GraphRNN/main_DeepGMG.py

# an implementation for "Learning Deep Generative Models of Graphs"
import os

import random
from statistics import mean

import networkx as nx
import numpy as np
from sklearn.metrics import roc_auc_score, average_precision_score

from main import *


class Args_DGMG():
    def __init__(self):
        ### CUDA
        self.cuda = 0

        ### model type
        self.note = 'Baseline_DGMG'  # do GCN after adding each edge
        # self.note = 'Baseline_DGMG_fast' # do GCN only after adding each node

        ### data config
        # self.graph_type = 'caveman_small'
        self.graph_type = 'grid_small'
        # self.graph_type = 'ladder_small'
        # self.graph_type = 'enzymes_small'
        # self.graph_type = 'barabasi_small'
        # self.graph_type = 'citeseer_small'

        self.max_num_node = 20

        ### network config
        self.node_embedding_size = 64
        self.test_graph_num = 200

        ### training config
        self.epochs = 100  # now one epoch means self.batch_ratio x batch_size
        self.load_epoch = 100
        self.epochs_test_start = 10
        self.epochs_test = 10
        self.epochs_log = 10
        self.epochs_save = 10
        if 'fast' in self.note:
            self.is_fast = True
        else:
            self.is_fast = False

        self.lr = 0.001
        self.milestones = [300, 600, 1000]
        self.lr_rate = 0.3

        ### output config
        self.model_save_path = 'model_save/'
        self.graph_save_path = 'graphs/'
        self.figure_save_path = 'figures/'
        self.timing_save_path = 'timing/'
        self.figure_prediction_save_path = 'figures_prediction/'
        self.nll_save_path = 'nll/'

        self.fname = self.note + '_' + self.graph_type + '_' + str(self.node_embedding_size)
        self.fname_pred = self.note + '_' + self.graph_type + '_' + str(self.node_embedding_size) + '_pred_'
        self.fname_train = self.note + '_' + self.graph_type + '_' + str(self.node_embedding_size) + '_train_'
        self.fname_test = self.note + '_' + self.graph_type + '_' + str(self.node_embedding_size) + '_test_'

        self.load = False
        self.save = True


def train_DGMG_epoch(epoch, args, model, dataset, optimizer, scheduler, is_fast=False):
    model.train()
    graph_num = len(dataset)
    order = list(range(graph_num))
    shuffle(order)

    loss_addnode = 0
    loss_addedge = 0
    loss_node = 0
    for i in order:
        model.zero_grad()

        graph = dataset[i]
        # do random ordering: relabel nodes
        node_order = list(range(graph.number_of_nodes()))
        shuffle(node_order)
        order_mapping = dict(zip(graph.nodes(), node_order))
        graph = nx.relabel_nodes(graph, order_mapping, copy=True)

        # NOTE: when starting loop, we assume a node has already been generated
        node_count = 1
        node_embedding = [
            Variable(torch.ones(1, args.node_embedding_size)).cuda()]  # list of torch tensors, each size: 1*hidden

        loss = 0
        while node_count <= graph.number_of_nodes():
            node_neighbor = graph.subgraph(
                list(range(node_count))).adjacency_list()  # list of lists (first node is zero)
            node_neighbor_new = graph.subgraph(list(range(node_count + 1))).adjacency_list()[
                -1]  # list of new node's neighbors

            # 1 message passing
            # do 2 times message passing
            node_embedding = message_passing(node_neighbor, node_embedding, model)

            # 2 graph embedding and new node embedding
            node_embedding_cat = torch.cat(node_embedding, dim=0)
            graph_embedding = calc_graph_embedding(node_embedding_cat, model)
            init_embedding = calc_init_embedding(node_embedding_cat, model)

