cogdl1/cogdl/models/emb/gatne.py
import numpy as np
import networkx as nx
from collections import defaultdict
from gensim.models.keyedvectors import Vocab
import random
import math
import tqdm
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter

from .. import BaseModel


class GATNE(BaseModel):
    r"""The GATNE model from the `"Representation Learning for Attributed Multiplex Heterogeneous Network"
    <https://dl.acm.org/doi/10.1145/3292500.3330964>`_ paper

    Args:
        walk_length (int) : The walk length.
        walk_num (int) : The number of walks to sample for each node.
        window_size (int) : The actual context size which is considered in language model.
        worker (int) : The number of workers for word2vec.
        epochs (int) : The number of training epochs.
        batch_size (int) : The size of each training batch.
        edge_dim (int) : Number of edge embedding dimensions.
        att_dim (int) : Number of attention dimensions.
        negative_samples (int) : Negative samples for optimization.
        neighbor_samples (int) : Neighbor samples for aggregation
        schema (str) : The metapath schema used in model. Metapaths are splited with ",",
        while each node type are connected with "-" in each metapath. For example:"0-1-0,0-1-2-1-0"
    """

    @staticmethod
    def add_args(parser):
        """Add model-specific arguments to the parser."""
        # fmt: off
        parser.add_argument('--walk-length', type=int, default=10,
                            help='Length of walk per source. Default is 10.')
        parser.add_argument('--walk-num', type=int, default=10,
                            help='Number of walks per source. Default is 10.')
        parser.add_argument('--window-size', type=int, default=5,
                            help='Window size of skip-gram model. Default is 5.')
        parser.add_argument('--worker', type=int, default=10,
                            help='Number of parallel workers. Default is 10.')
        parser.add_argument('--epochs', type=int, default=20,
                            help='Number of epochs. Default is 20.')
        parser.add_argument('--batch-size', type=int, default=256,
                            help='Number of batch_size. Default is 256.')
        parser.add_argument('--edge-dim', type=int, default=10,
                            help='Number of edge embedding dimensions. Default is 10.')
        parser.add_argument('--att-dim', type=int, default=20,
                            help='Number of attention dimensions. Default is 20.')
        parser.add_argument('--negative-samples', type=int, default=5,
                            help='Negative samples for optimization. Default is 5.')
        parser.add_argument('--neighbor-samples', type=int, default=10,
                            help='Neighbor samples for aggregation. Default is 10.')
        parser.add_argument('--schema', type=str, default=None,
                            help="Input schema for metapath random walk.")
        # fmt: on

    @classmethod
    def build_model_from_args(cls, args):
        return cls(
            args.hidden_size,
            args.walk_length,
            args.walk_num,
            args.window_size,
            args.worker,
            args.epochs,
            args.batch_size,
            args.edge_dim,
            args.att_dim,
            args.negative_samples,
            args.neighbor_samples,
            args.schema,
        )

    def __init__(
        self,
        dimension,
        walk_length,
        walk_num,
        window_size,
        worker,
        epochs,
        batch_size,
        edge_dim,
        att_dim,
        negative_samples,
        neighbor_samples,
        schema,
    ):
        super(GATNE, self).__init__()
        self.embedding_size = dimension
        self.walk_length = walk_length
        self.walk_num = walk_num
        self.window_size = window_size
        self.worker = worker
        self.epochs = epochs
        self.batch_size = batch_size
        self.embedding_u_size = edge_dim
        self.dim_att = att_dim
        self.num_sampled = negative_samples
        self.neighbor_samples = neighbor_samples
        self.schema = schema
        self.multiplicity = True

    def forward(self, network_data):
        device = "cpu" if not torch.cuda.is_available() else "cuda"
        all_walks = generate_walks(network_data, self.walk_num, self.walk_length, schema=self.schema)
        vocab, index2word = generate_vocab(all_walks)
        train_pairs = generate_pairs(all_walks, vocab)

        edge_types = list(network_data.keys())

        num_nodes = len(index2word)
        edge_type_count = len(edge_types)

        epochs = self.epochs
        batch_size = self.batch_size
        embedding_size = self.embedding_size
        embedding_u_size = self.embedding_u_size
        num_sampled = self.num_sampled
        dim_att = self.dim_att
        neighbor_samples = self.neighbor_samples

        neighbors = [[[] for __ in range(edge_type_count)] for _ in range(num_nodes)]
        for r in range(edge_type_count):
            g = network_data[edge_types[r]]
            for (x, y) in g:
                ix = vocab[x].index
                iy = vocab[y].index
                neighbors[ix][r].append(iy)
                neighbors[iy][r].append(ix)
            for i in range(num_nodes):
                if len(neighbors[i][r]) == 0:
                    neighbors[i][r] = [i] * neighbor_samples
                elif len(neighbors[i][r]) < neighbor_samples:
                    neighbors[i][r].extend(
                        list(
                            np.random.choice(
                                neighbors[i][r],
                                size=neighbor_samples - len(neighbors[i][r]),
                            )
                        )
                    )
                elif len(neighbors[i][r]) > neighbor_samples:
                    neighbors[i][r] = list(np.random.choice(neighbors[i][r], size=neighbor_samples))

