图

Hint

• 作者: 丰一帆

• 翻译: 颜杰龙

• 校对： 丰一帆 、张欣炜

在如下的例子中，我们使用四种典型图/超图神经网络在图关联结构中执行节点分类任务。

模型

• GCN (dhg.models.GCN), Semi-Supervised Classification with Graph Convolutional Networks 论文 (ICLR 2017).

• GAT (dhg.models.GAT), Graph Attention Networks 论文 (ICLR 2018).

• HGNN (dhg.models.HGNN), Hypergraph Neural Networks 论文 (AAAI 2019).

• HGNN+ (dhg.models.HGNNP), HGNN+: General Hypergraph Neural Networks 论文 (IEEE T-PAMI 2022).

数据集

Cora数据集 (dhg.data.Cora) 为节点分类任务中使用的引用网络数据集。 更多细节可以参考此 网页.

结果汇总

模型	验证集的Accuracy	测试集的Accuracy	测试集的F1 score
GCN	0.800	0.823	0.814
GAT	0.804	0.824	0.817
HGNN	0.804	0.820	0.811
HGNN+	0.802	0.827	0.820

Cora上使用GCN

导入依赖包

import time
from copy import deepcopy

import torch
import torch.optim as optim
import torch.nn.functional as F

from dhg import Graph
from dhg.data import Cora
from dhg.models import GCN
from dhg.random import set_seed
from dhg.metrics import GraphVertexClassificationEvaluator as Evaluator

定义函数

def train(net, X, A, lbls, train_idx, optimizer, epoch):
    net.train()

    st = time.time()
    optimizer.zero_grad()
    outs = net(X, A)
    outs, lbls = outs[train_idx], lbls[train_idx]
    loss = F.cross_entropy(outs, lbls)
    loss.backward()
    optimizer.step()
    print(f"Epoch: {epoch}, Time: {time.time()-st:.5f}s, Loss: {loss.item():.5f}")
    return loss.item()


@torch.no_grad()
def infer(net, X, A, lbls, idx, test=False):
    net.eval()
    outs = net(X, A)
    outs, lbls = outs[idx], lbls[idx]
    if not test:
        res = evaluator.validate(lbls, outs)
    else:
        res = evaluator.test(lbls, outs)
    return res

主函数

Note

更多关于评测器 Evaluator 的细节可以参照 构建指标评测器 章节。

if __name__ == "__main__":
    set_seed(2022)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cora()
    X, lbl = data["features"], data["labels"]
    G = Graph(data["num_vertices"], data["edge_list"])
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = GCN(data["dim_features"], 16, data["num_classes"])
    optimizer = optim.Adam(net.parameters(), lr=0.01, weight_decay=5e-4)

    X, lbl = X.to(device), lbl.to(device)
    G = G.to(device)
    net = net.to(device)

    best_state = None
    best_epoch, best_val = 0, 0
    for epoch in range(300):
        # train
        train(net, X, G, lbl, train_mask, optimizer, epoch)
        # validation
        if epoch % 1 == 0:
            with torch.no_grad():
                val_res = infer(net, X, G, lbl, val_mask)
            if val_res > best_val:
                print(f"update best: {val_res:.5f}")
                best_epoch = epoch
                best_val = val_res
                best_state = deepcopy(net.state_dict())
    print("\ntrain finished!")
    print(f"best val: {best_val:.5f}")
    # test
    print("test...")
    net.load_state_dict(best_state)
    res = infer(net, X, G, lbl, test_mask, test=True)
    print(f"final result: epoch: {best_epoch}")
    print(res)