            # 3 f_addnode
            p_addnode = model.f_an(graph_embedding)
            if node_count < graph.number_of_nodes():
                # add node
                node_neighbor.append([])
                node_embedding.append(init_embedding)
                if is_fast:
                    node_embedding_cat = torch.cat(node_embedding, dim=0)
                # calc loss
                loss_addnode_step = F.binary_cross_entropy(p_addnode, Variable(torch.ones((1, 1))).cuda())
                # loss_addnode_step.backward(retain_graph=True)
                loss += loss_addnode_step
                loss_addnode += loss_addnode_step.data
            else:
                # calc loss
                loss_addnode_step = F.binary_cross_entropy(p_addnode, Variable(torch.zeros((1, 1))).cuda())
                # loss_addnode_step.backward(retain_graph=True)
                loss += loss_addnode_step
                loss_addnode += loss_addnode_step.data
                break

            edge_count = 0
            while edge_count <= len(node_neighbor_new):
                if not is_fast:
                    node_embedding = message_passing(node_neighbor, node_embedding, model)
                    node_embedding_cat = torch.cat(node_embedding, dim=0)
                    graph_embedding = calc_graph_embedding(node_embedding_cat, model)

                # 4 f_addedge
                p_addedge = model.f_ae(graph_embedding)

                if edge_count < len(node_neighbor_new):
                    # calc loss
                    loss_addedge_step = F.binary_cross_entropy(p_addedge, Variable(torch.ones((1, 1))).cuda())
                    # loss_addedge_step.backward(retain_graph=True)
                    loss += loss_addedge_step
                    loss_addedge += loss_addedge_step.data

                    # 5 f_nodes
                    # excluding the last node (which is the new node)
                    node_new_embedding_cat = node_embedding_cat[-1, :].expand(node_embedding_cat.size(0) - 1,
                                                                              node_embedding_cat.size(1))
                    s_node = model.f_s(torch.cat((node_embedding_cat[0:-1, :], node_new_embedding_cat), dim=1))
                    p_node = F.softmax(s_node.permute(1, 0))
                    # get ground truth
                    a_node = torch.zeros((1, p_node.size(1)))
                    # print('node_neighbor_new',node_neighbor_new, edge_count)
                    a_node[0, node_neighbor_new[edge_count]] = 1
                    a_node = Variable(a_node).cuda()
                    # add edge
                    node_neighbor[-1].append(node_neighbor_new[edge_count])
                    node_neighbor[node_neighbor_new[edge_count]].append(len(node_neighbor) - 1)
                    # calc loss
                    loss_node_step = F.binary_cross_entropy(p_node, a_node)
                    # loss_node_step.backward(retain_graph=True)
                    loss += loss_node_step
                    loss_node += loss_node_step.data

                else:
                    # calc loss
                    loss_addedge_step = F.binary_cross_entropy(p_addedge, Variable(torch.zeros((1, 1))).cuda())
                    # loss_addedge_step.backward(retain_graph=True)
                    loss += loss_addedge_step
                    loss_addedge += loss_addedge_step.data
                    break

                edge_count += 1
            node_count += 1

        # update deterministic and lstm
        loss.backward()
        optimizer.step()
        scheduler.step()

    loss_all = loss_addnode + loss_addedge + loss_node

    if epoch % args.epochs_log == 0:
        print('Epoch: {}/{}, train loss: {:.6f}, graph type: {}, hidden: {}'.format(
            epoch, args.epochs, loss_all, args.graph_type, args.node_embedding_size))

    # loss_sum += loss.data[0]*x.size(0)
    # return loss_sum


def test_DGMG_epoch(args, model, is_fast=False):
    model.eval()
    graph_num = args.test_graph_num

    graphs_generated = []
    for i in range(graph_num):
        # NOTE: when starting loop, we assume a node has already been generated
        node_neighbor = [[]]  # list of lists (first node is zero)
        node_embedding = [
            Variable(torch.ones(1, args.node_embedding_size)).cuda()]  # list of torch tensors, each size: 1*hidden

        node_count = 1
        while node_count <= args.max_num_node:
            # 1 message passing
            # do 2 times message passing
            node_embedding = message_passing(node_neighbor, node_embedding, model)

            # 2 graph embedding and new node embedding
            node_embedding_cat = torch.cat(node_embedding, dim=0)
            graph_embedding = calc_graph_embedding(node_embedding_cat, model)
            init_embedding = calc_init_embedding(node_embedding_cat, model)

