encoder.py

import torch

from .layernorm import LayerNorm
from .multi_head_attention import MultiHeadAttention
from .projection import FeedForwardNetwork


class EncoderLayer(torch.nn.Module):
    

    def __init__(self, d_model: int, num_heads: int, d_ff: int = 2048):
        """
        Initializes the EncoderLayer module.

        Args:
            d_model (int): The dimensionality of the input and output.
            num_heads (int): The number of heads in the multi-head attention.
            d_ff (int, optional): The dimensionality of the inner-layer of the
                feed-forward network. Defaults to 2048.
        """
        super().__init__()
        self.mha = MultiHeadAttention(
            d_model=d_model, num_heads=num_heads
        )
        self.layer_norm_1 = LayerNorm(d_model)
        self.ffn = FeedForwardNetwork(d_model=d_model, d_ff=d_ff)
        self.layer_norm_2 = LayerNorm(d_model)

    def forward(self, in_embeddings: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for the EncoderLayer.

        Args:
            in_embeddings (torch.Tensor): The input tensor of shape
                (batch_size, seq_len, d_model).

        Returns:
            torch.Tensor: The output tensor of the same shape as the input.
        """
        # Multi-Head Attention sub-layer
        mha_out = self.mha(in_embeddings, in_embeddings, in_embeddings)
        # print(mha_out.min(), mha_out.max())
        # Residual connection and layer normalization
        norm_out_1 = self.layer_norm_1(mha_out + in_embeddings)

        # Feed-Forward Network sub-layer
        ffn_out = self.ffn(norm_out_1)
        # Residual connection and layer normalization
        output = self.layer_norm_2(ffn_out + norm_out_1)

        return output