EasyVtuber/models.py at main · yuyuyzl/EasyVtuber · GitHub

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
import time

import torch
import torch.nn as nn

import tha2.poser.modes.mode_20
import tha3.poser.modes.standard_float
import tha3.poser.modes.separable_float
import tha3.poser.modes.standard_half
import tha3.poser.modes.separable_half
from torch.nn.functional import interpolate

from args import args

from collections import OrderedDict

# THA4 适配器导入
try:
    from tha4_adapter import THA4Wrapper
    THA4_AVAILABLE = True
except ImportError:
    THA4_AVAILABLE = False
    print("Warning: THA4 adapter not available")


class TalkingAnimeLight(nn.Module):
    def __init__(self):
        super(TalkingAnimeLight, self).__init__()
        self.face_morpher = tha2.poser.modes.mode_20.load_face_morpher('pretrained/face_morpher.pt')
        self.two_algo_face_rotator = tha2.poser.modes.mode_20.load_face_rotater('pretrained/two_algo_face_rotator.pt')
        self.combiner = tha2.poser.modes.mode_20.load_combiner('pretrained/combiner.pt')
        self.face_cache = OrderedDict()
        self.tot = 0
        self.hit = 0

    def forward(self, image, mouth_eye_vector, pose_vector, mouth_eye_vector_c, ratio=None):
        x = image.clone()
        if args.perf == 'model':
            tic = time.perf_counter()
        input_hash = hash(tuple(mouth_eye_vector_c))
        cached = self.face_cache.get(input_hash)
        self.tot += 1
        if cached is None:
            mouth_eye_morp_image = self.face_morpher(image[:, :, 32:224, 32:224], mouth_eye_vector)
            self.face_cache[input_hash] = mouth_eye_morp_image.detach()
            if len(self.face_cache) > args.max_gpu_cache_len:
                self.face_cache.popitem(last=False)
        else:
            self.hit += 1
            mouth_eye_morp_image = cached
            self.face_cache.move_to_end(input_hash)
        if args.debug and ratio is not None:
            ratio.value = self.hit / self.tot
        if args.perf == 'model':
            print(" - face_morpher", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        x[:, :, 32:224, 32:224] = mouth_eye_morp_image
        rotate_image = self.two_algo_face_rotator(x, pose_vector)[:2]
        if args.perf == 'model':
            print(" - rotator", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        output_image = self.combiner(rotate_image[0], rotate_image[1], pose_vector)
        if args.perf == 'model':
            print(" - combiner", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        return output_image


class TalkingAnime3(nn.Module):
    def __init__(self):
        super(TalkingAnime3, self).__init__()
        if args.model == "standard_float":
            if args.eyebrow:
                self.eyebrow_decomposer = tha3.poser.modes.standard_float.load_eyebrow_decomposer(
                    'data/models/standard_float/eyebrow_decomposer.pt')
                self.eyebrow_morphing_combiner = tha3.poser.modes.standard_float.load_eyebrow_morphing_combiner(
                    'data/models/standard_float/eyebrow_morphing_combiner.pt')
            self.face_morpher = tha3.poser.modes.standard_float.load_face_morpher(
                'data/models/standard_float/face_morpher.pt')
            self.two_algo_face_body_rotator = tha3.poser.modes.standard_float.load_two_algo_generator(
                'data/models/standard_float/two_algo_face_body_rotator.pt')
            self.editor = tha3.poser.modes.standard_float.load_editor('data/models/standard_float/editor.pt')
        elif args.model == "standard_half":
            if args.eyebrow:
                self.eyebrow_decomposer = tha3.poser.modes.standard_half.load_eyebrow_decomposer(
                    'data/models/standard_half/eyebrow_decomposer.pt')
                self.eyebrow_morphing_combiner = tha3.poser.modes.standard_half.load_eyebrow_morphing_combiner(
                    'data/models/standard_half/eyebrow_morphing_combiner.pt')

            self.face_morpher = tha3.poser.modes.standard_half.load_face_morpher(
                'data/models/standard_half/face_morpher.pt')
            self.two_algo_face_body_rotator = tha3.poser.modes.standard_half.load_two_algo_generator(
                'data/models/standard_half/two_algo_face_body_rotator.pt')
            self.editor = tha3.poser.modes.standard_half.load_editor('data/models/standard_half/editor.pt')
        elif args.model == "separable_float":
            if args.eyebrow:
                self.eyebrow_decomposer = tha3.poser.modes.separable_float.load_eyebrow_decomposer(
                    'data/models/separable_float/eyebrow_decomposer.pt')
                self.eyebrow_morphing_combiner = tha3.poser.modes.separable_float.load_eyebrow_morphing_combiner(
                    'data/models/separable_float/eyebrow_morphing_combiner.pt')

            self.face_morpher = tha3.poser.modes.separable_float.load_face_morpher(
                'data/models/separable_float/face_morpher.pt')
            self.two_algo_face_body_rotator = tha3.poser.modes.separable_float.load_two_algo_generator(
                'data/models/separable_float/two_algo_face_body_rotator.pt')
            self.editor = tha3.poser.modes.separable_float.load_editor('data/models/separable_float/editor.pt')
        elif args.model == "separable_half":
            if args.eyebrow:
                self.eyebrow_decomposer = tha3.poser.modes.separable_half.load_eyebrow_decomposer(
                    'data/models/separable_half/eyebrow_decomposer.pt')
                self.eyebrow_morphing_combiner = tha3.poser.modes.separable_half.load_eyebrow_morphing_combiner(
                    'data/models/separable_half/eyebrow_morphing_combiner.pt')

