-
Notifications
You must be signed in to change notification settings - Fork 41
/
data_utils.py
185 lines (135 loc) · 5.84 KB
/
data_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
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
import torch
import numpy as np
from torch import optim
import math
import torch.nn.functional as F
def get_similarity_transform_matrix(scale, rotation, translation):
eye_matrix = torch.eye(3).type(scale.type()).to(scale.device)
# Scale transform
S = eye_matrix.clone()
S[0, 0] = scale
S[1, 1] = scale
# Rotation transform
R = eye_matrix.clone()
rotation = rotation.clamp(-math.pi / 2, math.pi)
rotation_cos = rotation.cos()
rotation_sin = rotation.sin()
R[0, 0] = rotation_cos
R[0, 1] = -rotation_sin
R[1, 0] = rotation_sin
R[1, 1] = rotation_cos
# Translation transform
T = eye_matrix.clone()
T[0, 2] = translation[0]
T[1, 2] = translation[1]
theta = (S @ R @ T)
theta = theta[:2]
return theta
def estimate_similarity_transform(source_points, target_points):
# Params
scale = torch.FloatTensor([1]).to(source_points.device).requires_grad_()
rotation = torch.FloatTensor([0]).to(source_points.device).requires_grad_()
translation = torch.FloatTensor([0, 0]).to(source_points.device).requires_grad_()
params = [scale, rotation, translation]
opt = optim.LBFGS(params)
transform_args = params
def closure():
opt.zero_grad()
transform_matrix = get_similarity_transform_matrix(*transform_args)
pred_aligned_points = source_points @ transform_matrix.transpose(0, 1)
loss = ((pred_aligned_points - target_points) ** 2).mean()
loss.backward()
return loss
for i in range(5):
opt.step(closure)
inv_theta = get_similarity_transform_matrix(*transform_args)
# Align input images using theta
eye_vector = torch.zeros(1, 3)
eye_vector[:, 2] = 1
eye_vector = eye_vector.to(source_points.device)
inv_theta_ = torch.cat([inv_theta, eye_vector], dim=0)
theta = inv_theta_.inverse()
theta = theta[:2]
inv_theta = inv_theta[:2]
transform = {
'theta': theta.detach().cpu().numpy(),
'inv_theta': inv_theta.detach().cpu().numpy(),
'scale': scale.detach().cpu().numpy(),
'rotation': rotation.detach().cpu().numpy(),
'translation': translation.detach().cpu().numpy()}
return transform
def calc_ffhq_alignment(lm, size=512, device='cpu'):
lm_chin = lm[0: 17] # left-right
lm_eyebrow_left = lm[17: 22] # left-right
lm_eyebrow_right = lm[22: 27] # left-right
lm_nose = lm[27: 31] # top-down
lm_nostrils = lm[31: 36] # top-down
lm_eye_left = lm[36: 42] # left-clockwise
lm_eye_right = lm[42: 48] # left-clockwise
lm_mouth_outer = lm[48: 60] # left-clockwise
lm_mouth_inner = lm[60: 68] # left-clockwise
# Calculate auxiliary vectors.
eye_left = np.mean(lm_eye_left, axis=0)
eye_right = np.mean(lm_eye_right, axis=0)
eye_avg = (eye_left + eye_right) * 0.5
eye_to_eye = eye_right - eye_left
mouth_left = lm_mouth_outer[0]
mouth_right = lm_mouth_outer[6]
mouth_avg = (mouth_left + mouth_right) * 0.5
eye_to_mouth = mouth_avg - eye_avg
# Choose oriented crop rectangle.
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
x /= np.hypot(*x)
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
y = np.flipud(x) * [-1, 1]
c = eye_avg + eye_to_mouth * 0.1
bbox = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
bbox = (torch.from_numpy(bbox).float() / size - 0.5) * 2
bbox = torch.cat([bbox, torch.ones(4, 1)], dim=1)
gt_bbox = torch.FloatTensor([[-1, -1], [-1, 1], [1, 1], [1, -1]])
bbox = bbox.to(device)
gt_bbox = gt_bbox.to(device)
return estimate_similarity_transform(bbox, gt_bbox)
def tensor2image(tensor):
image = tensor.detach().cpu().numpy()
image = image * 255.
image = np.maximum(np.minimum(image, 255), 0)
image = image.transpose(1, 2, 0)[:, :, [2, 1, 0]]
return image.astype(np.uint8).copy()
def preprocess_dict(data_dict, fa, image_size, align_source, align_target, align_scale, device):
image_size_ = data_dict['source_img'].shape[-1]
image_size = image_size
imgs = data_dict['source_img'].cpu()
masks = data_dict['source_mask'].cpu()
lm_2d = data_dict['source_keypoints'][0].detach().cpu().numpy()
transform_ffhq = calc_ffhq_alignment(lm_2d, size=imgs.shape[2])
theta = torch.FloatTensor(transform_ffhq['theta'])[None]
if align_source:
grid = torch.linspace(-1, 1, image_size)
v, u = torch.meshgrid(grid, grid)
identity_grid = torch.stack([u, v, torch.ones_like(u)], dim=2).view(1, -1, 3)
if align_source or align_target:
# Align input images using theta
eye_vector = torch.zeros(theta.shape[0], 1, 3)
eye_vector[:, :, 2] = 1
theta_ = torch.cat([theta, eye_vector], dim=1).float()
# Perform 2x zoom-in compared to default theta
scale = torch.zeros_like(theta_)
scale[:, [0, 1], [0, 1]] = align_scale
scale[:, 2, 2] = 1
theta_ = torch.bmm(theta_, scale)[:, :2]
align_warp = identity_grid.repeat_interleave(theta_.shape[0], dim=0)
align_warp = align_warp.bmm(theta_.transpose(1, 2)).view(theta_.shape[0], image_size, image_size, 2)
if align_source:
source_imgs_aligned = F.grid_sample(imgs, align_warp)
source_masks_aligned = F.grid_sample(masks, align_warp)
source_keypoints = torch.from_numpy(fa.get_landmarks_from_image(tensor2image(source_imgs_aligned[0]))[0])[
None]
output_data_dict = {
'source_img': source_imgs_aligned[0].to(device) if align_source else
F.interpolate(imgs, size=image_size, mode='bilinear')[0],
'source_mask': source_masks_aligned[0].to(device) if align_source else
F.interpolate(masks, size=image_size, mode='bilinear')[0],
'source_keypoints': (source_keypoints.to(device) / (image_size / 2) - 1)[0],
}
return output_data_dict