-
Notifications
You must be signed in to change notification settings - Fork 175
/
utils.py
82 lines (61 loc) · 2.08 KB
/
utils.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
import torch
import numpy as np
import yaml
def get_configs(dataset):
data = yaml.load(open('data/dataset_cfg.yaml'))
return data[dataset]
def get_actionness_configs(dataset):
data = yaml.load(open('data/dataset_actionness_cfg.yaml'))
return data[dataset]
def get_reference_model_url(dataset, modality, init, arch):
data = yaml.load(open('data/reference_models.yaml'))
return data[dataset][init][arch][modality]
def get_grad_hook(name):
def hook(m, grad_in, grad_out):
print(len(grad_in), len(grad_out))
print((name, grad_out[0].data.abs().mean(), grad_in[0].data.abs().mean()))
print((grad_out[0].size()))
print((grad_in[0].size()))
print((grad_in[1].size()))
print((grad_in[2].size()))
# print((grad_out[0]))
# print((grad_in[0]))
return hook
def softmax(scores):
es = np.exp(scores - scores.max(axis=-1)[..., None])
return es / es.sum(axis=-1)[..., None]
def temporal_iou(span_A, span_B):
"""
Calculates the intersection over union of two temporal "bounding boxes"
span_A: (start, end)
span_B: (start, end)
"""
union = min(span_A[0], span_B[0]), max(span_A[1], span_B[1])
inter = max(span_A[0], span_B[0]), min(span_A[1], span_B[1])
if inter[0] >= inter[1]:
return 0
else:
return float(inter[1] - inter[0]) / float(union[1] - union[0])
def temporal_nms(bboxes, thresh):
"""
One-dimensional non-maximal suppression
:param bboxes: [[st, ed, score, ...], ...]
:param thresh:
:return:
"""
t1 = bboxes[:, 0]
t2 = bboxes[:, 1]
scores = bboxes[:, 2]
durations = t2 - t1
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
tt1 = np.maximum(t1[i], t1[order[1:]])
tt2 = np.minimum(t2[i], t2[order[1:]])
intersection = tt2 - tt1
IoU = intersection / (durations[i] + durations[order[1:]] - intersection).astype(float)
inds = np.where(IoU <= thresh)[0]
order = order[inds + 1]
return bboxes[keep, :]