|
| 1 | +import time |
| 2 | +from tqdm import tqdm |
| 3 | +import numpy as np |
| 4 | +import torch |
| 5 | +from torch.nn import BCEWithLogitsLoss |
| 6 | +from dgl import NID, EID |
| 7 | +from dgl.dataloading import GraphDataLoader |
| 8 | +from utils import parse_arguments |
| 9 | +from utils import load_ogb_dataset, evaluate_hits |
| 10 | +from sampler import SEALData |
| 11 | +from model import GCN, DGCNN |
| 12 | +from logger import LightLogging |
| 13 | + |
| 14 | +''' |
| 15 | +Part of the code are adapted from |
| 16 | +https://github.com/facebookresearch/SEAL_OGB |
| 17 | +''' |
| 18 | + |
| 19 | + |
| 20 | +def train(model, dataloader, loss_fn, optimizer, device, num_graphs=32, total_graphs=None): |
| 21 | + model.train() |
| 22 | + |
| 23 | + total_loss = 0 |
| 24 | + for g, labels in tqdm(dataloader, ncols=100): |
| 25 | + g = g.to(device) |
| 26 | + labels = labels.to(device) |
| 27 | + optimizer.zero_grad() |
| 28 | + logits = model(g, g.ndata['z'], g.ndata[NID], g.edata[EID]) |
| 29 | + loss = loss_fn(logits, labels) |
| 30 | + loss.backward() |
| 31 | + optimizer.step() |
| 32 | + total_loss += loss.item() * num_graphs |
| 33 | + |
| 34 | + return total_loss / total_graphs |
| 35 | + |
| 36 | + |
| 37 | +@torch.no_grad() |
| 38 | +def evaluate(model, dataloader, device): |
| 39 | + model.eval() |
| 40 | + |
| 41 | + y_pred, y_true = [], [] |
| 42 | + for g, labels in tqdm(dataloader, ncols=100): |
| 43 | + g = g.to(device) |
| 44 | + logits = model(g, g.ndata['z'], g.ndata[NID], g.edata[EID]) |
| 45 | + y_pred.append(logits.view(-1).cpu()) |
| 46 | + y_true.append(labels.view(-1).cpu().to(torch.float)) |
| 47 | + |
| 48 | + y_pred, y_true = torch.cat(y_pred), torch.cat(y_true) |
| 49 | + pos_pred = y_pred[y_true == 1] |
| 50 | + neg_pred = y_pred[y_true == 0] |
| 51 | + |
| 52 | + return pos_pred, neg_pred |
| 53 | + |
| 54 | + |
| 55 | +def main(args, print_fn=print): |
| 56 | + print_fn("Experiment arguments: {}".format(args)) |
| 57 | + |
| 58 | + if args.random_seed: |
| 59 | + torch.manual_seed(args.random_seed) |
| 60 | + else: |
| 61 | + torch.manual_seed(123) |
| 62 | + # Load dataset |
| 63 | + if args.dataset.startswith('ogbl'): |
| 64 | + graph, split_edge = load_ogb_dataset(args.dataset) |
| 65 | + else: |
| 66 | + raise NotImplementedError |
| 67 | + |
| 68 | + num_nodes = graph.num_nodes() |
| 69 | + |
| 70 | + # set gpu |
| 71 | + if args.gpu_id >= 0 and torch.cuda.is_available(): |
| 72 | + device = 'cuda:{}'.format(args.gpu_id) |
| 73 | + else: |
| 74 | + device = 'cpu' |
| 75 | + |
| 76 | + if args.dataset == 'ogbl-collab': |
| 77 | + # ogbl-collab dataset is multi-edge graph |
| 78 | + use_coalesce = True |
| 79 | + else: |
| 80 | + use_coalesce = False |
| 81 | + |
| 82 | + # Generate positive and negative edges and corresponding labels |
| 83 | + # Sampling subgraphs and generate node labeling features |
| 84 | + seal_data = SEALData(g=graph, split_edge=split_edge, hop=args.hop, neg_samples=args.neg_samples, |
| 85 | + subsample_ratio=args.subsample_ratio, use_coalesce=use_coalesce, prefix=args.dataset, |
| 86 | + save_dir=args.save_dir, num_workers=args.num_workers, print_fn=print_fn) |
| 87 | + node_attribute = seal_data.ndata['feat'] |
| 88 | + edge_weight = seal_data.edata['edge_weight'].float() |
| 89 | + |
| 90 | + train_data = seal_data('train') |
| 91 | + val_data = seal_data('valid') |
| 92 | + test_data = seal_data('test') |
| 93 | + |
| 94 | + train_graphs = len(train_data.graph_list) |
