-
Notifications
You must be signed in to change notification settings - Fork 21
/
utils.py
171 lines (151 loc) · 6.61 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
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
import os
import cv2
import json
import torch
import tempfile
import subprocess
import numpy as np
import os.path as osp
from munch import munchify
import audio
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def read_wav(wav_path, fps=25, mel_step_size=16):
temp_audio_file = tempfile.NamedTemporaryFile(suffix=".wav")
if not wav_path.endswith('.wav'):
print('Extracting raw audio...')
audio_name = temp_audio_file.name
command = 'ffmpeg -i %s -loglevel error -y -f wav -acodec pcm_s16le -ar 16000 %s' % (wav_path, audio_name)
subprocess.call(command, shell=True)
wav_path = audio_name
wav = audio.load_wav(wav_path, 16000)
mel = audio.melspectrogram(wav)
if np.isnan(mel.reshape(-1)).sum() > 0:
raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
mel_chunks = []
mel_idx_multiplier = 80. / fps
i = 0
while 1:
start_idx = int(i * mel_idx_multiplier)
if start_idx + mel_step_size > len(mel[0]):
break
mel_chunks.append(mel[:, start_idx:start_idx + mel_step_size])
i += 1
temp_audio_file.close()
return mel_chunks, mel
def transformation_from_points(points1, points0, smooth=True, p_bias=None):
points2 = np.array(points0)
points2 = points2.astype(np.float64)
points1 = points1.astype(np.float64)
c1 = np.mean(points1, axis=0)
c2 = np.mean(points2, axis=0)
points1 -= c1
points2 -= c2
s1 = np.std(points1)
s2 = np.std(points2)
points1 /= s1
points2 /= s2
U, S, Vt = np.linalg.svd(np.matmul(points1.T, points2))
R = (np.matmul(U, Vt)).T
sR = (s2 / s1) * R
T = c2.reshape(2, 1) - (s2 / s1) * np.matmul(R, c1.reshape(2, 1))
M = np.concatenate((sR, T), axis=1)
if smooth:
bias = points2[2] - points1[2]
if p_bias is None:
p_bias = bias
else:
bias = p_bias * 0.2 + bias * 0.8
p_bias = bias
M[:, 2] = M[:, 2] + bias
return M, p_bias
class AlignRestore(object):
def __init__(self, align_points=3):
if align_points == 3:
self.upscale_factor = 1
self.crop_ratio = (2.8, 2.8)
self.face_template = np.array([[19 - 2, 30 - 10], [56 + 2, 30 - 10], [37.5, 45 - 5]])
self.face_template = self.face_template * 2.8
self.face_size = (int(75 * self.crop_ratio[0]), int(100 * self.crop_ratio[1]))
self.p_bias = None
def process(self, img, lmk_align=None, smooth=True, align_points=3):
aligned_face, affine_matrix = self.align_warp_face(img, lmk_align, smooth)
restored_img = self.restore_img(img, aligned_face, affine_matrix)
cv2.imwrite("restored.jpg", restored_img)
cv2.imwrite("aligned.jpg", aligned_face)
return aligned_face, restored_img
def align_warp_face(self, img, lmks3, smooth=True, border_mode='constant'):
affine_matrix, self.p_bias = transformation_from_points(lmks3, self.face_template, smooth, self.p_bias)
if border_mode == 'constant':
border_mode = cv2.BORDER_CONSTANT
elif border_mode == 'reflect101':
border_mode = cv2.BORDER_REFLECT101
elif border_mode == 'reflect':
border_mode = cv2.BORDER_REFLECT
cropped_face = cv2.warpAffine(img, affine_matrix, self.face_size, borderMode=border_mode, borderValue=[127, 127, 127])
return cropped_face, affine_matrix
def align_warp_face2(self, img, landmark, border_mode='constant'):
affine_matrix = cv2.estimateAffinePartial2D(landmark, self.face_template)[0]
if border_mode == 'constant':
border_mode = cv2.BORDER_CONSTANT
elif border_mode == 'reflect101':
border_mode = cv2.BORDER_REFLECT101
elif border_mode == 'reflect':
border_mode = cv2.BORDER_REFLECT
cropped_face = cv2.warpAffine(img, affine_matrix, self.face_size, borderMode=border_mode, borderValue=(135, 133, 132))
return cropped_face, affine_matrix
def restore_img(self, input_img, face, affine_matrix):
h, w, _, = input_img.shape
h_up, w_up = int(h * self.upscale_factor), int(w * self.upscale_factor)
upsample_img = cv2.resize(input_img, (w_up, h_up), interpolation=cv2.INTER_LANCZOS4)
inverse_affine = cv2.invertAffineTransform(affine_matrix)
inverse_affine *= self.upscale_factor
if self.upscale_factor > 1:
extra_offset = 0.5 * self.upscale_factor
else:
extra_offset = 0
inverse_affine[:, 2] += extra_offset
inv_restored = cv2.warpAffine(face, inverse_affine, (w_up, h_up))
mask = np.ones((self.face_size[1], self.face_size[0]), dtype=np.float32)
inv_mask = cv2.warpAffine(mask, inverse_affine, (w_up, h_up))
inv_mask_erosion = cv2.erode(inv_mask, np.ones((int(2 * self.upscale_factor), int(2 * self.upscale_factor)), np.uint8))
pasted_face = inv_mask_erosion[:, :, None] * inv_restored
total_face_area = np.sum(inv_mask_erosion)
w_edge = int(total_face_area**0.5) // 20
erosion_radius = w_edge * 2
inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
blur_size = w_edge * 2
inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
inv_soft_mask = inv_soft_mask[:, :, None]
upsample_img = inv_soft_mask * pasted_face + (1 - inv_soft_mask) * upsample_img
if np.max(upsample_img) > 256:
upsample_img = upsample_img.astype(np.uint16)
else:
upsample_img = upsample_img.astype(np.uint8)
return upsample_img
class laplacianSmooth(object):
def __init__(self, smoothAlpha=0.3):
self.smoothAlpha = smoothAlpha
self.pts_last = None
def smooth(self, pts_cur):
if self.pts_last is None:
self.pts_last = pts_cur.copy()
return pts_cur.copy()
x1 = min(pts_cur[:, 0])
x2 = max(pts_cur[:, 0])
y1 = min(pts_cur[:, 1])
y2 = max(pts_cur[:, 1])
width = x2 - x1
pts_update = []
for i in range(len(pts_cur)):
x_new, y_new = pts_cur[i]
x_old, y_old = self.pts_last[i]
tmp = (x_new - x_old)**2 + (y_new - y_old)**2
w = np.exp(-tmp / (width * self.smoothAlpha))
x = x_old * w + x_new * (1 - w)
y = y_old * w + y_new * (1 - w)
pts_update.append([x, y])
pts_update = np.array(pts_update)
self.pts_last = pts_update.copy()
return pts_update