1]
net = nn.DataParallel(model().cuda())
net.eval()
net.load_state_dict(torch.load('models/{}.pth'.format(name)))
pred = predict(dataloader=test_dataloader, net=net)
preds = preds + pred
preds = preds/len(models)
pred_csv(predictions=preds, threshold=t, name='ensembles')
if __name__ == '__main__':
valid_dataloader = get_validation_loader()
test_dataloader = get_test_dataloader()
# save results to files
probabilities = probs(valid_dataloader)
# get threshold
# file_names = get_files()
# t = do_thresholding(file_names, valid_dataloader.dataset.labels)
# testing
predict_test(threshold)