File size: 8,669 Bytes
d4cbafd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-np.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)

        self.register_buffer('pe', pe)

    def forward(self, x):
        # not used in the final model
        x = x + self.pe[:x.shape[0], :]
        return self.dropout(x)


# only for ablation / not used in the final model
class TimeEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(TimeEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

    def forward(self, x, mask, lengths):
        time = mask * 1/(lengths[..., None]-1)
        time = time[:, None] * torch.arange(time.shape[1], device=x.device)[None, :]
        time = time[:, 0].T
        # add the time encoding
        x = x + time[..., None]
        return self.dropout(x)


class Encoder_TRANSFORMER(nn.Module):
    def __init__(self, modeltype, njoints, nfeats, num_frames, num_classes, translation, pose_rep, glob, glob_rot,
                 latent_dim=256, ff_size=1024, num_layers=4, num_heads=4, dropout=0.1,
                 ablation=None, activation="gelu", **kargs):
        super().__init__()

        self.modeltype = modeltype
        self.njoints = njoints
        self.nfeats = nfeats
        self.num_frames = num_frames
        self.num_classes = num_classes

        self.pose_rep = pose_rep
        self.glob = glob
        self.glob_rot = glob_rot
        self.translation = translation

        self.latent_dim = latent_dim

        self.ff_size = ff_size
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.dropout = dropout

        self.ablation = ablation
        self.activation = activation

        self.input_feats = self.njoints*self.nfeats

        if self.ablation == "average_encoder":
            self.mu_layer = nn.Linear(self.latent_dim, self.latent_dim)
            self.sigma_layer = nn.Linear(self.latent_dim, self.latent_dim)
        else:
            self.muQuery = nn.Parameter(torch.randn(self.num_classes, self.latent_dim))
            self.sigmaQuery = nn.Parameter(torch.randn(self.num_classes, self.latent_dim))

        self.skelEmbedding = nn.Linear(self.input_feats, self.latent_dim)

        self.sequence_pos_encoder = PositionalEncoding(self.latent_dim, self.dropout)

        # self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))

        seqTransEncoderLayer = nn.TransformerEncoderLayer(d_model=self.latent_dim,
                                                          nhead=self.num_heads,
                                                          dim_feedforward=self.ff_size,
                                                          dropout=self.dropout,
                                                          activation=self.activation)
        self.seqTransEncoder = nn.TransformerEncoder(seqTransEncoderLayer,
                                                     num_layers=self.num_layers)

    def forward(self, batch):
        x, y, mask = batch["x"], batch["y"], batch["mask"]
        bs, njoints, nfeats, nframes = x.shape
        x = x.permute((3, 0, 1, 2)).reshape(nframes, bs, njoints*nfeats)

        # embedding of the skeleton
        x = self.skelEmbedding(x)

        # only for ablation / not used in the final model
        if self.ablation == "average_encoder":
            # add positional encoding
            x = self.sequence_pos_encoder(x)

            # transformer layers
            final = self.seqTransEncoder(x, src_key_padding_mask=~mask)
            # get the average of the output
            z = final.mean(axis=0)

            # extract mu and logvar
            mu = self.mu_layer(z)
            logvar = self.sigma_layer(z)
        else:
            # adding the mu and sigma queries
            xseq = torch.cat((self.muQuery[y][None], self.sigmaQuery[y][None], x), axis=0)

            # add positional encoding
            xseq = self.sequence_pos_encoder(xseq)

            # create a bigger mask, to allow attend to mu and sigma
            muandsigmaMask = torch.ones((bs, 2), dtype=bool, device=x.device)
            maskseq = torch.cat((muandsigmaMask, mask), axis=1)

            final = self.seqTransEncoder(xseq, src_key_padding_mask=~maskseq)
            mu = final[0]
            logvar = final[1]

        return {"mu": mu, "logvar": logvar}


class Decoder_TRANSFORMER(nn.Module):
    def __init__(self, modeltype, njoints, nfeats, num_frames, num_classes, translation, pose_rep, glob, glob_rot,
                 latent_dim=256, ff_size=1024, num_layers=4, num_heads=4, dropout=0.1, activation="gelu",
                 ablation=None, **kargs):
        super().__init__()

        self.modeltype = modeltype
        self.njoints = njoints
        self.nfeats = nfeats
        self.num_frames = num_frames
        self.num_classes = num_classes

        self.pose_rep = pose_rep
        self.glob = glob
        self.glob_rot = glob_rot
        self.translation = translation

        self.latent_dim = latent_dim

        self.ff_size = ff_size
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.dropout = dropout

        self.ablation = ablation

        self.activation = activation

        self.input_feats = self.njoints*self.nfeats

        # only for ablation / not used in the final model
        if self.ablation == "zandtime":
            self.ztimelinear = nn.Linear(self.latent_dim + self.num_classes, self.latent_dim)
        else:
            self.actionBiases = nn.Parameter(torch.randn(self.num_classes, self.latent_dim))

        # only for ablation / not used in the final model
        if self.ablation == "time_encoding":
            self.sequence_pos_encoder = TimeEncoding(self.dropout)
        else:
            self.sequence_pos_encoder = PositionalEncoding(self.latent_dim, self.dropout)

        seqTransDecoderLayer = nn.TransformerDecoderLayer(d_model=self.latent_dim,
                                                          nhead=self.num_heads,
                                                          dim_feedforward=self.ff_size,
                                                          dropout=self.dropout,
                                                          activation=activation)
        self.seqTransDecoder = nn.TransformerDecoder(seqTransDecoderLayer,
                                                     num_layers=self.num_layers)

        self.finallayer = nn.Linear(self.latent_dim, self.input_feats)

    def forward(self, batch):
        z, y, mask, lengths = batch["z"], batch["y"], batch["mask"], batch["lengths"]

        latent_dim = z.shape[1]
        bs, nframes = mask.shape
        njoints, nfeats = self.njoints, self.nfeats

        # only for ablation / not used in the final model
        if self.ablation == "zandtime":
            yoh = F.one_hot(y, self.num_classes)
            z = torch.cat((z, yoh), axis=1)
            z = self.ztimelinear(z)
            z = z[None]  # sequence of size 1
        else:
            # only for ablation / not used in the final model
            if self.ablation == "concat_bias":
                # sequence of size 2
                z = torch.stack((z, self.actionBiases[y]), axis=0)
            else:
                # shift the latent noise vector to be the action noise
                z = z + self.actionBiases[y]
                z = z[None]  # sequence of size 1

        timequeries = torch.zeros(nframes, bs, latent_dim, device=z.device)

        # only for ablation / not used in the final model
        if self.ablation == "time_encoding":
            timequeries = self.sequence_pos_encoder(timequeries, mask, lengths)
        else:
            timequeries = self.sequence_pos_encoder(timequeries)

        output = self.seqTransDecoder(tgt=timequeries, memory=z,
                                      tgt_key_padding_mask=~mask)

        output = self.finallayer(output).reshape(nframes, bs, njoints, nfeats)

        # zero for padded area
        output[~mask.T] = 0
        output = output.permute(1, 2, 3, 0)

        batch["output"] = output
        return batch