-
Notifications
You must be signed in to change notification settings - Fork 16
/
predict_single.py
240 lines (196 loc) · 9.47 KB
/
predict_single.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
import argparse
import os
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"
os.environ["OMP_NUM_THREADS"] = "1"
from collections import namedtuple
import cv2
import numpy as np
import skimage.io
import torch
from albumentations.pytorch.transforms import img_to_tensor
from skimage import measure
from skimage.morphology import watershed
import models
from tools.config import load_config
ModelConfig = namedtuple("ModelConfig", "config_path weight_path type weight")
weight_path = "weights"
configs = [
ModelConfig("configs/d92_loc.json", "localization_dpn_unet_dpn92_0_best_dice", "localization", 1),
ModelConfig("configs/d92_loc.json", "localization_dpn_unet_dpn92_1_best_dice", "localization", 1),
ModelConfig("configs/d92_loc.json", "localization_dpn_unet_dpn92_2_best_dice", "localization", 1),
ModelConfig("configs/d92_loc.json", "localization_dpn_unet_dpn92_3_best_dice", "localization", 1),
ModelConfig("configs/d161_loc.json", "localization_densenet_unet_densenet161_3_0_best_dice", "localization", 1),
ModelConfig("configs/d161_loc.json", "localization_densenet_unet_densenet161_3_1_best_dice", "localization", 1),
ModelConfig("configs/d92_softmax.json", "softmax_dpn_seamese_unet_shared_dpn92_0_best_xview", "damage", 1),
ModelConfig("configs/d92_softmax.json", "softmax_dpn_seamese_unet_shared_dpn92_2_best_xview", "damage", 1),
ModelConfig("configs/d92_softmax.json", "pseudo_dpn_seamese_unet_shared_dpn92_0_best_xview", "damage", 1),
ModelConfig("configs/d92_softmax.json", "pseudo_dpn_seamese_unet_shared_dpn92_2_best_xview", "damage", 1),
ModelConfig("configs/d161_softmax.json", "softmax_densenet_seamese_unet_shared_densenet161_0_best_xview",
"damage", 1),
ModelConfig("configs/d161_softmax.json", "softmax_densenet_seamese_unet_shared_densenet161_2_best_xview",
"damage", 1),
ModelConfig("configs/d161_softmax.json", "pseudo_densenet_seamese_unet_shared_densenet161_0_best_xview",
"damage", 1),
ModelConfig("configs/d161_softmax.json", "pseudo_densenet_seamese_unet_shared_densenet161_2_best_xview",
"damage", 1),
ModelConfig("configs/se50_softmax.json", "pseudo_scseresnext_seamese_unet_shared_seresnext50_0_best_xview",
"damage", 1),
ModelConfig("configs/se50_softmax.json", "pseudo_scseresnext_seamese_unet_shared_seresnext50_1_best_xview",
"damage", 1),
ModelConfig("configs/se50_softmax.json", "pseudo_scseresnext_seamese_unet_shared_seresnext50_2_best_xview",
"damage", 1),
ModelConfig("configs/se50_softmax.json", "pseudo_scseresnext_seamese_unet_shared_seresnext50_3_best_xview",
"damage", 1),
ModelConfig("configs/b2_softmax.json",
"softmax_sampling_efficient_seamese_unet_shared_efficientnet-b2_0_best_xview", "damage", 1),
ModelConfig("configs/b2_softmax.json",
"softmax_sampling_efficient_seamese_unet_shared_efficientnet-b2_1_best_xview", "damage", 1),
ModelConfig("configs/r101_softmax_sgd.json", "sgd_resnext_seamese_unet_shared_resnext101_0_best_xview", "damage", 2)
]
def predict_localization(image, config: ModelConfig):
conf = load_config(config.config_path)
model = models.__dict__[conf['network']](seg_classes=1, backbone_arch=conf['encoder'])
checkpoint_path = os.path.join(weight_path, config.weight_path)
print("=> loading checkpoint '{}'".format(checkpoint_path))
checkpoint = torch.load(checkpoint_path, map_location="cpu")
model.load_state_dict({k.replace("module.", ""): v for k, v in checkpoint['state_dict'].items()})
model.eval()
model = model.cpu()
print("predicting", config)
with torch.no_grad():
image = img_to_tensor(image, conf["input"]["normalize"]).cpu().numpy()
if "dpn" in config.weight_path:
image = np.pad(image, [(0, 0), (16, 16), (16, 16)], mode='reflect')
images = np.array(
[
image,
image[:, ::-1, :],
image[:, :, ::-1],
image[:, ::-1, ::-1],
])
images = torch.from_numpy(images).cpu().float()
prediction_masks = []
for i in range(4):
logits = model(images[i:i + 1])
preds = torch.sigmoid(logits).cpu().numpy()
pred = preds[0]
if i == 1:
pred = preds[0].copy()[:, ::-1, :]
if i == 2:
pred = preds[0].copy()[:, :, ::-1]
if i == 3:
pred = preds[0].copy()[:, ::-1, ::-1]
