超图

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，我们使用三种典型图/超图神经网络在超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).
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

Cooking 200 数据集 (dhg.data.Cooking200) 从 Yummly.com 收集并用于节点分类任务。其为超图数据集，顶点代表菜式，超边代表配料。每种菜式与一种分类信息关联，为该菜式的菜系（中餐、日本菜、法国菜、俄罗斯菜）。

Note

数据集为超图数据集，不能直接用于GCN模型。因此，使用 clique expansion 将超图转为图。

Note

数据集不包含顶点特征，因此我们生成一个单位矩阵代表顶点特征矩阵。

Warning

生成单位矩阵作为顶点特征会导致训练阶段参数不稳定。因此，batch_norm会用于以下示例中的GCN、HGNN和HGNN+模型中。

结果汇总

模型	验证集的Accuracy	测试集的Accuracy	测试集F1 score
GCN	0.500	0.434	0.356
HGNN	0.485	0.495	0.376
HGNN+	0.475	0.520	0.391

Cooking200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, Hypergraph
from dhg.data import Cooking200
from dhg.models import GCN
from dhg.random import set_seed
from dhg.metrics import HypergraphVertexClassificationEvaluator 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(2021)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cooking200()

    X, lbl = torch.eye(data["num_vertices"]), data["labels"]
    ft_dim = X.shape[1]
    HG = Hypergraph(data["num_vertices"], data["edge_list"])
    G = Graph.from_hypergraph_clique(HG, weighted=True)
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = GCN(ft_dim, 32, data["num_classes"], use_bn=True)
    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(200):
        # 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: 7.29884s, Loss: 3.02374
update best: 0.05000
Epoch: 1, Time: 0.02545s, Loss: 2.47223
Epoch: 2, Time: 0.02411s, Loss: 2.41279
update best: 0.05500
Epoch: 3, Time: 0.02656s, Loss: 2.36803
update best: 0.07500
Epoch: 4, Time: 0.02486s, Loss: 2.33794
Epoch: 5, Time: 0.02224s, Loss: 2.30590
Epoch: 6, Time: 0.02089s, Loss: 2.28631
Epoch: 7, Time: 0.02136s, Loss: 2.25775
Epoch: 8, Time: 0.02186s, Loss: 2.24081
update best: 0.08000
Epoch: 9, Time: 0.02203s, Loss: 2.22660
update best: 0.09500
Epoch: 10, Time: 0.02155s, Loss: 2.20722
update best: 0.14500
Epoch: 11, Time: 0.02141s, Loss: 2.19497
Epoch: 12, Time: 0.02263s, Loss: 2.17880
Epoch: 13, Time: 0.02199s, Loss: 2.16433
Epoch: 14, Time: 0.02258s, Loss: 2.15038
Epoch: 15, Time: 0.02230s, Loss: 2.13811
Epoch: 16, Time: 0.02135s, Loss: 2.12440
Epoch: 17, Time: 0.02217s, Loss: 2.11146
Epoch: 18, Time: 0.02183s, Loss: 2.10333
Epoch: 19, Time: 0.03591s, Loss: 2.09031
Epoch: 20, Time: 0.02081s, Loss: 2.07710
Epoch: 21, Time: 0.02111s, Loss: 2.06423
Epoch: 22, Time: 0.02114s, Loss: 2.05410
Epoch: 23, Time: 0.02137s, Loss: 2.04545
update best: 0.15500
Epoch: 24, Time: 0.02159s, Loss: 2.03412
update best: 0.16000
Epoch: 25, Time: 0.02189s, Loss: 2.01589
update best: 0.17500
Epoch: 26, Time: 0.02204s, Loss: 2.01508
Epoch: 27, Time: 0.02206s, Loss: 1.99630
Epoch: 28, Time: 0.02180s, Loss: 1.98635
update best: 0.18500
Epoch: 29, Time: 0.02168s, Loss: 1.97526
update best: 0.20000
Epoch: 30, Time: 0.02155s, Loss: 1.96057
update best: 0.21000
Epoch: 31, Time: 0.02147s, Loss: 1.95878
update best: 0.21500
Epoch: 32, Time: 0.02174s, Loss: 1.94054
Epoch: 33, Time: 0.02147s, Loss: 1.93238
Epoch: 34, Time: 0.02176s, Loss: 1.92268
update best: 0.23000
Epoch: 35, Time: 0.02169s, Loss: 1.91224
update best: 0.24000
Epoch: 36, Time: 0.02141s, Loss: 1.89593
update best: 0.25000
Epoch: 37, Time: 0.02133s, Loss: 1.89175
update best: 0.25500
Epoch: 38, Time: 0.02230s, Loss: 1.88137
Epoch: 39, Time: 0.02201s, Loss: 1.87121
Epoch: 40, Time: 0.02050s, Loss: 1.85513
Epoch: 41, Time: 0.02120s, Loss: 1.85149
Epoch: 42, Time: 0.02102s, Loss: 1.83702
update best: 0.27000
