File size: 6,072 Bytes
e68ca33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn.functional as F
from torch import nn
from src.audio2pose_models.res_unet import ResUnet

def class2onehot(idx, class_num):

    assert torch.max(idx).item() < class_num
    onehot = torch.zeros(idx.size(0), class_num).to(idx.device)
    onehot.scatter_(1, idx, 1)
    return onehot

class CVAE(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        encoder_layer_sizes = cfg.MODEL.CVAE.ENCODER_LAYER_SIZES
        decoder_layer_sizes = cfg.MODEL.CVAE.DECODER_LAYER_SIZES
        latent_size = cfg.MODEL.CVAE.LATENT_SIZE
        num_classes = cfg.DATASET.NUM_CLASSES
        audio_emb_in_size = cfg.MODEL.CVAE.AUDIO_EMB_IN_SIZE
        audio_emb_out_size = cfg.MODEL.CVAE.AUDIO_EMB_OUT_SIZE
        seq_len = cfg.MODEL.CVAE.SEQ_LEN

        self.latent_size = latent_size

        self.encoder = ENCODER(encoder_layer_sizes, latent_size, num_classes,
                                audio_emb_in_size, audio_emb_out_size, seq_len)
        self.decoder = DECODER(decoder_layer_sizes, latent_size, num_classes,
                                audio_emb_in_size, audio_emb_out_size, seq_len)
    def reparameterize(self, mu, logvar):
        std = torch.exp(0.5 * logvar)
        eps = torch.randn_like(std)
        return mu + eps * std

    def forward(self, batch):
        batch = self.encoder(batch)
        mu = batch['mu']
        logvar = batch['logvar']
        z = self.reparameterize(mu, logvar)
        batch['z'] = z
        return self.decoder(batch)

    def test(self, batch):
        '''
        class_id = batch['class']
        z = torch.randn([class_id.size(0), self.latent_size]).to(class_id.device)
        batch['z'] = z
        '''
        return self.decoder(batch)

class ENCODER(nn.Module):
    def __init__(self, layer_sizes, latent_size, num_classes, 
                audio_emb_in_size, audio_emb_out_size, seq_len):
        super().__init__()

        self.resunet = ResUnet()
        self.num_classes = num_classes
        self.seq_len = seq_len

        self.MLP = nn.Sequential()
        layer_sizes[0] += latent_size + seq_len*audio_emb_out_size + 6
        for i, (in_size, out_size) in enumerate(zip(layer_sizes[:-1], layer_sizes[1:])):
            self.MLP.add_module(
                name="L{:d}".format(i), module=nn.Linear(in_size, out_size))
            self.MLP.add_module(name="A{:d}".format(i), module=nn.ReLU())

        self.linear_means = nn.Linear(layer_sizes[-1], latent_size)
        self.linear_logvar = nn.Linear(layer_sizes[-1], latent_size)
        self.linear_audio = nn.Linear(audio_emb_in_size, audio_emb_out_size)

        self.classbias = nn.Parameter(torch.randn(self.num_classes, latent_size))

    def forward(self, batch):
        class_id = batch['class']
        pose_motion_gt = batch['pose_motion_gt']                             #bs seq_len 6
        ref = batch['ref']                             #bs 6
        bs = pose_motion_gt.shape[0]
        audio_in = batch['audio_emb']                          # bs seq_len audio_emb_in_size

        #pose encode
        pose_emb = self.resunet(pose_motion_gt.unsqueeze(1))          #bs 1 seq_len 6 
        pose_emb = pose_emb.reshape(bs, -1)                    #bs seq_len*6

        #audio mapping
        print(audio_in.shape)
        audio_out = self.linear_audio(audio_in)                # bs seq_len audio_emb_out_size
        audio_out = audio_out.reshape(bs, -1)

        class_bias = self.classbias[class_id]                  #bs latent_size
        x_in = torch.cat([ref, pose_emb, audio_out, class_bias], dim=-1) #bs seq_len*(audio_emb_out_size+6)+latent_size
        x_out = self.MLP(x_in)

        mu = self.linear_means(x_out)
        logvar = self.linear_means(x_out)                      #bs latent_size 

        batch.update({'mu':mu, 'logvar':logvar})
        return batch

class DECODER(nn.Module):
    def __init__(self, layer_sizes, latent_size, num_classes, 
                audio_emb_in_size, audio_emb_out_size, seq_len):
        super().__init__()

        self.resunet = ResUnet()
        self.num_classes = num_classes
        self.seq_len = seq_len

        self.MLP = nn.Sequential()
        input_size = latent_size + seq_len*audio_emb_out_size + 6
        for i, (in_size, out_size) in enumerate(zip([input_size]+layer_sizes[:-1], layer_sizes)):
            self.MLP.add_module(
                name="L{:d}".format(i), module=nn.Linear(in_size, out_size))
            if i+1 < len(layer_sizes):
                self.MLP.add_module(name="A{:d}".format(i), module=nn.ReLU())
            else:
                self.MLP.add_module(name="sigmoid", module=nn.Sigmoid())
        
        self.pose_linear = nn.Linear(6, 6)
        self.linear_audio = nn.Linear(audio_emb_in_size, audio_emb_out_size)

        self.classbias = nn.Parameter(torch.randn(self.num_classes, latent_size))

    def forward(self, batch):

        z = batch['z']                                          #bs latent_size
        bs = z.shape[0]
        class_id = batch['class']
        ref = batch['ref']                             #bs 6
        audio_in = batch['audio_emb']                           # bs seq_len audio_emb_in_size
        #print('audio_in: ', audio_in[:, :, :10])

        audio_out = self.linear_audio(audio_in)                 # bs seq_len audio_emb_out_size
        #print('audio_out: ', audio_out[:, :, :10])
        audio_out = audio_out.reshape([bs, -1])                 # bs seq_len*audio_emb_out_size
        class_bias = self.classbias[class_id]                   #bs latent_size

        z = z + class_bias
        x_in = torch.cat([ref, z, audio_out], dim=-1)
        x_out = self.MLP(x_in)                                  # bs layer_sizes[-1]
        x_out = x_out.reshape((bs, self.seq_len, -1))

        #print('x_out: ', x_out)

        pose_emb = self.resunet(x_out.unsqueeze(1))             #bs 1 seq_len 6

        pose_motion_pred = self.pose_linear(pose_emb.squeeze(1))       #bs seq_len 6

        batch.update({'pose_motion_pred':pose_motion_pred})
        return batch