Model.py

"""
Implementation of "Efficient Message-passing Transformer for Error Correcting Codes" (EfficientMPT)
Information and Intelligence Lab (IIL)
Department of Electrical Engineering, Graduate School of Artificial Intelligence
Pohang University of Science and Technology (POSTECH), South Korea.
@author: Seong-Joon Park, joonpark2247@gmail.com
"""

from torch.nn import LayerNorm
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import copy
import logging
from Codes import sign_to_bin
import numpy as np


def clones(module, N):
    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])


class Encoder(nn.Module):
    def __init__(self, layer, N):
        super(Encoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)
        if N > 1:
            self.norm2 = LayerNorm(layer.size)

    def forward(self, x, x2, mask):
        for idx, layer in enumerate(self.layers, start=1):
            x = layer(x, x2, mask.transpose(0,1))
            x2 = layer(x2, x, mask)
            if idx == len(self.layers)//2 and len(self.layers) > 1:
                x = self.norm2(x)
                x2 = self.norm2(x2)
        return self.norm(x), self.norm(x2)


class SublayerConnection(nn.Module):
    def __init__(self, size, dropout):
        super(SublayerConnection, self).__init__()
        self.norm = LayerNorm(size)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, sublayer):
        return x + self.dropout(sublayer(self.norm(x)))


class EncoderLayer(nn.Module):
    def __init__(self, size, self_attn, feed_forward, dropout):
        super(EncoderLayer, self).__init__()
        self.self_attn = self_attn
        self.feed_forward = feed_forward
        self.sublayer = clones(SublayerConnection(size, dropout), 2)
        self.size = size
        self.norm = LayerNorm(size)

    def forward(self, x, x2, mask):
        x = self.sublayer[0](x, lambda x: self.self_attn(x2, x2, mask))
        return self.sublayer[1](x, self.feed_forward)


class MultiHeadedAttention(nn.Module):
    def __init__(self, h, d_model, dropout=0.1):
        super(MultiHeadedAttention, self).__init__()
        assert d_model % h == 0
        self.d_k = d_model // h
        self.h = h
        
        self.linears = clones(nn.Linear(d_model, d_model), 3)
        self.attn = None
        self.dropout = nn.Dropout(p=dropout)
        self.sftmax = nn.Softmax(dim=-1)
    def forward(self, key, value, mask=None):
        nbatches = key.size(0)
        
        key = self.linears[0](key).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
        value = self.linears[1](value).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)

        x, self.attn = self.attention(key, value, mask=mask)

        x = x.transpose(1, 2).contiguous() \
            .view(nbatches, -1, self.h * self.d_k)
        return self.linears[-1](x)

    def attention(self, key, value, mask=None):
        key = torch.mean(key, dim=2).unsqueeze(2)

        v_p = value.permute(0,1,3,2)
        v_p = torch.matmul(v_p, mask)
        v_out = v_p.permute(0,1,3,2)
        
        scores = key / v_out.size(-2)
        scores = self.sftmax(scores)

        output = scores * v_out

        return output, scores


class PositionwiseFeedForward(nn.Module):
    def __init__(self, d_model, d_ff, dropout=0):
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = nn.Linear(d_model, d_ff)
        self.w_2 = nn.Linear(d_ff, d_model)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        return self.w_2(self.dropout(F.gelu(self.w_1(x))))

############################################################


class ECC_Transformer(nn.Module):
    def __init__(self, args, dropout=0):
        super(ECC_Transformer, self).__init__()
        ####
        code = args.code
        c = copy.deepcopy
        attn = MultiHeadedAttention(args.h, args.d_model)
        ff = PositionwiseFeedForward(args.d_model, args.d_model*4, dropout)
        
        self.register_buffer('pc_matrix', args.code.pc_matrix.transpose(0, 1).float())

        self.src_embed_VN = torch.nn.Parameter(torch.empty(
            (1, args.d_model)))
        self.src_embed_CN = torch.nn.Parameter(torch.empty(
            (1, args.d_model)))
        
        self.decoder = Encoder(EncoderLayer(
            args.d_model, c(attn), c(ff), dropout), args.N_dec)
        
        self.oned_final_embed = torch.nn.Sequential(
            *[nn.Linear(args.d_model, 1)])

        for p in self.parameters():
            if p.dim() > 1:
                nn.init.xavier_uniform_(p)

    def forward(self, magnitude, syndrome):
        
        VN = magnitude.unsqueeze(-1)
        CN = syndrome.unsqueeze(-1)

        VN = self.src_embed_VN.unsqueeze(0) * VN
        CN = self.src_embed_CN.unsqueeze(0) * CN
        
        emb1, emb2 = self.decoder(VN, CN, self.pc_matrix)

        emb2_per = emb2.permute(0,2,1)
        emb2_per = torch.matmul(emb2_per,self.pc_matrix.transpose(0,1))
        emb2_final = emb2_per.permute(0,2,1)

        emb = emb1 + emb2_final
        return self.oned_final_embed(emb).squeeze(-1)

    def loss(self, z_pred, z2, y):
        loss = F.binary_cross_entropy_with_logits(
            z_pred, sign_to_bin(torch.sign(z2)))
        x_pred = sign_to_bin(torch.sign(-z_pred * torch.sign(y)))
        return loss, x_pred


############################################################
############################################################

if __name__ == '__main__':
    pass