Epoch: 43, Time: 0.02095s, Loss: 1.82509
update best: 0.27500
Epoch: 44, Time: 0.02139s, Loss: 1.81752
update best: 0.29000
Epoch: 45, Time: 0.02115s, Loss: 1.80817
Epoch: 46, Time: 0.02119s, Loss: 1.79938
update best: 0.29500
Epoch: 47, Time: 0.02088s, Loss: 1.78561
update best: 0.33000
Epoch: 48, Time: 0.02106s, Loss: 1.78137
update best: 0.34000
Epoch: 49, Time: 0.02088s, Loss: 1.76117
update best: 0.34500
Epoch: 50, Time: 0.02143s, Loss: 1.75598
update best: 0.36000
Epoch: 51, Time: 0.02129s, Loss: 1.74965
Epoch: 52, Time: 0.02177s, Loss: 1.73695
Epoch: 53, Time: 0.02160s, Loss: 1.72132
update best: 0.36500
Epoch: 54, Time: 0.02177s, Loss: 1.71943
update best: 0.37000
Epoch: 55, Time: 0.02115s, Loss: 1.71475
update best: 0.37500
Epoch: 56, Time: 0.02157s, Loss: 1.69237
update best: 0.38500
Epoch: 57, Time: 0.02164s, Loss: 1.68571
update best: 0.39500
Epoch: 58, Time: 0.02150s, Loss: 1.67695
update best: 0.40000
Epoch: 59, Time: 0.02156s, Loss: 1.66385
Epoch: 60, Time: 0.02155s, Loss: 1.65498
Epoch: 61, Time: 0.02102s, Loss: 1.65138
update best: 0.41000
Epoch: 62, Time: 0.02167s, Loss: 1.63215
update best: 0.42000
Epoch: 63, Time: 0.02174s, Loss: 1.62920
update best: 0.43500
Epoch: 64, Time: 0.02154s, Loss: 1.61913
update best: 0.44000
Epoch: 65, Time: 0.02159s, Loss: 1.61141
Epoch: 66, Time: 0.02195s, Loss: 1.60337
Epoch: 67, Time: 0.02069s, Loss: 1.58908
update best: 0.45500
Epoch: 68, Time: 0.02115s, Loss: 1.57248
Epoch: 69, Time: 0.02138s, Loss: 1.57386
update best: 0.46500
Epoch: 70, Time: 0.02106s, Loss: 1.56231
Epoch: 71, Time: 0.02118s, Loss: 1.55329
Epoch: 72, Time: 0.02242s, Loss: 1.54713
Epoch: 73, Time: 0.02136s, Loss: 1.53178
Epoch: 74, Time: 0.02172s, Loss: 1.52513
Epoch: 75, Time: 0.02200s, Loss: 1.51584
Epoch: 76, Time: 0.02123s, Loss: 1.50966
update best: 0.47000
Epoch: 77, Time: 0.02147s, Loss: 1.50546
update best: 0.47500
Epoch: 78, Time: 0.02270s, Loss: 1.49482
Epoch: 79, Time: 0.02264s, Loss: 1.47653
Epoch: 80, Time: 0.02349s, Loss: 1.46740
Epoch: 81, Time: 0.02231s, Loss: 1.46205
Epoch: 82, Time: 0.02251s, Loss: 1.44632
Epoch: 83, Time: 0.02184s, Loss: 1.44394
Epoch: 84, Time: 0.02175s, Loss: 1.43398
Epoch: 85, Time: 0.02109s, Loss: 1.43450
Epoch: 86, Time: 0.02110s, Loss: 1.41855
Epoch: 87, Time: 0.02112s, Loss: 1.41488
Epoch: 88, Time: 0.02119s, Loss: 1.40113
Epoch: 89, Time: 0.02133s, Loss: 1.38627
Epoch: 90, Time: 0.02178s, Loss: 1.38061
Epoch: 91, Time: 0.02106s, Loss: 1.38012
Epoch: 92, Time: 0.02245s, Loss: 1.36612
Epoch: 93, Time: 0.02165s, Loss: 1.36384
Epoch: 94, Time: 0.02169s, Loss: 1.35315
Epoch: 95, Time: 0.02287s, Loss: 1.33591
Epoch: 96, Time: 0.02321s, Loss: 1.33441
Epoch: 97, Time: 0.02267s, Loss: 1.32461
Epoch: 98, Time: 0.02246s, Loss: 1.31650
Epoch: 99, Time: 0.02192s, Loss: 1.30920
Epoch: 100, Time: 0.02145s, Loss: 1.29616
Epoch: 101, Time: 0.02106s, Loss: 1.28773
Epoch: 102, Time: 0.02128s, Loss: 1.28913
Epoch: 103, Time: 0.02125s, Loss: 1.27793
Epoch: 104, Time: 0.02174s, Loss: 1.27127
Epoch: 105, Time: 0.02135s, Loss: 1.26090
Epoch: 106, Time: 0.02187s, Loss: 1.25673
Epoch: 107, Time: 0.02137s, Loss: 1.23971
Epoch: 108, Time: 0.02163s, Loss: 1.23427
Epoch: 109, Time: 0.02173s, Loss: 1.23829
Epoch: 110, Time: 0.02228s, Loss: 1.21614
Epoch: 111, Time: 0.02190s, Loss: 1.22033
Epoch: 112, Time: 0.02146s, Loss: 1.21155
update best: 0.48000
Epoch: 113, Time: 0.02183s, Loss: 1.19760
Epoch: 114, Time: 0.02472s, Loss: 1.20577
Epoch: 115, Time: 0.02249s, Loss: 1.18268
Epoch: 116, Time: 0.02274s, Loss: 1.17723
Epoch: 117, Time: 0.02290s, Loss: 1.16582