输出

Epoch: 0, Time: 0.51907s, Loss: 1.95010
update best: 0.31600
Epoch: 1, Time: 0.00182s, Loss: 1.94601
Epoch: 2, Time: 0.00164s, Loss: 1.94383
Epoch: 3, Time: 0.00167s, Loss: 1.93907
Epoch: 4, Time: 0.00164s, Loss: 1.93350
update best: 0.38000
Epoch: 5, Time: 0.00166s, Loss: 1.92899
Epoch: 6, Time: 0.00323s, Loss: 1.92461
Epoch: 7, Time: 0.00164s, Loss: 1.91764
Epoch: 8, Time: 0.00341s, Loss: 1.91163
Epoch: 9, Time: 0.00167s, Loss: 1.90064
Epoch: 10, Time: 0.00154s, Loss: 1.89617
Epoch: 11, Time: 0.00159s, Loss: 1.88721
Epoch: 12, Time: 0.00223s, Loss: 1.87626
Epoch: 13, Time: 0.00178s, Loss: 1.86941
Epoch: 14, Time: 0.00167s, Loss: 1.86202
Epoch: 15, Time: 0.00316s, Loss: 1.85207
Epoch: 16, Time: 0.00215s, Loss: 1.84285
Epoch: 17, Time: 0.00289s, Loss: 1.83392
Epoch: 18, Time: 0.00203s, Loss: 1.82120
Epoch: 19, Time: 0.00202s, Loss: 1.80663
Epoch: 20, Time: 0.00246s, Loss: 1.79340
Epoch: 21, Time: 0.00201s, Loss: 1.77829
Epoch: 22, Time: 0.00203s, Loss: 1.76851
update best: 0.38800
Epoch: 23, Time: 0.00162s, Loss: 1.75592
update best: 0.40200
Epoch: 24, Time: 0.00159s, Loss: 1.74545
update best: 0.43000
Epoch: 25, Time: 0.00175s, Loss: 1.72373
update best: 0.45000
Epoch: 26, Time: 0.00157s, Loss: 1.71025
update best: 0.46000
Epoch: 27, Time: 0.00164s, Loss: 1.68904
update best: 0.46400
Epoch: 28, Time: 0.00211s, Loss: 1.67401
update best: 0.46600
Epoch: 29, Time: 0.00168s, Loss: 1.67025
update best: 0.48400
Epoch: 30, Time: 0.00176s, Loss: 1.65349
update best: 0.49200
Epoch: 31, Time: 0.00250s, Loss: 1.61911
update best: 0.49800
Epoch: 32, Time: 0.00177s, Loss: 1.61325
update best: 0.51400
Epoch: 33, Time: 0.00192s, Loss: 1.56832
update best: 0.52600
Epoch: 34, Time: 0.00173s, Loss: 1.55827
update best: 0.55000
Epoch: 35, Time: 0.00172s, Loss: 1.55186
update best: 0.56200
Epoch: 36, Time: 0.00183s, Loss: 1.53794
update best: 0.57400
Epoch: 37, Time: 0.00222s, Loss: 1.50345
update best: 0.58600
Epoch: 38, Time: 0.00169s, Loss: 1.49760
update best: 0.59600
Epoch: 39, Time: 0.00164s, Loss: 1.47143
update best: 0.60200
Epoch: 40, Time: 0.00171s, Loss: 1.43501
update best: 0.62800
Epoch: 41, Time: 0.00170s, Loss: 1.42085
update best: 0.64800
Epoch: 42, Time: 0.00360s, Loss: 1.38769
update best: 0.65400
Epoch: 43, Time: 0.00156s, Loss: 1.36689
update best: 0.66200
Epoch: 44, Time: 0.00152s, Loss: 1.36428
update best: 0.66800
Epoch: 45, Time: 0.00167s, Loss: 1.32395
Epoch: 46, Time: 0.00153s, Loss: 1.29274
update best: 0.67600
Epoch: 47, Time: 0.00164s, Loss: 1.30380
Epoch: 48, Time: 0.00439s, Loss: 1.26099
update best: 0.68800
Epoch: 49, Time: 0.00186s, Loss: 1.25379
Epoch: 50, Time: 0.00175s, Loss: 1.23854
update best: 0.69800
Epoch: 51, Time: 0.00171s, Loss: 1.20378
update best: 0.72200
Epoch: 52, Time: 0.00170s, Loss: 1.16979
update best: 0.73200
Epoch: 53, Time: 0.00326s, Loss: 1.15275
update best: 0.74800
Epoch: 54, Time: 0.00183s, Loss: 1.11128
update best: 0.75200
Epoch: 55, Time: 0.00183s, Loss: 1.12654
update best: 0.75600
Epoch: 56, Time: 0.00172s, Loss: 1.12641
update best: 0.76400
Epoch: 57, Time: 0.00171s, Loss: 1.08093
update best: 0.76600
Epoch: 58, Time: 0.00228s, Loss: 1.06145
Epoch: 59, Time: 0.00163s, Loss: 1.03330
Epoch: 60, Time: 0.00240s, Loss: 1.02479
Epoch: 61, Time: 0.00179s, Loss: 1.01496
Epoch: 62, Time: 0.00187s, Loss: 0.93007
Epoch: 63, Time: 0.00176s, Loss: 0.97366
Epoch: 64, Time: 0.00296s, Loss: 0.92534
Epoch: 65, Time: 0.00230s, Loss: 0.91500
update best: 0.77400
Epoch: 66, Time: 0.00169s, Loss: 0.93400
update best: 0.77800
Epoch: 67, Time: 0.00161s, Loss: 0.86869
update best: 0.78000
Epoch: 68, Time: 0.00162s, Loss: 0.89109
Epoch: 69, Time: 0.00177s, Loss: 0.89371
Epoch: 70, Time: 0.00259s, Loss: 0.87362
update best: 0.78200
Epoch: 71, Time: 0.00159s, Loss: 0.80287
Epoch: 72, Time: 0.00155s, Loss: 0.88049
Epoch: 73, Time: 0.00160s, Loss: 0.78692
Epoch: 74, Time: 0.00163s, Loss: 0.79204
Epoch: 75, Time: 0.00152s, Loss: 0.81149
update best: 0.78400
Epoch: 76, Time: 0.00288s, Loss: 0.79278
Epoch: 77, Time: 0.00183s, Loss: 0.75974
update best: 0.78600
Epoch: 78, Time: 0.00155s, Loss: 0.74237
Epoch: 79, Time: 0.00162s, Loss: 0.72129
update best: 0.78800
Epoch: 80, Time: 0.00154s, Loss: 0.72252
update best: 0.79000
Epoch: 81, Time: 0.00170s, Loss: 0.69306
update best: 0.79200
Epoch: 82, Time: 0.00274s, Loss: 0.64976
Epoch: 83, Time: 0.00157s, Loss: 0.66782
Epoch: 84, Time: 0.00155s, Loss: 0.68008
Epoch: 85, Time: 0.00160s, Loss: 0.70714
Epoch: 86, Time: 0.00164s, Loss: 0.64139
Epoch: 87, Time: 0.00159s, Loss: 0.66335
Epoch: 88, Time: 0.00223s, Loss: 0.65881
Epoch: 89, Time: 0.00248s, Loss: 0.65215
Epoch: 90, Time: 0.00151s, Loss: 0.57064
Epoch: 91, Time: 0.00155s, Loss: 0.64725
Epoch: 92, Time: 0.00157s, Loss: 0.58507
Epoch: 93, Time: 0.00174s, Loss: 0.62494
Epoch: 94, Time: 0.00158s, Loss: 0.58289
Epoch: 95, Time: 0.00157s, Loss: 0.56591
Epoch: 96, Time: 0.00289s, Loss: 0.59959
Epoch: 97, Time: 0.00157s, Loss: 0.62588
Epoch: 98, Time: 0.00154s, Loss: 0.58035
Epoch: 99, Time: 0.00156s, Loss: 0.58727
Epoch: 100, Time: 0.00158s, Loss: 0.56111
Epoch: 101, Time: 0.00152s, Loss: 0.54035
Epoch: 102, Time: 0.00151s, Loss: 0.56815
Epoch: 103, Time: 0.00233s, Loss: 0.50579
Epoch: 104, Time: 0.00150s, Loss: 0.53285
Epoch: 105, Time: 0.00147s, Loss: 0.56204
Epoch: 106, Time: 0.00153s, Loss: 0.51602
Epoch: 107, Time: 0.00160s, Loss: 0.52320
Epoch: 108, Time: 0.00150s, Loss: 0.53845
Epoch: 109, Time: 0.00151s, Loss: 0.55428
Epoch: 110, Time: 0.00307s, Loss: 0.52966
Epoch: 111, Time: 0.00150s, Loss: 0.56845
Epoch: 112, Time: 0.00148s, Loss: 0.52385
update best: 0.79400
Epoch: 113, Time: 0.00155s, Loss: 0.52051
Epoch: 114, Time: 0.00178s, Loss: 0.51860
Epoch: 115, Time: 0.00159s, Loss: 0.48878
Epoch: 116, Time: 0.00375s, Loss: 0.50367
Epoch: 117, Time: 0.00160s, Loss: 0.49782
Epoch: 118, Time: 0.00153s, Loss: 0.51155
Epoch: 119, Time: 0.00153s, Loss: 0.47739
Epoch: 120, Time: 0.00178s, Loss: 0.50645
Epoch: 121, Time: 0.00157s, Loss: 0.49175
Epoch: 122, Time: 0.00157s, Loss: 0.47638
Epoch: 123, Time: 0.00345s, Loss: 0.46064
Epoch: 124, Time: 0.00159s, Loss: 0.44845
Epoch: 125, Time: 0.00153s, Loss: 0.44286
Epoch: 126, Time: 0.00151s, Loss: 0.46044
Epoch: 127, Time: 0.00156s, Loss: 0.45707
Epoch: 128, Time: 0.00177s, Loss: 0.50700
Epoch: 129, Time: 0.00153s, Loss: 0.46442
Epoch: 130, Time: 0.00345s, Loss: 0.44911
Epoch: 131, Time: 0.00153s, Loss: 0.46168
Epoch: 132, Time: 0.00153s, Loss: 0.47634
Epoch: 133, Time: 0.00152s, Loss: 0.41177
Epoch: 134, Time: 0.00162s, Loss: 0.42612
Epoch: 135, Time: 0.00160s, Loss: 0.46436
Epoch: 136, Time: 0.00153s, Loss: 0.42374
Epoch: 137, Time: 0.00380s, Loss: 0.42290
Epoch: 138, Time: 0.00181s, Loss: 0.43096
Epoch: 139, Time: 0.00166s, Loss: 0.43386
Epoch: 140, Time: 0.00170s, Loss: 0.47472
Epoch: 141, Time: 0.00175s, Loss: 0.40687
Epoch: 142, Time: 0.00170s, Loss: 0.43927
Epoch: 143, Time: 0.00347s, Loss: 0.39323
Epoch: 144, Time: 0.00174s, Loss: 0.42356
Epoch: 145, Time: 0.00168s, Loss: 0.44625
Epoch: 146, Time: 0.00165s, Loss: 0.38619
Epoch: 147, Time: 0.00171s, Loss: 0.40754
Epoch: 148, Time: 0.00169s, Loss: 0.38543
Epoch: 149, Time: 0.00166s, Loss: 0.39466
Epoch: 150, Time: 0.00280s, Loss: 0.43009
Epoch: 151, Time: 0.00165s, Loss: 0.38695
Epoch: 152, Time: 0.00166s, Loss: 0.41950
Epoch: 153, Time: 0.00166s, Loss: 0.41095
Epoch: 154, Time: 0.00174s, Loss: 0.40313
Epoch: 155, Time: 0.00167s, Loss: 0.43876
Epoch: 156, Time: 0.00384s, Loss: 0.40152
Epoch: 157, Time: 0.00170s, Loss: 0.39797
update best: 0.80000
Epoch: 158, Time: 0.00165s, Loss: 0.35990
Epoch: 159, Time: 0.00168s, Loss: 0.40668
Epoch: 160, Time: 0.00161s, Loss: 0.39737
Epoch: 161, Time: 0.00153s, Loss: 0.42709
Epoch: 162, Time: 0.00174s, Loss: 0.40306
Epoch: 163, Time: 0.00262s, Loss: 0.44195
Epoch: 164, Time: 0.00150s, Loss: 0.35434
Epoch: 165, Time: 0.00154s, Loss: 0.39269
Epoch: 166, Time: 0.00159s, Loss: 0.32633
Epoch: 167, Time: 0.00154s, Loss: 0.38579
Epoch: 168, Time: 0.00155s, Loss: 0.38941
Epoch: 169, Time: 0.00150s, Loss: 0.38425
Epoch: 170, Time: 0.00250s, Loss: 0.39287
Epoch: 171, Time: 0.00153s, Loss: 0.36239
Epoch: 172, Time: 0.00153s, Loss: 0.37962
Epoch: 173, Time: 0.00154s, Loss: 0.35394
Epoch: 174, Time: 0.00159s, Loss: 0.34589
Epoch: 175, Time: 0.00161s, Loss: 0.38056
Epoch: 176, Time: 0.00156s, Loss: 0.37199
Epoch: 177, Time: 0.00309s, Loss: 0.36108
Epoch: 178, Time: 0.00181s, Loss: 0.37211
Epoch: 179, Time: 0.00153s, Loss: 0.35234
Epoch: 180, Time: 0.00155s, Loss: 0.33577
Epoch: 181, Time: 0.00153s, Loss: 0.37541
Epoch: 182, Time: 0.00156s, Loss: 0.30629
Epoch: 183, Time: 0.00149s, Loss: 0.36643
Epoch: 184, Time: 0.00346s, Loss: 0.34131
Epoch: 185, Time: 0.00153s, Loss: 0.35421
Epoch: 186, Time: 0.00146s, Loss: 0.33999
Epoch: 187, Time: 0.00149s, Loss: 0.36365
Epoch: 188, Time: 0.00152s, Loss: 0.36926
Epoch: 189, Time: 0.00152s, Loss: 0.31029
Epoch: 190, Time: 0.00155s, Loss: 0.32959
Epoch: 191, Time: 0.00247s, Loss: 0.35637
Epoch: 192, Time: 0.00208s, Loss: 0.30936
Epoch: 193, Time: 0.00154s, Loss: 0.32842
Epoch: 194, Time: 0.00154s, Loss: 0.31046
Epoch: 195, Time: 0.00156s, Loss: 0.34217
Epoch: 196, Time: 0.00169s, Loss: 0.35384
Epoch: 197, Time: 0.00157s, Loss: 0.31096
Epoch: 198, Time: 0.00307s, Loss: 0.31790
Epoch: 199, Time: 0.00160s, Loss: 0.29574

train finished!
best val: 0.80000
test...
final result: epoch: 157
{'accuracy': 0.8230000138282776, 'f1_score': 0.8135442845966843, 'f1_score -> average@micro': 0.823}

Cora上使用GAT

导入依赖包

import time
from copy import deepcopy

import torch
import torch.optim as optim
import torch.nn.functional as F

from dhg import Graph
from dhg.data import Cora
from dhg.models import GAT
from dhg.random import set_seed
from dhg.metrics import GraphVertexClassificationEvaluator as Evaluator

定义函数

def train(net, X, A, lbls, train_idx, optimizer, epoch):
    net.train()

    st = time.time()
    optimizer.zero_grad()
    outs = net(X, A)
    outs, lbls = outs[train_idx], lbls[train_idx]
    loss = F.cross_entropy(outs, lbls)
    # loss = F.nll_loss(outs, lbls)
    loss.backward()
    optimizer.step()
    print(f"Epoch: {epoch}, Time: {time.time()-st:.5f}s, Loss: {loss.item():.5f}")
    return loss.item()


