-
Notifications
You must be signed in to change notification settings - Fork 0
/
face_align.py
515 lines (416 loc) · 18 KB
/
face_align.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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
"""
StarGAN v2
Copyright (c) 2020-present NAVER Corp.
This work is licensed under the Creative Commons Attribution-NonCommercial
4.0 International License. To view a copy of this license, visit
http://creativecommons.org/licenses/by-nc/4.0/ or send a letter to
Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
Lines (19 to 80) were adapted from https://github.com/1adrianb/face-alignment
Lines (83 to 235) were adapted from https://github.com/protossw512/AdaptiveWingLoss
"""
from collections import namedtuple
from copy import deepcopy
from functools import partial
from munch import Munch
import numpy as np
import cv2
from skimage.filters import gaussian
import torch
import torch.nn as nn
import torch.nn.functional as F
def get_preds_fromhm(hm):
max, idx = torch.max(
hm.view(hm.size(0), hm.size(1), hm.size(2) * hm.size(3)), 2)
idx += 1
preds = idx.view(idx.size(0), idx.size(1), 1).repeat(1, 1, 2).float()
preds[..., 0].apply_(lambda x: (x - 1) % hm.size(3) + 1)
preds[..., 1].add_(-1).div_(hm.size(2)).floor_().add_(1)
for i in range(preds.size(0)):
for j in range(preds.size(1)):
hm_ = hm[i, j, :]
pX, pY = int(preds[i, j, 0]) - 1, int(preds[i, j, 1]) - 1
if 0 < pX < 63 and 0 < pY < 63:
diff = torch.FloatTensor(
[hm_[pY, pX + 1] - hm_[pY, pX - 1],
hm_[pY + 1, pX] - hm_[pY - 1, pX]])
preds[i, j].add_(diff.sign_().mul_(.25))
preds.add_(-0.5)
return preds
class HourGlass(nn.Module):
def __init__(self, num_modules, depth, num_features, first_one=False):
super(HourGlass, self).__init__()
self.num_modules = num_modules
self.depth = depth
self.features = num_features
self.coordconv = CoordConvTh(64, 64, True, True, 256, first_one,
out_channels=256,
kernel_size=1, stride=1, padding=0)
self._generate_network(self.depth)
def _generate_network(self, level):
self.add_module('b1_' + str(level), ConvBlock(256, 256))
self.add_module('b2_' + str(level), ConvBlock(256, 256))
if level > 1:
self._generate_network(level - 1)
else:
self.add_module('b2_plus_' + str(level), ConvBlock(256, 256))
self.add_module('b3_' + str(level), ConvBlock(256, 256))
def _forward(self, level, inp):
up1 = inp
up1 = self._modules['b1_' + str(level)](up1)
low1 = F.avg_pool2d(inp, 2, stride=2)
low1 = self._modules['b2_' + str(level)](low1)
if level > 1:
low2 = self._forward(level - 1, low1)
else:
low2 = low1
low2 = self._modules['b2_plus_' + str(level)](low2)
low3 = low2
low3 = self._modules['b3_' + str(level)](low3)
up2 = F.interpolate(low3, scale_factor=2, mode='nearest')
return up1 + up2
def forward(self, x, heatmap):
x, last_channel = self.coordconv(x, heatmap)
return self._forward(self.depth, x), last_channel
class AddCoordsTh(nn.Module):
def __init__(self, height=64, width=64, with_r=False, with_boundary=False):
super(AddCoordsTh, self).__init__()
