NanoSocrates/Project_Model/Libs/Transformer/Classes/Decoder.py

from typing import Optional
import torch
import torch.nn as nn
from .FeedForwardNetwork import FeedForwardNetwork
from .TorchMultiHeadAttention import TorchMultiHeadAttention as MultiHeadAttention
from ..Utils.attention_mask import get_causal_attention_mask, get_prefix_causal_mask_from_padding_mask

# B, L(T), E_D


class Decoder(nn.Module):

    def __init__(
        self,
        embedding_dimension: int,
        feed_forward_hidden_layer_dimension: int,
        number_of_attention_heads: int,
    ) -> None:
        self.__attention_heads = number_of_attention_heads
        super().__init__()


        self.__masked_attention = MultiHeadAttention(
            embedding_dimension, number_of_attention_heads, dropout=0.1
        )

        self.__layer_norm_1 = nn.LayerNorm(embedding_dimension)

        self.__cross_attention = MultiHeadAttention(
            embedding_dimension, number_of_attention_heads, dropout=0.1
        )
        self.__layer_norm_2 = nn.LayerNorm(embedding_dimension)

        self.__dropout = nn.Dropout(0.1)

        self.__feed_forward_network = FeedForwardNetwork(
            embedding_dimension, feed_forward_hidden_layer_dimension
        )
        self.__layer_norm_3 = nn.LayerNorm(embedding_dimension)

    def forward(
        self,
        args: tuple[
            torch.Tensor,
            torch.Tensor,
            torch.Tensor,
            torch.Tensor,
            torch.Tensor,
            Optional[bool]
        ]
    ):  # -> list[torch.Tensor]:  # k_x = v_x . While x_q = x
        # WARNING: args is needed to have sequential
        if len(args) < 6:
            args = args + (False)
        x, k_x, v_x, src_padding_mask, tgt_padding_mask, decoder_only = args

        # build of attention mask
        # TODO: create a prefix causal mask if needed
        if decoder_only:
            attention_mask = get_prefix_causal_mask_from_padding_mask(x.size(1),src_padding_mask,self.__attention_heads) # the correct is tgt however ...
        else:
            attention_mask = get_causal_attention_mask(x.size(1))

        # 1) Masked Attention
        MASKED_ATTENTION = self.__masked_attention(
            x, x, x, key_padding_mask=tgt_padding_mask, attention_mask=attention_mask
        )

        # 2) Dropout
        DROPPED_MASKED_ATTENTION = self.__dropout(MASKED_ATTENTION)
        del MASKED_ATTENTION

        # 3) Residual Connection
        x = x + DROPPED_MASKED_ATTENTION
        del DROPPED_MASKED_ATTENTION

        # 4) Layer Normalization
        x = self.__layer_norm_1(x)


        if not decoder_only:
            # 5) Encoder–decoder (cross) attention
            CROSS_ATTENTION = self.__cross_attention(
                x, k_x, v_x, key_padding_mask=src_padding_mask
            )

            # 6) Dropout
            DROPPED_CROSS_ATTENTION = self.__dropout(CROSS_ATTENTION)
            del CROSS_ATTENTION

            # 7) Residual Connection
            x = x + DROPPED_CROSS_ATTENTION
            del DROPPED_CROSS_ATTENTION

            # 8) Layer Normalization
            x = self.__layer_norm_2(x)

        # 9) Position-wise feed-forward
        FEED_FORWARD = self.__feed_forward_network(x)

        # 10) Dropout
        DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
        del FEED_FORWARD

        # 11) Residual Connection
        x = x + DROPPED_FEED_FORWARD
        del DROPPED_FEED_FORWARD

        # 12) Layer Normalization
        x = self.__layer_norm_3(x)

        return (x, k_x, v_x, src_padding_mask, tgt_padding_mask, decoder_only)


# use eval to disable dropout ecc