forked from rishikksh20/ResUnet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_cd_4_channels.py
261 lines (204 loc) · 10.3 KB
/
eval_cd_4_channels.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
259
260
261
import os
import cv2
import torch
import argparse
import numpy as np
import albumentations as albums
import torchmetrics
from albumentations.pytorch import ToTensorV2
from core.siamese_resunet_cd_segmenation_based import SiameseResUnetSegmentationBased
from utils import metrics
from dataset.s2looking_allmask import S2LookingAllMask
from dataset.utils import *
from PIL import Image, ImagePalette
from torch.utils.data import DataLoader
from tqdm import tqdm
from utils.hparams import HParam
from utils.augmentation import *
from utils.images import *
from utils.hungarian import *
def main(hp, mode, weights, device, split, trained_path, saved_path, threshold=0.5, batch_size=8, save_sub_mask=False,
cm_weights=None):
if cm_weights is None:
cm_weights = [0.3, 0.7]
assert (0 <= mode < 3)
assert os.path.isfile(trained_path)
img1_save_path = os.path.join(saved_path, "img1")
img2_save_path = os.path.join(saved_path, "img2")
prob_img1_save_path = os.path.join(saved_path, "prob_img1")
prob_img2_save_path = os.path.join(saved_path, "prob_img2")
cd_save_path = os.path.join(saved_path, "cd")
hungarian_cd_save_path = os.path.join(saved_path, "hungarian_cd")
final_cd_path = os.path.join(saved_path, "final_cd")
if not os.path.exists(saved_path):
os.makedirs(saved_path)
if not os.path.exists(img1_save_path):
os.makedirs(img1_save_path)
if not os.path.exists(img2_save_path):
os.makedirs(img2_save_path)
if not os.path.exists(prob_img1_save_path):
os.makedirs(prob_img1_save_path)
if not os.path.exists(prob_img2_save_path):
os.makedirs(prob_img2_save_path)
if not os.path.exists(cd_save_path):
os.makedirs(cd_save_path)
if not os.path.exists(hungarian_cd_save_path):
os.makedirs(hungarian_cd_save_path)
if not os.path.exists(final_cd_path):
os.makedirs(final_cd_path)
model = SiameseResUnetSegmentationBased().to(device)
checkpoint = torch.load(trained_path, map_location=device)
model.load_segmentation_weight(checkpoint["state_dict"])
model.eval()
training_metrics = torchmetrics.MetricCollection(
{
"Dice": torchmetrics.Dice(average='none', num_classes=2),
},
prefix='test_'
)
if save_sub_mask:
n_masks = mode + 1
else:
n_masks = 1
augment_transform = albums.Compose([
albums.Resize(256, 256),
PerImageStandazation(),
ToTensorV2()
],
additional_targets={
'image0': 'image',
'mask1': 'mask',
'mask2': 'mask',
'border_mask1': 'mask',
'border_mask2': 'mask'
}
)
dataset = S2LookingAllMask(hp.cd_dset_dir, split, augment_transform=augment_transform)
dataloader = DataLoader(
dataset, batch_size=batch_size, num_workers=2, shuffle=False
)
cd_branch_acc = metrics.MetricTracker()
hungarian_branch_acc = metrics.MetricTracker()
final_acc = metrics.MetricTracker()
criterion = metrics.BCEDiceLoss(weight=[0.1, 0.9])
loader = tqdm(dataloader, desc="Evaluating")
nuclei_palette = ImagePalette.random()
img_height, img_width = dataset.get_full_resized_shape()
full_cm = np.zeros((img_height, img_width))
full_label = np.zeros((img_height, img_width))
full_x = np.zeros((2, img_height, img_width))
full_y = np.zeros((2, img_height, img_width))
full_x_probs = np.zeros((img_height, img_width), dtype=np.float64)
full_y_probs = np.zeros((img_height, img_width), dtype=np.float64)
full_cm_probs = np.zeros((img_height, img_width), dtype=np.float64)
with torch.no_grad():
for (idx, data) in enumerate(loader):
cd_i1 = data['x'].to(device)
cd_i2 = data['y'].to(device)
cd_labels = data['mask'].to(device)
outputs = model(cd_i1, cd_i2)
#cd_branch_acc.update(metrics.dice_coeff(outputs['cm'], cd_labels), outputs['cm'].size(0))
cm_probs = outputs['cm'].cpu().numpy()
x_probs = outputs['x'].cpu().numpy()
y_probs = outputs['y'].cpu().numpy()
cm = (cm_probs >= threshold) * 1
x = (x_probs >= threshold) * 1
y = (y_probs >= threshold) * 1
hg_probs = []
final_probs = []
for i in range(cm.shape[0]):
# Get file name
files = dataset.files[idx * batch_size + i]
divide = files['divide']
x1, x2, y1, y2 = dataset.get_resized_coord(divide)
filename = os.path.basename(files['image1'])[:-4]
full_x[:, x1:x2, y1:y2] = x[i, :, ...]
full_y[:, x1:x2, y1:y2] = y[i, :, ...]
full_cm[x1:x2, y1:y2] = cm[i, 0, ...]
