model.py

import copy
import math

import torch
import torch.nn as nn
import torch.nn.functional as F

import numpy as np

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# device = torch.device("cpu")

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


def clone_params(param, N):
    return nn.ParameterList([copy.deepcopy(param) for _ in range(N)])


# TODO: replaced with https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html?
class LayerNorm(nn.Module):
    def __init__(self, features, eps=1e-6):
        super(LayerNorm, self).__init__()
        self.a_2 = nn.Parameter(torch.ones(features))
        self.b_2 = nn.Parameter(torch.zeros(features))
        self.eps = eps

    def forward(self, x):
        mean = x.mean(-1, keepdim=True)
        std = x.std(-1, keepdim=True)
        return self.a_2 * (x - mean) / (std + self.eps) + self.b_2


class GraphLayer(nn.Module):

    def __init__(self, in_features, hidden_features, out_features, num_of_nodes,
                 num_of_heads, dropout, alpha, concat=True):
        super(GraphLayer, self).__init__()
        self.in_features = in_features              # MyNote: Embedding size
        self.hidden_features = hidden_features      # MyNote: Embedding size
        self.out_features = out_features            # MyNote: Embedding size (ngoại trừ Decoder Graph, khác chỗ này)
        self.alpha = alpha                          # MyNote: hardcoded 0.1
        self.concat = concat                        # MyNote: Encoder graph ->True; Decoder Graph -> False.
        self.num_of_nodes = num_of_nodes            # MyNote: Số node trong Graph.
        self.num_of_heads = num_of_heads            # MyNote: Số attention head. -> là 1 (VGNN/Mimic)
        
        # MyNote: gọi clones() nhưng List chỉ có 1 phần tử vì num_of_heads=1 (ghi trong paper).
        self.W = clones(nn.Linear(in_features, hidden_features), num_of_heads)
        self.a = clone_params(nn.Parameter(torch.rand(size=(1, 2 * hidden_features)), requires_grad=True), num_of_heads)
        self.ffn = nn.Sequential(
            nn.Linear(out_features, out_features),
            nn.ReLU()
        )
        
        if not concat:
            self.V = nn.Linear(hidden_features, out_features)
        else:
            self.V = nn.Linear(num_of_heads * hidden_features, out_features)
            
        self.dropout = nn.Dropout(dropout)
        self.leakyrelu = nn.LeakyReLU(self.alpha)
        
        if concat:  # MyNote: Ko hiểu khác nhau chỗ nào?
            self.norm = LayerNorm(hidden_features)
        else:
            self.norm = LayerNorm(hidden_features)

    def initialize(self):
        for i in range(len(self.W)):
            nn.init.xavier_normal_(self.W[i].weight.data)
        for i in range(len(self.a)):
            nn.init.xavier_normal_(self.a[i].data)
        if not self.concat:
            nn.init.xavier_normal_(self.V.weight.data)
            nn.init.xavier_normal_(self.out_layer.weight.data)

    def attention(self, linear, a, N, data, edge):
        """MyNote: _summary_

        Args:
            linear (_type_): weights (R^(dxd))
            a (_type_): bias (R^(1x(2*d)))
            N (_type_): number of nodes
            data (_type_): h_prime = Toàn bộ Nodes & Embedding của nó.
            edge (_type_): Vd: edge -> input_edges = 2x11664
                                        108x108=11664 -> 108 lab-value/procedure... (one-hot encoding)

        Returns:
            _type_: _description_
        """
        data = linear(data).unsqueeze(0)
        assert not torch.isnan(data).any()
        # edge: 2*D x E
        h = torch.cat((data[:, edge[0, :], :], data[:, edge[1, :], :]), 
                      dim=0)
        data = data.squeeze(0)
        # h: N x out
        assert not torch.isnan(h).any()
        # edge_h: 2*D x E
        edge_h = torch.cat((h[0, :, :], h[1, :, :]), dim=1).transpose(0, 1)
        # edge: 2*D x E
        edge_e = torch.exp(self.leakyrelu(a.mm(edge_h).squeeze()) / np.sqrt(self.hidden_features * self.num_of_heads))
        assert not torch.isnan(edge_e).any()
        # edge_e: E
        edge_e = torch.sparse_coo_tensor(edge, edge_e, torch.Size([N, N]))
        e_rowsum = torch.sparse.mm(edge_e, torch.ones(size=(N, 1)).to(device))
        # e_rowsum: N x 1
        row_check = (e_rowsum == 0) 
        e_rowsum[row_check] = 1
        zero_idx = row_check.nonzero()[:, 0]
        edge_e = edge_e.add(
            torch.sparse.FloatTensor(zero_idx.repeat(2, 1), torch.ones(len(zero_idx)).to(device), torch.Size([N, N])))  # type: ignore
        # edge_e: E
        h_prime = torch.sparse.mm(edge_e, data)
        assert not torch.isnan(h_prime).any()
        # h_prime: N x out
        h_prime.div_(e_rowsum)
        # h_prime: N x out
        assert not torch.isnan(h_prime).any()
        return h_prime

