/
ensemble.py
187 lines (158 loc) · 6.68 KB
/
ensemble.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
"""
Ensembling methods for object detection.
"""
import argparse
import os
import pandas as pd
import numpy as np
import pydicom
import PIL.Image as Image
from tqdm import tqdm
from object_detection.utils import visualization_utils as visutil
from collections import OrderedDict
from calculate_map import parse_submission, calculate_iou
"""
Ensemble - find overlapping boxes of the same class and average their positions
while adding their confidences. Can weigh different detectors with different weights.
No real learning here, although the weights and iou_thresh can be optimized.
Input:
- dets : List of detections. Each detection is all the output from one detector, and
should be a list of boxes, where each box should be on the format
[box_x, box_y, box_w, box_h, class, confidence] where box_x and box_y
are the center coordinates, box_w and box_h are width and height resp.
The values should be floats, except the class which should be an integer.
- iou_thresh: Threshold in terms of IOU where two boxes are considered the same,
if they also belong to the same class.
- weights: A list of weights, describing how much more some detectors should
be trusted compared to others. The list should be as long as the
number of detections. If this is set to None, then all detectors
will be considered equally reliable. The sum of weights does not
necessarily have to be 1.
Output:
A list of boxes, on the same format as the input. Confidences are in range 0-1.
"""
def ensemble(dets, conf_thresh=0.5, iou_thresh=0.1, weights=None):
assert(type(iou_thresh) == float)
ndets = len(dets)
if weights is None:
w = 1/float(ndets)
weights = [w]*ndets
else:
assert(len(weights) == ndets)
s = sum(weights)
for i in range(0, len(weights)):
weights[i] /= s
out = list()
used = {}
for idet in range(0,ndets):
det = dets[idet]
for box in det:
if tuple(box) in used:
continue
used[tuple(box)] = True
# Search the other detectors for overlapping box of same class
found = []
for iodet in range(0, ndets):
odet = dets[iodet]
if odet == det:
continue
bestbox = None
bestiou = iou_thresh
for obox in odet:
if not tuple(obox) in used:
# Not already used
if box[4] == obox[4]:
# Same class
iou = calculate_iou(box[:4], obox[:4])
if iou > bestiou:
bestiou = iou
bestbox = obox
if not bestbox is None:
w = weights[iodet]
found.append((bestbox,w))
used[tuple(bestbox)] = True
# Now we've gone through all other detectors
if len(found) == 0:
new_box = list(box)
new_box[5] /= ndets
if new_box[5] >= conf_thresh:
out.append(new_box)
else:
allboxes = [(box, weights[idet])]
allboxes.extend(found)
xc = 0.0
yc = 0.0
bw = 0.0
bh = 0.0
conf = 0.0
wsum = 0.0
for bb in allboxes:
w = bb[1]
wsum += w
b = bb[0]
xc += w*b[0]
yc += w*b[1]
bw += w*b[2]
bh += w*b[3]
conf += w*b[5]
xc /= wsum
yc /= wsum
bw /= wsum
bh /= wsum
new_box = [xc, yc, bw, bh, box[4], conf]
if new_box[5] >= conf_thresh:
out.append(new_box)
return out
if __name__=="__main__":
parser = argparse.ArgumentParser()
parser.add_argument('dicom_dir', help='path to the directory with DICOM images')
parser.add_argument('out_images_dir', help='path to the directory for the output images')
parser.add_argument('ensembled_submission', help='ensembled submission file')
parser.add_argument('submissions', nargs='+', help='subsmission files to ensemble')
parser.add_argument('--weights', nargs='+', type=float, help='weight for each submission', default=None)
parser.add_argument('--conf-thresh', type=float, help='confidence threshold. default 0.5', default=0.5)
parser.add_argument('--iou-thresh', type=float, help='iou threshold. default 0.1', default=0.1)
args = parser.parse_args()
if not args.submissions:
print('no submissions')
exit()
if not os.path.exists(args.out_images_dir):
os.mkdir(args.out_images_dir)
submissions = OrderedDict([(s, parse_submission(s)) for s in args.submissions])
patient_ids = list(submissions[args.submissions[0]].keys())
submit_dict = {'patientId': [], 'PredictionString': []}
for pid in tqdm(patient_ids):
dets = []
for _, parsed in submissions.items():
sub_dets = []
for box, score in zip(parsed[pid]['boxes'], parsed[pid]['scores']):
sub_dets.append(box + [1, score])
dets.append(sub_dets)
ensembled = ensemble(dets, conf_thresh=args.conf_thresh, iou_thresh=args.iou_thresh, weights=args.weights)
print_boxes = []
display_boxes = []
display_classes = []
display_scores = []
for e in ensembled:
x, y, w, h, c, conf = e
print_boxes.append('{0} {1} {2} {3} {4}'.format(conf, x, y, w, h))
display_boxes.append([y, x, y + h, x + w])
display_classes.append(c)
display_scores.append(conf)
submit_dict['patientId'].append(pid)
submit_dict['PredictionString'].append(' '.join(print_boxes))
im = pydicom.read_file(os.path.join(args.dicom_dir, pid + '.dcm'))
im = im.pixel_array
im = np.stack((im, ) * 3, -1)
if display_boxes:
visutil.visualize_boxes_and_labels_on_image_array(
im,
np.array(display_boxes),
display_classes,
display_scores,
{1: {'id': 1, 'name': 'pneumonia'}},
use_normalized_coordinates=False,
min_score_thresh=args.conf_thresh,
)
Image.fromarray(im).save(os.path.join(args.out_images_dir, pid + '.jpg'))
pd.DataFrame(submit_dict).to_csv(args.ensembled_submission, index=False)