full_x_probs[x1:x2, y1:y2] = x_probs[i, 0, ...]
full_y_probs[x1:x2, y1:y2] = y_probs[i, 0, ...]
full_cm_probs[x1:x2, y1:y2] = cm_probs[i, 0, ...]
if divide >= dataset.divide*dataset.divide - 1:
o_full_x_probs = cv2.resize(full_x_probs, (dataset.width, dataset.height), interpolation=cv2.INTER_LINEAR)
o_full_y_probs = cv2.resize(full_y_probs, (dataset.width, dataset.height), interpolation=cv2.INTER_LINEAR)
np.savez_compressed(os.path.join(prob_img1_save_path, "prob_{filename}".format(filename=filename)), a=o_full_x_probs)
np.savez_compressed(os.path.join(prob_img2_save_path, "prob_{filename}".format(filename=filename)), a=o_full_y_probs)
# Colorize instance segmentation map and save
masks1 = save_mask_and_contour(
full_x[0, ...], full_x[1, ...], nuclei_palette,
os.path.join(img1_save_path, "mask1_{filename}.png".format(filename=filename)),
(dataset.width, dataset.height)
).astype(int)
masks2 = save_mask_and_contour(
full_y[0, ...], full_y[1, ...], nuclei_palette,
os.path.join(img2_save_path, "mask2_{filename}.png".format(filename=filename)),
(dataset.width, dataset.height)
).astype(int)
masks1 = torch.from_numpy(masks1)
masks2 = torch.from_numpy(masks2)
# Hungarian algorithm
hg_map = change_detection_map(masks1, masks2, 1024, 1024)
hg_img = (hg_map * 255).astype(np.uint8)
# mask_color_1 = convert_to_color_map(masks1, w=1024, h=1024)
# mask_color_2 = convert_to_color_map(masks2, w=1024, h=1024)
# plot_and_save(mask_color_1, mask_color_2, hg_img, os.path.join(hungarian_cd_save_path, "{filename}.png".format(filename=filename)))
# cm_img = Image.fromarray(full_cm, mode='1')
# cm_img = np.asarray(cm_img.resize((dataset.width, dataset.height)))*1
gt_cd = (np.array(Image.open(files["mask"])) / 255.0).astype('int')
training_metrics(
target=torch.from_numpy(gt_cd),
preds=torch.from_numpy(hg_map.astype(np.uint8))
)
cd_branch_acc.update(
metrics.np_dice_coeff(hg_img[np.newaxis, :, :], gt_cd[np.newaxis, :, :]), 1)
# Save CM from CD branch
cv2.imwrite(
os.path.join(cd_save_path, "cd_{filename}.png".format(filename=filename)),
hg_img*255
)
# cm_im = Image.fromarray((full_cm * 255).astype(np.uint8), mode='P')
# cm_im = cm_im.resize((dataset.width, dataset.height))
# cm_im.save(os.path.join(cd_save_path, "cd_{filename}.png".format(filename=filename)))
# # Calculate final CM
# cm_x_probs = np.multiply(full_x_probs, hg_map)
# cm_y_probs = np.multiply(full_y_probs, hg_map)
#
# hg_prob = np.maximum(cm_x_probs, cm_y_probs)
# hg_probs.append(hg_prob)
#
# # final_prob = hg_prob * cm_weights[0] + cm_probs[i, 0, ...] * cm_weights[1]
# final_prob = np.maximum(hg_prob, full_cm_probs)
# final_probs.append(final_prob)
#
# final_map = (final_prob >= threshold) * 255
# final_map = Image.fromarray(final_map.astype(np.uint8), mode='P')
# final_map = final_map.resize((dataset.width, dataset.height))
# final_map.save(os.path.join(final_cd_path, "final_{filename}.png".format(filename=filename)))
#hg_probs = np.array(hg_probs)
#final_probs = np.array(final_probs)
#hungarian_branch_acc.update(metrics.np_dice_coeff((hg_probs >= threshold)*1, cd_labels.cpu().numpy()), hg_probs.shape[0])
#final_acc.update(metrics.np_dice_coeff((final_probs >= threshold)*1, cd_labels.cpu().numpy()), final_probs.shape[0])
values = training_metrics.compute()
print(values)
print("CD Branch dice: {:.4f}".format(cd_branch_acc.avg))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Validate ResUnet")
parser.add_argument("--mode", default=1)
parser.add_argument("--pretrain", type=str)
parser.add_argument("--savepath", type=str)
parser.add_argument("--threshold", default=0.5, type=float)
parser.add_argument("--batchsize", default=8, type=int)
parser.add_argument("--split", default="test", type=str)
parser.add_argument(
"-c", "--config", type=str, required=True, help="yaml file for configuration"
)
parser.add_argument("--device", default="cuda:0", type=str, help="Device ID")
args = parser.parse_args()
device = torch.device(args.device)
weights = [1.0]
if int(args.mode) == 1:
weights = [1.0, 0.1]
elif int(args.mode) == 2:
weights = [1.0, 0.1, 0.05]
hp = HParam(args.config)
main(hp, int(args.mode), weights, device, args.split, args.pretrain, args.savepath, args.threshold, args.batchsize)