### Point Transformer 代码实现与解析 #### 输入数据准备 Point Transformer接收点云中的每个点的位置信息和特征信息作为输入。位置信息通常由三维坐标组成，而特征信息可能来自先前的预处理步骤。 ```python import torch import torch.nn as nn import torch.nn.functional as F class PositionEmbedding(nn.Module): """Position Embedding模块用于将点的位置信息转换为高维空间""" def __init__(self, input_dim=3, output_dim=64): super(PositionEmbedding, self).__init__() self.fc = nn.Linear(input_dim, output_dim) def forward(self, pos): return self.fc(pos) # 将位置信息映射到更高维度的空间中[^2] ``` #### 多头自注意力机制层设计 多头自注意机制允许模型并行关注不同子空间内的信息交互。对于每一个查询点`xi`及其邻域内所有点`xj`，计算它们之间的重要性分数，并据此调整特征表达。 ```python class MultiHeadAttention(nn.Module): """Multi-head Attention Layer""" def __init__(self, d_model=64, num_heads=8): super(MultiHeadAttention, self).__init__() assert d_model % num_heads == 0 self.d_k = d_model // num_heads self.num_heads = num_heads self.q_linear = nn.Linear(d_model, d_model) self.k_linear = nn.Linear(d_model, d_model) self.v_linear = nn.Linear(d_model, d_model) self.out_proj = nn.Linear(d_model, d_model) def forward(self, query, key, value, delta): batch_size = query.size(0) q = self.q_linear(query).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2) k = self.k_linear(key).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2) v = self.v_linear(value).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2) delta_expanded = delta.unsqueeze(-2).expand_as(k) scores = (q @ k.transpose(-2, -1)) / math.sqrt(self.d_k) + gamma(delta_expanded) attn_weights = F.softmax(scores, dim=-1) context = attn_weights @ v concat_context = context.transpose(1, 2).contiguous().view(batch_size, -1, self.num_heads * self.d_k) output = self.out_proj(concat_context) return output # 输出新的特征表示[^4] ``` #### 整合Transformer结构 最终构建完整的Point Transformer框架，包括但不限于上述组件以及必要的前馈神经网络（Feed Forward Network）、残差连接（Residual Connection）等辅助部分。 ```python class PointTransformerLayer(nn.Module): def __init__(self, embed_dim=64, mlp_ratio=2., drop_rate=.1, num_heads=8): super(PointTransformerLayer, self).__init__() hidden_features = int(embed_dim * mlp_ratio) self.attn = MultiHeadAttention(embed_dim, num_heads=num_heads) self.norm1 = nn.LayerNorm(embed_dim) self.drop_path = DropPath(drop_rate) if drop_rate > 0. else nn.Identity() self.mlp = Mlp(in_features=embed_dim, hidden_features=hidden_features, out_features=embed_dim, act_layer=GELU, drop=drop_rate) self.norm2 = nn.LayerNorm(embed_dim) def forward(self, x, pos_embd): residual_x = x x = self.norm1(x) x = self.attn(x, x, x, pos_embd) x = residual_x + self.drop_path(x) residual_x = x x = self.norm2(x) x = self.mlp(x) x = residual_x + self.drop_path(x) return x ```