-
Notifications
You must be signed in to change notification settings - Fork 12
/
utils.py
130 lines (104 loc) · 4.55 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
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
import math
import random
import torch
import numpy as np
def box_muller_transform(x: torch.FloatTensor):
r"""Box-Muller transform"""
shape = x.shape
x = x.view(shape[:-1] + (-1, 2))
z = torch.zeros_like(x, device=x.device)
z[..., 0] = (-2 * x[..., 0].log()).sqrt() * (2 * np.pi * x[..., 1]).cos()
z[..., 1] = (-2 * x[..., 0].log()).sqrt() * (2 * np.pi * x[..., 1]).sin()
return z.view(shape)
def inv_box_muller_transform(z: torch.FloatTensor):
r"""Inverse Box-Muller transform"""
shape = z.shape
z = z.view(shape[:-1] + (-1, 2))
x = torch.zeros_like(z, device=z.device)
x[..., 0] = z.square().sum(dim=-1).div(-2).exp()
x[..., 1] = torch.atan2(z[..., 1], z[..., 0]).div(2 * np.pi).add(0.5)
return x.view(shape)
def generate_statistics_matrices(V):
r"""generate mean and covariance matrices from the network output."""
mu = V[:, :, 0:2]
sx = V[:, :, 2].exp()
sy = V[:, :, 3].exp()
corr = V[:, :, 4].tanh()
cov = torch.zeros(V.size(0), V.size(1), 2, 2, device=V.device)
cov[:, :, 0, 0] = sx * sx
cov[:, :, 0, 1] = corr * sx * sy
cov[:, :, 1, 0] = corr * sx * sy
cov[:, :, 1, 1] = sy * sy
return mu, cov
def compute_batch_metric(pred, gt):
"""Get ADE, FDE, TCC scores for each pedestrian"""
# Calculate ADEs and FDEs
temp = (pred - gt).norm(p=2, dim=-1)
ADEs = temp.mean(dim=1).min(dim=0)[0]
FDEs = temp[:, -1, :].min(dim=0)[0]
# Calculate TCCs
pred_best = pred[temp[:, -1, :].argmin(dim=0), :, range(pred.size(2)), :]
pred_gt_stack = torch.stack([pred_best, gt.permute(1, 0, 2)], dim=0)
pred_gt_stack = pred_gt_stack.permute(3, 1, 0, 2)
covariance = pred_gt_stack - pred_gt_stack.mean(dim=-1, keepdim=True)
factor = 1 / (covariance.shape[-1] - 1)
covariance = factor * covariance @ covariance.transpose(-1, -2)
variance = covariance.diagonal(offset=0, dim1=-2, dim2=-1)
stddev = variance.sqrt()
corrcoef = covariance / stddev.unsqueeze(-1) / stddev.unsqueeze(-2)
corrcoef = corrcoef.clamp(-1, 1)
corrcoef[torch.isnan(corrcoef)] = 0
TCCs = corrcoef[:, :, 0, 1].mean(dim=0)
return ADEs, FDEs, TCCs
def evaluate_tcc(pred, gt):
"""Get ADE, FDE, TCC scores for each pedestrian"""
pred, gt = torch.FloatTensor(pred).permute(1, 0, 2), torch.FloatTensor(gt).permute(1, 0, 2)
pred_best = pred
pred_gt_stack = torch.stack([pred_best.permute(1, 0, 2), gt.permute(1, 0, 2)], dim=0)
pred_gt_stack = pred_gt_stack.permute(3, 1, 0, 2)
covariance = pred_gt_stack - pred_gt_stack.mean(dim=-1, keepdim=True)
factor = 1 / (covariance.shape[-1] - 1)
covariance = factor * covariance @ covariance.transpose(-1, -2)
variance = covariance.diagonal(offset=0, dim1=-2, dim2=-1)
stddev = variance.sqrt()
corrcoef = covariance / stddev.unsqueeze(-1) / stddev.unsqueeze(-2)
corrcoef.clip_(-1, 1)
corrcoef[torch.isnan(corrcoef)] = 0
TCCs = corrcoef[:, :, 0, 1].mean(dim=0)
return TCCs
def data_sampler(V_obs, V_tr, A_obs=None, A_tr=None, scale=True, rotation=True, flip=True):
if scale:
V_obs, V_tr, A_obs, A_tr = random_scale(V_obs, V_tr, A_obs, A_tr)
if rotation:
V_obs, V_tr, A_obs, A_tr = random_rotation(V_obs, V_tr, A_obs, A_tr)
if flip:
V_obs, V_tr, A_obs, A_tr = random_flip(V_obs, V_tr, A_obs, A_tr)
return V_obs, V_tr, A_obs, A_tr
def random_scale(V_obs, V_tr, A_obs, A_tr, min=0.8, max=1.2):
scale = random.uniform(min, max)
V_obs[..., -2:] = V_obs[..., -2:] * scale
V_tr = V_tr * scale
return V_obs, V_tr, A_obs, A_tr
def random_rotation(V_obs, V_tr, A_obs, A_tr):
theta = random.uniform(-math.pi, math.pi)
theta = (theta // (math.pi / 2)) * (math.pi / 2)
r_mat = [[math.cos(theta), -math.sin(theta)],
[math.sin(theta), math.cos(theta)]]
r = torch.tensor(r_mat, dtype=torch.float, requires_grad=False).cuda()
V_obs[..., -2:] = torch.einsum('rc,ntvc->ntvr', r, V_obs[..., -2:])
V_tr = torch.einsum('rc,ntvc->ntvr', r, V_tr)
return V_obs, V_tr, A_obs, A_tr
def random_flip(V_obs, V_tr, A_obs, A_tr):
if random.random() > 0.5:
flip = torch.cat([V_obs[..., -2:], V_tr], dim=1)
flip = torch.flip(flip, dims=[1])
V_obs[..., -2:] = flip[:, :8]
V_tr = flip[:, 8:]
if A_obs is not None:
flip = torch.cat([A_obs, A_tr], dim=-3)
flip = torch.flip(flip, dims=[-3])
A_obs = flip[..., :8, :, :]
A_tr = flip[..., 8:, :, :]
return V_obs, V_tr, A_obs, A_tr
def count_parameters(model):
return sum(p.numel() for p in model.parameters())