File size: 8,887 Bytes
16ee39a
 
 
 
 
8d04433
16ee39a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd6fe9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16ee39a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd6fe9d
16ee39a
 
 
 
 
 
 
 
 
 
 
 
 
 
7ab7ce4
3008a78
 
 
 
7ab7ce4
3008a78
7ab7ce4
 
 
 
3008a78
 
7ab7ce4
 
 
 
 
3008a78
7ab7ce4
3008a78
 
7ab7ce4
3008a78
 
 
7ab7ce4
3008a78
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ab7ce4
ca16892
 
 
 
 
7ab7ce4
8d04433
3008a78
 
8d04433
3008a78
 
 
8d04433
 
 
3008a78
ca16892
 
a815574
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
923c799
ca16892
8b72e71
3008a78
 
cf2e13b
3008a78
cf2e13b
 
16ee39a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
04c1e71
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
from scipy.spatial import ConvexHull
import torch
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm 
import time

def normalize_kp(kp_source, kp_driving, kp_driving_initial, adapt_movement_scale=False,
                 use_relative_movement=False, use_relative_jacobian=False):
    if adapt_movement_scale:
        source_area = ConvexHull(kp_source['value'][0].data.cpu().numpy()).volume
        driving_area = ConvexHull(kp_driving_initial['value'][0].data.cpu().numpy()).volume
        adapt_movement_scale = np.sqrt(source_area) / np.sqrt(driving_area)
    else:
        adapt_movement_scale = 1

    kp_new = {k: v for k, v in kp_driving.items()}

    if use_relative_movement:
        kp_value_diff = (kp_driving['value'] - kp_driving_initial['value'])
        kp_value_diff *= adapt_movement_scale
        kp_new['value'] = kp_value_diff + kp_source['value']

        if use_relative_jacobian:
            jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian']))
            kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian'])

    return kp_new

def headpose_pred_to_degree(pred):
    device = pred.device
    idx_tensor = [idx for idx in range(66)]
    idx_tensor = torch.FloatTensor(idx_tensor).to(device)
    pred = F.softmax(pred)
    degree = torch.sum(pred*idx_tensor, 1) * 3 - 99
    return degree

def get_rotation_matrix(yaw, pitch, roll):
    yaw = yaw / 180 * 3.14
    pitch = pitch / 180 * 3.14
    roll = roll / 180 * 3.14

    roll = roll.unsqueeze(1)
    pitch = pitch.unsqueeze(1)
    yaw = yaw.unsqueeze(1)

    pitch_mat = torch.cat([torch.ones_like(pitch), torch.zeros_like(pitch), torch.zeros_like(pitch), 
                          torch.zeros_like(pitch), torch.cos(pitch), -torch.sin(pitch),
                          torch.zeros_like(pitch), torch.sin(pitch), torch.cos(pitch)], dim=1)
    pitch_mat = pitch_mat.view(pitch_mat.shape[0], 3, 3)

    yaw_mat = torch.cat([torch.cos(yaw), torch.zeros_like(yaw), torch.sin(yaw), 
                           torch.zeros_like(yaw), torch.ones_like(yaw), torch.zeros_like(yaw),
                           -torch.sin(yaw), torch.zeros_like(yaw), torch.cos(yaw)], dim=1)
    yaw_mat = yaw_mat.view(yaw_mat.shape[0], 3, 3)

    roll_mat = torch.cat([torch.cos(roll), -torch.sin(roll), torch.zeros_like(roll),  
                         torch.sin(roll), torch.cos(roll), torch.zeros_like(roll),
                         torch.zeros_like(roll), torch.zeros_like(roll), torch.ones_like(roll)], dim=1)
    roll_mat = roll_mat.view(roll_mat.shape[0], 3, 3)


    rot_mat = torch.einsum('bij,bjk,bkm->bim', pitch_mat, yaw_mat, roll_mat)

    return rot_mat

def keypoint_transformation(kp_canonical, he, wo_exp=False):
    kp = kp_canonical['value']    # (bs, k, 3) 
    yaw, pitch, roll= he['yaw'], he['pitch'], he['roll']      
    yaw = headpose_pred_to_degree(yaw) 
    pitch = headpose_pred_to_degree(pitch)
    roll = headpose_pred_to_degree(roll)

    if 'yaw_in' in he:
        yaw = he['yaw_in']
    if 'pitch_in' in he:
        pitch = he['pitch_in']
    if 'roll_in' in he:
        roll = he['roll_in']

    rot_mat = get_rotation_matrix(yaw, pitch, roll)    # (bs, 3, 3)

    t, exp = he['t'], he['exp']
    if wo_exp:
        exp =  exp*0  
    
    # keypoint rotation
    kp_rotated = torch.einsum('bmp,bkp->bkm', rot_mat, kp)

    # keypoint translation
    t[:, 0] = t[:, 0]*0
    t[:, 2] = t[:, 2]*0
    t = t.unsqueeze(1).repeat(1, kp.shape[1], 1)
    kp_t = kp_rotated + t