@torch.no_grad()
def infer(net, X, A, lbls, idx, test=False):
    net.eval()
    outs = net(X, A)
    outs, lbls = outs[idx], lbls[idx]
    if not test:
        res = evaluator.validate(lbls, outs)
    else:
        res = evaluator.test(lbls, outs)
    return res

主函数

Note

更多关于评测器 Evaluator 的细节可以参照 构建指标评测器 章节。

if __name__ == "__main__":
    set_seed(2022)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cora()
    X, lbl = data["features"], data["labels"]
    G = Graph(data["num_vertices"], data["edge_list"])
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = GAT(data["dim_features"], 8, data["num_classes"], num_heads=8, drop_rate=0.6)
    optimizer = optim.Adam(net.parameters(), lr=0.005, weight_decay=5e-4)

    X, lbl = X.cuda(), lbl.cuda()
    G = G.to(device)
    net = net.cuda()

    best_state = None
    best_epoch, best_val = 0, 0
    for epoch in range(300):
        # train
        train(net, X, G, lbl, train_mask, optimizer, epoch)
        # validation
        if epoch % 1 == 0:
            with torch.no_grad():
                val_res = infer(net, X, G, lbl, val_mask)
            if val_res > best_val:
                print(f"update best: {val_res:.5f}")
                best_epoch = epoch
                best_val = val_res
                best_state = deepcopy(net.state_dict())
    print("\ntrain finished!")
    print(f"best val: {best_val:.5f}")
    # test
    print("test...")
    net.load_state_dict(best_state)
    res = infer(net, X, G, lbl, test_mask, test=True)
    print(f"final result: epoch: {best_epoch}")
    print(res)