            # 3 f_addnode
            p_addnode = model.f_an(graph_embedding)
            a_addnode = sample_tensor(p_addnode)
            # print(a_addnode.data[0][0])
            if a_addnode.data[0][0] == 1:
                # print('add node')
                # add node
                node_neighbor.append([])
                node_embedding.append(init_embedding)
                if is_fast:
                    node_embedding_cat = torch.cat(node_embedding, dim=0)
            else:
                break

            edge_count = 0
            while edge_count < args.max_num_node:
                if not is_fast:
                    node_embedding = message_passing(node_neighbor, node_embedding, model)
                    node_embedding_cat = torch.cat(node_embedding, dim=0)
                    graph_embedding = calc_graph_embedding(node_embedding_cat, model)

                # 4 f_addedge
                p_addedge = model.f_ae(graph_embedding)
                a_addedge = sample_tensor(p_addedge)
                # print(a_addedge.data[0][0])

                if a_addedge.data[0][0] == 1:
                    # print('add edge')
                    # 5 f_nodes
                    # excluding the last node (which is the new node)
                    node_new_embedding_cat = node_embedding_cat[-1, :].expand(node_embedding_cat.size(0) - 1,
                                                                              node_embedding_cat.size(1))
                    s_node = model.f_s(torch.cat((node_embedding_cat[0:-1, :], node_new_embedding_cat), dim=1))
                    p_node = F.softmax(s_node.permute(1, 0))
                    a_node = gumbel_softmax(p_node, temperature=0.01)
                    _, a_node_id = a_node.topk(1)
                    a_node_id = int(a_node_id.data[0][0])
                    # add edge
                    node_neighbor[-1].append(a_node_id)
                    node_neighbor[a_node_id].append(len(node_neighbor) - 1)
                else:
                    break

                edge_count += 1
            node_count += 1
        # save graph
        node_neighbor_dict = dict(zip(list(range(len(node_neighbor))), node_neighbor))
        graph = nx.from_dict_of_lists(node_neighbor_dict)
        graphs_generated.append(graph)

    return graphs_generated


def test_DGMG_2(args, model, test_graph, is_fast=False):
    model.eval()
    graph_num = args.test_graph_num

    graphs_generated = []
    # for i in range(graph_num):
    # NOTE: when starting loop, we assume a node has already been generated
    node_neighbor = [[]]  # list of lists (first node is zero)
    node_embedding = [
        Variable(torch.ones(1, args.node_embedding_size)).cuda()]  # list of torch tensors, each size: 1*hidden

    node_max = len(test_graph.nodes())
    node_count = 1
    while node_count <= node_max:
        # 1 message passing
        # do 2 times message passing
        node_embedding = message_passing(node_neighbor, node_embedding, model)

        # 2 graph embedding and new node embedding
        node_embedding_cat = torch.cat(node_embedding, dim=0)
        graph_embedding = calc_graph_embedding(node_embedding_cat, model)
        init_embedding = calc_init_embedding(node_embedding_cat, model)

        # 3 f_addnode
        p_addnode = model.f_an(graph_embedding)
        a_addnode = sample_tensor(p_addnode)

        if a_addnode.data[0][0] == 1:
            # add node
            node_neighbor.append([])
            node_embedding.append(init_embedding)
            if is_fast:
                node_embedding_cat = torch.cat(node_embedding, dim=0)
        else:
            break

        edge_count = 0
        while edge_count < args.max_num_node:
            if not is_fast:
                node_embedding = message_passing(node_neighbor, node_embedding, model)
                node_embedding_cat = torch.cat(node_embedding, dim=0)
                graph_embedding = calc_graph_embedding(node_embedding_cat, model)

            # 4 f_addedge
            p_addedge = model.f_ae(graph_embedding)
            a_addedge = sample_tensor(p_addedge)

            if a_addedge.data[0][0] == 1:
                # 5 f_nodes
                # excluding the last node (which is the new node)
                node_new_embedding_cat = node_embedding_cat[-1, :].expand(node_embedding_cat.size(0) - 1,
                                                                          node_embedding_cat.size(1))
                s_node = model.f_s(torch.cat((node_embedding_cat[0:-1, :], node_new_embedding_cat), dim=1))
                p_node = F.softmax(s_node.permute(1, 0))
                a_node = gumbel_softmax(p_node, temperature=0.01)
                _, a_node_id = a_node.topk(1)
                a_node_id = int(a_node_id.data[0][0])
                # add edge

                node_neighbor[-1].append(a_node_id)
                node_neighbor[a_node_id].append(len(node_neighbor) - 1)
            else:
                break

            edge_count += 1
        node_count += 1

    # clear node_neighbor and build it again
    node_neighbor = []
    for n in range(node_max):
        temp_neighbor = [k for k in test_graph.edge[n]]
        node_neighbor.append(temp_neighbor)