Epoch: 118, Time: 0.02262s, Loss: 1.16943
Epoch: 119, Time: 0.02180s, Loss: 1.16023
Epoch: 120, Time: 0.02193s, Loss: 1.14612
update best: 0.48500
Epoch: 121, Time: 0.02191s, Loss: 1.14254
Epoch: 122, Time: 0.02162s, Loss: 1.13199
Epoch: 123, Time: 0.02136s, Loss: 1.12077
Epoch: 124, Time: 0.02165s, Loss: 1.11500
Epoch: 125, Time: 0.02177s, Loss: 1.11730
Epoch: 126, Time: 0.02150s, Loss: 1.10626
Epoch: 127, Time: 0.02119s, Loss: 1.09788
Epoch: 128, Time: 0.02119s, Loss: 1.09148
Epoch: 129, Time: 0.02130s, Loss: 1.08841
Epoch: 130, Time: 0.02211s, Loss: 1.08878
Epoch: 131, Time: 0.02171s, Loss: 1.08039
Epoch: 132, Time: 0.02172s, Loss: 1.06337
Epoch: 133, Time: 0.02185s, Loss: 1.05798
Epoch: 134, Time: 0.02197s, Loss: 1.05995
Epoch: 135, Time: 0.02310s, Loss: 1.04716
Epoch: 136, Time: 0.02271s, Loss: 1.03834
update best: 0.49000
Epoch: 137, Time: 0.02218s, Loss: 1.03407
Epoch: 138, Time: 0.02329s, Loss: 1.02641
Epoch: 139, Time: 0.02310s, Loss: 1.02540
Epoch: 140, Time: 0.02245s, Loss: 1.02152
Epoch: 141, Time: 0.02171s, Loss: 1.01990
Epoch: 142, Time: 0.02151s, Loss: 1.00520
Epoch: 143, Time: 0.02128s, Loss: 1.01225
Epoch: 144, Time: 0.02179s, Loss: 1.00302
Epoch: 145, Time: 0.02164s, Loss: 0.98153
Epoch: 146, Time: 0.02117s, Loss: 0.97740
Epoch: 147, Time: 0.02110s, Loss: 0.97149
Epoch: 148, Time: 0.02131s, Loss: 0.97149
Epoch: 149, Time: 0.02128s, Loss: 0.97657
Epoch: 150, Time: 0.02155s, Loss: 0.95241
Epoch: 151, Time: 0.02171s, Loss: 0.96010
Epoch: 152, Time: 0.02174s, Loss: 0.94509
Epoch: 153, Time: 0.02167s, Loss: 0.94987
Epoch: 154, Time: 0.02262s, Loss: 0.94258
Epoch: 155, Time: 0.02226s, Loss: 0.93526
Epoch: 156, Time: 0.02236s, Loss: 0.93201
Epoch: 157, Time: 0.02148s, Loss: 0.92291
Epoch: 158, Time: 0.02158s, Loss: 0.93494
Epoch: 159, Time: 0.02159s, Loss: 0.91413
Epoch: 160, Time: 0.02150s, Loss: 0.91853
Epoch: 161, Time: 0.02143s, Loss: 0.90566
Epoch: 162, Time: 0.02117s, Loss: 0.90713
Epoch: 163, Time: 0.02124s, Loss: 0.89651
Epoch: 164, Time: 0.02103s, Loss: 0.89034
Epoch: 165, Time: 0.02168s, Loss: 0.88661
Epoch: 166, Time: 0.02163s, Loss: 0.88348
Epoch: 167, Time: 0.02174s, Loss: 0.87290
Epoch: 168, Time: 0.02185s, Loss: 0.87435
Epoch: 169, Time: 0.02155s, Loss: 0.86458
Epoch: 170, Time: 0.02088s, Loss: 0.87389
Epoch: 171, Time: 0.02264s, Loss: 0.86114
Epoch: 172, Time: 0.02286s, Loss: 0.84979
Epoch: 173, Time: 0.02272s, Loss: 0.85025
Epoch: 174, Time: 0.02237s, Loss: 0.85343
Epoch: 175, Time: 0.02243s, Loss: 0.84297
Epoch: 176, Time: 0.02235s, Loss: 0.84274
Epoch: 177, Time: 0.02185s, Loss: 0.83616
Epoch: 178, Time: 0.02188s, Loss: 0.83237
Epoch: 179, Time: 0.02110s, Loss: 0.83829
Epoch: 180, Time: 0.02102s, Loss: 0.83292
Epoch: 181, Time: 0.02157s, Loss: 0.82355
Epoch: 182, Time: 0.02148s, Loss: 0.82146
Epoch: 183, Time: 0.02148s, Loss: 0.82488
Epoch: 184, Time: 0.02128s, Loss: 0.81608
Epoch: 185, Time: 0.02128s, Loss: 0.81082
Epoch: 186, Time: 0.02121s, Loss: 0.81338
Epoch: 187, Time: 0.02183s, Loss: 0.81301
Epoch: 188, Time: 0.02234s, Loss: 0.79188
Epoch: 189, Time: 0.02182s, Loss: 0.79709
update best: 0.50000
Epoch: 190, Time: 0.02134s, Loss: 0.78706
Epoch: 191, Time: 0.02183s, Loss: 0.77257
Epoch: 192, Time: 0.02276s, Loss: 0.77896
Epoch: 193, Time: 0.02326s, Loss: 0.77773
Epoch: 194, Time: 0.02287s, Loss: 0.76515
Epoch: 195, Time: 0.02281s, Loss: 0.76747
Epoch: 196, Time: 0.02164s, Loss: 0.76833
Epoch: 197, Time: 0.02182s, Loss: 0.75029
Epoch: 198, Time: 0.02136s, Loss: 0.76452
Epoch: 199, Time: 0.02135s, Loss: 0.75916

