File size: 9,605 Bytes
de2377a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
# motion_generation/lib/model/unet1d.py

import torch
import torch.nn as nn
import math

class PositionalEncoding(nn.Module):
    """
    用于编码时间步 t 的标准 Transformer 位置编码
    """
    def __init__(self, dim):
        super().__init__()
        self.dim = dim
        self.register_buffer('inv_freq', 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim)))

    def forward(self, x):
        # x 形状: (B,) 时间步索引
        sinusoid_inp = torch.einsum('i, j -> i j', x.float(), self.inv_freq)
        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
        return emb

class ResBlock1D(nn.Module):
    """
    一维残差块,包含 time embedding 的融合
    """
    def __init__(self, in_channels, out_channels, time_dim, kernel_size=3):
        super().__init__()
        padding = kernel_size // 2

        self.conv1 = nn.Conv1d(in_channels, out_channels, kernel_size, padding=padding)
        self.bn1 = nn.BatchNorm1d(out_channels)
        self.act1 = nn.GELU()

        self.conv2 = nn.Conv1d(out_channels, out_channels, kernel_size, padding=padding)
        self.bn2 = nn.BatchNorm1d(out_channels)
        self.act2 = nn.GELU()

        # 时间步嵌入层
        self.time_proj = nn.Linear(time_dim, out_channels)

        # 确保输入/输出通道匹配
        self.residual_conv = nn.Conv1d(in_channels, out_channels, 1) if in_channels != out_channels else nn.Identity()

    def forward(self, x, time_emb):
        # x: (B, C_in, L) L=72, C_in=通道数

        h = self.conv1(x)
        h = self.bn1(h)
        h = self.act1(h)

        # 融合时间嵌入:沿特征维度广播并相加
        time_emb_proj = self.time_proj(time_emb).unsqueeze(-1) # (B, C_out) -> (B, C_out, 1)
        h = h + time_emb_proj

        h = self.conv2(h)
        h = self.bn2(h)
        h = self.act2(h)

        return h + self.residual_conv(x)


class SelfAttention1D(nn.Module):
    """
    一维自注意力模块,包含 time embedding 的融合
    """
    def __init__(self, channels, time_dim, num_heads=4):
        super().__init__()
        self.channels = channels
        self.num_heads = num_heads
        assert channels % num_heads == 0, "channels must be divisible by num_heads"

        # Group normalization for better stability
        self.norm = nn.GroupNorm(num_groups=8, num_channels=channels)

        # Multi-head attention components
        self.qkv = nn.Conv1d(channels, channels * 3, 1)
        self.proj = nn.Conv1d(channels, channels, 1)

        # Time embedding projection
        self.time_proj = nn.Linear(time_dim, channels)

    def forward(self, x, time_emb):
        # x: (B, C, L)
        B, C, L = x.shape

        # Normalize input
        h = self.norm(x)

        # Add time embedding
        time_emb_proj = self.time_proj(time_emb).unsqueeze(-1)  # (B, C, 1)
        h = h + time_emb_proj

        # Compute Q, K, V
        qkv = self.qkv(h)  # (B, 3*C, L)
        qkv = qkv.reshape(B, 3, self.num_heads, C // self.num_heads, L)
        qkv = qkv.permute(1, 0, 2, 4, 3)  # (3, B, num_heads, L, head_dim)
        q, k, v = qkv[0], qkv[1], qkv[2]

        # Attention
        scale = (C // self.num_heads) ** -0.5
        attn = torch.matmul(q, k.transpose(-2, -1)) * scale  # (B, num_heads, L, L)
        attn = torch.softmax(attn, dim=-1)

        # Apply attention to values
        out = torch.matmul(attn, v)  # (B, num_heads, L, head_dim)
        out = out.permute(0, 1, 3, 2).reshape(B, C, L)  # (B, C, L)

        # Project and add residual
        out = self.proj(out)
        return out + x


class UNet1D(nn.Module):
    def __init__(self, pose_dim=72, base_channels=128, channel_multipliers=[1, 2, 4],
                 time_emb_dim=256, mid_structure='conv', mid_num_heads=4):
        """
        UNet1D model for 1D sequence processing

        Args:
            pose_dim: Dimension of input pose (default: 72)
            base_channels: Base number of channels (default: 128)
            channel_multipliers: Channel multipliers for each level (default: [1, 2, 4])
            time_emb_dim: Time embedding dimension (default: 256)
            mid_structure: Structure for middle layer, either 'conv' or 'attention' (default: 'conv')
            mid_num_heads: Number of attention heads for mid layer when using attention (default: 4)
        """
        super().__init__()

        self.mid_structure = mid_structure
        assert mid_structure in ['conv', 'attention'], "mid_structure must be 'conv' or 'attention'"

        # 1. Time Embedding
        self.time_mlp = nn.Sequential(
            PositionalEncoding(base_channels),
            nn.Linear(base_channels, time_emb_dim),
            nn.GELU(),
            nn.Linear(time_emb_dim, time_emb_dim)
        )