输出

Epoch: 0, Time: 0.56263s, Loss: 1.94867
update best: 0.12200
Epoch: 1, Time: 0.03209s, Loss: 1.94319
Epoch: 2, Time: 0.03641s, Loss: 1.94076
Epoch: 3, Time: 0.03197s, Loss: 1.93497
Epoch: 4, Time: 0.03536s, Loss: 1.92976
Epoch: 5, Time: 0.03239s, Loss: 1.92725
update best: 0.18200
Epoch: 6, Time: 0.03332s, Loss: 1.91903
update best: 0.38200
Epoch: 7, Time: 0.03125s, Loss: 1.91507
update best: 0.49200
Epoch: 8, Time: 0.02778s, Loss: 1.91092
update best: 0.50400
Epoch: 9, Time: 0.03188s, Loss: 1.90407
update best: 0.51800
Epoch: 10, Time: 0.02726s, Loss: 1.89345
update best: 0.54000
Epoch: 11, Time: 0.03213s, Loss: 1.88819
update best: 0.56800
Epoch: 12, Time: 0.03265s, Loss: 1.88074
update best: 0.58800
Epoch: 13, Time: 0.03181s, Loss: 1.87387
update best: 0.61200
Epoch: 14, Time: 0.02740s, Loss: 1.86807
update best: 0.63600
Epoch: 15, Time: 0.02897s, Loss: 1.85900
update best: 0.68200
Epoch: 16, Time: 0.02701s, Loss: 1.84736
update best: 0.69800
Epoch: 17, Time: 0.02716s, Loss: 1.83908
update best: 0.72200
Epoch: 18, Time: 0.02782s, Loss: 1.83323
update best: 0.74800
Epoch: 19, Time: 0.02795s, Loss: 1.81363
update best: 0.77400
Epoch: 20, Time: 0.02841s, Loss: 1.80020
update best: 0.78200
Epoch: 21, Time: 0.02796s, Loss: 1.79673
update best: 0.79400
Epoch: 22, Time: 0.02665s, Loss: 1.77684
update best: 0.79600
Epoch: 23, Time: 0.02657s, Loss: 1.75575
Epoch: 24, Time: 0.02713s, Loss: 1.74837
Epoch: 25, Time: 0.02716s, Loss: 1.74580
Epoch: 26, Time: 0.02909s, Loss: 1.71996
Epoch: 27, Time: 0.02656s, Loss: 1.70845
Epoch: 28, Time: 0.02675s, Loss: 1.69779
Epoch: 29, Time: 0.02614s, Loss: 1.66469
Epoch: 30, Time: 0.02627s, Loss: 1.66196
Epoch: 31, Time: 0.02743s, Loss: 1.65097
Epoch: 32, Time: 0.02776s, Loss: 1.62630
Epoch: 33, Time: 0.02752s, Loss: 1.60284
Epoch: 34, Time: 0.02749s, Loss: 1.58056
Epoch: 35, Time: 0.02549s, Loss: 1.57601
Epoch: 36, Time: 0.02724s, Loss: 1.55081
Epoch: 37, Time: 0.02836s, Loss: 1.53101
Epoch: 38, Time: 0.02641s, Loss: 1.53054
Epoch: 39, Time: 0.02638s, Loss: 1.51172
Epoch: 40, Time: 0.02669s, Loss: 1.45463
Epoch: 41, Time: 0.02674s, Loss: 1.43432
Epoch: 42, Time: 0.02729s, Loss: 1.39888
Epoch: 43, Time: 0.02715s, Loss: 1.40660
Epoch: 44, Time: 0.02975s, Loss: 1.41301
Epoch: 45, Time: 0.02658s, Loss: 1.32990
Epoch: 46, Time: 0.02753s, Loss: 1.31327
Epoch: 47, Time: 0.02823s, Loss: 1.30501
Epoch: 48, Time: 0.02904s, Loss: 1.28125
Epoch: 49, Time: 0.02605s, Loss: 1.23469
Epoch: 50, Time: 0.02684s, Loss: 1.25209
Epoch: 51, Time: 0.02576s, Loss: 1.24679
Epoch: 52, Time: 0.02693s, Loss: 1.20283
Epoch: 53, Time: 0.02735s, Loss: 1.16539
Epoch: 54, Time: 0.02733s, Loss: 1.16182
Epoch: 55, Time: 0.02691s, Loss: 1.12086
Epoch: 56, Time: 0.02706s, Loss: 1.09962
Epoch: 57, Time: 0.02628s, Loss: 1.09911
Epoch: 58, Time: 0.02716s, Loss: 1.05156
Epoch: 59, Time: 0.02729s, Loss: 1.03817
Epoch: 60, Time: 0.03020s, Loss: 0.99580
Epoch: 61, Time: 0.02628s, Loss: 0.98298
Epoch: 62, Time: 0.02804s, Loss: 0.95318
Epoch: 63, Time: 0.02650s, Loss: 0.94846
Epoch: 64, Time: 0.02753s, Loss: 0.94741
Epoch: 65, Time: 0.02678s, Loss: 0.92977
Epoch: 66, Time: 0.02639s, Loss: 0.85785
Epoch: 67, Time: 0.02938s, Loss: 0.87859
Epoch: 68, Time: 0.02816s, Loss: 0.81501
Epoch: 69, Time: 0.02799s, Loss: 0.82868
Epoch: 70, Time: 0.02577s, Loss: 0.83454
Epoch: 71, Time: 0.03040s, Loss: 0.81279
Epoch: 72, Time: 0.02764s, Loss: 0.80267
Epoch: 73, Time: 0.02707s, Loss: 0.77012
Epoch: 74, Time: 0.02769s, Loss: 0.75785
Epoch: 75, Time: 0.02844s, Loss: 0.70275
Epoch: 76, Time: 0.02718s, Loss: 0.73779
Epoch: 77, Time: 0.02707s, Loss: 0.75283
Epoch: 78, Time: 0.02642s, Loss: 0.71528
Epoch: 79, Time: 0.02563s, Loss: 0.65665
Epoch: 80, Time: 0.02572s, Loss: 0.72648
Epoch: 81, Time: 0.02690s, Loss: 0.64160
Epoch: 82, Time: 0.02741s, Loss: 0.67890
Epoch: 83, Time: 0.03295s, Loss: 0.66671
Epoch: 84, Time: 0.02697s, Loss: 0.68267
Epoch: 85, Time: 0.02802s, Loss: 0.62096
Epoch: 86, Time: 0.02694s, Loss: 0.59566
Epoch: 87, Time: 0.02695s, Loss: 0.61715
Epoch: 88, Time: 0.02584s, Loss: 0.56823
Epoch: 89, Time: 0.02680s, Loss: 0.58922
Epoch: 90, Time: 0.02628s, Loss: 0.62176
Epoch: 91, Time: 0.02630s, Loss: 0.56168
Epoch: 92, Time: 0.02729s, Loss: 0.59730
Epoch: 93, Time: 0.03309s, Loss: 0.54350
Epoch: 94, Time: 0.02711s, Loss: 0.52554
Epoch: 95, Time: 0.03073s, Loss: 0.55863
Epoch: 96, Time: 0.03009s, Loss: 0.54187
Epoch: 97, Time: 0.02847s, Loss: 0.51606
Epoch: 98, Time: 0.02721s, Loss: 0.58703
Epoch: 99, Time: 0.02683s, Loss: 0.45709
Epoch: 100, Time: 0.02546s, Loss: 0.48065
Epoch: 101, Time: 0.02661s, Loss: 0.47521
Epoch: 102, Time: 0.02708s, Loss: 0.49044
Epoch: 103, Time: 0.02877s, Loss: 0.54857
Epoch: 104, Time: 0.02891s, Loss: 0.49147
Epoch: 105, Time: 0.02831s, Loss: 0.51098
Epoch: 106, Time: 0.02855s, Loss: 0.47384
Epoch: 107, Time: 0.02663s, Loss: 0.44903
Epoch: 108, Time: 0.02739s, Loss: 0.48902
Epoch: 109, Time: 0.02786s, Loss: 0.47107
Epoch: 110, Time: 0.02680s, Loss: 0.44998
Epoch: 111, Time: 0.02667s, Loss: 0.45758
Epoch: 112, Time: 0.02677s, Loss: 0.48968
Epoch: 113, Time: 0.03363s, Loss: 0.47052
Epoch: 114, Time: 0.02720s, Loss: 0.42302
Epoch: 115, Time: 0.02691s, Loss: 0.46022
Epoch: 116, Time: 0.02800s, Loss: 0.44152
Epoch: 117, Time: 0.02809s, Loss: 0.41619
Epoch: 118, Time: 0.02747s, Loss: 0.42209
Epoch: 119, Time: 0.02731s, Loss: 0.39555
Epoch: 120, Time: 0.02757s, Loss: 0.41737
Epoch: 121, Time: 0.02572s, Loss: 0.43961
Epoch: 122, Time: 0.02781s, Loss: 0.45638
Epoch: 123, Time: 0.03219s, Loss: 0.40218
Epoch: 124, Time: 0.02912s, Loss: 0.39478
Epoch: 125, Time: 0.02836s, Loss: 0.42770
Epoch: 126, Time: 0.02821s, Loss: 0.44723
Epoch: 127, Time: 0.02668s, Loss: 0.44981
Epoch: 128, Time: 0.02659s, Loss: 0.36467
Epoch: 129, Time: 0.02790s, Loss: 0.41371
Epoch: 130, Time: 0.02687s, Loss: 0.43008
Epoch: 131, Time: 0.02749s, Loss: 0.39013
Epoch: 132, Time: 0.02737s, Loss: 0.38068
Epoch: 133, Time: 0.02744s, Loss: 0.41307
Epoch: 134, Time: 0.02709s, Loss: 0.37499
Epoch: 135, Time: 0.03620s, Loss: 0.38330
Epoch: 136, Time: 0.03489s, Loss: 0.36262
Epoch: 137, Time: 0.03187s, Loss: 0.37654
Epoch: 138, Time: 0.03120s, Loss: 0.39200
Epoch: 139, Time: 0.03104s, Loss: 0.38622
Epoch: 140, Time: 0.03423s, Loss: 0.40245
Epoch: 141, Time: 0.02714s, Loss: 0.42246
Epoch: 142, Time: 0.02613s, Loss: 0.38597
Epoch: 143, Time: 0.02614s, Loss: 0.33846
Epoch: 144, Time: 0.02727s, Loss: 0.35218
Epoch: 145, Time: 0.02886s, Loss: 0.34761
Epoch: 146, Time: 0.02711s, Loss: 0.36396
Epoch: 147, Time: 0.02971s, Loss: 0.36457
Epoch: 148, Time: 0.02699s, Loss: 0.34745
Epoch: 149, Time: 0.02773s, Loss: 0.35060
Epoch: 150, Time: 0.02763s, Loss: 0.33626
Epoch: 151, Time: 0.02665s, Loss: 0.31920
Epoch: 152, Time: 0.02700s, Loss: 0.35494
Epoch: 153, Time: 0.02631s, Loss: 0.32023
Epoch: 154, Time: 0.02521s, Loss: 0.33341
Epoch: 155, Time: 0.02761s, Loss: 0.33163
Epoch: 156, Time: 0.03211s, Loss: 0.37067
Epoch: 157, Time: 0.02632s, Loss: 0.31185
Epoch: 158, Time: 0.02799s, Loss: 0.32024
Epoch: 159, Time: 0.02868s, Loss: 0.33890
Epoch: 160, Time: 0.02777s, Loss: 0.34390
Epoch: 161, Time: 0.02628s, Loss: 0.34751
Epoch: 162, Time: 0.02660s, Loss: 0.34165
Epoch: 163, Time: 0.02635s, Loss: 0.32915
Epoch: 164, Time: 0.02783s, Loss: 0.34125
Epoch: 165, Time: 0.02822s, Loss: 0.35261
Epoch: 166, Time: 0.02855s, Loss: 0.31803
Epoch: 167, Time: 0.02532s, Loss: 0.34157
Epoch: 168, Time: 0.02748s, Loss: 0.36173
Epoch: 169, Time: 0.02843s, Loss: 0.29295
Epoch: 170, Time: 0.02735s, Loss: 0.32935
Epoch: 171, Time: 0.02742s, Loss: 0.32463
Epoch: 172, Time: 0.02704s, Loss: 0.34419
Epoch: 173, Time: 0.02737s, Loss: 0.32393
Epoch: 174, Time: 0.02667s, Loss: 0.32464
Epoch: 175, Time: 0.02750s, Loss: 0.32668
Epoch: 176, Time: 0.02771s, Loss: 0.33835
Epoch: 177, Time: 0.02783s, Loss: 0.32610
Epoch: 178, Time: 0.03027s, Loss: 0.31611
Epoch: 179, Time: 0.02945s, Loss: 0.31614
Epoch: 180, Time: 0.02750s, Loss: 0.33912
Epoch: 181, Time: 0.02655s, Loss: 0.29072
Epoch: 182, Time: 0.02566s, Loss: 0.33455
Epoch: 183, Time: 0.02669s, Loss: 0.29251
Epoch: 184, Time: 0.02900s, Loss: 0.32722
Epoch: 185, Time: 0.02738s, Loss: 0.29612
Epoch: 186, Time: 0.02708s, Loss: 0.30084
Epoch: 187, Time: 0.02681s, Loss: 0.28315
Epoch: 188, Time: 0.02847s, Loss: 0.31396
Epoch: 189, Time: 0.02638s, Loss: 0.31683
Epoch: 190, Time: 0.02819s, Loss: 0.33803
Epoch: 191, Time: 0.02756s, Loss: 0.31791
Epoch: 192, Time: 0.02695s, Loss: 0.35256
Epoch: 193, Time: 0.02624s, Loss: 0.30407
Epoch: 194, Time: 0.02629s, Loss: 0.30797
Epoch: 195, Time: 0.02591s, Loss: 0.29365
Epoch: 196, Time: 0.02655s, Loss: 0.28897
Epoch: 197, Time: 0.02585s, Loss: 0.31783
Epoch: 198, Time: 0.02900s, Loss: 0.28889
Epoch: 199, Time: 0.02735s, Loss: 0.31066
Epoch: 200, Time: 0.02652s, Loss: 0.31168
Epoch: 201, Time: 0.02635s, Loss: 0.26849
Epoch: 202, Time: 0.02685s, Loss: 0.29419
Epoch: 203, Time: 0.02794s, Loss: 0.31236
update best: 0.79800
Epoch: 204, Time: 0.02748s, Loss: 0.29655
Epoch: 205, Time: 0.02772s, Loss: 0.32185
update best: 0.80000
Epoch: 206, Time: 0.03271s, Loss: 0.28461
Epoch: 207, Time: 0.02841s, Loss: 0.28718
Epoch: 208, Time: 0.02810s, Loss: 0.28859
Epoch: 209, Time: 0.02825s, Loss: 0.33484
Epoch: 210, Time: 0.02748s, Loss: 0.25476
Epoch: 211, Time: 0.02689s, Loss: 0.31217
Epoch: 212, Time: 0.02616s, Loss: 0.30048
Epoch: 213, Time: 0.02599s, Loss: 0.25396
Epoch: 214, Time: 0.02509s, Loss: 0.25659
Epoch: 215, Time: 0.02558s, Loss: 0.27736
Epoch: 216, Time: 0.02744s, Loss: 0.29813
Epoch: 217, Time: 0.02797s, Loss: 0.26633
Epoch: 218, Time: 0.02972s, Loss: 0.26556
Epoch: 219, Time: 0.02468s, Loss: 0.26812
Epoch: 220, Time: 0.02691s, Loss: 0.27502
Epoch: 221, Time: 0.02941s, Loss: 0.27201
Epoch: 222, Time: 0.03062s, Loss: 0.24750
Epoch: 223, Time: 0.02580s, Loss: 0.25536
Epoch: 224, Time: 0.02601s, Loss: 0.24400
Epoch: 225, Time: 0.02609s, Loss: 0.26673
Epoch: 226, Time: 0.02816s, Loss: 0.28496
Epoch: 227, Time: 0.02798s, Loss: 0.27348
Epoch: 228, Time: 0.02800s, Loss: 0.30068
Epoch: 229, Time: 0.02711s, Loss: 0.25621
Epoch: 230, Time: 0.02845s, Loss: 0.28133
Epoch: 231, Time: 0.02709s, Loss: 0.26263
Epoch: 232, Time: 0.02776s, Loss: 0.28019
Epoch: 233, Time: 0.02760s, Loss: 0.24621
Epoch: 234, Time: 0.02652s, Loss: 0.25726
Epoch: 235, Time: 0.02607s, Loss: 0.27996
Epoch: 236, Time: 0.02545s, Loss: 0.26172
Epoch: 237, Time: 0.02611s, Loss: 0.28643
update best: 0.80200
Epoch: 238, Time: 0.02843s, Loss: 0.27893
Epoch: 239, Time: 0.02436s, Loss: 0.23068
Epoch: 240, Time: 0.02698s, Loss: 0.26539
Epoch: 241, Time: 0.02526s, Loss: 0.26346
Epoch: 242, Time: 0.02636s, Loss: 0.25852
Epoch: 243, Time: 0.02681s, Loss: 0.24250
Epoch: 244, Time: 0.02879s, Loss: 0.26560
Epoch: 245, Time: 0.02841s, Loss: 0.24397
Epoch: 246, Time: 0.02649s, Loss: 0.22487
Epoch: 247, Time: 0.02529s, Loss: 0.28920
Epoch: 248, Time: 0.02598s, Loss: 0.25361
Epoch: 249, Time: 0.02651s, Loss: 0.23220
Epoch: 250, Time: 0.02981s, Loss: 0.24851
Epoch: 251, Time: 0.02647s, Loss: 0.26154
Epoch: 252, Time: 0.02915s, Loss: 0.28003
Epoch: 253, Time: 0.02627s, Loss: 0.27142
Epoch: 254, Time: 0.02771s, Loss: 0.24000
Epoch: 255, Time: 0.02807s, Loss: 0.22970
Epoch: 256, Time: 0.02778s, Loss: 0.25055
Epoch: 257, Time: 0.02756s, Loss: 0.25298
Epoch: 258, Time: 0.02604s, Loss: 0.25399
Epoch: 259, Time: 0.02515s, Loss: 0.23506
Epoch: 260, Time: 0.02584s, Loss: 0.27011
Epoch: 261, Time: 0.02733s, Loss: 0.27896
Epoch: 262, Time: 0.03368s, Loss: 0.27697
Epoch: 263, Time: 0.02622s, Loss: 0.25122
Epoch: 264, Time: 0.02557s, Loss: 0.22288
Epoch: 265, Time: 0.02677s, Loss: 0.24788
Epoch: 266, Time: 0.02789s, Loss: 0.25024
Epoch: 267, Time: 0.02766s, Loss: 0.24291
Epoch: 268, Time: 0.02734s, Loss: 0.23501
Epoch: 269, Time: 0.02628s, Loss: 0.22473
update best: 0.80400
Epoch: 270, Time: 0.02710s, Loss: 0.23869
Epoch: 271, Time: 0.02704s, Loss: 0.23497
Epoch: 272, Time: 0.02797s, Loss: 0.27661
Epoch: 273, Time: 0.02528s, Loss: 0.22743
Epoch: 274, Time: 0.02586s, Loss: 0.27344
Epoch: 275, Time: 0.02527s, Loss: 0.24526
Epoch: 276, Time: 0.02694s, Loss: 0.23004
Epoch: 277, Time: 0.02799s, Loss: 0.26727
Epoch: 278, Time: 0.02743s, Loss: 0.24816
Epoch: 279, Time: 0.02808s, Loss: 0.24808
Epoch: 280, Time: 0.02596s, Loss: 0.21776
Epoch: 281, Time: 0.02563s, Loss: 0.21926
Epoch: 282, Time: 0.02653s, Loss: 0.22270
Epoch: 283, Time: 0.02805s, Loss: 0.24317
Epoch: 284, Time: 0.02826s, Loss: 0.26508
Epoch: 285, Time: 0.02821s, Loss: 0.27642
Epoch: 286, Time: 0.02656s, Loss: 0.28210
Epoch: 287, Time: 0.02595s, Loss: 0.21376
Epoch: 288, Time: 0.02581s, Loss: 0.22294
Epoch: 289, Time: 0.02792s, Loss: 0.22761
Epoch: 290, Time: 0.02788s, Loss: 0.21223
Epoch: 291, Time: 0.02840s, Loss: 0.25497
Epoch: 292, Time: 0.02945s, Loss: 0.25667
Epoch: 293, Time: 0.02686s, Loss: 0.28930
Epoch: 294, Time: 0.02824s, Loss: 0.27815
Epoch: 295, Time: 0.02799s, Loss: 0.29124
Epoch: 296, Time: 0.02615s, Loss: 0.23398
Epoch: 297, Time: 0.02607s, Loss: 0.21476
Epoch: 298, Time: 0.02598s, Loss: 0.22739
Epoch: 299, Time: 0.02830s, Loss: 0.26215