| 95 | + |
| 96 | + # Set data loader |
| 97 | + |
| 98 | + train_loader = GraphDataLoader(train_data, batch_size=args.batch_size, num_workers=args.num_workers) |
| 99 | + val_loader = GraphDataLoader(val_data, batch_size=args.batch_size, num_workers=args.num_workers) |
| 100 | + test_loader = GraphDataLoader(test_data, batch_size=args.batch_size, num_workers=args.num_workers) |
| 101 | + |
| 102 | + # set model |
| 103 | + if args.model == 'gcn': |
| 104 | + model = GCN(num_layers=args.num_layers, |
| 105 | + hidden_units=args.hidden_units, |
| 106 | + gcn_type=args.gcn_type, |
| 107 | + pooling_type=args.pooling, |
| 108 | + node_attributes=node_attribute, |
| 109 | + edge_weights=edge_weight, |
| 110 | + node_embedding=None, |
| 111 | + use_embedding=True, |
| 112 | + num_nodes=num_nodes, |
| 113 | + dropout=args.dropout) |
| 114 | + elif args.model == 'dgcnn': |
| 115 | + model = DGCNN(num_layers=args.num_layers, |
| 116 | + hidden_units=args.hidden_units, |
| 117 | + k=args.sort_k, |
| 118 | + gcn_type=args.gcn_type, |
| 119 | + node_attributes=node_attribute, |
| 120 | + edge_weights=edge_weight, |
| 121 | + node_embedding=None, |
| 122 | + use_embedding=True, |
| 123 | + num_nodes=num_nodes, |
| 124 | + dropout=args.dropout) |
| 125 | + else: |
| 126 | + raise ValueError('Model error') |
| 127 | + |
| 128 | + model = model.to(device) |
| 129 | + parameters = model.parameters() |
| 130 | + optimizer = torch.optim.Adam(parameters, lr=args.lr) |
| 131 | + loss_fn = BCEWithLogitsLoss() |
| 132 | + print_fn("Total parameters: {}".format(sum([p.numel() for p in model.parameters()]))) |
| 133 | + |
| 134 | + # train and evaluate loop |
| 135 | + summary_val = [] |
| 136 | + summary_test = [] |
| 137 | + for epoch in range(args.epochs): |
| 138 | + start_time = time.time() |
| 139 | + loss = train(model=model, |
| 140 | + dataloader=train_loader, |
| 141 | + loss_fn=loss_fn, |
| 142 | + optimizer=optimizer, |
| 143 | + device=device, |
| 144 | + num_graphs=args.batch_size, |
| 145 | + total_graphs=train_graphs) |
| 146 | + train_time = time.time() |
| 147 | + if epoch % args.eval_steps == 0: |
| 148 | + val_pos_pred, val_neg_pred = evaluate(model=model, |
| 149 | + dataloader=val_loader, |
| 150 | + device=device) |
| 151 | + test_pos_pred, test_neg_pred = evaluate(model=model, |
| 152 | + dataloader=test_loader, |
| 153 | + device=device) |
| 154 | + |
| 155 | + val_metric = evaluate_hits(args.dataset, val_pos_pred, val_neg_pred, args.hits_k) |
| 156 | + test_metric = evaluate_hits(args.dataset, test_pos_pred, test_neg_pred, args.hits_k) |
| 157 | + evaluate_time = time.time() |
| 158 | + print_fn("Epoch-{}, train loss: {:.4f}, hits@{}: val-{:.4f}, test-{:.4f}, " |
| 159 | + "cost time: train-{:.1f}s, total-{:.1f}s".format(epoch, loss, args.hits_k, val_metric, test_metric, |
| 160 | + train_time - start_time, |
| 161 | + evaluate_time - start_time)) |
| 162 | + summary_val.append(val_metric) |
| 163 | + summary_test.append(test_metric) |
| 164 | + |
| 165 | + summary_test = np.array(summary_test) |
| 166 | + |
| 167 | + print_fn("Experiment Results:") |
| 168 | + print_fn("Best hits@{}: {:.4f}, epoch: {}".format(args.hits_k, np.max(summary_test), np.argmax(summary_test))) |
| 169 | + |
| 170 | + |
| 171 | +if __name__ == '__main__': |
| 172 | + args = parse_arguments() |
| 173 | + logger = LightLogging(log_name='SEAL', log_path='./logs') |
| 174 | + main(args, logger.info) |
0 commit comments