train finished!
best val: 0.50000
test...
final result: epoch: 189
{'accuracy': 0.4340996742248535, 'f1_score': 0.35630662515488015, 'f1_score -> average@micro': 0.43409967156932744}

Cooking200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 Hypergraph
from dhg.data import Cooking200
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, 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(2021)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cooking200()

    X, lbl = torch.eye(data["num_vertices"]), data["labels"]
    G = Hypergraph(data["num_vertices"], data["edge_list"])
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = HGNN(X.shape[1], 32, data["num_classes"], use_bn=True)
    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(200):
        # 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: 0.57807s, Loss: 2.99290
update best: 0.10000
Epoch: 1, Time: 0.02624s, Loss: 2.28624
Epoch: 2, Time: 0.02707s, Loss: 2.15988
Epoch: 3, Time: 0.02373s, Loss: 2.05894
Epoch: 4, Time: 0.02545s, Loss: 1.99918
Epoch: 5, Time: 0.02619s, Loss: 1.92948
Epoch: 6, Time: 0.02215s, Loss: 1.88097
Epoch: 7, Time: 0.02229s, Loss: 1.83393
Epoch: 8, Time: 0.02181s, Loss: 1.79070
Epoch: 9, Time: 0.02256s, Loss: 1.75345
Epoch: 10, Time: 0.02264s, Loss: 1.70969
Epoch: 11, Time: 0.02248s, Loss: 1.68242
Epoch: 12, Time: 0.02248s, Loss: 1.64419
Epoch: 13, Time: 0.02257s, Loss: 1.60876
Epoch: 14, Time: 0.02238s, Loss: 1.58108
Epoch: 15, Time: 0.02194s, Loss: 1.54466
Epoch: 16, Time: 0.02172s, Loss: 1.52140
Epoch: 17, Time: 0.02130s, Loss: 1.48225
Epoch: 18, Time: 0.02156s, Loss: 1.46237
Epoch: 19, Time: 0.02133s, Loss: 1.43527
Epoch: 20, Time: 0.02148s, Loss: 1.40451
Epoch: 21, Time: 0.02133s, Loss: 1.39555
Epoch: 22, Time: 0.02182s, Loss: 1.36368
Epoch: 23, Time: 0.02151s, Loss: 1.33732
Epoch: 24, Time: 0.02178s, Loss: 1.32686
Epoch: 25, Time: 0.02232s, Loss: 1.30681
Epoch: 26, Time: 0.02289s, Loss: 1.28287
Epoch: 27, Time: 0.02245s, Loss: 1.28563
Epoch: 28, Time: 0.02210s, Loss: 1.24644
Epoch: 29, Time: 0.02195s, Loss: 1.22813
Epoch: 30, Time: 0.02205s, Loss: 1.20336
Epoch: 31, Time: 0.02245s, Loss: 1.20308
Epoch: 32, Time: 0.02129s, Loss: 1.16802
Epoch: 33, Time: 0.02144s, Loss: 1.17182
Epoch: 34, Time: 0.02215s, Loss: 1.14047
Epoch: 35, Time: 0.02195s, Loss: 1.13377
Epoch: 36, Time: 0.02233s, Loss: 1.09250
Epoch: 37, Time: 0.02283s, Loss: 1.09588
Epoch: 38, Time: 0.02356s, Loss: 1.09042
Epoch: 39, Time: 0.02211s, Loss: 1.08532
Epoch: 40, Time: 0.02340s, Loss: 1.04074
update best: 0.11000
Epoch: 41, Time: 0.02125s, Loss: 1.05056
update best: 0.13500
Epoch: 42, Time: 0.02302s, Loss: 1.02834
update best: 0.14000
Epoch: 43, Time: 0.02278s, Loss: 0.99903
update best: 0.14500
Epoch: 44, Time: 0.02238s, Loss: 1.01756
update best: 0.15000
Epoch: 45, Time: 0.02286s, Loss: 0.99652
update best: 0.17500
Epoch: 46, Time: 0.02251s, Loss: 0.97935
update best: 0.21500
Epoch: 47, Time: 0.02234s, Loss: 0.97873
update best: 0.24500
Epoch: 48, Time: 0.02245s, Loss: 0.95888
update best: 0.26000
Epoch: 49, Time: 0.02228s, Loss: 0.95761
update best: 0.28000
Epoch: 50, Time: 0.02254s, Loss: 0.94229
Epoch: 51, Time: 0.02264s, Loss: 0.92833
update best: 0.29000
Epoch: 52, Time: 0.02238s, Loss: 0.92601
update best: 0.30000
Epoch: 53, Time: 0.02311s, Loss: 0.90252
update best: 0.31000
Epoch: 54, Time: 0.02189s, Loss: 0.89501
update best: 0.32500
Epoch: 55, Time: 0.02193s, Loss: 0.89724
Epoch: 56, Time: 0.02246s, Loss: 0.87068
update best: 0.33500
Epoch: 57, Time: 0.02181s, Loss: 0.87531
update best: 0.34000
Epoch: 58, Time: 0.02287s, Loss: 0.84288
update best: 0.34500
Epoch: 59, Time: 0.02227s, Loss: 0.84243
update best: 0.36500
Epoch: 60, Time: 0.02149s, Loss: 0.83892
update best: 0.38500
Epoch: 61, Time: 0.02253s, Loss: 0.83062
update best: 0.40000
Epoch: 62, Time: 0.02271s, Loss: 0.82245
update best: 0.42000
Epoch: 63, Time: 0.02195s, Loss: 0.81214
update best: 0.43000
Epoch: 64, Time: 0.02162s, Loss: 0.80847
update best: 0.44000
Epoch: 65, Time: 0.02136s, Loss: 0.78325
Epoch: 66, Time: 0.02245s, Loss: 0.79052
update best: 0.45500
Epoch: 67, Time: 0.02248s, Loss: 0.78128
Epoch: 68, Time: 0.02295s, Loss: 0.77049
Epoch: 69, Time: 0.02315s, Loss: 0.75469
Epoch: 70, Time: 0.02331s, Loss: 0.74771
Epoch: 71, Time: 0.02317s, Loss: 0.73701
Epoch: 72, Time: 0.02307s, Loss: 0.74350
Epoch: 73, Time: 0.02176s, Loss: 0.73698
Epoch: 74, Time: 0.02164s, Loss: 0.72565
Epoch: 75, Time: 0.02148s, Loss: 0.70553
update best: 0.46500
Epoch: 76, Time: 0.02136s, Loss: 0.71696
Epoch: 77, Time: 0.02111s, Loss: 0.72410
Epoch: 78, Time: 0.02111s, Loss: 0.71131
update best: 0.47000
Epoch: 79, Time: 0.02180s, Loss: 0.68748
Epoch: 80, Time: 0.02095s, Loss: 0.68774
Epoch: 81, Time: 0.02147s, Loss: 0.70136
Epoch: 82, Time: 0.02122s, Loss: 0.66882
Epoch: 83, Time: 0.02164s, Loss: 0.64563
Epoch: 84, Time: 0.02149s, Loss: 0.66794