train finished!
best val: 0.80400
test...
final result: epoch: 269
{'accuracy': 0.8240000009536743, 'f1_score': 0.8174891298012773, 'f1_score -> average@micro': 0.824}

Cora上使用HGNN

导入依赖包

import time
from copy import deepcopy

import torch
import torch.optim as optim
import torch.nn.functional as F

from dhg import Graph, Hypergraph
from dhg.data import Cora
from dhg.models import HGNN
from dhg.random import set_seed
from dhg.metrics import HypergraphVertexClassificationEvaluator as Evaluator

定义函数

def train(net, X, G, lbls, train_idx, optimizer, epoch):
    net.train()

    st = time.time()
    optimizer.zero_grad()
    outs = net(X, G)
    outs, lbls = outs[train_idx], lbls[train_idx]
    loss = F.cross_entropy(outs, lbls)
    loss.backward()
    optimizer.step()
    print(f"Epoch: {epoch}, Time: {time.time()-st:.5f}s, Loss: {loss.item():.5f}")
    return loss.item()


@torch.no_grad()
def infer(net, X, G, lbls, idx, test=False):
    net.eval()
    outs = net(X, G)
    outs, lbls = outs[idx], lbls[idx]
    if not test:
        res = evaluator.validate(lbls, outs)
    else:
        res = evaluator.test(lbls, outs)
    return res

主函数

Note

更多关于评测器 Evaluator 的细节可以参照 构建指标评测器 章节。

if __name__ == "__main__":
    set_seed(2022)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cora()
    X, lbl = data["features"], data["labels"]
    G = Graph(data["num_vertices"], data["edge_list"])
    HG = Hypergraph.from_graph_kHop(G, k=1)
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = HGNN(data["dim_features"], 16, data["num_classes"])
    optimizer = optim.Adam(net.parameters(), lr=0.01, weight_decay=5e-4)

    X, lbl = X.to(device), lbl.to(device)
    HG = HG.to(device)
    net = net.to(device)

    best_state = None
    best_epoch, best_val = 0, 0
    for epoch in range(200):
        # train
        train(net, X, HG, lbl, train_mask, optimizer, epoch)
        # validation
        if epoch % 1 == 0:
            with torch.no_grad():
                val_res = infer(net, X, HG, lbl, val_mask)
            if val_res > best_val:
                print(f"update best: {val_res:.5f}")
                best_epoch = epoch
                best_val = val_res
                best_state = deepcopy(net.state_dict())
    print("\ntrain finished!")
    print(f"best val: {best_val:.5f}")
    # test
    print("test...")
    net.load_state_dict(best_state)
    res = infer(net, X, HG, lbl, test_mask, test=True)
    print(f"final result: epoch: {best_epoch}")
    print(res)