    def forward(self, edge, data=None):
        # MyNote: input: (input_edges, h_prime)
        # MyNote: Vd: edge -> input_edges = 2x11881
        # MyNote: data -> h_prime = Toàn bộ Nodes & Embedding của nó.
        N = self.num_of_nodes
        
        if self.concat: # MyNote: hardcoded True
            # MyNote: Zip nhưng thực ra chỉ có 1 element vì Attention head là 1 (ghi trong paper).
            h_prime = torch.cat([self.attention(l, a, N, data, edge) for l, a in zip(self.W, self.a)], dim=1)
        else:
            h_prime = torch.stack([self.attention(l, a, N, data, edge) for l, a in zip(self.W, self.a)], dim=0).mean(
                dim=0)
        
        h_prime = self.dropout(h_prime)
        
        if self.concat:
            return F.elu(self.norm(h_prime))
        else:
            return self.V(F.relu(self.norm(h_prime)))


class VariationalGNN(nn.Module):

    def __init__(self, 
                 in_features, 
                 out_features, 
                 num_of_nodes, 
                 n_heads, 
                 n_layers,
                 dropout, 
                 alpha,                 # MyNote: hardcoded 0.1
                 variational=True, 
                 none_graph_features=0, 
                 concat=True):
        
        # Save input parameters for later convenient restoration of the object for inference.
        self.kwargs = {'in_features': in_features, 
                       'out_features': out_features, 
                       'num_of_nodes': num_of_nodes,
                       'n_heads': n_heads,
                       'n_layers': n_layers,
                       'dropout': dropout,
                       'alpha': alpha,
                       'variational': variational,
                       'none_graph_features': none_graph_features,
                       'concat': concat}
        
        super(VariationalGNN, self).__init__()
        self.variational = variational
        # Add two more nodes: the 1st indicates the patient is normal; the last node is used to absorb features from specific nodes of specific patients, to make prediction.
        self.num_of_nodes = num_of_nodes + 2 - none_graph_features
        # MyNote: this is the lookup embedding in paper. (Patient)
        self.embed = nn.Embedding(self.num_of_nodes, in_features, padding_idx=0)

        self.in_att = clones(
            GraphLayer(in_features, in_features, in_features, self.num_of_nodes,
                       n_heads, dropout, alpha, concat=True), n_layers)
        self.out_features = out_features
        self.out_att = GraphLayer(in_features, in_features, out_features, self.num_of_nodes,
                                  n_heads, dropout, alpha, concat=False)
        self.n_heads = n_heads
        self.dropout = nn.Dropout(dropout)
        self.parameterize = nn.Linear(out_features, out_features * 2)
        self.out_layer = nn.Sequential(
            nn.Linear(out_features, out_features),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(out_features, 1))
        self.none_graph_features = none_graph_features
        #region none_graph_features > 0
        if none_graph_features > 0:
            self.features_ffn = nn.Sequential(
                nn.Linear(none_graph_features, out_features//2),
                nn.ReLU(),
                nn.Dropout(dropout))
            self.out_layer = nn.Sequential(
                nn.Linear(out_features + out_features//2, out_features),
                nn.ReLU(),
                nn.Dropout(dropout),
                nn.Linear(out_features, 1))
        #endregion
        for i in range(n_layers):
            self.in_att[i].initialize()

    """MyNote: Hàm này để chi? -> data là 1 patient sample với multihot encoding (chỉ bệnh).
    Cần trả về các Edges nối các bệnh này với nhau. Nhớ rằng: mặc định tất cả các bệnh Connect với nhau.
    """
    def data_to_edges(self, data):
        """MyNote: Must return (input_edges, output_edges)"""
        length = data.size()[0]
        nonzero = data.nonzero()    # MyNote: return indices indicating non-zero values.
        if nonzero.size()[0] == 0:  # MyNote: case mà Patient bình thường! (ko có chẩn đoán, xét nghiệm gì!)
            # MyNote: Why return so? shape(2, 1), shape(2, 1) Why length + 1? -> Khi bệnh nhân bình thường, vector bệnh của họ toàn là 0 -> cũng phải trả
            # ra cái gì đó (vậy là chọn Node đầu và node cuối)
            # MyNote: Right side: should include also torch.LongTensor([[0], [0]]) -> ám chỉ là "bình thường" (ko bệnh tật)???
            return torch.LongTensor([[0], [0]]), torch.LongTensor([[length + 1], [length + 1]])
        if self.training:
            mask = torch.rand(nonzero.size()[0])
            mask = mask > 0.05
            nonzero = nonzero[mask]
            if nonzero.size()[0] == 0:
                # MyNote: có phải ý là ngay cả khi Patient có issue, 5% trong số đó ta sẽ đối xử như là ko có issue???
                return torch.LongTensor([[0], [0]]), torch.LongTensor([[length + 1], [length + 1]])
        
