File size: 10,155 Bytes
73dfb75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
238
239
240
241
242
243
## Our PoseFormer model was revised from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# Written by Ce Zheng (cezheng@knights.ucf.edu)
# Modified by Qitao Zhao (qitaozhao@mail.sdu.edu.cn)

import math
import logging
from functools import partial
from einops import rearrange

import torch
import torch_dct as dct
import torch.nn as nn
import torch.nn.functional as F

import numpy as np
from timm.models.layers import DropPath


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class FreqMlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        b, f, _ = x.shape
        x = dct.dct(x.permute(0, 2, 1)).permute(0, 2, 1).contiguous()
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        x = dct.idct(x.permute(0, 2, 1)).permute(0, 2, 1).contiguous()
        return x


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
        self.scale = qk_scale or head_dim ** -0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Block(nn.Module):

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x):
        x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class MixedBlock(nn.Module):

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp1 = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.norm3 = norm_layer(dim)
        self.mlp2 = FreqMlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x):
        b, f, c = x.shape
        x = x + self.drop_path(self.attn(self.norm1(x)))
        x1 = x[:, :f//2] + self.drop_path(self.mlp1(self.norm2(x[:, :f//2])))
        x2 = x[:, f//2:] + self.drop_path(self.mlp2(self.norm3(x[:, f//2:])))
        return torch.cat((x1, x2), dim=1)


class PoseTransformerV2(nn.Module):
    def __init__(self, num_frame=9, num_joints=17, in_chans=2,
                 num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.2,  norm_layer=None, args=None):
        """    ##########hybrid_backbone=None, representation_size=None,
        Args:
            num_frame (int, tuple): input frame number
            num_joints (int, tuple): joints number
            in_chans (int): number of input channels, 2D joints have 2 channels: (x,y)
            embed_dim_ratio (int): embedding dimension ratio
            depth (int): depth of transformer
            num_heads (int): number of attention heads
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            qkv_bias (bool): enable bias for qkv if True
            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            norm_layer: (nn.Module): normalization layer
        """
        super().__init__()

        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        embed_dim_ratio = args.embed_dim_ratio
        depth = args.depth
        embed_dim = embed_dim_ratio * num_joints   #### temporal embed_dim is num_joints * spatial embedding dim ratio
        out_dim = num_joints * 3    #### output dimension is num_joints * 3
        self.num_frame_kept = args.number_of_kept_frames
        self.num_coeff_kept = args.number_of_kept_coeffs if args.number_of_kept_coeffs else self.num_frame_kept

        ### spatial patch embedding
        self.Joint_embedding = nn.Linear(in_chans, embed_dim_ratio)
        self.Freq_embedding = nn.Linear(in_chans*num_joints, embed_dim)

        self.Spatial_pos_embed = nn.Parameter(torch.zeros(1, num_joints, embed_dim_ratio))
        self.Temporal_pos_embed = nn.Parameter(torch.zeros(1, self.num_frame_kept, embed_dim))
        self.Temporal_pos_embed_ = nn.Parameter(torch.zeros(1, self.num_coeff_kept, embed_dim))
        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule

        self.Spatial_blocks = nn.ModuleList([
            Block(
                dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
            for i in range(depth)])

        self.blocks = nn.ModuleList([
            MixedBlock(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
            for i in range(depth)])

        self.Spatial_norm = norm_layer(embed_dim_ratio)
        self.Temporal_norm = norm_layer(embed_dim)

        ####### A easy way to implement weighted mean
        self.weighted_mean = torch.nn.Conv1d(in_channels=self.num_coeff_kept, out_channels=1, kernel_size=1)
        self.weighted_mean_ = torch.nn.Conv1d(in_channels=self.num_frame_kept, out_channels=1, kernel_size=1)

        self.head = nn.Sequential(
            nn.LayerNorm(embed_dim*2),
            nn.Linear(embed_dim*2, out_dim),
        )

    def Spatial_forward_features(self, x):
        b, f, p, _ = x.shape  ##### b is batch size, f is number of frames, p is number of joints
        num_frame_kept = self.num_frame_kept

        index = torch.arange((f-1)//2-num_frame_kept//2, (f-1)//2+num_frame_kept//2+1)

        x = self.Joint_embedding(x[:, index].view(b*num_frame_kept, p, -1))
        x += self.Spatial_pos_embed
        x = self.pos_drop(x)

        for blk in self.Spatial_blocks:
            x = blk(x)

        x = self.Spatial_norm(x)
        x = rearrange(x, '(b f) p c -> b f (p c)', f=num_frame_kept)
        return x

    def forward_features(self, x, Spatial_feature):
        b, f, p, _ = x.shape
        num_coeff_kept = self.num_coeff_kept

        x = dct.dct(x.permute(0, 2, 3, 1))[:, :, :, :num_coeff_kept]
        x = x.permute(0, 3, 1, 2).contiguous().view(b, num_coeff_kept, -1)
        x = self.Freq_embedding(x) 
        
        Spatial_feature += self.Temporal_pos_embed
        x += self.Temporal_pos_embed_
        x = torch.cat((x, Spatial_feature), dim=1)

        for blk in self.blocks:
            x = blk(x)

        x = self.Temporal_norm(x)
        return x

    def forward(self, x):
        b, f, p, _ = x.shape
        x_ = x.clone()

        Spatial_feature = self.Spatial_forward_features(x)
        x = self.forward_features(x_, Spatial_feature)
        x = torch.cat((self.weighted_mean(x[:, :self.num_coeff_kept]), self.weighted_mean_(x[:, self.num_coeff_kept:])), dim=-1)

        x = self.head(x).view(b, 1, p, -1)
        return x