输出

Epoch: 0, Time: 0.50315s, Loss: 1.94993
update best: 0.31600
Epoch: 1, Time: 0.00196s, Loss: 1.94627
Epoch: 2, Time: 0.00175s, Loss: 1.94413
Epoch: 3, Time: 0.00200s, Loss: 1.93941
Epoch: 4, Time: 0.00180s, Loss: 1.93488
Epoch: 5, Time: 0.00174s, Loss: 1.92980
update best: 0.32000
Epoch: 6, Time: 0.00184s, Loss: 1.92559
update best: 0.36400
Epoch: 7, Time: 0.00256s, Loss: 1.91934
update best: 0.46400
Epoch: 8, Time: 0.00198s, Loss: 1.91385
Epoch: 9, Time: 0.00177s, Loss: 1.90319
Epoch: 10, Time: 0.00248s, Loss: 1.89834
Epoch: 11, Time: 0.00248s, Loss: 1.89127
Epoch: 12, Time: 0.00173s, Loss: 1.87880
Epoch: 13, Time: 0.00247s, Loss: 1.87467
Epoch: 14, Time: 0.00194s, Loss: 1.86688
Epoch: 15, Time: 0.00181s, Loss: 1.85891
Epoch: 16, Time: 0.00266s, Loss: 1.85094
Epoch: 17, Time: 0.00289s, Loss: 1.84161
Epoch: 18, Time: 0.00179s, Loss: 1.82744
Epoch: 19, Time: 0.00239s, Loss: 1.81666
Epoch: 20, Time: 0.00198s, Loss: 1.80902
Epoch: 21, Time: 0.00177s, Loss: 1.78956
Epoch: 22, Time: 0.00252s, Loss: 1.78221
update best: 0.49000
Epoch: 23, Time: 0.00191s, Loss: 1.76655
update best: 0.50200
Epoch: 24, Time: 0.00174s, Loss: 1.76185
update best: 0.51600
Epoch: 25, Time: 0.00253s, Loss: 1.74321
update best: 0.51800
Epoch: 26, Time: 0.00187s, Loss: 1.72027
update best: 0.52200
Epoch: 27, Time: 0.00369s, Loss: 1.70986
update best: 0.52600
Epoch: 28, Time: 0.00241s, Loss: 1.69354
update best: 0.53000
Epoch: 29, Time: 0.00309s, Loss: 1.69100
update best: 0.53800
Epoch: 30, Time: 0.00232s, Loss: 1.66968
update best: 0.54400
Epoch: 31, Time: 0.00313s, Loss: 1.65087
update best: 0.54600
Epoch: 32, Time: 0.00224s, Loss: 1.64182
update best: 0.56000
Epoch: 33, Time: 0.00277s, Loss: 1.60257
update best: 0.57800
Epoch: 34, Time: 0.00208s, Loss: 1.58798
update best: 0.59200
Epoch: 35, Time: 0.00176s, Loss: 1.58344
update best: 0.60000
Epoch: 36, Time: 0.00200s, Loss: 1.56942
update best: 0.63200
Epoch: 37, Time: 0.00206s, Loss: 1.53224
update best: 0.64800
Epoch: 38, Time: 0.00215s, Loss: 1.53036
update best: 0.67000
Epoch: 39, Time: 0.00200s, Loss: 1.50875
update best: 0.68000
Epoch: 40, Time: 0.00209s, Loss: 1.46828
update best: 0.69200
Epoch: 41, Time: 0.00243s, Loss: 1.45782
update best: 0.69400
Epoch: 42, Time: 0.00208s, Loss: 1.42179
Epoch: 43, Time: 0.00267s, Loss: 1.40893
Epoch: 44, Time: 0.00176s, Loss: 1.40358
update best: 0.69800
Epoch: 45, Time: 0.00175s, Loss: 1.37788
Epoch: 46, Time: 0.00274s, Loss: 1.34310
Epoch: 47, Time: 0.00173s, Loss: 1.32779
update best: 0.70200
Epoch: 48, Time: 0.00175s, Loss: 1.30572
update best: 0.71200
Epoch: 49, Time: 0.00221s, Loss: 1.28909
update best: 0.71800
Epoch: 50, Time: 0.00184s, Loss: 1.28903
update best: 0.72400
Epoch: 51, Time: 0.00345s, Loss: 1.25486
update best: 0.73200
Epoch: 52, Time: 0.00176s, Loss: 1.22994
update best: 0.74200
Epoch: 53, Time: 0.00173s, Loss: 1.20690
update best: 0.75000
Epoch: 54, Time: 0.00241s, Loss: 1.17115
Epoch: 55, Time: 0.00198s, Loss: 1.18836
update best: 0.75600
Epoch: 56, Time: 0.00279s, Loss: 1.17722
update best: 0.75800
Epoch: 57, Time: 0.00204s, Loss: 1.13414
Epoch: 58, Time: 0.00173s, Loss: 1.12058
update best: 0.76200
Epoch: 59, Time: 0.00228s, Loss: 1.09260
update best: 0.77400
Epoch: 60, Time: 0.00188s, Loss: 1.07260
Epoch: 61, Time: 0.00256s, Loss: 1.09610
Epoch: 62, Time: 0.00280s, Loss: 1.02422
Epoch: 63, Time: 0.00221s, Loss: 1.03871
update best: 0.77800
Epoch: 64, Time: 0.00311s, Loss: 1.00255
Epoch: 65, Time: 0.00226s, Loss: 0.99640
update best: 0.78000
Epoch: 66, Time: 0.00296s, Loss: 0.99191
update best: 0.78200
Epoch: 67, Time: 0.00235s, Loss: 0.95631
update best: 0.78600
Epoch: 68, Time: 0.00255s, Loss: 0.94336
Epoch: 69, Time: 0.00183s, Loss: 0.92673
update best: 0.79000
Epoch: 70, Time: 0.00165s, Loss: 0.92654
update best: 0.79600
Epoch: 71, Time: 0.00188s, Loss: 0.86986
update best: 0.80000
Epoch: 72, Time: 0.00170s, Loss: 0.90749
Epoch: 73, Time: 0.00164s, Loss: 0.86787
Epoch: 74, Time: 0.00218s, Loss: 0.86549
Epoch: 75, Time: 0.00182s, Loss: 0.86944
Epoch: 76, Time: 0.00189s, Loss: 0.83897
Epoch: 77, Time: 0.00167s, Loss: 0.82139
Epoch: 78, Time: 0.00168s, Loss: 0.81658
Epoch: 79, Time: 0.00198s, Loss: 0.78883
Epoch: 80, Time: 0.00207s, Loss: 0.78880
Epoch: 81, Time: 0.00209s, Loss: 0.77039
Epoch: 82, Time: 0.00170s, Loss: 0.74785
Epoch: 83, Time: 0.00185s, Loss: 0.74238
Epoch: 84, Time: 0.00293s, Loss: 0.73360
Epoch: 85, Time: 0.00164s, Loss: 0.76029
Epoch: 86, Time: 0.00163s, Loss: 0.71382
Epoch: 87, Time: 0.00162s, Loss: 0.72503
Epoch: 88, Time: 0.00202s, Loss: 0.70878
Epoch: 89, Time: 0.00172s, Loss: 0.71945
Epoch: 90, Time: 0.00180s, Loss: 0.65032
Epoch: 91, Time: 0.00302s, Loss: 0.71030
Epoch: 92, Time: 0.00157s, Loss: 0.67237
Epoch: 93, Time: 0.00161s, Loss: 0.68624
Epoch: 94, Time: 0.00161s, Loss: 0.65738
Epoch: 95, Time: 0.00203s, Loss: 0.65683
Epoch: 96, Time: 0.00171s, Loss: 0.63819
Epoch: 97, Time: 0.00177s, Loss: 0.66612
Epoch: 98, Time: 0.00231s, Loss: 0.64060
Epoch: 99, Time: 0.00161s, Loss: 0.63596
Epoch: 100, Time: 0.00161s, Loss: 0.62215
Epoch: 101, Time: 0.00195s, Loss: 0.59992
Epoch: 102, Time: 0.00184s, Loss: 0.63610
Epoch: 103, Time: 0.00168s, Loss: 0.60803
Epoch: 104, Time: 0.00174s, Loss: 0.60519
Epoch: 105, Time: 0.00203s, Loss: 0.61317
update best: 0.80200
Epoch: 106, Time: 0.00163s, Loss: 0.56701
Epoch: 107, Time: 0.00160s, Loss: 0.58649
Epoch: 108, Time: 0.00202s, Loss: 0.60864
Epoch: 109, Time: 0.00171s, Loss: 0.59734
Epoch: 110, Time: 0.00174s, Loss: 0.58395
Epoch: 111, Time: 0.00262s, Loss: 0.59959
Epoch: 112, Time: 0.00166s, Loss: 0.57178
Epoch: 113, Time: 0.00162s, Loss: 0.57493
Epoch: 114, Time: 0.00166s, Loss: 0.56720
Epoch: 115, Time: 0.00207s, Loss: 0.57864
Epoch: 116, Time: 0.00174s, Loss: 0.55171
Epoch: 117, Time: 0.00201s, Loss: 0.56022
Epoch: 118, Time: 0.00295s, Loss: 0.54393
Epoch: 119, Time: 0.00162s, Loss: 0.54266
Epoch: 120, Time: 0.00162s, Loss: 0.54640
Epoch: 121, Time: 0.00165s, Loss: 0.51695
Epoch: 122, Time: 0.00193s, Loss: 0.53059
Epoch: 123, Time: 0.00175s, Loss: 0.49817
Epoch: 124, Time: 0.00168s, Loss: 0.49963
Epoch: 125, Time: 0.00280s, Loss: 0.50499
Epoch: 126, Time: 0.00165s, Loss: 0.51792
Epoch: 127, Time: 0.00162s, Loss: 0.48759
Epoch: 128, Time: 0.00188s, Loss: 0.52524
Epoch: 129, Time: 0.00192s, Loss: 0.49752
Epoch: 130, Time: 0.00182s, Loss: 0.48539
Epoch: 131, Time: 0.00178s, Loss: 0.51904
Epoch: 132, Time: 0.00210s, Loss: 0.51619
Epoch: 133, Time: 0.00164s, Loss: 0.46799
Epoch: 134, Time: 0.00168s, Loss: 0.47253
Epoch: 135, Time: 0.00220s, Loss: 0.50235
Epoch: 136, Time: 0.00179s, Loss: 0.48068
Epoch: 137, Time: 0.00181s, Loss: 0.48230
Epoch: 138, Time: 0.00311s, Loss: 0.47752
Epoch: 139, Time: 0.00165s, Loss: 0.46344
Epoch: 140, Time: 0.00168s, Loss: 0.50513
Epoch: 141, Time: 0.00175s, Loss: 0.45315
Epoch: 142, Time: 0.00234s, Loss: 0.45984
Epoch: 143, Time: 0.00184s, Loss: 0.45598
Epoch: 144, Time: 0.00181s, Loss: 0.48745
Epoch: 145, Time: 0.00208s, Loss: 0.47391
Epoch: 146, Time: 0.00167s, Loss: 0.42658
Epoch: 147, Time: 0.00164s, Loss: 0.44139
Epoch: 148, Time: 0.00211s, Loss: 0.44337
Epoch: 149, Time: 0.00174s, Loss: 0.43854
Epoch: 150, Time: 0.00194s, Loss: 0.45141
Epoch: 151, Time: 0.00337s, Loss: 0.43659
Epoch: 152, Time: 0.00223s, Loss: 0.45104
Epoch: 153, Time: 0.00217s, Loss: 0.45788
Epoch: 154, Time: 0.00256s, Loss: 0.44208
Epoch: 155, Time: 0.00216s, Loss: 0.47642
Epoch: 156, Time: 0.00289s, Loss: 0.41826
Epoch: 157, Time: 0.00219s, Loss: 0.44075
Epoch: 158, Time: 0.00212s, Loss: 0.39873
Epoch: 159, Time: 0.00235s, Loss: 0.43970
Epoch: 160, Time: 0.00170s, Loss: 0.41875
Epoch: 161, Time: 0.00185s, Loss: 0.42697
Epoch: 162, Time: 0.00185s, Loss: 0.44240
Epoch: 163, Time: 0.00165s, Loss: 0.45397
Epoch: 164, Time: 0.00217s, Loss: 0.38061
Epoch: 165, Time: 0.00187s, Loss: 0.40102
Epoch: 166, Time: 0.00194s, Loss: 0.39496
Epoch: 167, Time: 0.00208s, Loss: 0.41661
Epoch: 168, Time: 0.00187s, Loss: 0.41864
Epoch: 169, Time: 0.00262s, Loss: 0.41757
Epoch: 170, Time: 0.00188s, Loss: 0.41356
Epoch: 171, Time: 0.00180s, Loss: 0.38835
Epoch: 172, Time: 0.00213s, Loss: 0.42775
Epoch: 173, Time: 0.00187s, Loss: 0.39169
Epoch: 174, Time: 0.00164s, Loss: 0.41415
Epoch: 175, Time: 0.00290s, Loss: 0.39668
update best: 0.80400
Epoch: 176, Time: 0.00161s, Loss: 0.42034
Epoch: 177, Time: 0.00164s, Loss: 0.40507
Epoch: 178, Time: 0.00206s, Loss: 0.39741
Epoch: 179, Time: 0.00181s, Loss: 0.40042
Epoch: 180, Time: 0.00163s, Loss: 0.37404
Epoch: 181, Time: 0.00167s, Loss: 0.40175
Epoch: 182, Time: 0.00217s, Loss: 0.35673
Epoch: 183, Time: 0.00162s, Loss: 0.39076
Epoch: 184, Time: 0.00157s, Loss: 0.39327
Epoch: 185, Time: 0.00208s, Loss: 0.38354
Epoch: 186, Time: 0.00172s, Loss: 0.36611
Epoch: 187, Time: 0.00174s, Loss: 0.38952
Epoch: 188, Time: 0.00276s, Loss: 0.39074
Epoch: 189, Time: 0.00160s, Loss: 0.36561
Epoch: 190, Time: 0.00164s, Loss: 0.37361
Epoch: 191, Time: 0.00162s, Loss: 0.37590
Epoch: 192, Time: 0.00188s, Loss: 0.36160
Epoch: 193, Time: 0.00173s, Loss: 0.37451
Epoch: 194, Time: 0.00170s, Loss: 0.36310
Epoch: 195, Time: 0.00285s, Loss: 0.39782
Epoch: 196, Time: 0.00160s, Loss: 0.36185
Epoch: 197, Time: 0.00161s, Loss: 0.35991
Epoch: 198, Time: 0.00191s, Loss: 0.37487
Epoch: 199, Time: 0.00219s, Loss: 0.36310

