Delete src/facerender/modules/make_animation.py
Browse files
src/facerender/modules/make_animation.py
DELETED
@@ -1,221 +0,0 @@
|
|
1 |
-
from scipy.spatial import ConvexHull
|
2 |
-
import torch
|
3 |
-
import torch.nn.functional as F
|
4 |
-
import numpy as np
|
5 |
-
from tqdm import tqdm
|
6 |
-
|
7 |
-
def normalize_kp(kp_source, kp_driving, kp_driving_initial, adapt_movement_scale=False,
|
8 |
-
use_relative_movement=False, use_relative_jacobian=False):
|
9 |
-
if adapt_movement_scale:
|
10 |
-
source_area = ConvexHull(kp_source['value'][0].data.cpu().numpy()).volume
|
11 |
-
driving_area = ConvexHull(kp_driving_initial['value'][0].data.cpu().numpy()).volume
|
12 |
-
adapt_movement_scale = np.sqrt(source_area) / np.sqrt(driving_area)
|
13 |
-
else:
|
14 |
-
adapt_movement_scale = 1
|
15 |
-
|
16 |
-
kp_new = {k: v for k, v in kp_driving.items()}
|
17 |
-
|
18 |
-
if use_relative_movement:
|
19 |
-
kp_value_diff = (kp_driving['value'] - kp_driving_initial['value'])
|
20 |
-
kp_value_diff *= adapt_movement_scale
|
21 |
-
kp_new['value'] = kp_value_diff + kp_source['value']
|
22 |
-
|
23 |
-
if use_relative_jacobian:
|
24 |
-
jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian']))
|
25 |
-
kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian'])
|
26 |
-
|
27 |
-
return kp_new
|
28 |
-
|
29 |
-
def headpose_pred_to_degree(pred):
|
30 |
-
device = pred.device
|
31 |
-
idx_tensor = [idx for idx in range(66)]
|
32 |
-
idx_tensor = torch.FloatTensor(idx_tensor).to(device)
|
33 |
-
pred = F.softmax(pred)
|
34 |
-
degree = torch.sum(pred*idx_tensor, 1) * 3 - 99
|
35 |
-
return degree
|
36 |
-
|
37 |
-
def get_rotation_matrix(yaw, pitch, roll):
|
38 |
-
yaw = yaw / 180 * 3.14
|
39 |
-
pitch = pitch / 180 * 3.14
|
40 |
-
roll = roll / 180 * 3.14
|
41 |
-
|
42 |
-
roll = roll.unsqueeze(1)
|
43 |
-
pitch = pitch.unsqueeze(1)
|
44 |
-
yaw = yaw.unsqueeze(1)
|
45 |
-
|
46 |
-
pitch_mat = torch.cat([torch.ones_like(pitch), torch.zeros_like(pitch), torch.zeros_like(pitch),
|
47 |
-
torch.zeros_like(pitch), torch.cos(pitch), -torch.sin(pitch),
|
48 |
-
torch.zeros_like(pitch), torch.sin(pitch), torch.cos(pitch)], dim=1)
|
49 |
-
pitch_mat = pitch_mat.view(pitch_mat.shape[0], 3, 3)
|
50 |
-
|
51 |
-
yaw_mat = torch.cat([torch.cos(yaw), torch.zeros_like(yaw), torch.sin(yaw),
|
52 |
-
torch.zeros_like(yaw), torch.ones_like(yaw), torch.zeros_like(yaw),
|
53 |
-
-torch.sin(yaw), torch.zeros_like(yaw), torch.cos(yaw)], dim=1)
|
54 |
-
yaw_mat = yaw_mat.view(yaw_mat.shape[0], 3, 3)
|
55 |
-
|
56 |
-
roll_mat = torch.cat([torch.cos(roll), -torch.sin(roll), torch.zeros_like(roll),
|
57 |
-
torch.sin(roll), torch.cos(roll), torch.zeros_like(roll),
|
58 |
-
torch.zeros_like(roll), torch.zeros_like(roll), torch.ones_like(roll)], dim=1)
|
59 |
-
roll_mat = roll_mat.view(roll_mat.shape[0], 3, 3)
|
60 |
-
|
61 |
-
|
62 |
-
rot_mat = torch.einsum('bij,bjk,bkm->bim', pitch_mat, yaw_mat, roll_mat)
|
63 |
-
|
64 |
-
return rot_mat
|
65 |
-
|
66 |
-
def keypoint_transformation(kp_canonical, he, wo_exp=False):
|
67 |
-
kp = kp_canonical['value'] # (bs, k, 3)
|
68 |
-