        model = GATNEModel(num_nodes, embedding_size, embedding_u_size, edge_type_count, dim_att)
        nsloss = NSLoss(num_nodes, num_sampled, embedding_size)

        model.to(device)
        nsloss.to(device)

        optimizer = torch.optim.Adam([{"params": model.parameters()}, {"params": nsloss.parameters()}], lr=1e-4)

        for epoch in range(epochs):
            random.shuffle(train_pairs)
            batches = get_batches(train_pairs, neighbors, batch_size)

            data_iter = tqdm.tqdm(
                batches,
                desc="epoch %d" % (epoch),
                total=(len(train_pairs) + (batch_size - 1)) // batch_size,
                bar_format="{l_bar}{r_bar}",
            )
            avg_loss = 0.0

            for i, data in enumerate(data_iter):
                optimizer.zero_grad()
                embs = model(
                    data[0].to(device),
                    data[2].to(device),
                    data[3].to(device),
                )
                loss = nsloss(data[0].to(device), embs, data[1].to(device))
                loss.backward()
                optimizer.step()

                avg_loss += loss.item()

                if i % 5000 == 0:
                    post_fix = {
                        "epoch": epoch,
                        "iter": i,
                        "avg_loss": avg_loss / (i + 1),
                        "loss": loss.item(),
                    }
                    data_iter.write(str(post_fix))

        final_model = dict(zip(edge_types, [dict() for _ in range(edge_type_count)]))
        for i in range(num_nodes):
            train_inputs = torch.tensor([i for _ in range(edge_type_count)]).to(device)
            train_types = torch.tensor(list(range(edge_type_count))).to(device)
            node_neigh = torch.tensor([neighbors[i] for _ in range(edge_type_count)]).to(device)
            node_emb = model(train_inputs, train_types, node_neigh)
            for j in range(edge_type_count):
                final_model[edge_types[j]][index2word[i]] = node_emb[j].cpu().detach().numpy()
        return final_model


class GATNEModel(nn.Module):
    def __init__(self, num_nodes, embedding_size, embedding_u_size, edge_type_count, dim_a):
        super(GATNEModel, self).__init__()
        self.num_nodes = num_nodes
        self.embedding_size = embedding_size
        self.embedding_u_size = embedding_u_size
        self.edge_type_count = edge_type_count
        self.dim_a = dim_a

        self.node_embeddings = Parameter(torch.FloatTensor(num_nodes, embedding_size))
        self.node_type_embeddings = Parameter(torch.FloatTensor(num_nodes, edge_type_count, embedding_u_size))
        self.trans_weights = Parameter(torch.FloatTensor(edge_type_count, embedding_u_size, embedding_size))
        self.trans_weights_s1 = Parameter(torch.FloatTensor(edge_type_count, embedding_u_size, dim_a))
        self.trans_weights_s2 = Parameter(torch.FloatTensor(edge_type_count, dim_a, 1))

        self.reset_parameters()

    def reset_parameters(self):
        self.node_embeddings.data.uniform_(-1.0, 1.0)
        self.node_type_embeddings.data.uniform_(-1.0, 1.0)
        self.trans_weights.data.normal_(std=1.0 / math.sqrt(self.embedding_size))
        self.trans_weights_s1.data.normal_(std=1.0 / math.sqrt(self.embedding_size))
        self.trans_weights_s2.data.normal_(std=1.0 / math.sqrt(self.embedding_size))

    def forward(self, train_inputs, train_types, node_neigh):
        node_embed = self.node_embeddings[train_inputs]
        node_embed_neighbors = self.node_type_embeddings[node_neigh]
        node_embed_tmp = torch.cat(
            [node_embed_neighbors[:, i, :, i, :].unsqueeze(1) for i in range(self.edge_type_count)],
            dim=1,
        )
        node_type_embed = torch.sum(node_embed_tmp, dim=2)

        trans_w = self.trans_weights[train_types]
        trans_w_s1 = self.trans_weights_s1[train_types]
        trans_w_s2 = self.trans_weights_s2[train_types]

        attention = F.softmax(
            torch.matmul(F.tanh(torch.matmul(node_type_embed, trans_w_s1)), trans_w_s2).squeeze()
        ).unsqueeze(1)
        node_type_embed = torch.matmul(attention, node_type_embed)
        node_embed = node_embed + torch.matmul(node_type_embed, trans_w).squeeze()

        last_node_embed = F.normalize(node_embed, dim=1)

        return last_node_embed


class NSLoss(nn.Module):
    def __init__(self, num_nodes, num_sampled, embedding_size):
        super(NSLoss, self).__init__()
        self.num_nodes = num_nodes
        self.num_sampled = num_sampled
        self.embedding_size = embedding_size
        self.weights = Parameter(torch.FloatTensor(num_nodes, embedding_size))
        self.sample_weights = F.normalize(
            torch.Tensor([(math.log(k + 2) - math.log(k + 1)) / math.log(num_nodes + 1) for k in range(num_nodes)]),
            dim=0,
        )

