在Cooking200上使用HGNN+

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，我们介绍在超图节点分类任务中使用Auto-ML。关于如何使用 dhg.experiments 进行自动模型调优的细节可以参考自 <Training with Auto ML>。

配置

模型: HGNN+ (dhg.models.HGNNP): HGNN+: General Hypergraph Neural Networks 论文 (IEEE T-PAMI 2022)。
数据集: Cooking 200 (dhg.data.Cooking200): 一个用于顶点分类任务的烹饪食谱超图数据集，收集自 Yummly.com 。

导入依赖包

import torch
import torch.nn as nn
import torch.optim as optim

from dhg import Hypergraph
from dhg.data import Cooking200
from dhg.models import HGNNP
from dhg.random import set_seed
from dhg.experiments import HypergraphVertexClassificationTask as Task
from dhg.metrics import HypergraphVertexClassificationEvaluator as Evaluator

定义函数

def structure_builder(trial):
    global hg_base, g
    cur_hg: Hypergraph = hg_base.clone()
    return cur_hg


def model_builder(trial):
    return HGNNP(dim_features, trial.suggest_int("hidden_dim", 10, 20), num_classes, use_bn=True)


def train_builder(trial, model):
    optimizer = optim.Adam(
        model.parameters(),
        lr=trial.suggest_loguniform("lr", 1e-4, 1e-2),
        weight_decay=trial.suggest_loguniform("weight_decay", 1e-4, 1e-2),
    )
    criterion = nn.CrossEntropyLoss()
    return {
        "optimizer": optimizer,
        "criterion": criterion,
    }

主函数

Important

您需要修改 work_root 变量为您的工作目录。

if __name__ == "__main__":
    work_root = "/home/fengyifan/OS3D/toolbox/exp_cache/tmp"
    set_seed(2022)
    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    data = Cooking200()
    dim_features = data["num_vertices"]
    num_classes = data["num_classes"]
    hg_base = Hypergraph(data["num_vertices"], data["edge_list"])
    input_data = {
        "features": torch.eye(data["num_vertices"]),
        "labels": data["labels"],
        "train_mask": data["train_mask"],
        "val_mask": data["val_mask"],
        "test_mask": data["test_mask"],
    }
    evaluator = Evaluator(["accuracy", "f1_score", {"f1_score": {"average": "micro"}}])
    task = Task(
        work_root, input_data, model_builder, train_builder, evaluator, device, structure_builder=structure_builder,
    )
    task.run(200, 50, "maximize")

