import torch import torch.nn as nn import torch.nn.functional as F class RMSNorm(nn.Module): """兼容 PyTorch < 2.4 的 RMSNorm 实现""" def __init__(self, dim: int, eps: float = 1e-8): super().__init__() self.scale = nn.Parameter(torch.ones(dim)) self.eps = eps def forward(self, x: torch.Tensor) -> torch.Tensor: rms = x.pow(2).mean(-1, keepdim=True).add(self.eps).sqrt() return x / rms * self.scale class SwiGLU(nn.Module): def __init__(self, in_features, out_features, expansion_factor=2.67, bias=True, dropout=0.3): super(SwiGLU, self).__init__() hidden_features = int(out_features * expansion_factor) self.W1 = nn.Linear(in_features, hidden_features, bias=bias) self.W2 = nn.Linear(in_features, hidden_features, bias=bias) self.W3 = nn.Linear(hidden_features, out_features, bias=bias) self.dropout = nn.Dropout(dropout) def forward(self, x): x1 = self.W1(x) x2 = self.W2(x) x = F.silu(x1) * x2 x = self.dropout(x) x = self.W3(x) return x class FFN(nn.Module): def __init__(self, n_layers=3, model_dim=256, expansion_factor=2.67, bias=True, dropout=0.3): super(FFN, self).__init__() self.n_layers = n_layers self.layers = nn.ModuleList([ SwiGLU(model_dim, model_dim, expansion_factor, bias, dropout) for _ in range(n_layers) ]) self.norms = nn.ModuleList([ RMSNorm(model_dim) for _ in range(n_layers) ]) def forward(self, x): for layer, norm in zip(self.layers, self.norms): x = norm(layer(x) + x) return x class Hierarchical_Decoder(nn.Module): def __init__(self, input_dim=20, hidden_dim=256, output_num=[3, 24, 137], is_hierarchical=True, dropout=0.1, apply_softmax=False): """ Args: input_dim: Encoder 传过来的隐向量维度 (Latent dim) hidden_dim: Decoder 内部 MLPs 的隐藏层维度 output_num: 分类树各层级的类别数量列表 is_hierarchical: 开关。True 为串联分层+残差，False 为完全并列 apply_softmax: 是否在最后一层应用 Softmax (注: 若使用 nn.CrossEntropyLoss，此处应保持 False 输出 Logits) """ super(Hierarchical_Decoder, self).__init__() self.output_num = output_num self.is_hierarchical = is_hierarchical self.apply_softmax = apply_softmax # 1. 维度投影层：将输入的隐向量对齐到隐藏层维度 self.input_proj = nn.Linear(input_dim, hidden_dim) if input_dim != hidden_dim else nn.Identity() # 2. 核心网络：N 个 MLP 块组成的 ModuleList self.decoders = nn.ModuleList([ SwiGLU(hidden_dim, hidden_dim, dropout=dropout) for _ in range(len(output_num)) ]) # 3. 分类头：N 个 Linear 层组成的 ModuleList self.heads = nn.ModuleList([ nn.Linear(hidden_dim, out_classes) for out_classes in output_num ]) def forward(self, x): # 投影到一致的 hidden_dim x = self.input_proj(x) outputs = [] # 初始化当前特征为原始输入 curr_feat = x for i in range(len(self.output_num)): if self.is_hierarchical: # --------------------------------------------------------- # 【分层模式 (Hierarchical + Residual)】 # 当前特征进入第 i 层的 MLP mlp_out = self.decoders[i](curr_feat) # 核心设计：残差连接 (Residual) # 新特征 = 提取的层级特征 + 原始/上一层特征 # 这样既有分层的深度概念，又并行保留了原始信息 curr_feat = curr_feat + mlp_out # 将融合后的特征输入到分类头 head_input = curr_feat # --------------------------------------------------------- else: # --------------------------------------------------------- # 【完全并行模式 (Flat / Parallel)】 # 所有的 MLP 都只看最初始的投影输入 x，互不干扰 mlp_out = self.decoders[i](x) head_input = mlp_out # --------------------------------------------------------- # 分类头输出 (Logits) logits = self.heads[i](head_input) # 根据需求决定是否加 Softmax if self.apply_softmax: logits = torch.softmax(logits, dim=-1) outputs.append(logits) return outputs class mjm(nn.Module): def __init__(self, input_dim=180, latent_dim=20, e_layers=3, d_layers=1, enc_hidden_dim=256, dec_hidden_dim=256, expansion_factor=2.67, dropout=0.3, output_num=[3, 24, 137], is_hierarchical=True, ): super().__init__() self.encoder = nn.Sequential( nn.Linear(input_dim, enc_hidden_dim), FFN(n_layers=e_layers, model_dim=enc_hidden_dim, expansion_factor=expansion_factor, dropout=dropout), nn.Linear(enc_hidden_dim, latent_dim), ) self.recon_decoder = SwiGLU(latent_dim, input_dim) self.decoder = Hierarchical_Decoder( input_dim=latent_dim, hidden_dim=dec_hidden_dim, output_num=output_num, is_hierarchical=is_hierarchical, dropout=dropout, ) def forward(self, x): z = self.encoder(x) recon = self.recon_decoder(z) class_outputs = self.decoder(z) return recon, class_outputs, z # ============================================================ # mjm_1: 三级级联编解码器 + 跨层残差分类 # ============================================================ # 内部工具函数：构建一个「线性投影 + FFN」解码块 def _make_dec_block(in_dim: int, out_dim: int, n_layers: int, expansion_factor: float, dropout: float) -> nn.Sequential: return nn.Sequential( nn.Linear(in_dim, out_dim), FFN(n_layers=n_layers, model_dim=out_dim, expansion_factor=expansion_factor, dropout=dropout), ) class mjm_1(nn.Module): """ mjm_1：三级级联编解码器，层间残差分类 架构示意 ───────────────────────────────────────────────────────────── X [B, input_dim] │ ├─ E1 (Linear + FFN×e_layers) ──────────────────────────────┐ │ h1 [B, enc_hidden_dim] D1 (Linear + FFN×d_layers) ├─ E2 (FFN×e_layers) ───────────────────────────────────┐ C1 = D1(h1) → logits1 │ h2 [B, enc_hidden_dim] D2 └─ E3 (FFN×e_layers + Linear) ─> H [B, latent_dim] C2 = D2(h2) ⊕ residual(C1/logits1) │ → logits2 ├─ recon_decoder (SwiGLU) ─> x_hat D3 C3 = D3(H) ⊕ residual(C2/logits2) → logits3 ───────────────────────────────────────────────────────────── 残差模式 (residual_mode)： 'feature' : C_{i+1} = D_{i+1}(...) + C_i 在 Decoder 隐层特征空间直接残差（维度恒为 dec_hidden_dim） 'logit' : C_{i+1} = D_{i+1}(...) + proj(logits_i) 将上一级 logits 投影回 dec_hidden_dim 后残差； 上级预测信号显式注入下级，信息约束更强 'none' : 无残差，三路 Decoder 完全独立 参数说明： input_dim: 输入基因表达维度 latent_dim: 隐空间维度（E3 输出） e_layers: 每个 Encoder FFN 块的层数（Encoder 宜重） d_layers: 每个 Decoder FFN 块的层数（Decoder 宜轻） enc_hidden_dim: E1/E2 的隐藏层维度（建议 ≥ dec_hidden_dim） dec_hidden_dim: D1/D2/D3 的输出特征维度（C 的维度） expansion_factor: FFN SwiGLU 扩张比 dropout: Dropout 概率 output_num: 三级类别数 [n1, n2, n3] 顺序：[C1(E1出口), C2(E2出口), C3(E3/H出口)] 示例：[3, 24, 137] → Class / Subclass / Supertype residual_mode: 层间残差方式，见上方说明 返回 (forward)： recon: 重构基因表达 [B, input_dim] class_outputs: 三级 logits 列表 [[B,n1], [B,n2], [B,n3]] H: 隐向量 [B, latent_dim] """ def __init__(self, input_dim=180, latent_dim=20, e_layers=3, d_layers=1, enc_hidden_dim=256, dec_hidden_dim=128, expansion_factor=2.67, dropout=0.3, output_num=[3, 24, 137], residual_mode='feature', spatial_dim=0, ): super().