prediction_masks.append(pred)
preds = np.average(prediction_masks, axis=0)
if "dpn" in config.weight_path:
preds = preds[:, 16:-16, 16:-16]
return preds
def predict_localization_ensemble(pre_path):
image = cv2.imread(pre_path, cv2.IMREAD_COLOR)[:, :, ::-1]
preds = []
for model_config in configs:
if model_config.type == "localization":
preds.append((predict_localization(image, model_config) * 255).astype(np.uint8))
return np.average(preds, axis=0)
def predict_damage_ensemble(pre_path, post_path):
image_pre = cv2.imread(pre_path, cv2.IMREAD_COLOR)[:, :, ::-1]
image_post = cv2.imread(post_path, cv2.IMREAD_COLOR)[:, :, ::-1]
image = np.concatenate([image_pre, image_post], axis=-1)
preds = []
for model_config in configs:
if model_config.type == "damage":
damage = (predict_damage(image, model_config) * 255).astype(np.uint8)
preds.append(damage)
if model_config.weight == 2:
preds.append(damage)
return np.average(preds, axis=0)
def predict_damage(image, config: ModelConfig):
conf = load_config(config.config_path)
model = models.__dict__[conf['network']](seg_classes=5, backbone_arch=conf['encoder'])
checkpoint_path = os.path.join(weight_path, config.weight_path)
print("=> loading checkpoint '{}'".format(checkpoint_path))
checkpoint = torch.load(checkpoint_path, map_location="cpu")
model.load_state_dict({k[7:]: v for k, v in checkpoint['state_dict'].items()})
model.eval()
model = model.cpu()
print("predicting", config)
with torch.no_grad():
image = img_to_tensor(image, conf["input"]["normalize"]).cpu().numpy()
images = np.array(
[
image,
image[:, ::-1, :],
image[:, :, ::-1],
image[:, ::-1, ::-1],
])
images = torch.from_numpy(images).cpu().float()
prediction_masks = []
for i in range(4):
logits = model(images[i:i + 1])
preds = torch.softmax(logits, dim=1).cpu().numpy()
pred = preds[0]
if i == 1:
pred = preds[0].copy()[:, ::-1, :]
if i == 2:
pred = preds[0].copy()[:, :, ::-1]
if i == 3:
pred = preds[0].copy()[:, ::-1, ::-1]
prediction_masks.append(pred)
preds = np.average(prediction_masks, axis=0)
return preds
def label_mask(loc, labels, intensity, mask, seed_threshold=0.8):
av_pred = 1 * (loc > seed_threshold)
av_pred = av_pred.astype(np.uint8)
y_pred = measure.label(av_pred, neighbors=8, background=0)
nucl_msk = (1 - loc)
nucl_msk = nucl_msk.astype('uint8')
y_pred = watershed(nucl_msk, y_pred, mask=mask, watershed_line=False)
props = measure.regionprops(y_pred)
for i in range(1, np.max(y_pred)):
reg_labels = labels[y_pred == i]
unique, counts = np.unique(reg_labels, return_counts=True)
max_idx = np.argmax(counts)
out_label = unique[max_idx]
if out_label > 0:
prop = props[i - 1]
if counts[max_idx] > 0.6 * sum(counts) \
and prop.eccentricity < 1.5 \
and prop.euler_number == 1:
labels[(y_pred == i) & (intensity < 0.6)] = out_label
return y_pred
def post_process(loc, damage, out_loc, out_damage):
localization = loc/255.
damage = damage/255.
first = np.zeros((1024, 1024, 1))
first[:, :, 0] = 1 - np.sum(damage, axis=2)
first[:, :, :] *= 0.8
damage_pred = np.concatenate([first, damage], axis=-1)
damage_pred[:, :, 2] *= 2.
damage_pred[:, :, 3] *= 2.
damage_pred[:, :, 4] *= 2.
argmax = np.argmax(damage_pred, axis=-1)
loc = 1 * ((localization > 0.25) | (argmax > 0))
argmax = np.argmax(damage, axis=-1) + 1
max = np.max(damage, axis=-1)
label_mask(localization, argmax, max, loc)
cv2.imwrite(out_loc, loc)
cv2.imwrite(out_damage, argmax)
def main():
parser = argparse.ArgumentParser("Xview Predictor")
arg = parser.add_argument
arg('--pre', type=str, help='Path to pre test image')
arg('--post', type=str, help='Path to post test image')
arg('--out-loc', type=str, help='Path to output localization image')
arg('--out-damage', type=str, help='Path to output damage image')
args = parser.parse_args()
localization = predict_localization_ensemble(args.pre)
skimage.io.imsave("_localization.png", localization[0, :, :].astype(np.uint8))
damage = predict_damage_ensemble(args.pre, args.post)
preds = (np.moveaxis(damage, 0, -1)).astype(np.uint8)
cv2.imwrite("_damage.png", cv2.cvtColor(preds[:, :, 1:], cv2.COLOR_RGBA2BGRA))
damage = skimage.io.imread("_damage.png")
localization = cv2.imread( "_localization.png", cv2.IMREAD_GRAYSCALE)
post_process(localization, damage, args.out_loc, args.out_damage)
if __name__ == '__main__':
main()