train finished!
best val: 0.80400
test...
final result: epoch: 175
{'accuracy': 0.8209999799728394, 'f1_score': 0.8113491851888245, 'f1_score -> average@micro': 0.821}

Cora上使用HGNN+

导入依赖包

import time
from copy import deepcopy

import torch
import torch.optim as optim
import torch.nn.functional as F

from dhg import Graph, Hypergraph
from dhg.data import Cora
from dhg.models import HGNNP
from dhg.random import set_seed
from dhg.metrics import HypergraphVertexClassificationEvaluator as Evaluator

定义函数

def train(net, X, G, lbls, train_idx, optimizer, epoch):
    net.train()

    st = time.time()
    optimizer.zero_grad()
    outs = net(X, G)
    outs, lbls = outs[train_idx], lbls[train_idx]
    loss = F.cross_entropy(outs, lbls)
    loss.backward()
    optimizer.step()
    print(f"Epoch: {epoch}, Time: {time.time()-st:.5f}s, Loss: {loss.item():.5f}")
    return loss.item()


@torch.no_grad()
def infer(net, X, G, lbls, idx, test=False):
    net.eval()
    outs = net(X, G)
    outs, lbls = outs[idx], lbls[idx]
    if not test:
        res = evaluator.validate(lbls, outs)
    else:
        res = evaluator.test(lbls, outs)
    return res

主函数

Note

更多关于评测器 Evaluator 的细节可以参照 构建指标评测器 章节。

if __name__ == "__main__":
    set_seed(2022)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cora()
    X, lbl = data["features"], data["labels"]
    G = Graph(data["num_vertices"], data["edge_list"])
    HG = Hypergraph.from_graph(G)
    HG.add_hyperedges_from_graph_kHop(G, k=1)
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = HGNNP(data["dim_features"], 16, data["num_classes"])
    optimizer = optim.Adam(net.parameters(), lr=0.01, weight_decay=5e-4)

    X, lbl = X.to(device), lbl.to(device)
    HG = HG.to(device)
    net = net.to(device)

    best_state = None
    best_epoch, best_val = 0, 0
    for epoch in range(200):
        # train
        train(net, X, HG, lbl, train_mask, optimizer, epoch)
        # validation
        if epoch % 1 == 0:
            with torch.no_grad():
                val_res = infer(net, X, HG, lbl, val_mask)
            if val_res > best_val:
                print(f"update best: {val_res:.5f}")
                best_epoch = epoch
                best_val = val_res
                best_state = deepcopy(net.state_dict())
    print("\ntrain finished!")
    print(f"best val: {best_val:.5f}")
    # test
    print("test...")
    net.load_state_dict(best_state)
    res = infer(net, X, HG, lbl, test_mask, test=True)
    print(f"final result: epoch: {best_epoch}")
    print(res)

