def test(adj): ''' test on GCN ''' adj = normalize_adj_tensor(adj) gcn = GCN(nfeat=features.shape[1], nhid=args.hidden, nclass=labels.max().item() + 1, dropout=0.5) if device != 'cpu': gcn = gcn.to(device) optimizer = optim.Adam(gcn.parameters(), lr=args.lr, weight_decay=5e-4) gcn.train() for epoch in range(args.epochs): optimizer.zero_grad() output = gcn(features, adj) loss_train = F.nll_loss(output[idx_train], labels[idx_train]) acc_train = accuracy(output[idx_train], labels[idx_train]) loss_train.backward() optimizer.step() gcn.eval() output = gcn(features, adj) loss_test = F.nll_loss(output[idx_test], labels[idx_test]) acc_test = accuracy(output[idx_test], labels[idx_test]) # print("Test set results:", # "loss= {:.4f}".format(loss_test.item()), # "accuracy= {:.4f}".format(acc_test.item())) return acc_test.item()
def main(args): # convert boolean type for args assert args.self_loop in ['True', 'False'], [ "Only True or False for self_loop, get ", args.self_loop ] assert args.use_layernorm in ['True', 'False'], [ "Only True or False for use_layernorm, get ", args.use_layernorm ] self_loop = (args.self_loop == 'True') use_layernorm = (args.use_layernorm == 'True') global t0 if args.dataset in {'cora', 'citeseer', 'pubmed'}: data = load_data(args) else: raise NotImplementedError(f'{args.dataset} is not a valid dataset') features = torch.FloatTensor(data.features) labels = torch.LongTensor(data.labels) train_mask = torch.ByteTensor(data.train_mask) val_mask = torch.ByteTensor(data.val_mask) test_mask = torch.ByteTensor(data.test_mask) in_feats = features.shape[1] n_classes = data.num_labels n_edges = data.graph.number_of_edges() print("""----Data statistics------' #Edges %d #Classes %d #Train samples %d #Val samples %d #Test samples %d""" % (n_edges, n_classes, train_mask.sum().item(), val_mask.sum().item(), test_mask.sum().item())) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') features = features.to(device) labels = labels.to(device) train_mask = train_mask.to(device) val_mask = val_mask.to(device) test_mask = test_mask.to(device) # graph preprocess and calculate normalization factor g = data.graph # add self loop if self_loop: g.remove_edges_from(nx.selfloop_edges(g)) g.add_edges_from(zip(g.nodes(), g.nodes())) g = DGLGraph(g) g = g.to(device) n_edges = g.number_of_edges() # normalization degs = g.in_degrees().float() norm = torch.pow(degs, -0.5) norm[torch.isinf(norm)] = 0 norm = norm.to(device) g.ndata['norm'] = norm.unsqueeze(1) # create GCN model model = GCN(g, in_feats, args.n_hidden, n_classes, args.n_layers, F.relu, args.dropout, use_layernorm) model = model.to(device) loss_fcn = torch.nn.CrossEntropyLoss() # use optimizer optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay) # initialize graph record = [] dur = [] for epoch in range(args.n_epochs): if args.lr_scheduler: if epoch == int(0.5 * args.n_epochs): for pg in optimizer.param_groups: pg['lr'] = pg['lr'] / 10 elif epoch == int(0.75 * args.n_epochs): for pg in optimizer.param_groups: pg['lr'] = pg['lr'] / 10 model.train() if epoch >= 3: t0 = time.time() # forward optimizer.zero_grad() logits = model(features) loss = loss_fcn(logits[train_mask], labels[train_mask]) loss.backward() optimizer.step() if epoch >= 3: dur.append(time.time() - t0) acc_val = evaluate(model, features, labels, val_mask) acc_test = evaluate(model, features, labels, test_mask) record.append([acc_val, acc_test]) all_test_acc = [v[1] for v in record] all_val_acc = [v[0] for v in record] acc = evaluate(model, features, labels, test_mask) print(f"Final Test Accuracy: {acc:.4f}") print(f"Best Val Accuracy: {max(all_val_acc):.4f}") print(f"Best Test Accuracy: {max(all_test_acc):.4f}")