            self.face_morpher = tha3.poser.modes.separable_half.load_face_morpher(
                'data/models/separable_half/face_morpher.pt')
            self.two_algo_face_body_rotator = tha3.poser.modes.separable_half.load_two_algo_generator(
                'data/models/separable_half/two_algo_face_body_rotator.pt')
            self.editor = tha3.poser.modes.separable_half.load_editor('data/models/separable_half/editor.pt')
        else:
            raise RuntimeError("Invalid model: '%s'" % args.model)
        self.face_cache = OrderedDict()
        self.tot = 0
        self.hit = 0

    def forward(self, image, mouth_eye_vector, pose_vector, eyebrow_vector, mouth_eye_vector_c, eyebrow_vector_c,
                ratio=None):
        if args.perf == 'model':
            tic = time.perf_counter()
        x = image.clone()
        if args.eyebrow:
            input_hash = hash(tuple(eyebrow_vector_c + mouth_eye_vector_c))
        else:
            input_hash = hash(tuple(mouth_eye_vector_c))
        cached = self.face_cache.get(input_hash)
        self.tot += 1
        if cached is None:
            face_image = x[:, :, 32:32 + 192, (32 + 128):(32 + 192 + 128)].clone()
            if args.eyebrow:
                eyebrow_morp_image = self.eyebrow_decomposer(x[:, :, 64:192, 64 + 128:192 + 128].clone())
                eyebrow_morp_image = \
                self.eyebrow_morphing_combiner(eyebrow_morp_image[3], eyebrow_morp_image[0], eyebrow_vector)[2]
                face_image[:, :, 32:32 + 128, 32:32 + 128] = eyebrow_morp_image
            mouth_eye_morp_image = self.face_morpher(face_image, mouth_eye_vector)[0]
            self.face_cache[input_hash] = mouth_eye_morp_image.detach()
            if len(self.face_cache) > args.max_gpu_cache_len:
                self.face_cache.popitem(last=False)
        else:
            self.hit += 1
            mouth_eye_morp_image = cached
            self.face_cache.move_to_end(input_hash)
        if args.debug and ratio is not None:
            ratio.value = self.hit / self.tot
        if args.perf == 'model':
            print(" - face_morpher", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        x[:, :, 32:32 + 192, (32 + 128):(32 + 192 + 128)] = mouth_eye_morp_image
        x_half = interpolate(x, size=(256, 256), mode='bilinear', align_corners=False)
        rotate_image = self.two_algo_face_body_rotator(x_half, pose_vector)
        if args.perf == 'model':
            print(" - rotator", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        output_image = self.editor(x,
                                   interpolate(rotate_image[1], size=(512, 512), mode='bilinear', align_corners=False),
                                   interpolate(rotate_image[2], size=(512, 512), mode='bilinear', align_corners=False),
                                   pose_vector)[0]
        if args.perf == 'model':
            print(" - editor", (time.perf_counter() - tic) * 1000)
            tic = time.perf_counter()
        return output_image


class TalkingAnime(nn.Module):
    def __init__(self):
        super(TalkingAnime, self).__init__()

    def forward(self, image, mouth_eye_vector, pose_vector):
        x = image.clone()
        mouth_eye_morp_image = self.face_morpher(image[:, :, 32:224, 32:224], mouth_eye_vector)
        x[:, :, 32:224, 32:224] = mouth_eye_morp_image
        rotate_image = self.two_algo_face_rotator(x, pose_vector)[:2]
        output_image = self.combiner(rotate_image[0], rotate_image[1], pose_vector)
        return output_image


class TalkingAnime4(nn.Module):
    """
    THA4 model class, compatible with THA3 TalkingAnime3 interface
    Uses THA4's Siren architecture for inference
    PyTorch only, float32 precision
    """
    def __init__(self, device):
        super(TalkingAnime4, self).__init__()
        if not THA4_AVAILABLE:
            raise RuntimeError(
                "THA4 adapter not available. Check tha4_adapter.py"
            )

        # Create THA4 wrapper
        self.wrapper = THA4Wrapper(device=device)

        print("THA4 model initialized (float32)")

    def forward(self, image, mouth_eye_vector, pose_vector, eyebrow_vector,
                mouth_eye_vector_c, eyebrow_vector_c, ratio=None):
        """
        Forward inference, compatible with THA3 interface

        Args:
            image: [batch, 4, 512, 512]
            mouth_eye_vector: [batch, 27]
            pose_vector: [batch, 6]
            eyebrow_vector: [batch, 12]
            mouth_eye_vector_c: compressed (for caching)
            eyebrow_vector_c: compressed (for caching)
            ratio: GPU cache hit ratio

        Returns:
            output_image: [batch, 4, 512, 512]
        """
        if args.perf == 'model':
            tic = time.perf_counter()

        # Call THA4 wrapper
        output_image = self.wrapper.forward(
            image, mouth_eye_vector, pose_vector, eyebrow_vector,
            mouth_eye_vector_c, eyebrow_vector_c, ratio
        )

        if args.perf == 'model':
            print(" - tha4_inference", (time.perf_counter() - tic) * 1000)

        return output_image

    def to(self, device):
        """Move model to specified device"""
        self.wrapper.to(device)
        return super().to(device)