yaw, pitch, roll= he['yaw'], he['pitch'], he['roll']
|
69 |
-
yaw = headpose_pred_to_degree(yaw)
|
70 |
-
pitch = headpose_pred_to_degree(pitch)
|
71 |
-
roll = headpose_pred_to_degree(roll)
|
72 |
-
|
73 |
-
if 'yaw_in' in he:
|
74 |
-
yaw = he['yaw_in']
|
75 |
-
if 'pitch_in' in he:
|
76 |
-
pitch = he['pitch_in']
|
77 |
-
if 'roll_in' in he:
|
78 |
-
roll = he['roll_in']
|
79 |
-
|
80 |
-
rot_mat = get_rotation_matrix(yaw, pitch, roll) # (bs, 3, 3)
|
81 |
-
|
82 |
-
t, exp = he['t'], he['exp']
|
83 |
-
if wo_exp:
|
84 |
-
exp = exp*0
|
85 |
-
|
86 |
-
# keypoint rotation
|
87 |
-
kp_rotated = torch.einsum('bmp,bkp->bkm', rot_mat, kp)
|
88 |
-
|
89 |
-
# keypoint translation
|
90 |
-
t[:, 0] = t[:, 0]*0
|
91 |
-
t[:, 2] = t[:, 2]*0
|
92 |
-
t = t.unsqueeze(1).repeat(1, kp.shape[1], 1)
|
93 |
-
kp_t = kp_rotated + t
|
94 |
-
|
95 |
-
# add expression deviation
|
96 |
-
exp = exp.view(exp.shape[0], -1, 3)
|
97 |
-
kp_transformed = kp_t + exp
|
98 |
-
|
99 |
-
return {'value': kp_transformed}
|
100 |
-
|
101 |
-
|
102 |
-
# def make_animation(source_image, source_semantics, target_semantics,
|
103 |
-
# generator, kp_detector, he_estimator, mapping,
|
104 |
-
# yaw_c_seq=None, pitch_c_seq=None, roll_c_seq=None,
|
105 |
-
# use_exp=True):
|
106 |
-
# with torch.no_grad():
|
107 |
-
# predictions = []
|
108 |
-
|
109 |
-
# kp_canonical = kp_detector(source_image)
|
110 |
-
# he_source = mapping(source_semantics)
|
111 |
-
# kp_source = keypoint_transformation(kp_canonical, he_source)
|
112 |
-
|
113 |
-
|
114 |
-
# for frame_idx in tqdm(range(target_semantics.shape[1]), 'Face Renderer:'):
|
115 |
-
# target_semantics_frame = target_semantics[:, frame_idx]
|
116 |
-
# he_driving = mapping(target_semantics_frame)
|
117 |
-
# if yaw_c_seq is not None:
|
118 |
-
# he_driving['yaw_in'] = yaw_c_seq[:, frame_idx]
|
119 |
-
# if pitch_c_seq is not None:
|
120 |
-
# he_driving['pitch_in'] = pitch_c_seq[:, frame_idx]
|
121 |
-
# if roll_c_seq is not None:
|
122 |
-
# he_driving['roll_in'] = roll_c_seq[:, frame_idx]
|
123 |
-
|
124 |
-
# kp_driving = keypoint_transformation(kp_canonical, he_driving)
|
125 |
-
|
126 |
-
# #kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
|
127 |
-
# #kp_driving_initial=kp_driving_initial)
|
128 |
-
# kp_norm = kp_driving
|
129 |
-
# out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
|
130 |
-
# '''
|
131 |
-
# source_image_new = out['prediction'].squeeze(1)
|
132 |
-
# kp_canonical_new = kp_detector(source_image_new)
|
133 |
-
# he_source_new = he_estimator(source_image_new)
|
134 |
-
# kp_source_new = keypoint_transformation(kp_canonical_new, he_source_new, wo_exp=True)
|
135 |
-
# kp_driving_new = keypoint_transformation(kp_canonical_new, he_driving, wo_exp=True)
|
136 |
-
# out = generator(source_image_new, kp_source=kp_source_new, kp_driving=kp_driving_new)
|
137 |
-
# '''
|
138 |
-
# predictions.append(out['prediction'])
|
139 |
-
# torch.cuda.empty_cache()
|
140 |
-
# predictions_ts = torch.stack(predictions, dim=1)
|
141 |
-
# return predictions_ts
|
142 |
-
|
143 |
-
import torch
|
144 |
-