self.with_r = with_r
self.with_boundary = with_boundary
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
with torch.no_grad():
x_coords = torch.arange(height).unsqueeze(1).expand(height, width).float()
y_coords = torch.arange(width).unsqueeze(0).expand(height, width).float()
x_coords = (x_coords / (height - 1)) * 2 - 1
y_coords = (y_coords / (width - 1)) * 2 - 1
coords = torch.stack([x_coords, y_coords], dim=0) # (2, height, width)
if self.with_r:
rr = torch.sqrt(torch.pow(x_coords, 2) + torch.pow(y_coords, 2)) # (height, width)
rr = (rr / torch.max(rr)).unsqueeze(0)
coords = torch.cat([coords, rr], dim=0)
self.coords = coords.unsqueeze(0).to(device) # (1, 2 or 3, height, width)
self.x_coords = x_coords.to(device)
self.y_coords = y_coords.to(device)
def forward(self, x, heatmap=None):
"""
x: (batch, c, x_dim, y_dim)
"""
coords = self.coords.repeat(x.size(0), 1, 1, 1)
if self.with_boundary and heatmap is not None:
boundary_channel = torch.clamp(heatmap[:, -1:, :, :], 0.0, 1.0)
zero_tensor = torch.zeros_like(self.x_coords)
xx_boundary_channel = torch.where(boundary_channel > 0.05, self.x_coords, zero_tensor).to(zero_tensor.device)
yy_boundary_channel = torch.where(boundary_channel > 0.05, self.y_coords, zero_tensor).to(zero_tensor.device)
coords = torch.cat([coords, xx_boundary_channel, yy_boundary_channel], dim=1)
x_and_coords = torch.cat([x, coords], dim=1)
return x_and_coords
class CoordConvTh(nn.Module):
"""CoordConv layer as in the paper."""
def __init__(self, height, width, with_r, with_boundary,
in_channels, first_one=False, *args, **kwargs):
super(CoordConvTh, self).__init__()
self.addcoords = AddCoordsTh(height, width, with_r, with_boundary)
in_channels += 2
if with_r:
in_channels += 1
if with_boundary and not first_one:
in_channels += 2
self.conv = nn.Conv2d(in_channels=in_channels, *args, **kwargs)
def forward(self, input_tensor, heatmap=None):
ret = self.addcoords(input_tensor, heatmap)
last_channel = ret[:, -2:, :, :]
ret = self.conv(ret)
return ret, last_channel
class ConvBlock(nn.Module):
def __init__(self, in_planes, out_planes):
super(ConvBlock, self).__init__()
self.bn1 = nn.BatchNorm2d(in_planes)
conv3x3 = partial(nn.Conv2d, kernel_size=3, stride=1, padding=1, bias=False, dilation=1)
self.conv1 = conv3x3(in_planes, int(out_planes / 2))
self.bn2 = nn.BatchNorm2d(int(out_planes / 2))
self.conv2 = conv3x3(int(out_planes / 2), int(out_planes / 4))
self.bn3 = nn.BatchNorm2d(int(out_planes / 4))
self.conv3 = conv3x3(int(out_planes / 4), int(out_planes / 4))
self.downsample = None
if in_planes != out_planes:
self.downsample = nn.Sequential(nn.BatchNorm2d(in_planes),
nn.ReLU(True),
nn.Conv2d(in_planes, out_planes, 1, 1, bias=False))
def forward(self, x):
residual = x
out1 = self.bn1(x)
out1 = F.relu(out1, True)
out1 = self.conv1(out1)
out2 = self.bn2(out1)
out2 = F.relu(out2, True)
out2 = self.conv2(out2)
out3 = self.bn3(out2)
out3 = F.relu(out3, True)
out3 = self.conv3(out3)
out3 = torch.cat((out1, out2, out3), 1)