all_train_features = [] for _, _, _, features, _, _ in gcn_train_data: all_train_features.append(features) all_train_features = torch.cat(all_train_features) train_mean = torch.mean(all_train_features, dim=0) train_std = torch.std(all_train_features, dim=0) # train_mean[-10:] = 0 # train_std[-10:] = 1 # training GCN # model initiailization device = 'cuda' lr = 0.0005 model = GCN(raw_feature_size=full_feature_size) model = model.to(device=device) optimizer = torch.optim.Adam(model.parameters(), lr=lr) # loss functions mse_loss = nn.MSELoss() mae_loss = nn.L1Loss() loss_fn = mse_loss f_train_result = open('results/gcn/train.csv', 'a') f_test_result = open('results/gcn/test.csv', 'a') # training for epoch in range(2001): train_r2_list, train_mae_list, train_mse_list = [], [], [] test_r2_list, test_mae_list, test_mse_list = [], [], []
def main(training_file, dev_file, test_file, epochs=None, patience=None, num_heads=None, num_out_heads=None, num_layers=None, num_hidden=None, residual=None, in_drop=None, attn_drop=None, lr=None, weight_decay=None, alpha=None, batch_size=None, graph_type=None, net=None, freeze=None, cuda=None, fw=None): # number of training epochs if epochs is None: epochs = 400 print('EPOCHS', epochs) # used for early stop if patience is None: patience = 15 print('PATIENCE', patience) # number of hidden attention heads if num_heads is None: num_heads_ch = [4, 5, 6, 7] else: num_heads_ch = flattenList(num_heads) print('NUM HEADS', num_heads_ch) # number of output attention heads if num_out_heads is None: num_out_heads_ch = [4, 5, 6, 7] else: num_out_heads_ch = flattenList(num_out_heads) print('NUM OUT HEADS', num_out_heads_ch) # number of hidden layers if num_layers is None: num_layers_ch = [2, 3, 4, 5, 6] else: num_layers_ch = flattenList(num_layers) print('NUM LAYERS', num_layers_ch) # number of hidden units if num_hidden is None: num_hidden_ch = [32, 64, 96, 128, 256, 350, 512] else: num_hidden_ch = flattenList(num_hidden) print('NUM HIDDEN', num_hidden_ch) # use residual connection if residual is None: residual_ch = [True, False] else: residual_ch = flattenList(residual) print('RESIDUAL', residual_ch) # input feature dropout if in_drop is None: in_drop_ch = [0., 0.001, 0.0001, 0.00001] else: in_drop_ch = flattenList(in_drop) print('IN DROP', in_drop_ch) # attention dropout if attn_drop is None: attn_drop_ch = [0., 0.001, 0.0001, 0.00001] else: attn_drop_ch = flattenList(attn_drop) print('ATTENTION DROP', attn_drop_ch) # learning rate if lr is None: lr_ch = [0.0000005, 0.0000015, 0.00001, 0.00005, 0.0001] else: lr_ch = flattenList(lr) print('LEARNING RATE', lr_ch) # weight decay if weight_decay is None: weight_decay_ch = [0.0001, 0.001, 0.005] else: weight_decay_ch = flattenList(weight_decay) print('WEIGHT DECAY', weight_decay_ch) # the negative slop of leaky relu if alpha is None: alpha_ch = [0.1, 0.15, 0.2] else: alpha_ch = flattenList(alpha) print('ALPHA', alpha_ch) # batch size used for training, validation and test if batch_size is None: batch_size_ch = [175, 256, 350, 450, 512, 800, 1600] else: batch_size_ch = flattenList(batch_size) print('BATCH SIZE', batch_size_ch) # net type if net is None: net_ch = [GCN, GAT, RGCN, PGCN, PRGCN, GGN, PGAT] else: net_ch_raw = flattenList(net) net_ch = [] for ch in net_ch_raw: if ch.lower() == 'gcn': if fw == 'dgl': net_ch.append(GCN) else: net_ch.append(PGCN) elif ch.lower() == 'gat': if fw == 'dgl': net_ch.append(GAT) else: net_ch.append(PGAT) elif ch.lower() == 'rgcn': if fw == 'dgl': net_ch.append(RGCN) else: net_ch.append(PRGCN) elif ch.lower() == 'ggn': net_ch.append(GGN) elif ch.lower() == 'rgat': net_ch.append(PRGAT) else: print('Network type {} is not recognised.'