输出

Epoch: 0, Time: 0.50397s, Loss: 1.95489
update best: 0.31600
Epoch: 1, Time: 0.00688s, Loss: 1.95044
Epoch: 2, Time: 0.00594s, Loss: 1.94790
Epoch: 3, Time: 0.00777s, Loss: 1.94277
Epoch: 4, Time: 0.00543s, Loss: 1.93662
Epoch: 5, Time: 0.00805s, Loss: 1.93121
Epoch: 6, Time: 0.00549s, Loss: 1.92640
update best: 0.31800
Epoch: 7, Time: 0.00687s, Loss: 1.91871
update best: 0.37600
Epoch: 8, Time: 0.00722s, Loss: 1.91161
update best: 0.41000
Epoch: 9, Time: 0.00553s, Loss: 1.90009
update best: 0.50400
Epoch: 10, Time: 0.00770s, Loss: 1.89464
update best: 0.57000
Epoch: 11, Time: 0.00566s, Loss: 1.88557
Epoch: 12, Time: 0.00769s, Loss: 1.87337
Epoch: 13, Time: 0.00549s, Loss: 1.86598
Epoch: 14, Time: 0.00767s, Loss: 1.85734
Epoch: 15, Time: 0.00546s, Loss: 1.84511
Epoch: 16, Time: 0.00752s, Loss: 1.83575
Epoch: 17, Time: 0.00545s, Loss: 1.82488
Epoch: 18, Time: 0.00840s, Loss: 1.80935
Epoch: 19, Time: 0.00536s, Loss: 1.79647
Epoch: 20, Time: 0.00756s, Loss: 1.78831
Epoch: 21, Time: 0.00538s, Loss: 1.76364
Epoch: 22, Time: 0.00797s, Loss: 1.75609
Epoch: 23, Time: 0.00601s, Loss: 1.74039
Epoch: 24, Time: 0.00737s, Loss: 1.73402
update best: 0.57200
Epoch: 25, Time: 0.00510s, Loss: 1.70649
Epoch: 26, Time: 0.00626s, Loss: 1.68333
update best: 0.57600
Epoch: 27, Time: 0.00489s, Loss: 1.67384
Epoch: 28, Time: 0.00637s, Loss: 1.64703
Epoch: 29, Time: 0.00569s, Loss: 1.65015
Epoch: 30, Time: 0.00616s, Loss: 1.61904
Epoch: 31, Time: 0.00482s, Loss: 1.60483
Epoch: 32, Time: 0.00657s, Loss: 1.58717
update best: 0.57800
Epoch: 33, Time: 0.00671s, Loss: 1.54870
update best: 0.58400
Epoch: 34, Time: 0.00547s, Loss: 1.53594
update best: 0.59800
Epoch: 35, Time: 0.00591s, Loss: 1.52464
update best: 0.61000
Epoch: 36, Time: 0.00569s, Loss: 1.50577
update best: 0.62800
Epoch: 37, Time: 0.00447s, Loss: 1.47224
update best: 0.64400
Epoch: 38, Time: 0.00566s, Loss: 1.46083
update best: 0.65800
Epoch: 39, Time: 0.00448s, Loss: 1.44008
update best: 0.67400
Epoch: 40, Time: 0.00560s, Loss: 1.39763
update best: 0.68800
Epoch: 41, Time: 0.00452s, Loss: 1.38902
update best: 0.69600
Epoch: 42, Time: 0.00592s, Loss: 1.34805
update best: 0.70600
Epoch: 43, Time: 0.00460s, Loss: 1.32505
update best: 0.71200
Epoch: 44, Time: 0.00575s, Loss: 1.32579
update best: 0.71600
Epoch: 45, Time: 0.00456s, Loss: 1.29263
update best: 0.72200
Epoch: 46, Time: 0.00590s, Loss: 1.25758
update best: 0.72800
Epoch: 47, Time: 0.00457s, Loss: 1.25460
update best: 0.73000
Epoch: 48, Time: 0.00577s, Loss: 1.21283
update best: 0.73200
Epoch: 49, Time: 0.00555s, Loss: 1.22506
update best: 0.73800
Epoch: 50, Time: 0.00590s, Loss: 1.20866
update best: 0.74200
Epoch: 51, Time: 0.00607s, Loss: 1.17283
update best: 0.75800
Epoch: 52, Time: 0.00558s, Loss: 1.14841
update best: 0.78000
Epoch: 53, Time: 0.00534s, Loss: 1.12203
update best: 0.78800
Epoch: 54, Time: 0.00525s, Loss: 1.07957
update best: 0.79000
Epoch: 55, Time: 0.00598s, Loss: 1.09576
update best: 0.79200
Epoch: 56, Time: 0.00518s, Loss: 1.08737
update best: 0.79400
Epoch: 57, Time: 0.00666s, Loss: 1.03506
Epoch: 58, Time: 0.00471s, Loss: 1.02326
Epoch: 59, Time: 0.00623s, Loss: 1.01210
Epoch: 60, Time: 0.00557s, Loss: 0.99087
Epoch: 61, Time: 0.00454s, Loss: 0.99048
Epoch: 62, Time: 0.00614s, Loss: 0.92911
Epoch: 63, Time: 0.00461s, Loss: 0.96758
Epoch: 64, Time: 0.00739s, Loss: 0.90397
Epoch: 65, Time: 0.00469s, Loss: 0.89135
Epoch: 66, Time: 0.00745s, Loss: 0.90936
Epoch: 67, Time: 0.00459s, Loss: 0.85870
Epoch: 68, Time: 0.00657s, Loss: 0.86560
Epoch: 69, Time: 0.00534s, Loss: 0.84675
Epoch: 70, Time: 0.00564s, Loss: 0.85727
Epoch: 71, Time: 0.00590s, Loss: 0.79680
Epoch: 72, Time: 0.00453s, Loss: 0.82477
Epoch: 73, Time: 0.00614s, Loss: 0.79762
Epoch: 74, Time: 0.00452s, Loss: 0.78480
Epoch: 75, Time: 0.00735s, Loss: 0.81077
Epoch: 76, Time: 0.00463s, Loss: 0.77174
Epoch: 77, Time: 0.00706s, Loss: 0.74386
Epoch: 78, Time: 0.00569s, Loss: 0.73486
Epoch: 79, Time: 0.00738s, Loss: 0.70369
update best: 0.79600
Epoch: 80, Time: 0.00563s, Loss: 0.70949
Epoch: 81, Time: 0.00649s, Loss: 0.68134
Epoch: 82, Time: 0.00542s, Loss: 0.65184
update best: 0.79800
Epoch: 83, Time: 0.00635s, Loss: 0.66273
Epoch: 84, Time: 0.00545s, Loss: 0.65232
Epoch: 85, Time: 0.00696s, Loss: 0.69817
Epoch: 86, Time: 0.00574s, Loss: 0.64078
Epoch: 87, Time: 0.00686s, Loss: 0.65521
Epoch: 88, Time: 0.00470s, Loss: 0.63180
Epoch: 89, Time: 0.00449s, Loss: 0.65444
Epoch: 90, Time: 0.00605s, Loss: 0.56861
Epoch: 91, Time: 0.00456s, Loss: 0.64074
Epoch: 92, Time: 0.00659s, Loss: 0.59132
update best: 0.80200
Epoch: 93, Time: 0.00465s, Loss: 0.62925
Epoch: 94, Time: 0.00662s, Loss: 0.60163
Epoch: 95, Time: 0.00453s, Loss: 0.58727
Epoch: 96, Time: 0.00693s, Loss: 0.57620
Epoch: 97, Time: 0.00481s, Loss: 0.60987
Epoch: 98, Time: 0.00702s, Loss: 0.57996
Epoch: 99, Time: 0.00462s, Loss: 0.56781
Epoch: 100, Time: 0.00570s, Loss: 0.54706
Epoch: 101, Time: 0.00507s, Loss: 0.54080
Epoch: 102, Time: 0.00444s, Loss: 0.57735
Epoch: 103, Time: 0.00613s, Loss: 0.52275
Epoch: 104, Time: 0.00452s, Loss: 0.53871
Epoch: 105, Time: 0.00667s, Loss: 0.54541
Epoch: 106, Time: 0.00565s, Loss: 0.51127
Epoch: 107, Time: 0.00738s, Loss: 0.52514
Epoch: 108, Time: 0.00540s, Loss: 0.54392
Epoch: 109, Time: 0.00604s, Loss: 0.54753
Epoch: 110, Time: 0.00465s, Loss: 0.53154
Epoch: 111, Time: 0.00629s, Loss: 0.53460
Epoch: 112, Time: 0.00568s, Loss: 0.52337
Epoch: 113, Time: 0.00587s, Loss: 0.52842
Epoch: 114, Time: 0.00562s, Loss: 0.50907
Epoch: 115, Time: 0.00454s, Loss: 0.51616
Epoch: 116, Time: 0.00561s, Loss: 0.50364
Epoch: 117, Time: 0.00459s, Loss: 0.49458
Epoch: 118, Time: 0.00545s, Loss: 0.49913
Epoch: 119, Time: 0.00529s, Loss: 0.48824
Epoch: 120, Time: 0.00519s, Loss: 0.52106
Epoch: 121, Time: 0.00555s, Loss: 0.46541
Epoch: 122, Time: 0.00459s, Loss: 0.47356
Epoch: 123, Time: 0.00539s, Loss: 0.44043
Epoch: 124, Time: 0.00468s, Loss: 0.44389
Epoch: 125, Time: 0.00569s, Loss: 0.45298
Epoch: 126, Time: 0.00500s, Loss: 0.46986
Epoch: 127, Time: 0.00551s, Loss: 0.45141
Epoch: 128, Time: 0.00533s, Loss: 0.48571
Epoch: 129, Time: 0.00460s, Loss: 0.43895
Epoch: 130, Time: 0.00600s, Loss: 0.44426
Epoch: 131, Time: 0.00457s, Loss: 0.47401
Epoch: 132, Time: 0.00579s, Loss: 0.46865
Epoch: 133, Time: 0.00464s, Loss: 0.41215
Epoch: 134, Time: 0.00528s, Loss: 0.42941
Epoch: 135, Time: 0.00642s, Loss: 0.46532
Epoch: 136, Time: 0.00538s, Loss: 0.42108
Epoch: 137, Time: 0.00690s, Loss: 0.41919
Epoch: 138, Time: 0.00617s, Loss: 0.44285
Epoch: 139, Time: 0.00577s, Loss: 0.42653
Epoch: 140, Time: 0.00548s, Loss: 0.45898
Epoch: 141, Time: 0.00539s, Loss: 0.41800
Epoch: 142, Time: 0.00467s, Loss: 0.40399
Epoch: 143, Time: 0.00487s, Loss: 0.38347
Epoch: 144, Time: 0.00509s, Loss: 0.42234
Epoch: 145, Time: 0.00721s, Loss: 0.42908
Epoch: 146, Time: 0.00489s, Loss: 0.37335
Epoch: 147, Time: 0.00664s, Loss: 0.40119
Epoch: 148, Time: 0.00465s, Loss: 0.38477
Epoch: 149, Time: 0.00451s, Loss: 0.40037
Epoch: 150, Time: 0.00553s, Loss: 0.40168
Epoch: 151, Time: 0.00454s, Loss: 0.38555
Epoch: 152, Time: 0.00729s, Loss: 0.40183
Epoch: 153, Time: 0.00465s, Loss: 0.40610
Epoch: 154, Time: 0.00669s, Loss: 0.39806
Epoch: 155, Time: 0.00463s, Loss: 0.43478
Epoch: 156, Time: 0.00641s, Loss: 0.37409
Epoch: 157, Time: 0.00509s, Loss: 0.39802
Epoch: 158, Time: 0.00453s, Loss: 0.34516
Epoch: 159, Time: 0.00563s, Loss: 0.39663
Epoch: 160, Time: 0.00456s, Loss: 0.37089
Epoch: 161, Time: 0.00711s, Loss: 0.39547
Epoch: 162, Time: 0.00455s, Loss: 0.41472
Epoch: 163, Time: 0.00645s, Loss: 0.40523
Epoch: 164, Time: 0.00465s, Loss: 0.33511
Epoch: 165, Time: 0.00565s, Loss: 0.35864
Epoch: 166, Time: 0.00575s, Loss: 0.33017
Epoch: 167, Time: 0.00785s, Loss: 0.36668
Epoch: 168, Time: 0.00604s, Loss: 0.36207
Epoch: 169, Time: 0.00650s, Loss: 0.37902
Epoch: 170, Time: 0.00473s, Loss: 0.38248
Epoch: 171, Time: 0.00664s, Loss: 0.34953
Epoch: 172, Time: 0.00556s, Loss: 0.38132
Epoch: 173, Time: 0.00686s, Loss: 0.34698
Epoch: 174, Time: 0.00619s, Loss: 0.36063
Epoch: 175, Time: 0.00468s, Loss: 0.34594
Epoch: 176, Time: 0.00545s, Loss: 0.37555
Epoch: 177, Time: 0.00457s, Loss: 0.35946
Epoch: 178, Time: 0.00718s, Loss: 0.35694
Epoch: 179, Time: 0.00458s, Loss: 0.34922
Epoch: 180, Time: 0.00693s, Loss: 0.30437
Epoch: 181, Time: 0.00461s, Loss: 0.34730
Epoch: 182, Time: 0.00632s, Loss: 0.31228
Epoch: 183, Time: 0.00509s, Loss: 0.36002
Epoch: 184, Time: 0.00454s, Loss: 0.36114
Epoch: 185, Time: 0.00546s, Loss: 0.34812
Epoch: 186, Time: 0.00456s, Loss: 0.33244
Epoch: 187, Time: 0.00696s, Loss: 0.34411
Epoch: 188, Time: 0.00459s, Loss: 0.35262
Epoch: 189, Time: 0.00628s, Loss: 0.32643
Epoch: 190, Time: 0.00472s, Loss: 0.32591
Epoch: 191, Time: 0.00451s, Loss: 0.33036
Epoch: 192, Time: 0.00594s, Loss: 0.31552
Epoch: 193, Time: 0.00559s, Loss: 0.32376
Epoch: 194, Time: 0.00627s, Loss: 0.31232
Epoch: 195, Time: 0.00550s, Loss: 0.33725
Epoch: 196, Time: 0.00570s, Loss: 0.34083
Epoch: 197, Time: 0.00508s, Loss: 0.30638
Epoch: 198, Time: 0.00559s, Loss: 0.33905
Epoch: 199, Time: 0.00603s, Loss: 0.30302

train finished!
best val: 0.80200
test...
final result: epoch: 92
{'accuracy': 0.8270000219345093, 'f1_score': 0.8198394539104813, 'f1_score -> average@micro': 0.827}