        # MyNote: case: when (testing/validating/infering) OR 95% probability bệnh nhân có ít nhất 1 issue nào đó.
        nonzero = nonzero.transpose(0, 1) + 1   # MyNote: Why +1? -> Cộng để tăng Index vì có 2 Node giả đầu (là node chỉ bình thường) và cuối (là node absorb các node khác cho predict)
        lengths = nonzero.size()[1]
        input_edges = torch.cat((nonzero.repeat(1, lengths),
                                 nonzero.repeat(lengths, 1).transpose(0, 1)
                                 .contiguous().view((1, lengths ** 2))), dim=0)

        nonzero = torch.cat((nonzero, torch.LongTensor([[length + 1]]).to(device)), dim=1)
        lengths = nonzero.size()[1]
        output_edges = torch.cat((nonzero.repeat(1, lengths),
                                  nonzero.repeat(lengths, 1).transpose(0, 1)
                                  .contiguous().view((1, lengths ** 2))), dim=0)
        return input_edges.to(device), output_edges.to(device)

    def reparameterise(self, mu, logvar):
        if self.training:
            # Assume log_variation (NOT log_standard_deviation!)
            std = logvar.mul(0.5).exp_()
            # MyNote: tensor.new() -> Constructs a new tensor of the same data type as self tensor.
            eps = std.data.new(std.size()).normal_()
            return eps.mul(std).add_(mu)
        else:
            return mu

    def encoder_decoder(self, data):
        """Given a patient data, encode it into the total graph, then decode to the last node.

        Args:
            data ([N]): multi-hot encoding (of diagnose codes). E.g. shape = [1309]

        Returns:
            Tuple[Tensor, Tensor]: The last node's features, plus KL Divergence
        """
        N = self.num_of_nodes
        input_edges, output_edges = self.data_to_edges(data)
        h_prime = self.embed(torch.arange(N).long().to(device))
        
        # Encoder:
        for attn in self.in_att:
            h_prime = attn(input_edges, h_prime)
            
        if self.variational:
            # Even given only a patient's data, this parameterization affects the total graph.
            h_prime = self.parameterize(h_prime).view(-1, 2, self.out_features)
            h_prime = self.dropout(h_prime)
            mu = h_prime[:, 0, :]
            logvar = h_prime[:, 1, :]
            h_prime = self.reparameterise(mu, logvar)   # h_prime.shape = [N, z_dim] e.g. (1311x256)
            
            # Essential variables (mu, ,logvar) for computing DL Divergence later.
            # Note: only consider the patient's graph (NOT the total graph).
            split = int(math.sqrt(len(input_edges[0])))
            pat_diag_code_idx = input_edges[0][0:split]
            mu = mu[pat_diag_code_idx, :]
            logvar = logvar[pat_diag_code_idx, :]
            
        # Decoder:
        h_prime = self.out_att(output_edges, h_prime)
        
        if self.variational:
            """
            Need to divide with mu.size()[0] because the original formula sums over all latent dimensions.
            """
            return (h_prime[-1],            # The last node's features.
                    0.5 * torch.sum(logvar.exp() - logvar - 1 + mu.pow(2)) / mu.size()[0]
                    )
        else:
            return (h_prime[-1], \
                    torch.tensor(0.0).to(device)
                    )

    def forward(self, data):
        # Concate batches
        batch_size = data.size()[0]
        # In eicu data the first feature whether have be admitted before is not included in the graph
        if self.none_graph_features == 0:   # MyNote: self.none_graph_features hardcoded = 0!!! -> cái này ko phải ám chỉ là ko dùng features cho nodes!
            # MyNote: for each Patient-Encounter, encode the graph specifically for that.
            outputs = [self.encoder_decoder(data[i, :]) for i in range(batch_size)]
            # MyNote: return logits (output of out_layer()) -> later use BCEWithLogitsLoss
            return self.out_layer(F.relu(torch.stack([out[0] for out in outputs]))), \
                   torch.sum(torch.stack([out[1] for out in outputs]))
        else:
            outputs = [(data[i, :self.none_graph_features],
                        self.encoder_decoder(data[i, self.none_graph_features:])) for i in range(batch_size)]
            return self.out_layer(F.relu(
                torch.stack([torch.cat((self.features_ffn(torch.FloatTensor([out[0]]).to(device)), out[1][0]))
                             for out in outputs]))), \
                   torch.sum(torch.stack([out[1][1] for out in outputs]), dim=-1)