.format(ch)) sys.exit(1) print('NET TYPE', net_ch) # graph type if net_ch in [GCN, GAT, PGCN, GGN, PGAT]: if graph_type is None: graph_type_ch = ['raw', '1', '2', '3', '4', 'relational'] else: graph_type_ch = flattenList(graph_type) else: if graph_type is None: graph_type_ch = ['relational'] else: graph_type_ch = flattenList(graph_type) print('GRAPH TYPE', graph_type_ch) # Freeze input neurons? if freeze is None: freeze_ch = [True, False] else: freeze_ch = flattenList(freeze) print('FREEZE', freeze_ch) # CUDA? if cuda is None: device = torch.device("cpu") elif cuda: device = torch.device("cuda") else: device = torch.device("cpu") print('DEVICE', device) if fw is None: fw = ['dgl', 'pg'] # define loss function # loss_fcn = torch.nn.BCEWithLogitsLoss() loss_fcn = torch.nn.MSELoss() for trial in range(10): trial_s = str(trial).zfill(6) num_heads = random.choice(num_heads_ch) num_out_heads = random.choice(num_out_heads_ch) num_layers = random.choice(num_layers_ch) num_hidden = random.choice(num_hidden_ch) residual = random.choice(residual_ch) in_drop = random.choice(in_drop_ch) attn_drop = random.choice(attn_drop_ch) lr = random.choice(lr_ch) weight_decay = random.choice(weight_decay_ch) alpha = random.choice(alpha_ch) batch_size = random.choice(batch_size_ch) graph_type = random.choice(graph_type_ch) net_class = random.choice(net_ch) freeze = random.choice(freeze_ch) fw = random.choice(fw) if freeze == False: freeze = 0 else: if graph_type == 'raw' or graph_type == '1' or graph_type == '2': freeze = 4 elif graph_type == '3' or graph_type == '4': freeze = 6 elif graph_type == 'relational': freeze = 5 else: sys.exit(1) print('=========================') print('TRIAL', trial_s) print('HEADS', num_heads) print('OUT_HEADS', num_out_heads) print('LAYERS', num_layers) print('HIDDEN', num_hidden) print('RESIDUAL', residual) print('inDROP', in_drop) print('atDROP', attn_drop) print('LR', lr) print('DECAY', weight_decay) print('ALPHA', alpha) print('BATCH', batch_size) print('GRAPH_ALT', graph_type) print('ARCHITECTURE', net_class) print('FREEZE', freeze) print('FRAMEWORK', fw) print('=========================') # create the dataset print('Loading training set...') train_dataset = SocNavDataset(training_file, mode='train', alt=graph_type) print('Loading dev set...') valid_dataset = SocNavDataset(dev_file, mode='valid', alt=graph_type) print('Loading test set...') test_dataset = SocNavDataset(test_file, mode='test', alt=graph_type) print('Done loading files') train_dataloader = DataLoader(train_dataset, batch_size=batch_size, collate_fn=collate) valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, collate_fn=collate) test_dataloader = DataLoader(test_dataset, batch_size=batch_size, collate_fn=collate) num_rels = train_dataset.data[0].num_rels cur_step = 0 best_loss = -1 n_classes = train_dataset.labels.shape[1] print('Number of classes: {}'.format(n_classes)) num_feats = train_dataset.features.shape[1] print('Number of features: {}'.format(num_feats)) g = train_dataset.graph heads = ([num_heads] * num_layers) + [num_out_heads] # define the model if fw == 'dgl': if net_class in [GCN]: model = GCN(g, num_feats, num_hidden, n_classes, num_layers, F.elu, in_drop) elif net_class in [GAT]: model = net_class(g, num_layers, num_feats, num_hidden, n_classes, heads, F.elu, in_drop, attn_drop, alpha, residual, freeze=freeze) else: # def __init__(self, g, in_dim, h_dim, out_dim, num_rels, num_hidden_layers=1): model = RGCN(g, in_dim=num_feats, h_dim=num_hidden, out_dim=n_classes, num_rels=num_rels, feat_drop=in_drop, num_hidden_layers=num_layers, freeze=freeze) else: if net_class in [PGCN]: model = PGCN( num_feats, n_classes, num_hidden, num_layers, in_drop, F.relu, improved=True, #Compute A-hat as A + 2I bias=True) elif net_class in [PRGCN]: model = PRGCN( num_feats, n_classes, num_rels, num_rels, #num_rels? # TODO: Add variable num_hidden, num_layers, in_drop, F.relu, bias=True) elif net_class in [PGAT]: model = PGAT(num_feats, n_classes, num_heads, in_drop, num_hidden, num_layers, F.relu, concat=True, neg_slope=alpha, bias=True) elif net_class in [PRGAT]: model = PRGAT( num_feats, n_classes, num_heads, num_rels, num_rels, #num_rels? # TODO: Add variable num_hidden, num_layers, num_layers, in_drop, F.relu, alpha, bias=True) else: model = GGN(num_feats, num_layers, aggr='mean', bias=True) #Describe the model #describe_model(model) # define the optimizer optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay) # for name, param in model.named_parameters(): # if param.requires_grad: # print(name, param.data.shape) model = model.to(device) for epoch in range(epochs): model.train() loss_list = [] for batch, data in enumerate(train_dataloader): subgraph, feats, labels = data subgraph.set_n_initializer(dgl.init.zero_initializer) subgraph.set_e_initializer(dgl.init.zero_initializer) feats = feats.to(device) labels = labels.to(device) if fw == 'dgl': model.g = subgraph for layer in model.layers: layer.g = subgraph logits = model(feats.float()) else: if net_class in [PGCN, PGAT, GGN]: data = Data(x=feats.float(), edge_index=torch.stack( subgraph.edges()).to(device)) else: data = Data( x=feats.float(), edge_index=torch.stack( subgraph.edges()).to(device), edge_type=subgraph.edata['rel_type'].squeeze().to( device)) logits = model(data) loss = loss_fcn(logits[getMaskForBatch(subgraph)], labels.float()) optimizer.zero_grad() a = list(model.parameters())[0].clone() loss.backward() optimizer.step() b = list(model.parameters())[0].clone() not_learning = torch.equal(a.data, b.data) if not_learning: import sys print('Not learning') # sys.exit(1) else: pass # print('Diff: ', (a.data-b.data).sum()) # print(loss.item()) loss_list.append(loss.item()) loss_data = np.array(loss_list).mean() print('Loss: {}'.format(loss_data)) if epoch % 5 == 0: if epoch % 5 == 0: print( "Epoch {:05d} | Loss: {:.4f} | Patience: {} | ".format( epoch, loss_data, cur_step), end='') score_list = [] val_loss_list = [] for batch, valid_data in enumerate(valid_dataloader): subgraph, feats, labels = valid_data subgraph.set_n_initializer(dgl.init.zero_initializer) subgraph.set_e_initializer(dgl.init.zero_initializer) feats = feats.to(device) labels = labels.to(device) score, val_loss = evaluate(feats.float(), model, subgraph, labels.float(), loss_fcn, fw, net_class) score_list.append(score) val_loss_list.append(val_loss) mean_score = np.array(score_list).mean() mean_val_loss = np.array(val_loss_list).mean() if epoch % 5 == 0: print("Score: {:.4f} MEAN: {:.4f} BEST: {:.4f}".format( mean_score, mean_val_loss, best_loss)) # early stop if best_loss > mean_val_loss or best_loss < 0: best_loss = mean_val_loss # Save the model # print('Writing to', trial_s) torch.save( model.state_dict(), fw + str(net) + '.tch' ) # 3 4 5 6 7 8 9 10 11 12 13 14 15 params = [ val_loss, graph_type, str(type(net_class)), g, num_layers, num_feats, num_hidden, n_classes, heads, F.elu, in_drop, attn_drop, alpha, residual, num_rels, freeze ] pickle.dump(params, open(fw + str(net) + '.prms', 'wb')) cur_step = 0 else: cur_step += 1 if cur_step >= patience: break torch.save(model, 'gattrial.pth') test_score_list = [] for batch, test_data in enumerate(test_dataloader): subgraph, feats, labels = test_data subgraph.set_n_initializer(dgl.init.zero_initializer) subgraph.set_e_initializer(dgl.init.zero_initializer) feats = feats.to(device) labels = labels.to(device) test_score_list.append( evaluate(feats, model, subgraph, labels.float(), loss_fcn, fw, net_class)[0]) print("F1-Score: {:.4f}".format(np.array(test_score_list).mean())) model.eval() return best_loss
adj_mtx = data_generator.get_adj_mat() # create model name and save modelname = "GCN" + \ "_bs_" + str(batch_size) + \ "_nemb_" + str(emb_dim) + \ "_layers_" + str(layers) + \ "_nodedr_" + str(node_dropout) + \ "_messdr_" + str(mess_dropout) + \ "_reg_" + str(reg) + \ "_lr_" + str(lr) # create GCN model model = GCN(data_generator.n_users, data_generator.n_items, emb_dim, layers, reg, node_dropout, mess_dropout, adj_mtx) model.to(device='cuda:0') model = torch.nn.DataParallel(model, device_ids=[0, 1]) # current best metric cur_best_metric = 0 # Adam optimizer # optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) optimizer = torch.optim.SGD(model.parameters(), lr=args.lr) # Set values for early stopping cur_best_loss, stopping_step, should_stop = 1e3, 0, False today = datetime.now() print("Start at " + str(today)) print("Using " + str(device) + " for computations") print("Params on CUDA: " + str(next(model.parameters()).is_cuda))