from torch.cuda.amp import autocast
|
145 |
-
|
146 |
-
def make_animation(source_image, source_semantics, target_semantics,
|
147 |
-
generator, kp_detector, he_estimator, mapping,
|
148 |
-
yaw_c_seq=None, pitch_c_seq=None, roll_c_seq=None,
|
149 |
-
use_exp=True):
|
150 |
-
|
151 |
-
# device='cuda'
|
152 |
-
# # Move inputs to GPU
|
153 |
-
# source_image = source_image.to(device)
|
154 |
-
# source_semantics = source_semantics.to(device)
|
155 |
-
# target_semantics = target_semantics.to(device)
|
156 |
-
|
157 |
-
with torch.no_grad(): # No gradients needed
|
158 |
-
predictions = []
|
159 |
-
kp_canonical = kp_detector(source_image)
|
160 |
-
he_source = mapping(source_semantics)
|
161 |
-
kp_source = keypoint_transformation(kp_canonical, he_source)
|
162 |
-
|
163 |
-
for frame_idx in tqdm(range(target_semantics.shape[1]), desc='Face Renderer:', unit='frame'):
|
164 |
-
target_semantics_frame = target_semantics[:, frame_idx]
|
165 |
-
he_driving = mapping(target_semantics_frame)
|
166 |
-
|
167 |
-
if yaw_c_seq is not None:
|
168 |
-
he_driving['yaw_in'] = yaw_c_seq[:, frame_idx]
|
169 |
-
if pitch_c_seq is not None:
|
170 |
-
he_driving['pitch_in'] = pitch_c_seq[:, frame_idx]
|
171 |
-
if roll_c_seq is not None:
|
172 |
-
he_driving['roll_in'] = roll_c_seq[:, frame_idx]
|
173 |
-
|
174 |
-
kp_driving = keypoint_transformation(kp_canonical, he_driving)
|
175 |
-
kp_norm = kp_driving
|
176 |
-
|
177 |
-
# Use mixed precision for faster computation
|
178 |
-
with autocast():
|
179 |
-
out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
|
180 |
-
|
181 |
-
predictions.append(out['prediction'])
|
182 |
-
|
183 |
-
# Optional: Explicitly synchronize (use only if necessary)
|
184 |
-
torch.cuda.synchronize()
|
185 |
-
|
186 |
-
# Stack predictions into a single tensor
|
187 |
-
predictions_ts = torch.stack(predictions, dim=1)
|
188 |
-
|
189 |
-
return predictions_ts
|
190 |
-
|
191 |
-
|
192 |
-
class AnimateModel(torch.nn.Module):
|
193 |
-
"""
|
194 |
-
Merge all generator related updates into single model for better multi-gpu usage
|
195 |
-
"""
|
196 |
-
|
197 |
-
def __init__(self, generator, kp_extractor, mapping):
|
198 |
-
super(AnimateModel, self).__init__()
|
199 |
-
self.kp_extractor = kp_extractor
|
200 |
-
self.generator = generator
|
201 |
-
self.mapping = mapping
|
202 |
-
|
203 |
-
self.kp_extractor.eval()
|
204 |
-
self.generator.eval()
|
205 |
-
self.mapping.eval()
|
206 |
-
|
207 |
-
def forward(self, x):
|
208 |
-
|
209 |
-
source_image = x['source_image']
|
210 |
-
source_semantics = x['source_semantics']
|
211 |
-
target_semantics = x['target_semantics']
|
212 |
-
yaw_c_seq = x['yaw_c_seq']
|
213 |
-
pitch_c_seq = x['pitch_c_seq']
|
214 |
-
roll_c_seq = x['roll_c_seq']
|
215 |
-
|
216 |
-
predictions_video = make_animation(source_image, source_semantics, target_semantics,
|
217 |
-
self.generator, self.kp_extractor,
|
218 |
-
self.mapping, use_exp = True,
|
219 |
-
yaw_c_seq=yaw_c_seq, pitch_c_seq=pitch_c_seq, roll_c_seq=roll_c_seq)
|
220 |
-
|
221 |
-
return predictions_video
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|