Epoch: 85, Time: 0.02194s, Loss: 0.65860
Epoch: 86, Time: 0.02157s, Loss: 0.66000
Epoch: 87, Time: 0.02267s, Loss: 0.65452
Epoch: 88, Time: 0.02250s, Loss: 0.64512
Epoch: 89, Time: 0.02169s, Loss: 0.64318
Epoch: 90, Time: 0.02175s, Loss: 0.63814
Epoch: 91, Time: 0.02177s, Loss: 0.62040
Epoch: 92, Time: 0.02108s, Loss: 0.61942
Epoch: 93, Time: 0.02111s, Loss: 0.61757
Epoch: 94, Time: 0.02118s, Loss: 0.60520
Epoch: 95, Time: 0.02112s, Loss: 0.58358
Epoch: 96, Time: 0.02129s, Loss: 0.58866
Epoch: 97, Time: 0.02171s, Loss: 0.58599
Epoch: 98, Time: 0.02220s, Loss: 0.59330
Epoch: 99, Time: 0.02243s, Loss: 0.56555
Epoch: 100, Time: 0.02262s, Loss: 0.57273
Epoch: 101, Time: 0.02240s, Loss: 0.57785
Epoch: 102, Time: 0.02086s, Loss: 0.56949
Epoch: 103, Time: 0.02111s, Loss: 0.55187
Epoch: 104, Time: 0.02136s, Loss: 0.55166
Epoch: 105, Time: 0.02119s, Loss: 0.54706
Epoch: 106, Time: 0.02107s, Loss: 0.55239
Epoch: 107, Time: 0.02136s, Loss: 0.53656
Epoch: 108, Time: 0.02115s, Loss: 0.53478
Epoch: 109, Time: 0.02146s, Loss: 0.52564
Epoch: 110, Time: 0.02189s, Loss: 0.52242
Epoch: 111, Time: 0.02248s, Loss: 0.52779
Epoch: 112, Time: 0.02191s, Loss: 0.50813
Epoch: 113, Time: 0.02182s, Loss: 0.51623
Epoch: 114, Time: 0.02143s, Loss: 0.51834
Epoch: 115, Time: 0.02220s, Loss: 0.49232
Epoch: 116, Time: 0.02117s, Loss: 0.51582
Epoch: 117, Time: 0.02116s, Loss: 0.49434
Epoch: 118, Time: 0.02110s, Loss: 0.49518
Epoch: 119, Time: 0.02147s, Loss: 0.49155
Epoch: 120, Time: 0.02122s, Loss: 0.48029
Epoch: 121, Time: 0.02153s, Loss: 0.49079
Epoch: 122, Time: 0.02151s, Loss: 0.48253
Epoch: 123, Time: 0.02170s, Loss: 0.46945
Epoch: 124, Time: 0.02259s, Loss: 0.47764
Epoch: 125, Time: 0.02228s, Loss: 0.47102
Epoch: 126, Time: 0.02196s, Loss: 0.45784
Epoch: 127, Time: 0.02184s, Loss: 0.46020
Epoch: 128, Time: 0.02245s, Loss: 0.45922
Epoch: 129, Time: 0.02191s, Loss: 0.46458
Epoch: 130, Time: 0.02215s, Loss: 0.46924
Epoch: 131, Time: 0.02222s, Loss: 0.45952
Epoch: 132, Time: 0.02226s, Loss: 0.44490
Epoch: 133, Time: 0.02174s, Loss: 0.44763
Epoch: 134, Time: 0.02143s, Loss: 0.45225
Epoch: 135, Time: 0.02149s, Loss: 0.42556
Epoch: 136, Time: 0.02141s, Loss: 0.42714
Epoch: 137, Time: 0.02150s, Loss: 0.43604
Epoch: 138, Time: 0.02171s, Loss: 0.42259
Epoch: 139, Time: 0.02168s, Loss: 0.41784
Epoch: 140, Time: 0.02149s, Loss: 0.41759
Epoch: 141, Time: 0.02125s, Loss: 0.41633
Epoch: 142, Time: 0.02220s, Loss: 0.42547
Epoch: 143, Time: 0.02271s, Loss: 0.41790
Epoch: 144, Time: 0.02280s, Loss: 0.39776
Epoch: 145, Time: 0.02264s, Loss: 0.41429
Epoch: 146, Time: 0.02128s, Loss: 0.39543
Epoch: 147, Time: 0.02141s, Loss: 0.39529
Epoch: 148, Time: 0.02100s, Loss: 0.41145
Epoch: 149, Time: 0.02103s, Loss: 0.40083
Epoch: 150, Time: 0.02170s, Loss: 0.39246
Epoch: 151, Time: 0.02154s, Loss: 0.39613
Epoch: 152, Time: 0.02188s, Loss: 0.38080
Epoch: 153, Time: 0.02213s, Loss: 0.39159
Epoch: 154, Time: 0.02236s, Loss: 0.38570
Epoch: 155, Time: 0.02209s, Loss: 0.38382
Epoch: 156, Time: 0.02146s, Loss: 0.37949
update best: 0.47500
Epoch: 157, Time: 0.02179s, Loss: 0.37078
Epoch: 158, Time: 0.02223s, Loss: 0.37063
Epoch: 159, Time: 0.02219s, Loss: 0.37556
Epoch: 160, Time: 0.02217s, Loss: 0.37468
Epoch: 161, Time: 0.02146s, Loss: 0.38581
update best: 0.48500
Epoch: 162, Time: 0.02278s, Loss: 0.36664
Epoch: 163, Time: 0.02172s, Loss: 0.35075
Epoch: 164, Time: 0.02139s, Loss: 0.35056
Epoch: 165, Time: 0.02156s, Loss: 0.36339
Epoch: 166, Time: 0.02149s, Loss: 0.36245
Epoch: 167, Time: 0.02133s, Loss: 0.34675
Epoch: 168, Time: 0.02141s, Loss: 0.36043
Epoch: 169, Time: 0.02148s, Loss: 0.34538
Epoch: 170, Time: 0.02128s, Loss: 0.34694
Epoch: 171, Time: 0.02138s, Loss: 0.33723
Epoch: 172, Time: 0.02260s, Loss: 0.34017
Epoch: 173, Time: 0.02259s, Loss: 0.33932
Epoch: 174, Time: 0.02307s, Loss: 0.33170
Epoch: 175, Time: 0.02290s, Loss: 0.31819
Epoch: 176, Time: 0.02261s, Loss: 0.33577
Epoch: 177, Time: 0.02269s, Loss: 0.34146
Epoch: 178, Time: 0.02284s, Loss: 0.33086
Epoch: 179, Time: 0.02215s, Loss: 0.34498
Epoch: 180, Time: 0.02317s, Loss: 0.33026
Epoch: 181, Time: 0.02228s, Loss: 0.32811
Epoch: 182, Time: 0.02216s, Loss: 0.33203
Epoch: 183, Time: 0.02248s, Loss: 0.31955
Epoch: 184, Time: 0.02239s, Loss: 0.34238
Epoch: 185, Time: 0.02253s, Loss: 0.30963
Epoch: 186, Time: 0.02240s, Loss: 0.31527
Epoch: 187, Time: 0.02199s, Loss: 0.31484
Epoch: 188, Time: 0.02200s, Loss: 0.32514
Epoch: 189, Time: 0.02171s, Loss: 0.32029
Epoch: 190, Time: 0.02169s, Loss: 0.32122
Epoch: 191, Time: 0.02157s, Loss: 0.30233
Epoch: 192, Time: 0.02125s, Loss: 0.30417
Epoch: 193, Time: 0.02159s, Loss: 0.30060
Epoch: 194, Time: 0.02142s, Loss: 0.29333
Epoch: 195, Time: 0.02155s, Loss: 0.29596
Epoch: 196, Time: 0.02158s, Loss: 0.30458
Epoch: 197, Time: 0.02204s, Loss: 0.29744
Epoch: 198, Time: 0.02227s, Loss: 0.29473
Epoch: 199, Time: 0.02259s, Loss: 0.30488