        # 调整输入通道:姿态向量 (B, 72) -> (B, 1, 72)
        # 我们将特征维度 (72) 视为长度 L,将 1 视为通道 C
        in_channels = 1

        # 2. 编码器 (Downsampling)
        channels = [in_channels] + [base_channels * m for m in channel_multipliers]
        self.downs = nn.ModuleList()

        for i in range(len(channel_multipliers)):
            in_c = channels[i]
            out_c = channels[i+1]

            self.downs.append(nn.ModuleList([
                ResBlock1D(in_c if i == 0 else in_c, out_c, time_emb_dim), # 输入是 1, L 或 C_in, L
                nn.MaxPool1d(2) if i < len(channel_multipliers) - 1 else nn.Identity()
            ]))

        # 3. 中间层 - 根据 mid_structure 选择使用卷积或自注意力
        mid_c = channels[-1]
        if mid_structure == 'conv':
            self.mid = ResBlock1D(mid_c, mid_c, time_emb_dim)
        elif mid_structure == 'attention':
            # 确保通道数能被注意力头数整除
            assert mid_c % mid_num_heads == 0, f"mid_c ({mid_c}) must be divisible by mid_num_heads ({mid_num_heads})"
            self.mid = SelfAttention1D(mid_c, time_emb_dim, num_heads=mid_num_heads)
        
        # 4. 解码器 (Upsampling)
        self.ups = nn.ModuleList()
        reversed_channels = list(reversed(channels))

        for i in range(len(channel_multipliers)):
            in_c = reversed_channels[i]  # 来自上一层的通道数
            # 最后一层输出 base_channels,而不是 in_channels (1)
            out_c = reversed_channels[i+1] if i < len(channel_multipliers) - 1 else base_channels
            skip_c = in_c  # 跳跃连接通道数(来自对应编码器层)

            self.ups.append(nn.ModuleList([
                # ResBlock 接收拼接后的通道: in_c(来自上层) + skip_c(来自编码器)
                # 输出为 out_c 通道
                ResBlock1D(in_c + skip_c, out_c, time_emb_dim),
                # 上采样到下一层的空间尺寸
                nn.ConvTranspose1d(out_c, out_c, kernel_size=2, stride=2) if i < len(channel_multipliers) - 1 else nn.Identity(),
            ]))

        # 5. 输出层 (回到 1 个通道)
        self.out_conv = nn.Conv1d(base_channels, in_channels, kernel_size=1)
        
    def forward(self, x, t):
        # import pdb; pdb.set_trace()
        # x: (B, 72) 归一化姿态,t: (B,) 时间步索引
        x = x.unsqueeze(1) # (B, 1, 72)

        # 1. Time Embedding
        time_emb = self.time_mlp(t)
        
        # 2. 编码器
        skips = []
        for resblock, downsample in self.downs:
            x = resblock(x, time_emb)
            skips.append(x)
            x = downsample(x)

        # 3. 中间层
        x = self.mid(x, time_emb)
        
        # 4. 解码器
        for i, (resblock, upsample) in enumerate(self.ups):
            skip = skips.pop()

            # 跳跃连接
            # 检查维度是否匹配,如果 MaxPool 导致了奇数/偶数长度不匹配,需要裁剪
            if x.shape[-1] != skip.shape[-1]:
                x = nn.functional.pad(x, (0, skip.shape[-1] - x.shape[-1]))

            # 拼接跳跃连接
            x = torch.cat((x, skip), dim=1) # 沿通道维度拼接

            # 处理拼接后的特征
            x = resblock(x, time_emb)

            # 上采样到下一层的空间尺寸(最后一层不上采样)
            if i < len(self.ups) - 1:
                x = upsample(x) # (B, C, L) -> (B, C, L*2)

        # 5. 输出层
        x = self.out_conv(x) # (B, 1, 72)
        
        return x.squeeze(1) # (B, 72)


if __name__ == "__main__":
    # 测试 UNet1D 模型
    print("Testing UNet1D with conv mid structure...")
    model_conv = UNet1D(mid_structure='conv')

    x = torch.randn(4, 72) # 模拟输入 (B, 72)
    t = torch.randint(0, 1000, (4,)) # 模拟时间步索引 (B,)

    output_conv = model_conv(x, t)
    print(f"Input shape: {x.shape}")
    print(f"Output shape (conv): {output_conv.shape}")
    print(f"Expected output shape: (4, 72)")
    print(f"Test passed (conv): {output_conv.shape == torch.Size([4, 72])}")

    print("\nTesting UNet1D with attention mid structure...")
    model_attn = UNet1D(mid_structure='attention', base_channels=128, mid_num_heads=8)

    output_attn = model_attn(x, t)
    print(f"Output shape (attention): {output_attn.shape}")
    print(f"Test passed (attention): {output_attn.shape == torch.Size([4, 72])}")

    # 统计参数量
    conv_params = sum(p.numel() for p in model_conv.parameters())
    attn_params = sum(p.numel() for p in model_attn.parameters())
    print(f"\nParameters (conv): {conv_params:,}")
    print(f"Parameters (attention): {attn_params:,}")
    print(f"Difference: {abs(attn_params - conv_params):,}")