if self.downsample is not None:
residual = self.downsample(residual)
out3 += residual
return out3
class FAN(nn.Module):
def __init__(self, num_modules=1, end_relu=False, num_landmarks=98, fname_pretrained=None):
super(FAN, self).__init__()
self.num_modules = num_modules
self.end_relu = end_relu
# Base part
self.conv1 = CoordConvTh(256, 256, True, False,
in_channels=3, out_channels=64,
kernel_size=7, stride=2, padding=3)
self.bn1 = nn.BatchNorm2d(64)
self.conv2 = ConvBlock(64, 128)
self.conv3 = ConvBlock(128, 128)
self.conv4 = ConvBlock(128, 256)
# Stacking part
self.add_module('m0', HourGlass(1, 4, 256, first_one=True))
self.add_module('top_m_0', ConvBlock(256, 256))
self.add_module('conv_last0', nn.Conv2d(256, 256, 1, 1, 0))
self.add_module('bn_end0', nn.BatchNorm2d(256))
self.add_module('l0', nn.Conv2d(256, num_landmarks+1, 1, 1, 0))
if fname_pretrained is not None:
self.load_pretrained_weights(fname_pretrained)
def load_pretrained_weights(self, fname):
if torch.cuda.is_available():
checkpoint = torch.load(fname)
else:
checkpoint = torch.load(fname, map_location=torch.device('cpu'))
model_weights = self.state_dict()
model_weights.update({k: v for k, v in checkpoint['state_dict'].items()
if k in model_weights})
self.load_state_dict(model_weights)
def forward(self, x):
x, _ = self.conv1(x)
x = F.relu(self.bn1(x), True)
x = F.avg_pool2d(self.conv2(x), 2, stride=2)
x = self.conv3(x)
x = self.conv4(x)
outputs = []
boundary_channels = []
tmp_out = None
ll, boundary_channel = self._modules['m0'](x, tmp_out)
ll = self._modules['top_m_0'](ll)
ll = F.relu(self._modules['bn_end0']
(self._modules['conv_last0'](ll)), True)
# Predict heatmaps
tmp_out = self._modules['l0'](ll)
if self.end_relu:
tmp_out = F.relu(tmp_out) # HACK: Added relu
outputs.append(tmp_out)
boundary_channels.append(boundary_channel)
return outputs, boundary_channels
@torch.no_grad()
def get_heatmap(self, x, b_preprocess=True):
""" outputs 0-1 normalized heatmap """
x = F.interpolate(x, size=256, mode='bilinear', align_corners=True)
x_01 = x*0.5 + 0.5
outputs, _ = self(x_01)
heatmaps = outputs[-1][:, :-1, :, :]
scale_factor = x.size(2) // heatmaps.size(2)
if b_preprocess:
heatmaps = F.interpolate(heatmaps, scale_factor=scale_factor,
mode='bilinear', align_corners=True)
heatmaps = preprocess(heatmaps)
return heatmaps
@torch.no_grad()
def get_landmark(self, x):
""" outputs landmarks of x.shape """
heatmaps = self.get_heatmap(x, b_preprocess=False)
landmarks = []
for i in range(x.size(0)):
pred_landmarks = get_preds_fromhm(heatmaps[i].cpu().unsqueeze(0))
landmarks.append(pred_landmarks)
scale_factor = x.size(2) // heatmaps.size(2)
landmarks = torch.cat(landmarks) * scale_factor
return landmarks
# ========================== #
# Align related functions #
# ========================== #
def tensor2numpy255(tensor):
"""Converts torch tensor to numpy array."""
return ((tensor.permute(1, 2, 0).cpu().numpy() * 0.5 + 0.5) * 255).astype('uint8')
def np2tensor(image):
"""Converts numpy array to torch tensor."""