        self.reset_parameters()

    def reset_parameters(self):
        self.weights.data.normal_(std=1.0 / math.sqrt(self.embedding_size))

    def forward(self, input, embs, label):
        n = input.shape[0]
        log_target = torch.log(torch.sigmoid(torch.sum(torch.mul(embs, self.weights[label]), 1)))
        negs = torch.multinomial(self.sample_weights, self.num_sampled * n, replacement=True).view(n, self.num_sampled)
        noise = torch.neg(self.weights[negs])
        sum_log_sampled = torch.sum(torch.log(torch.sigmoid(torch.bmm(noise, embs.unsqueeze(2)))), 1).squeeze()

        loss = log_target + sum_log_sampled
        return -loss.sum() / n


class RWGraph:
    def __init__(self, nx_G, node_type=None):
        self.G = nx_G
        self.node_type = node_type

    def walk(self, walk_length, start, schema=None):
        # Simulate a random walk starting from start node.
        G = self.G

        rand = random.Random()

        if schema:
            schema_items = schema.split("-")
            assert schema_items[0] == schema_items[-1]

        walk = [start]
        while len(walk) < walk_length:
            cur = walk[-1]
            candidates = []
            for node in G[cur].keys():
                if schema is None or self.node_type[node] == schema_items[len(walk) % (len(schema_items) - 1)]:
                    candidates.append(node)
            if candidates:
                walk.append(rand.choice(candidates))
            else:
                break
        return walk

    def simulate_walks(self, num_walks, walk_length, schema=None):
        G = self.G
        walks = []
        nodes = list(G.nodes())
        # print('Walk iteration:')
        if schema is not None:
            schema_list = schema.split(",")
        for walk_iter in range(num_walks):
            random.shuffle(nodes)
            for node in nodes:
                if schema is None:
                    walks.append(self.walk(walk_length=walk_length, start=node))
                else:
                    for schema_iter in schema_list:
                        if schema_iter.split("-")[0] == self.node_type[node]:
                            walks.append(
                                self.walk(
                                    walk_length=walk_length,
                                    start=node,
                                    schema=schema_iter,
                                )
                            )

        return walks


def get_G_from_edges(edges):
    edge_dict = dict()
    for edge in edges:
        edge_key = str(edge[0]) + "_" + str(edge[1])
        if edge_key not in edge_dict:
            edge_dict[edge_key] = 1
        else:
            edge_dict[edge_key] += 1
    tmp_G = nx.Graph()
    for edge_key in edge_dict:
        weight = edge_dict[edge_key]
        x = int(edge_key.split("_")[0])
        y = int(edge_key.split("_")[1])
        tmp_G.add_edge(x, y)
        tmp_G[x][y]["weight"] = weight
    return tmp_G


def generate_pairs(all_walks, vocab, window_size=5):
    pairs = []
    skip_window = window_size // 2
    for layer_id, walks in enumerate(all_walks):
        for walk in walks:
            for i in range(len(walk)):
                for j in range(1, skip_window + 1):
                    if i - j >= 0:
                        pairs.append((vocab[walk[i]].index, vocab[walk[i - j]].index, layer_id))
                    if i + j < len(walk):
                        pairs.append((vocab[walk[i]].index, vocab[walk[i + j]].index, layer_id))
    return pairs


def generate_vocab(all_walks):
    index2word = []
    raw_vocab = defaultdict(int)

    for walks in all_walks:
        for walk in walks:
            for word in walk:
                raw_vocab[word] += 1

    vocab = {}
    for word, v in raw_vocab.items():
        vocab[word] = Vocab(count=v, index=len(index2word))
        index2word.append(word)

    index2word.sort(key=lambda word: vocab[word].count, reverse=True)
    for i, word in enumerate(index2word):
        vocab[word].index = i

    return vocab, index2word


def get_batches(pairs, neighbors, batch_size):
    n_batches = (len(pairs) + (batch_size - 1)) // batch_size

    # result = []
    for idx in range(n_batches):
        x, y, t, neigh = [], [], [], []
        for i in range(batch_size):
            index = idx * batch_size + i
            if index >= len(pairs):
                break
            x.append(pairs[index][0])
            y.append(pairs[index][1])
            t.append(pairs[index][2])
            neigh.append(neighbors[pairs[index][0]])
        yield torch.tensor(x), torch.tensor(y), torch.tensor(t), torch.tensor(neigh)


def generate_walks(network_data, num_walks, walk_length, schema=None):
    # if schema is not None:
    #     pass
    # else:
    #     node_type = None

    all_walks = []
    for layer_id in network_data:
        tmp_data = network_data[layer_id]
        # start to do the random walk on a layer

        layer_walker = RWGraph(get_G_from_edges(tmp_data))
        layer_walks = layer_walker.simulate_walks(num_walks, walk_length, schema=schema)

        all_walks.append(layer_walks)

    return all_walks

