""" Conformer ConvModule. Captures local temporal patterns that Grassmann mixing misses. From Gulati et al. 2020 (arXiv 2005.08100): LayerNorm → Pointwise Conv (GLU) → Depthwise Conv → BatchNorm → SiLU → Pointwise Conv → Dropout This is the key component that makes Conformer superior to pure Transformer/Grassmann for speech — it captures local phonetic patterns within a convolutional receptive field. """ import torch import torch.nn as nn import torch.nn.functional as F class ConvModule(nn.Module): """Conformer convolution module with depthwise separable convolution.""" def __init__(self, d_model: int, kernel_size: int = 31, dropout: float = 0.1): super().__init__() self.ln = nn.LayerNorm(d_model) # Pointwise expansion with GLU gating self.pw1 = nn.Conv1d(d_model, 2 * d_model, kernel_size=1) # Depthwise convolution (each channel independently) self.dw = nn.Conv1d( d_model, d_model, kernel_size, padding=kernel_size // 2, groups=d_model, ) self.bn = nn.BatchNorm1d(d_model) # Pointwise projection back self.pw2 = nn.Conv1d(d_model, d_model, kernel_size=1) self.drop = nn.Dropout(dropout) def forward(self, x: torch.Tensor) -> torch.Tensor: """ Args: x: (batch, seq_len, d_model) Returns: (batch, seq_len, d_model) """ # LayerNorm → transpose for Conv1d y = self.ln(x).transpose(1, 2) # (B, d, T) # Pointwise expansion + GLU gate y = F.glu(self.pw1(y), dim=1) # (B, d, T) # Depthwise conv + BatchNorm + SiLU y = F.silu(self.bn(self.dw(y))) # (B, d, T) # Pointwise projection + dropout y = self.drop(self.pw2(y)) # (B, d, T) return y.transpose(1, 2) # (B, T, d)