train finished!
best val: 0.48500
test...
final result: epoch: 161
{'accuracy': 0.4949307441711426, 'f1_score': 0.37618299381063885, 'f1_score -> average@micro': 0.49493074396687137}

Cooking200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 Hypergraph
from dhg.data import Cooking200
from dhg.models import HGNN, HGNNP
from dhg.random import set_seed
from dhg.metrics import HypergraphVertexClassificationEvaluator 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(2021)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    data = Cooking200()

    X, lbl = torch.eye(data["num_vertices"]), data["labels"]
    G = Hypergraph(data["num_vertices"], data["edge_list"])
    train_mask = data["train_mask"]
    val_mask = data["val_mask"]
    test_mask = data["test_mask"]

    net = HGNNP(X.shape[1], 32, data["num_classes"], use_bn=True)
    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(200):
        # 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: 0.52802s, Loss: 2.98654
update best: 0.05000
Epoch: 1, Time: 0.00738s, Loss: 2.28235
Epoch: 2, Time: 0.00829s, Loss: 2.15288
Epoch: 3, Time: 0.00929s, Loss: 2.05343
Epoch: 4, Time: 0.00716s, Loss: 1.99081
Epoch: 5, Time: 0.00703s, Loss: 1.92390
Epoch: 6, Time: 0.01025s, Loss: 1.87569
Epoch: 7, Time: 0.01015s, Loss: 1.83000
Epoch: 8, Time: 0.00870s, Loss: 1.78668
update best: 0.06500
Epoch: 9, Time: 0.00811s, Loss: 1.75019
Epoch: 10, Time: 0.00792s, Loss: 1.70593
Epoch: 11, Time: 0.00855s, Loss: 1.68245
Epoch: 12, Time: 0.00940s, Loss: 1.64045
Epoch: 13, Time: 0.00667s, Loss: 1.60735
Epoch: 14, Time: 0.00808s, Loss: 1.58477
Epoch: 15, Time: 0.00863s, Loss: 1.54530
Epoch: 16, Time: 0.00839s, Loss: 1.52168
Epoch: 17, Time: 0.00863s, Loss: 1.48935
Epoch: 18, Time: 0.01009s, Loss: 1.46205
Epoch: 19, Time: 0.00998s, Loss: 1.43605
Epoch: 20, Time: 0.00808s, Loss: 1.40635
Epoch: 21, Time: 0.00765s, Loss: 1.39397
Epoch: 22, Time: 0.00749s, Loss: 1.36317
Epoch: 23, Time: 0.00791s, Loss: 1.34086
Epoch: 24, Time: 0.00627s, Loss: 1.32558
Epoch: 25, Time: 0.00784s, Loss: 1.30849
Epoch: 26, Time: 0.00752s, Loss: 1.27822
Epoch: 27, Time: 0.00628s, Loss: 1.28945
Epoch: 28, Time: 0.00731s, Loss: 1.24414
Epoch: 29, Time: 0.00741s, Loss: 1.22858
Epoch: 30, Time: 0.00677s, Loss: 1.20161
Epoch: 31, Time: 0.00777s, Loss: 1.19882
Epoch: 32, Time: 0.00707s, Loss: 1.16460
Epoch: 33, Time: 0.00730s, Loss: 1.16780
Epoch: 34, Time: 0.00787s, Loss: 1.13391
update best: 0.07000
Epoch: 35, Time: 0.00747s, Loss: 1.13935
update best: 0.08500
Epoch: 36, Time: 0.00683s, Loss: 1.08887
update best: 0.12000
Epoch: 37, Time: 0.00780s, Loss: 1.08907
Epoch: 38, Time: 0.00782s, Loss: 1.08394
Epoch: 39, Time: 0.00626s, Loss: 1.07832
Epoch: 40, Time: 0.00783s, Loss: 1.03877
update best: 0.12500
Epoch: 41, Time: 0.00795s, Loss: 1.03990
update best: 0.13500
Epoch: 42, Time: 0.00626s, Loss: 1.02008
update best: 0.14500
Epoch: 43, Time: 0.00709s, Loss: 0.99529
update best: 0.16000
Epoch: 44, Time: 0.00763s, Loss: 1.01162
update best: 0.17500
Epoch: 45, Time: 0.00749s, Loss: 0.99196
update best: 0.20500
Epoch: 46, Time: 0.00629s, Loss: 0.97237
update best: 0.21000
Epoch: 47, Time: 0.00754s, Loss: 0.97511
update best: 0.22500
Epoch: 48, Time: 0.00805s, Loss: 0.95078
update best: 0.23000
Epoch: 49, Time: 0.00745s, Loss: 0.94715
update best: 0.24500
Epoch: 50, Time: 0.00643s, Loss: 0.93461