    # add expression deviation 
    exp = exp.view(exp.shape[0], -1, 3)
    kp_transformed = kp_t + exp

    return {'value': kp_transformed}


# def make_animation(source_image, source_semantics, target_semantics,
#                             generator, kp_detector, he_estimator, mapping, 
#                             yaw_c_seq=None, pitch_c_seq=None, roll_c_seq=None,
#                             use_exp=True):
#     with torch.no_grad():
#         predictions = []

#         kp_canonical = kp_detector(source_image)
#         he_source = mapping(source_semantics)
#         kp_source = keypoint_transformation(kp_canonical, he_source)

    
#         for frame_idx in tqdm(range(target_semantics.shape[1]), 'Face Renderer:'):
#             target_semantics_frame = target_semantics[:, frame_idx]
#             he_driving = mapping(target_semantics_frame)
#             if yaw_c_seq is not None:
#                 he_driving['yaw_in'] = yaw_c_seq[:, frame_idx]
#             if pitch_c_seq is not None:
#                 he_driving['pitch_in'] = pitch_c_seq[:, frame_idx] 
#             if roll_c_seq is not None:
#                 he_driving['roll_in'] = roll_c_seq[:, frame_idx] 
            
#             kp_driving = keypoint_transformation(kp_canonical, he_driving)
                
#             #kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
#                                    #kp_driving_initial=kp_driving_initial)
#             kp_norm = kp_driving
#             out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
#             '''
#             source_image_new = out['prediction'].squeeze(1)
#             kp_canonical_new =  kp_detector(source_image_new)
#             he_source_new = he_estimator(source_image_new) 
#             kp_source_new = keypoint_transformation(kp_canonical_new, he_source_new, wo_exp=True)
#             kp_driving_new = keypoint_transformation(kp_canonical_new, he_driving, wo_exp=True)
#             out = generator(source_image_new, kp_source=kp_source_new, kp_driving=kp_driving_new)
#             '''
#             predictions.append(out['prediction'])
#             torch.cuda.empty_cache()
#         predictions_ts = torch.stack(predictions, dim=1)
#     return predictions_ts

import torch
from torch.cuda.amp import autocast

def make_animation(source_image, source_semantics, target_semantics,
                   generator, kp_detector, he_estimator, mapping, 
                   yaw_c_seq=None, pitch_c_seq=None, roll_c_seq=None,
                   use_exp=True):

    device='cuda'
    # Move inputs to GPU
    source_image = source_image.to(device)
    source_semantics = source_semantics.to(device)
    target_semantics = target_semantics.to(device)

    
    with torch.no_grad():  # No gradients needed
        predictions = []
        start = time.time()
        kp_canonical = kp_detector(source_image)
        he_source = mapping(source_semantics)
        kp_source = keypoint_transformation(kp_canonical, he_source)
        end = time.time()
        diff = end-start
        print("time taken for 3 values within make animation:", diff)

        with autocast():
            for frame_idx in tqdm(range(target_semantics.shape[1]), desc='Face Renderer:', unit='frame'):
                target_semantics_frame = target_semantics[:, frame_idx]
                he_driving = mapping(target_semantics_frame)
    
                if yaw_c_seq is not None:
                    he_driving['yaw_in'] = yaw_c_seq[:, frame_idx]
                if pitch_c_seq is not None:
                    he_driving['pitch_in'] = pitch_c_seq[:, frame_idx]
                if roll_c_seq is not None:
                    he_driving['roll_in'] = roll_c_seq[:, frame_idx]
    
                kp_driving = keypoint_transformation(kp_canonical, he_driving)
                kp_norm = kp_driving
                
                out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
    
                predictions.append(out['prediction'])
            
        torch.cuda.synchronize()

        # Stack predictions into a single tensor
        predictions_ts = torch.stack(predictions, dim=1)
    
    return predictions_ts


class AnimateModel(torch.nn.Module):
    """
    Merge all generator related updates into single model for better multi-gpu usage
    """

    def __init__(self, generator, kp_extractor, mapping):
        super(AnimateModel, self).__init__()
        self.kp_extractor = kp_extractor
        self.generator = generator
        self.mapping = mapping

        self.kp_extractor.eval()
        self.generator.eval()
        self.mapping.eval()

    def forward(self, x):
        
        source_image = x['source_image']
        source_semantics = x['source_semantics']
        target_semantics = x['target_semantics']
        yaw_c_seq = x['yaw_c_seq']
        pitch_c_seq = x['pitch_c_seq']
        roll_c_seq = x['roll_c_seq']

        predictions_video = make_animation(source_image, source_semantics, target_semantics,
                                        self.generator, self.kp_extractor,
                                        self.mapping, use_exp = True,
                                        yaw_c_seq=yaw_c_seq, pitch_c_seq=pitch_c_seq, roll_c_seq=roll_c_seq)
        
        return predictions_video