    # now add the last node for real
    # 1 message passing
    # do 2 times message passing
    try:
        node_embedding = message_passing(node_neighbor, node_embedding, model)

        # 2 graph embedding and new node embedding
        node_embedding_cat = torch.cat(node_embedding, dim=0)
        graph_embedding = calc_graph_embedding(node_embedding_cat, model)
        init_embedding = calc_init_embedding(node_embedding_cat, model)

        # 3 f_addnode
        p_addnode = model.f_an(graph_embedding)
        a_addnode = sample_tensor(p_addnode)

        if a_addnode.data[0][0] == 1:
            # add node
            node_neighbor.append([])
            node_embedding.append(init_embedding)
            if is_fast:
                node_embedding_cat = torch.cat(node_embedding, dim=0)

        edge_count = 0
        while edge_count < args.max_num_node:
            if not is_fast:
                node_embedding = message_passing(node_neighbor, node_embedding, model)
                node_embedding_cat = torch.cat(node_embedding, dim=0)
                graph_embedding = calc_graph_embedding(node_embedding_cat, model)

            # 4 f_addedge
            p_addedge = model.f_ae(graph_embedding)
            a_addedge = sample_tensor(p_addedge)

            if a_addedge.data[0][0] == 1:
                # 5 f_nodes
                # excluding the last node (which is the new node)
                node_new_embedding_cat = node_embedding_cat[-1, :].expand(node_embedding_cat.size(0) - 1,
                                                                          node_embedding_cat.size(1))
                s_node = model.f_s(torch.cat((node_embedding_cat[0:-1, :], node_new_embedding_cat), dim=1))
                p_node = F.softmax(s_node.permute(1, 0))
                a_node = gumbel_softmax(p_node, temperature=0.01)
                _, a_node_id = a_node.topk(1)
                a_node_id = int(a_node_id.data[0][0])
                # add edge

                node_neighbor[-1].append(a_node_id)
                node_neighbor[a_node_id].append(len(node_neighbor) - 1)
            else:
                break

            edge_count += 1
        node_count += 1
    except:
        print('error')
    # save graph
    node_neighbor_dict = dict(zip(list(range(len(node_neighbor))), node_neighbor))
    graph = nx.from_dict_of_lists(node_neighbor_dict)
    graphs_generated.append(graph)

    return graphs_generated


########### train function for LSTM + VAE
def train_DGMG(args, dataset_train, model):
    # check if load existing model
    if args.load:
        fname = args.model_save_path + args.fname + 'model_' + str(args.load_epoch) + '.dat'
        model.load_state_dict(torch.load(fname))

        args.lr = 0.00001
        epoch = args.load_epoch
        print('model loaded!, lr: {}'.format(args.lr))
    else:
        epoch = 1

    # initialize optimizer
    optimizer = optim.Adam(list(model.parameters()), lr=args.lr)

    scheduler = MultiStepLR(optimizer, milestones=args.milestones, gamma=args.lr_rate)

    # start main loop
    time_all = np.zeros(args.epochs)
    while epoch <= args.epochs:
        time_start = tm.time()
        # train
        train_DGMG_epoch(epoch, args, model, dataset_train, optimizer, scheduler, is_fast=args.is_fast)
        time_end = tm.time()
        time_all[epoch - 1] = time_end - time_start
        print('time used', time_all[epoch - 1])
        # test
        if epoch % args.epochs_test == 0 and epoch >= args.epochs_test_start:
            graphs = test_DGMG_epoch(args, model, is_fast=args.is_fast)
            fname = args.graph_save_path + args.fname_pred + str(epoch) + '.dat'
            save_graph_list(graphs, fname)
            # print('test done, graphs saved')