update best: 0.25500
Epoch: 51, Time: 0.00743s, Loss: 0.92102
update best: 0.27500
Epoch: 52, Time: 0.00772s, Loss: 0.91536
update best: 0.29500
Epoch: 53, Time: 0.00714s, Loss: 0.89386
update best: 0.30500
Epoch: 54, Time: 0.00722s, Loss: 0.88108
Epoch: 55, Time: 0.00777s, Loss: 0.88809
Epoch: 56, Time: 0.00717s, Loss: 0.85739
Epoch: 57, Time: 0.00724s, Loss: 0.86278
update best: 0.31000
Epoch: 58, Time: 0.00804s, Loss: 0.83276
update best: 0.32500
Epoch: 59, Time: 0.00786s, Loss: 0.83001
update best: 0.35000
Epoch: 60, Time: 0.00629s, Loss: 0.83385
update best: 0.37500
Epoch: 61, Time: 0.00712s, Loss: 0.82473
update best: 0.39500
Epoch: 62, Time: 0.00904s, Loss: 0.81101
update best: 0.41000
Epoch: 63, Time: 0.00745s, Loss: 0.80212
Epoch: 64, Time: 0.00715s, Loss: 0.79534
update best: 0.42000
Epoch: 65, Time: 0.00705s, Loss: 0.77077
Epoch: 66, Time: 0.00710s, Loss: 0.77775
update best: 0.43000
Epoch: 67, Time: 0.00717s, Loss: 0.77026
update best: 0.43500
Epoch: 68, Time: 0.00789s, Loss: 0.75978
Epoch: 69, Time: 0.00747s, Loss: 0.74209
Epoch: 70, Time: 0.00639s, Loss: 0.73636
Epoch: 71, Time: 0.00689s, Loss: 0.72454
Epoch: 72, Time: 0.00793s, Loss: 0.72910
Epoch: 73, Time: 0.00729s, Loss: 0.72512
Epoch: 74, Time: 0.00775s, Loss: 0.71034
update best: 0.44500
Epoch: 75, Time: 0.00766s, Loss: 0.69282
update best: 0.45000
Epoch: 76, Time: 0.00627s, Loss: 0.70622
update best: 0.46000
Epoch: 77, Time: 0.00706s, Loss: 0.70540
update best: 0.47500
Epoch: 78, Time: 0.00849s, Loss: 0.69790
Epoch: 79, Time: 0.00731s, Loss: 0.66718
Epoch: 80, Time: 0.00748s, Loss: 0.67149
Epoch: 81, Time: 0.00900s, Loss: 0.68492
Epoch: 82, Time: 0.00624s, Loss: 0.65467
Epoch: 83, Time: 0.00713s, Loss: 0.63049
Epoch: 84, Time: 0.00852s, Loss: 0.65693
Epoch: 85, Time: 0.00622s, Loss: 0.64821
Epoch: 86, Time: 0.00717s, Loss: 0.64481
Epoch: 87, Time: 0.00784s, Loss: 0.64284
Epoch: 88, Time: 0.00630s, Loss: 0.62653
Epoch: 89, Time: 0.00726s, Loss: 0.62808
Epoch: 90, Time: 0.00786s, Loss: 0.62135
Epoch: 91, Time: 0.00729s, Loss: 0.59833
Epoch: 92, Time: 0.00731s, Loss: 0.60561
Epoch: 93, Time: 0.00801s, Loss: 0.60091
Epoch: 94, Time: 0.00630s, Loss: 0.58819
Epoch: 95, Time: 0.00763s, Loss: 0.56774
Epoch: 96, Time: 0.00743s, Loss: 0.57335
Epoch: 97, Time: 0.00662s, Loss: 0.56947
Epoch: 98, Time: 0.00899s, Loss: 0.57430
Epoch: 99, Time: 0.00751s, Loss: 0.56189
Epoch: 100, Time: 0.00719s, Loss: 0.55171
Epoch: 101, Time: 0.00791s, Loss: 0.56934
Epoch: 102, Time: 0.00627s, Loss: 0.54815
Epoch: 103, Time: 0.00731s, Loss: 0.54027
Epoch: 104, Time: 0.00817s, Loss: 0.54291
Epoch: 105, Time: 0.00623s, Loss: 0.52773
Epoch: 106, Time: 0.00737s, Loss: 0.53735
Epoch: 107, Time: 0.00790s, Loss: 0.51841
Epoch: 108, Time: 0.00631s, Loss: 0.51548
Epoch: 109, Time: 0.00753s, Loss: 0.51153
Epoch: 110, Time: 0.00822s, Loss: 0.50702
Epoch: 111, Time: 0.00689s, Loss: 0.50974
Epoch: 112, Time: 0.00648s, Loss: 0.49094
Epoch: 113, Time: 0.00768s, Loss: 0.50044
Epoch: 114, Time: 0.00808s, Loss: 0.50632
Epoch: 115, Time: 0.00744s, Loss: 0.48155
Epoch: 116, Time: 0.00774s, Loss: 0.49875
Epoch: 117, Time: 0.00633s, Loss: 0.48650
Epoch: 118, Time: 0.00742s, Loss: 0.48026
Epoch: 119, Time: 0.00928s, Loss: 0.48162
Epoch: 120, Time: 0.00687s, Loss: 0.46713
Epoch: 121, Time: 0.00679s, Loss: 0.46894
Epoch: 122, Time: 0.00891s, Loss: 0.47300
Epoch: 123, Time: 0.00639s, Loss: 0.45836
