cogdl1/cogdl/models/emb/sdne.py
import time

import networkx as nx
import numpy as np
import scipy.sparse as sp
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm

from .. import BaseModel


class SDNE_layer(nn.Module):
    def __init__(self, num_node, hidden_size1, hidden_size2, droput, alpha, beta, nu1, nu2):
        super(SDNE_layer, self).__init__()
        self.num_node = num_node
        self.hidden_size1 = hidden_size1
        self.hidden_size2 = hidden_size2
        self.droput = droput
        self.alpha = alpha
        self.beta = beta
        self.nu1 = nu1
        self.nu2 = nu2

        self.encode0 = nn.Linear(self.num_node, self.hidden_size1)
        self.encode1 = nn.Linear(self.hidden_size1, self.hidden_size2)
        self.decode0 = nn.Linear(self.hidden_size2, self.hidden_size1)
        self.decode1 = nn.Linear(self.hidden_size1, self.num_node)

    def forward(self, adj_mat, l_mat):
        t0 = F.leaky_relu(self.encode0(adj_mat))
        t0 = F.leaky_relu(self.encode1(t0))
        self.embedding = t0
        t0 = F.leaky_relu(self.decode0(t0))
        t0 = F.leaky_relu(self.decode1(t0))

        L_1st = 2 * torch.trace(torch.mm(torch.mm(torch.t(self.embedding), l_mat), self.embedding))
        L_2nd = torch.sum(((adj_mat - t0) * adj_mat * self.beta) * ((adj_mat - t0) * adj_mat * self.beta))

        L_reg = 0
        for param in self.parameters():
            L_reg += self.nu1 * torch.sum(torch.abs(param)) + self.nu2 * torch.sum(param * param)
        return self.alpha * L_1st, L_2nd, self.alpha * L_1st + L_2nd, L_reg

    def get_emb(self, adj):
        t0 = self.encode0(adj)
        t0 = self.encode1(t0)
        return t0


class SDNE(BaseModel):
    r"""The SDNE model from the `"Structural Deep Network Embedding"
    <https://www.kdd.org/kdd2016/papers/files/rfp0191-wangAemb.pdf>`_ paper

    Args:
        hidden_size1 (int) : The size of the first hidden layer.
        hidden_size2 (int) : The size of the second hidden layer.
        droput (float) : Droput rate.
        alpha (float) : Trade-off parameter between 1-st and 2-nd order objective function in SDNE.
        beta (float) : Parameter of 2-nd order objective function in SDNE.
        nu1 (float) : Parameter of l1 normlization in SDNE.
        nu2 (float) : Parameter of l2 normlization in SDNE.
        epochs (int) : The max epoches in training step.
        lr (float) : Learning rate in SDNE.
        cpu (bool) : Use CPU or GPU to train hin2vec.
    """

    @staticmethod
    def add_args(parser):
        """Add model-specific arguments to the parser."""
        # fmt: off
        parser.add_argument("--hidden-size1", type=int, default=1000, help="Hidden size in first layer of Auto-Encoder")
        parser.add_argument("--hidden-size2", type=int, default=128, help="Hidden size in second layer of Auto-Encoder")
        parser.add_argument("--droput", type=float, default=0.5, help="Dropout rate")
        parser.add_argument("--alpha", type=float, default=1e-1, help="alhpa is a hyperparameter in SDNE")
        parser.add_argument("--beta", type=float, default=5, help="beta is a hyperparameter in SDNE")
        parser.add_argument("--nu1", type=float, default=1e-4, help="nu1 is a hyperparameter in SDNE")
        parser.add_argument("--nu2", type=float, default=1e-3, help="nu2 is a hyperparameter in SDNE")
        # fmt: on

    @classmethod
    def build_model_from_args(cls, args):
        return cls(
            args.hidden_size1,
            args.hidden_size2,
            args.droput,
            args.alpha,
            args.beta,
            args.nu1,
            args.nu2,
            args.epochs,
            args.lr,
            args.cpu,
        )

    def __init__(self, hidden_size1, hidden_size2, droput, alpha, beta, nu1, nu2, epochs, lr, cpu):
        super(SDNE, self).__init__()
        self.hidden_size1 = hidden_size1
        self.hidden_size2 = hidden_size2
        self.droput = droput
        self.alpha = alpha
        self.beta = beta
        self.nu1 = nu1
        self.nu2 = nu2
        self.epochs = epochs
        self.lr = lr
        self.cpu = cpu

    def forward(self, graph, return_dict=False):
        device = "cuda" if torch.cuda.is_available() and not self.cpu else "cpu"
        G = graph.to_networkx()
        num_node = G.number_of_nodes()
        model = SDNE_layer(
            num_node, self.hidden_size1, self.hidden_size2, self.droput, self.alpha, self.beta, self.nu1, self.nu2
        )

        A = torch.from_numpy(nx.adjacency_matrix(G).todense().astype(np.float32))
        L = torch.from_numpy(nx.laplacian_matrix(G).todense().astype(np.float32))

        A, L = A.to(device), L.to(device)
        model = model.to(device)

        opt = torch.optim.Adam(model.parameters(), lr=self.lr)
        epoch_iter = tqdm(range(self.epochs))
        for epoch in epoch_iter:
            opt.zero_grad()
            L_1st, L_2nd, L_all, L_reg = model.forward(A, L)
            Loss = L_all + L_reg
            Loss.backward()
            epoch_iter.set_description(
                f"Epoch: {epoch:03d}, L_1st: {L_1st:.4f}, L_2nd: {L_2nd:.4f}, L_reg: {L_reg:.4f}"
            )
            opt.step()
        embeddings = model.get_emb(A)
        embeddings = embeddings.detach().cpu().numpy()

        if return_dict:
            features_matrix = dict()
            for vid, node in enumerate(G.nodes()):
                features_matrix[node] = embeddings[vid]
        else:
            features_matrix = np.zeros((graph.num_nodes, embeddings.shape[1]))
            nx_nodes = G.nodes()
            features_matrix[nx_nodes] = embeddings[np.arange(graph.num_nodes)]
        return features_matrix