return torch.FloatTensor(image).permute(2, 0, 1) / 255 * 2 - 1
class FaceAligner:
def __init__(self, fname_wing, fname_celeba_mean, output_size):
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.fan = FAN(fname_pretrained=fname_wing).to(self.device).eval()
scale = output_size // 256
self.CELEB_REF = np.float32(np.load(fname_celeba_mean)['mean']) * scale
self.xaxis_ref = landmarks2xaxis(self.CELEB_REF)
self.output_size = output_size
def align(self, imgs):
""" imgs = torch.CUDATensor of BCHW """
imgs = imgs.to(self.device)
landmarkss = self.fan.get_landmark(imgs).cpu().numpy()
for i, (img, landmarks) in enumerate(zip(imgs, landmarkss)):
img_np = tensor2numpy255(img)
img_np, landmarks = pad_mirror(img_np, landmarks)
transform = self.landmarks2mat(landmarks)
rows, cols, _ = img_np.shape
rows = max(rows, self.output_size)
cols = max(cols, self.output_size)
aligned = cv2.warpPerspective(img_np, transform, (cols, rows), flags=cv2.INTER_LANCZOS4)
imgs[i] = np2tensor(aligned[:self.output_size, :self.output_size, :])
return imgs
def landmarks2mat(self, landmarks):
T_origin = points2T(landmarks, 'from')
xaxis_src = landmarks2xaxis(landmarks)
R = vecs2R(xaxis_src, self.xaxis_ref)
S = landmarks2S(landmarks, self.CELEB_REF)
T_ref = points2T(self.CELEB_REF, 'to')
matrix = np.dot(T_ref, np.dot(S, np.dot(R, T_origin)))
return matrix
def points2T(point, direction):
point_mean = point.mean(axis=0)
T = np.eye(3)
coef = -1 if direction == 'from' else 1
T[:2, 2] = coef * point_mean
return T
def landmarks2eyes(landmarks):
idx_left = np.array(list(range(60, 67+1)) + [96])
idx_right = np.array(list(range(68, 75+1)) + [97])
left = landmarks[idx_left]
right = landmarks[idx_right]
return left.mean(axis=0), right.mean(axis=0)
def landmarks2mouthends(landmarks):
left = landmarks[76]
right = landmarks[82]
return left, right
def rotate90(vec):
x, y = vec
return np.array([y, -x])
def landmarks2xaxis(landmarks):
eye_left, eye_right = landmarks2eyes(landmarks)
mouth_left, mouth_right = landmarks2mouthends(landmarks)
xp = eye_right - eye_left # x' in pggan
eye_center = (eye_left + eye_right) * 0.5
mouth_center = (mouth_left + mouth_right) * 0.5
yp = eye_center - mouth_center
xaxis = xp - rotate90(yp)
return xaxis / np.linalg.norm(xaxis)
def vecs2R(vec_x, vec_y):
vec_x = vec_x / np.linalg.norm(vec_x)
vec_y = vec_y / np.linalg.norm(vec_y)
c = np.dot(vec_x, vec_y)
s = np.sqrt(1 - c * c) * np.sign(np.cross(vec_x, vec_y))
R = np.array(((c, -s, 0), (s, c, 0), (0, 0, 1)))
return R
def landmarks2S(x, y):
x_mean = x.mean(axis=0).squeeze()
y_mean = y.mean(axis=0).squeeze()
# vectors = mean -> each point
x_vectors = x - x_mean
y_vectors = y - y_mean
x_norms = np.linalg.norm(x_vectors, axis=1)
y_norms = np.linalg.norm(y_vectors, axis=1)
indices = [96, 97, 76, 82] # indices for eyes, lips
scale = (y_norms / x_norms)[indices].mean()
S = np.eye(3)
S[0, 0] = S[1, 1] = scale
return S
def pad_mirror(img, landmarks):
H, W, _ = img.shape
img = np.pad(img, ((H//2, H//2), (W//2, W//2), (0, 0)), 'reflect')
small_blurred = gaussian(cv2.resize(img, (W, H)), H//100, multichannel=True)
blurred = cv2.resize(small_blurred, (W * 2, H * 2)) * 255
H, W, _ = img.shape
coords = np.meshgrid(np.arange(H), np.arange(W), indexing="ij")
weight_y = np.clip(coords[0] / (H//4), 0, 1)
weight_x = np.clip(coords[1] / (H//4), 0, 1)
weight_y = np.minimum(weight_y, np.flip(weight_y, axis=0))
weight_x = np.minimum(weight_x, np.flip(weight_x, axis=1))
weight = np.expand_dims(np.minimum(weight_y, weight_x), 2)**4
img = img * weight + blurred * (1 - weight)
landmarks += np.array([W//4, H//4])
return img, landmarks
# ========================== #
# Mask related functions #
# ========================== #
def normalize(x, eps=1e-6):
"""Apply min-max normalization."""