__init__() assert len(output_num) == 3, "output_num 须含 3 个元素，对应 C1/C2/C3" assert residual_mode in ('feature', 'logit', 'none'), \ "residual_mode 须为 'feature' | 'logit' | 'none'" self.residual_mode = residual_mode self.input_dim = input_dim self.spatial_dim = spatial_dim # ── Encoders（重）：逐级压缩基因特征 ────────────────────────────────── self.E1 = nn.Sequential( nn.Linear(input_dim + spatial_dim, enc_hidden_dim), FFN(n_layers=e_layers, model_dim=enc_hidden_dim, expansion_factor=expansion_factor, dropout=dropout), ) self.E2 = FFN(n_layers=e_layers, model_dim=enc_hidden_dim, expansion_factor=expansion_factor, dropout=dropout) self.E3 = nn.Sequential( FFN(n_layers=e_layers, model_dim=enc_hidden_dim, expansion_factor=expansion_factor, dropout=dropout), nn.Linear(enc_hidden_dim, latent_dim), ) # ── Decoders（轻）：各层级特征解码 ─────────────────────────────────── self.D1 = _make_dec_block(enc_hidden_dim, dec_hidden_dim, d_layers, expansion_factor, dropout) self.D2 = _make_dec_block(enc_hidden_dim, dec_hidden_dim, d_layers, expansion_factor, dropout) self.D3 = _make_dec_block(latent_dim, dec_hidden_dim, d_layers, expansion_factor, dropout) # ── 残差投影（仅 'logit' 模式）────────────────────────────────────── if residual_mode == 'logit': self.proj_c1_to_c2 = nn.Linear(output_num[0], dec_hidden_dim) self.proj_c2_to_c3 = nn.Linear(output_num[1], dec_hidden_dim) # ── Classification heads ──────────────────────────────────────────── self.head1 = nn.Linear(dec_hidden_dim, output_num[0]) self.head2 = nn.Linear(dec_hidden_dim, output_num[1]) self.head3 = nn.Linear(dec_hidden_dim, output_num[2]) # ── Reconstruction decoder ────────────────────────────────────────── self.recon_decoder = SwiGLU(latent_dim, input_dim, dropout=dropout) def forward(self, x): # ── 编码 ───────────────────────────────────────────────────────────── h1 = self.E1(x) # [B, enc_hidden_dim] h2 = self.E2(h1) # [B, enc_hidden_dim] H = self.E3(h2) # [B, latent_dim] # ── 解码 + 分类（含层间残差）───────────────────────────────────────── C1 = self.D1(h1) # [B, dec_hidden_dim] logits1 = self.head1(C1) # [B, n1] if self.residual_mode == 'feature': C2 = self.D2(h2) + C1 elif self.residual_mode == 'logit': C2 = self.D2(h2) + self.proj_c1_to_c2(logits1) else: C2 = self.D2(h2) logits2 = self.head2(C2) # [B, n2] if self.residual_mode == 'feature': C3 = self.D3(H) + C2 elif self.residual_mode == 'logit': C3 = self.D3(H) + self.proj_c2_to_c3(logits2) else: C3 = self.D3(H) logits3 = self.head3(C3) # [B, n3] # ── 重构 ───────────────────────────────────────────────────────────── recon = self.recon_decoder(H) # [B, input_dim] return recon, [logits1, logits2, logits3], H