Epoch: 124, Time: 0.00676s, Loss: 0.46030
Epoch: 125, Time: 0.00940s, Loss: 0.45373
Epoch: 126, Time: 0.00926s, Loss: 0.44894
Epoch: 127, Time: 0.00701s, Loss: 0.45110
Epoch: 128, Time: 0.00710s, Loss: 0.43749
Epoch: 129, Time: 0.00913s, Loss: 0.45104
Epoch: 130, Time: 0.00706s, Loss: 0.45284
Epoch: 131, Time: 0.00693s, Loss: 0.44452
Epoch: 132, Time: 0.00937s, Loss: 0.43088
Epoch: 133, Time: 0.00810s, Loss: 0.43557
Epoch: 134, Time: 0.00713s, Loss: 0.44251
Epoch: 135, Time: 0.00822s, Loss: 0.41227
Epoch: 136, Time: 0.00981s, Loss: 0.41414
Epoch: 137, Time: 0.00706s, Loss: 0.42148
Epoch: 138, Time: 0.00649s, Loss: 0.40822
Epoch: 139, Time: 0.00860s, Loss: 0.41343
Epoch: 140, Time: 0.00616s, Loss: 0.39754
Epoch: 141, Time: 0.00644s, Loss: 0.39057
Epoch: 142, Time: 0.00860s, Loss: 0.41271
Epoch: 143, Time: 0.00631s, Loss: 0.39916
Epoch: 144, Time: 0.00675s, Loss: 0.37878
Epoch: 145, Time: 0.00897s, Loss: 0.40234
Epoch: 146, Time: 0.00621s, Loss: 0.38136
Epoch: 147, Time: 0.00864s, Loss: 0.38960
Epoch: 148, Time: 0.00633s, Loss: 0.40494
Epoch: 149, Time: 0.00629s, Loss: 0.38099
Epoch: 150, Time: 0.00883s, Loss: 0.37809
Epoch: 151, Time: 0.00621s, Loss: 0.38888
Epoch: 152, Time: 0.00633s, Loss: 0.35971
Epoch: 153, Time: 0.00842s, Loss: 0.37553
Epoch: 154, Time: 0.00622s, Loss: 0.36924
Epoch: 155, Time: 0.00739s, Loss: 0.37269
Epoch: 156, Time: 0.00864s, Loss: 0.36131
Epoch: 157, Time: 0.00627s, Loss: 0.35630
Epoch: 158, Time: 0.00854s, Loss: 0.36315
Epoch: 159, Time: 0.00648s, Loss: 0.37506
Epoch: 160, Time: 0.00638s, Loss: 0.36177
Epoch: 161, Time: 0.00867s, Loss: 0.37122
Epoch: 162, Time: 0.00632s, Loss: 0.35660
Epoch: 163, Time: 0.00641s, Loss: 0.34108
Epoch: 164, Time: 0.00873s, Loss: 0.34228
Epoch: 165, Time: 0.00619s, Loss: 0.34731
Epoch: 166, Time: 0.00656s, Loss: 0.34604
Epoch: 167, Time: 0.00881s, Loss: 0.33136
Epoch: 168, Time: 0.00620s, Loss: 0.35096
Epoch: 169, Time: 0.00874s, Loss: 0.33567
Epoch: 170, Time: 0.00766s, Loss: 0.32705
Epoch: 171, Time: 0.00628s, Loss: 0.32490
Epoch: 172, Time: 0.00880s, Loss: 0.32892
Epoch: 173, Time: 0.00619s, Loss: 0.32556
Epoch: 174, Time: 0.00631s, Loss: 0.32410
Epoch: 175, Time: 0.00878s, Loss: 0.30940
Epoch: 176, Time: 0.00629s, Loss: 0.33027
Epoch: 177, Time: 0.00636s, Loss: 0.32709
Epoch: 178, Time: 0.00887s, Loss: 0.32104
Epoch: 179, Time: 0.00625s, Loss: 0.33687
Epoch: 180, Time: 0.00694s, Loss: 0.31593
Epoch: 181, Time: 0.00861s, Loss: 0.31409
Epoch: 182, Time: 0.00627s, Loss: 0.31477
Epoch: 183, Time: 0.00847s, Loss: 0.30355
Epoch: 184, Time: 0.00642s, Loss: 0.33237
Epoch: 185, Time: 0.00630s, Loss: 0.30555
Epoch: 186, Time: 0.00839s, Loss: 0.29973
Epoch: 187, Time: 0.00631s, Loss: 0.30695
Epoch: 188, Time: 0.00645s, Loss: 0.30313
Epoch: 189, Time: 0.00899s, Loss: 0.30699
Epoch: 190, Time: 0.00626s, Loss: 0.31283
Epoch: 191, Time: 0.00654s, Loss: 0.28851
Epoch: 192, Time: 0.00879s, Loss: 0.28803
Epoch: 193, Time: 0.00621s, Loss: 0.28213
Epoch: 194, Time: 0.00846s, Loss: 0.27823
Epoch: 195, Time: 0.00704s, Loss: 0.29048
Epoch: 196, Time: 0.00638s, Loss: 0.28898
Epoch: 197, Time: 0.00894s, Loss: 0.29096
Epoch: 198, Time: 0.00642s, Loss: 0.27857
Epoch: 199, Time: 0.00817s, Loss: 0.29117

train finished!
best val: 0.47500
test...
final result: epoch: 77
{'accuracy': 0.5203484296798706, 'f1_score': 0.39131907709452823, 'f1_score -> average@micro': 0.5203484221048122}