输出

[I 2022-08-25 17:46:08,399] Logs will be saved to /home/fengyifan/OS3D/toolbox/exp_cache/tmp/2022-08-25--17-46-08/log.txt
[I 2022-08-25 17:46:08,399] Files in training will be saved in /home/fengyifan/OS3D/toolbox/exp_cache/tmp/2022-08-25--17-46-08
[I 2022-08-25 17:46:09,904] Random seed is 2022
[I 2022-08-25 17:46:09,905] A new study created in memory with name: no-name-9e617917-a809-40dc-a6b9-87aeda5bb6ee
[I 2022-08-25 17:46:12,361] Trial 0 finished with value: 0.4000000059604645 and parameters: {'hidden_dim': 10, 'lr': 0.0009956704582324435, 'weight_decay': 0.00016856499028548418}. Best is trial 0 with value: 0.4000000059604645.
[I 2022-08-25 17:46:14,217] Trial 1 finished with value: 0.41999998688697815 and parameters: {'hidden_dim': 10, 'lr': 0.002348633160857829, 'weight_decay': 0.0009418378430920174}. Best is trial 1 with value: 0.41999998688697815.
[I 2022-08-25 17:46:16,074] Trial 2 finished with value: 0.48500001430511475 and parameters: {'hidden_dim': 19, 'lr': 0.0019719874263090698, 'weight_decay': 0.006221946114841155}. Best is trial 2 with value: 0.48500001430511475.
[I 2022-08-25 17:46:18,074] Trial 3 finished with value: 0.48500001430511475 and parameters: {'hidden_dim': 17, 'lr': 0.004599459949791714, 'weight_decay': 0.0045199760918655345}. Best is trial 2 with value: 0.48500001430511475.
[I 2022-08-25 17:46:20,060] Trial 4 finished with value: 0.4950000047683716 and parameters: {'hidden_dim': 19, 'lr': 0.008205190552892963, 'weight_decay': 0.0005446140912512398}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:20,551] Trial 5 pruned.
[I 2022-08-25 17:46:21,146] Trial 6 pruned.
[I 2022-08-25 17:46:21,751] Trial 7 pruned.
[I 2022-08-25 17:46:22,397] Trial 8 pruned.
[I 2022-08-25 17:46:22,720] Trial 9 pruned.
[I 2022-08-25 17:46:24,731] Trial 10 finished with value: 0.49000000953674316 and parameters: {'hidden_dim': 18, 'lr': 0.009112327540785461, 'weight_decay': 0.0002825142053930118}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:26,677] Trial 11 finished with value: 0.49000000953674316 and parameters: {'hidden_dim': 17, 'lr': 0.009700863338872084, 'weight_decay': 0.00024395653633063402}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:28,785] Trial 12 finished with value: 0.4950000047683716 and parameters: {'hidden_dim': 18, 'lr': 0.009506157011953582, 'weight_decay': 0.00034409703681570236}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:30,801] Trial 13 finished with value: 0.49000000953674316 and parameters: {'hidden_dim': 20, 'lr': 0.004245693592715978, 'weight_decay': 0.00046142123936015995}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:32,813] Trial 14 finished with value: 0.4950000047683716 and parameters: {'hidden_dim': 17, 'lr': 0.00494083746774663, 'weight_decay': 0.0001151901195440639}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:33,293] Trial 15 pruned.
[I 2022-08-25 17:46:33,810] Trial 16 pruned.
[I 2022-08-25 17:46:34,329] Trial 17 pruned.
[I 2022-08-25 17:46:34,840] Trial 18 pruned.
[I 2022-08-25 17:46:35,358] Trial 19 pruned.
[I 2022-08-25 17:46:35,902] Trial 20 pruned.
[I 2022-08-25 17:46:36,895] Trial 21 pruned.
[I 2022-08-25 17:46:37,406] Trial 22 pruned.
[I 2022-08-25 17:46:39,326] Trial 23 finished with value: 0.49000000953674316 and parameters: {'hidden_dim': 16, 'lr': 0.006943644200360305, 'weight_decay': 0.0006003049507614988}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:41,236] Trial 24 finished with value: 0.48500001430511475 and parameters: {'hidden_dim': 20, 'lr': 0.009971146065887018, 'weight_decay': 0.00035931897741066387}. Best is trial 4 with value: 0.4950000047683716.
[I 2022-08-25 17:46:41,732] Trial 25 pruned.
[I 2022-08-25 17:46:42,160] Trial 26 pruned.
[I 2022-08-25 17:46:42,589] Trial 27 pruned.
[I 2022-08-25 17:46:43,120] Trial 28 pruned.
[I 2022-08-25 17:46:43,631] Trial 29 pruned.
[I 2022-08-25 17:46:44,143] Trial 30 pruned.
[I 2022-08-25 17:46:44,612] Trial 31 pruned.
[I 2022-08-25 17:46:45,187] Trial 32 pruned.
[I 2022-08-25 17:46:45,906] Trial 33 pruned.
[I 2022-08-25 17:46:46,544] Trial 34 pruned.
[I 2022-08-25 17:46:46,965] Trial 35 pruned.
[I 2022-08-25 17:46:48,842] Trial 36 finished with value: 0.5049999952316284 and parameters: {'hidden_dim': 17, 'lr': 0.009648904316000167, 'weight_decay': 0.00013498962749734303}. Best is trial 36 with value: 0.5049999952316284.
[I 2022-08-25 17:46:49,339] Trial 37 pruned.
[I 2022-08-25 17:46:51,214] Trial 38 finished with value: 0.48500001430511475 and parameters: {'hidden_dim': 18, 'lr': 0.009528262435822034, 'weight_decay': 0.00013603318896175282}. Best is trial 36 with value: 0.5049999952316284.
[I 2022-08-25 17:46:51,612] Trial 39 pruned.
[I 2022-08-25 17:46:53,637] Trial 40 finished with value: 0.48500001430511475 and parameters: {'hidden_dim': 17, 'lr': 0.005722162043271019, 'weight_decay': 0.0003712595876989976}. Best is trial 36 with value: 0.5049999952316284.
[I 2022-08-25 17:46:54,125] Trial 41 pruned.
[I 2022-08-25 17:46:54,627] Trial 42 pruned.
[I 2022-08-25 17:46:55,069] Trial 43 pruned.
[I 2022-08-25 17:46:55,541] Trial 44 pruned.
[I 2022-08-25 17:46:57,467] Trial 45 finished with value: 0.5 and parameters: {'hidden_dim': 18, 'lr': 0.009996814276559166, 'weight_decay': 0.00030144984469652667}. Best is trial 36 with value: 0.5049999952316284.
[I 2022-08-25 17:46:58,015] Trial 46 pruned.
[I 2022-08-25 17:46:58,499] Trial 47 pruned.
[I 2022-08-25 17:46:58,970] Trial 48 pruned.
[I 2022-08-25 17:46:59,430] Trial 49 pruned.
[I 2022-08-25 17:46:59,483] Best trial:
[I 2022-08-25 17:46:59,483]     Value: 0.505
[I 2022-08-25 17:46:59,483]     Params:
[I 2022-08-25 17:46:59,484]             hidden_dim |-> 17
[I 2022-08-25 17:46:59,484]             lr |-> 0.009648904316000167
[I 2022-08-25 17:46:59,484]             weight_decay |-> 0.00013498962749734303
[I 2022-08-25 17:46:59,496] Final test results:
[I 2022-08-25 17:46:59,496]     accuracy |-> 0.526
[I 2022-08-25 17:46:59,497]     f1_score |-> 0.402
[I 2022-08-25 17:46:59,497]     f1_score -> average@micro |-> 0.526