x = x.contiguous()
N, C, H, W = x.size()
x_ = x.view(N*C, -1)
max_val = torch.max(x_, dim=1, keepdim=True)[0]
min_val = torch.min(x_, dim=1, keepdim=True)[0]
x_ = (x_ - min_val) / (max_val - min_val + eps)
out = x_.view(N, C, H, W)
return out
def truncate(x, thres=0.1):
"""Remove small values in heatmaps."""
return torch.where(x < thres, torch.zeros_like(x), x)
def resize(x, p=2):
"""Resize heatmaps."""
return x**p
def shift(x, N):
"""Shift N pixels up or down."""
up = N >= 0
N = abs(N)
_, _, H, W = x.size()
head = torch.arange(N)
tail = torch.arange(H-N)
if up:
head = torch.arange(H-N)+N
tail = torch.arange(N)
else:
head = torch.arange(N) + (H-N)
tail = torch.arange(H-N)
# permutation indices
perm = torch.cat([head, tail]).to(x.device)
out = x[:, :, perm, :]
return out
IDXPAIR = namedtuple('IDXPAIR', 'start end')
index_map = Munch(chin=IDXPAIR(0 + 8, 33 - 8),
eyebrows=IDXPAIR(33, 51),
eyebrowsedges=IDXPAIR(33, 46),
nose=IDXPAIR(51, 55),
nostrils=IDXPAIR(55, 60),
eyes=IDXPAIR(60, 76),
lipedges=IDXPAIR(76, 82),
lipupper=IDXPAIR(77, 82),
liplower=IDXPAIR(83, 88),
lipinner=IDXPAIR(88, 96))
OPPAIR = namedtuple('OPPAIR', 'shift resize')
def preprocess(x):
"""Preprocess 98-dimensional heatmaps."""
N, C, H, W = x.size()
x = truncate(x)
x = normalize(x)
sw = H // 256
operations = Munch(chin=OPPAIR(0, 3),
eyebrows=OPPAIR(-7*sw, 2),
nostrils=OPPAIR(8*sw, 4),
lipupper=OPPAIR(-8*sw, 4),
liplower=OPPAIR(8*sw, 4),
lipinner=OPPAIR(-2*sw, 3))
for part, ops in operations.items():
start, end = index_map[part]
x[:, start:end] = resize(shift(x[:, start:end], ops.shift), ops.resize)
zero_out = torch.cat([torch.arange(0, index_map.chin.start),
torch.arange(index_map.chin.end, 33),
torch.LongTensor([index_map.eyebrowsedges.start,
index_map.eyebrowsedges.end,
index_map.lipedges.start,
index_map.lipedges.end])])
x[:, zero_out] = 0
start, end = index_map.nose
x[:, start+1:end] = shift(x[:, start+1:end], 4*sw)
x[:, start:end] = resize(x[:, start:end], 1)
start, end = index_map.eyes
x[:, start:end] = resize(x[:, start:end], 1)
x[:, start:end] = resize(shift(x[:, start:end], -8), 3) + \
shift(x[:, start:end], -24)
# Second-level mask
x2 = deepcopy(x)
x2[:, index_map.chin.start:index_map.chin.end] = 0 # start:end was 0:33
x2[:, index_map.lipedges.start:index_map.lipinner.end] = 0 # start:end was 76:96
x2[:, index_map.eyebrows.start:index_map.eyebrows.end] = 0 # start:end was 33:51
x = torch.sum(x, dim=1, keepdim=True) # (N, 1, H, W)
x2 = torch.sum(x2, dim=1, keepdim=True) # mask without faceline and mouth
x[x != x] = 0 # set nan to zero
x2[x != x] = 0 # set nan to zero
return x.clamp_(0, 1), x2.clamp_(0, 1)