        # save model checkpoint
        if args.save:
            if epoch % args.epochs_save == 0:
                fname = args.model_save_path + args.fname + 'model_' + str(epoch) + '.dat'
                torch.save(model.state_dict(), fname)
        epoch += 1
    np.save(args.timing_save_path + args.fname, time_all)


def neigh_to_mat(neigh, size):
    ret_list = np.zeros(size)
    for i in neigh:
        ret_list[i] = 1
    return ret_list


def calc_lable_result(test_graphs, returned_graphs):
    labels = []
    results = []
    i = 0
    for test_graph in test_graphs:
        n = len(test_graph.nodes())
        returned_graph = returned_graphs[i]
        label = neigh_to_mat([k for k in test_graph.edge[n - 1]], n)
        try:
            result = neigh_to_mat([k for k in returned_graph.edge[n - 1]], n)
        except:
            result = np.zeros(n)
        labels.append(label)
        results.append(result)
        i += 1
    return labels, results


def evaluate(labels, results):
    mae_list = []
    roc_score_list = []
    ap_score_list = []
    precision_list = []
    recall_list = []
    iter = 0
    for result in results:
        label = labels[iter]
        iter += 1
        part1 = label[result == 1]
        part2 = part1[part1 == 1]
        part3 = part1[part1 == 0]
        part4 = label[result == 0]
        part5 = part4[part4 == 1]
        tp = len(part2)
        fp = len(part3)
        fn = part5.sum()
        if tp + fp > 0:
            precision = tp / (tp + fp)
        else:
            precision = 0
        recall = tp / (tp + fn)
        precision_list.append(precision)
        recall_list.append(recall)

        positive = result[label == 1]
        if len(positive) <= len(list(result[label == 0])):
            negative = random.sample(list(result[label == 0]), len(positive))
        else:
            negative = result[label == 0]
            positive = random.sample(list(result[label == 1]), len(negative))
        preds_all = np.hstack([positive, negative])
        labels_all = np.hstack([np.ones(len(positive)), np.zeros(len(positive))])

        if len(labels_all) > 0:
            roc_score = roc_auc_score(labels_all, preds_all)
            ap_score = average_precision_score(labels_all, preds_all)

        roc_score_list.append(roc_score)
        ap_score_list.append(ap_score)

        mae = 0
        for x in range(len(result)):
            if result[x] != label[x]:
                mae += 1

        mae = mae / len(label)
        mae_list.append(mae)

    mean_roc = mean(roc_score_list)
    mean_ap = mean(ap_score_list)
    mean_precision = mean(precision_list)
    mean_recall = mean(recall_list)
    mean_mae = mean(mae_list)
    print('roc_score ' + str(mean_roc))
    print('ap_score ' + str(mean_ap))
    print('precision ' + str(mean_precision))
    print('recall ' + str(mean_recall))
    print('mae ' + str(mean_mae))
    return mean_roc, mean_ap, mean_precision, mean_recall


if __name__ == '__main__':

    args = Args_DGMG()
    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda)
    print('CUDA', args.cuda)
    print('File name prefix', args.fname)

    graphs = []
    for i in range(4, 10):
        graphs.append(nx.ladder_graph(i))
    model = DGM_graphs(h_size=args.node_embedding_size).cuda()

    if args.graph_type == 'grid_small':
        graphs = []
        for i in range(2, 3):
            for j in range(2, 4):
                graphs.append(nx.grid_2d_graph(i, j))
        args.max_prev_node = 5

    # remove self loops
    for graph in graphs:
        edges_with_selfloops = graph.selfloop_edges()
        if len(edges_with_selfloops) > 0:
            graph.remove_edges_from(edges_with_selfloops)

    # split datasets
    random.seed(123)
    shuffle(graphs)
    # graphs_len = len(graphs)
    # graphs_test = graphs[int(0.8 * graphs_len):]
    # graphs_validate = graphs[int(0.7 * graphs_len):int(0.8 * graphs_len)]
    # graphs_train = graphs[0:int(0.7 * graphs_len)]

    args.max_num_node = max([graphs[i].number_of_nodes() for i in range(len(graphs))])

    print('max number node: {}'.format(args.max_num_node))
    print('max previous node: {}'.format(args.max_prev_node))
    test_graph = nx.grid_2d_graph(2, 3)
    test_graph.remove_node(test_graph.nodes()[5])
    train_DGMG(args, graphs, model)

    test_graph = nx.convert_node_labels_to_integers(test_graph)
    test_DGMG_2(args, model, test_graph)

    labels, results = calc_lable_result(test_graphs, eval_graphs)

    evaluate(labels, results)