获取TypeError:当试图训练神经网络回归者时，predict()得到了一个意外的关键字参数'return_std‘

Question

因此，我试图使用modAL创建一个贝叶斯优化工作流。我的任务是一个样本集上的回归任务，包含2048个特征和1个回归目标的样本。我目前正在使用圆周率获取函数，但是遇到了以下错误。

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
~\AppData\Local\Temp/ipykernel_13924/1162646393.py in <module>
     99                     pvg_row+=1
    100                 try:
--> 101                     query_idx,query_inst = learner.query(X, n_instances=20)
    102                 except AssertionError:
    103                     print("Encountered a case where the number of intances is lower than utility")

~\anaconda3\envs\fyp-al\lib\site-packages\modAL\models\base.py in query(self, X_pool, *query_args, **query_kwargs)
    251             labelled upon query synthesis.
    252         """
--> 253         query_result = self.query_strategy(self, X_pool, *query_args, **query_kwargs)
    254 
    255         if isinstance(query_result, tuple):

~\anaconda3\envs\fyp-al\lib\site-packages\modAL\acquisition.py in max_PI(optimizer, X, tradeoff, n_instances)
    118         The indices of the instances from X chosen to be labelled; the instances from X chosen to be labelled.
    119     """
--> 120     pi = optimizer_PI(optimizer, X, tradeoff=tradeoff)
    121     return multi_argmax(pi, n_instances=n_instances)
    122 

~\anaconda3\envs\fyp-al\lib\site-packages\modAL\acquisition.py in optimizer_PI(optimizer, X, tradeoff)
     47     """
     48     try:
---> 49         mean, std = optimizer.predict(X, return_std=True)
     50         mean, std = mean.reshape(-1, ), std.reshape(-1, )
     51     except NotFittedError:

~\anaconda3\envs\fyp-al\lib\site-packages\modAL\models\base.py in predict(self, X, **predict_kwargs)
    220             Estimator predictions for X.
    221         """
--> 222         return self.estimator.predict(X, **predict_kwargs)
    223 
    224     def predict_proba(self, X: modALinput, **predict_proba_kwargs) -> Any:

TypeError: predict() got an unexpected keyword argument 'return_std'

我的回归模型是

            reg = SkorchNNR(
               Torch_Model,
               lr=1e-3,
               criterion= nn.MSELoss,
               max_epochs = 50,
               batch_size = 128,
               train_split = skorch.dataset.CVSplit(5) if int(X.shape[0]*0.1)>5 else skorch.dataset.CVSplit(int(X.shape[0]*0.1)),
               optimizer =torch.optim.Adam,
               device = device,
               verbose = 1
            )

相应的Bayesian优化对象是

            learner = BayesianOptimizer(
                estimator=clf,
                query_strategy=gpr_model["acquisition_function"],
                X_training=x_initial, y_training=y_initial.reshape(-1,1)
            )

最起码的可重现性例子是：

# Importing Libraries
import numpy as np
import torch
from torch import nn
import skorch
from skorch import NeuralNetRegressor as SkorchNNR
from modAL.models import BayesianOptimizer
from modAL.acquisition import max_PI
import numpy as np

# Generating Dataset
row,col = 50,2048
X = np.random.randint(2, size=(row,col)).astype(np.float32)
y = np.random.randn(row,1).astype(np.float32) ** 2

# Picking out initial training samples
train_idx = np.random.choice(range(X.shape[0]), size=int(X.shape[0]*0.2), replace=False)

x_initial = X[train_idx]
y_initial = y[train_idx]

X = np.delete(X,train_idx, axis=0)
y = np.delete(y,train_idx)

# Displaying Shape after initial selection
print(X.shape, y.shape)
print(x_initial.shape, y_initial.shape)

# PyTorch Model
class Torch_Model_BasicRegressor(nn.Module):
    def __init__(self):
        super(Torch_Model_BasicRegressor, self).__init__()
        self.fcs = nn.Sequential(
            nn.Linear(2048,256),
            nn.ReLU(),
            nn.Dropout(0.25),
            nn.Linear(256,32),
            nn.ReLU(),
            nn.Dropout(0.25),
            nn.Linear(32,4),
            nn.ReLU(),
            nn.Dropout(0.25),
            nn.Linear(4,1)
        )
    def forward(self, x):
        out = self.fcs(x)
        return out

# SKORCH model
device = "cuda" if torch.cuda.is_available()==True else "cpu"
regressor = SkorchNNR(
    Torch_Model_BasicRegressor,
    lr=1e-5,
    criterion= nn.MSELoss,
    max_epochs = 25,
    batch_size = 128,
    train_split = skorch.dataset.ValidSplit(5) if int(X.shape[0]*0.2)>5 else skorch.dataset.ValidSplit(int(X.shape[0]*0.2)),
    optimizer =torch.optim.Adam,
    device = device,
    verbose = 1
)

# modAL model
learner = BayesianOptimizer(
    estimator=regressor,
    query_strategy=max_PI,
    X_training=x_initial, y_training=y_initial.reshape(-1,1)
)

# Querying & Teaching the model : Error over here
for n_query in range(15):
    query_idx,query_inst = learner.query(X, n_instances=5)
    learner.teach(X=query_inst,y=y[query_idx])

Answer 1

我看到您正在BayesianOptimizer估计器参数中声明一个分类器(clf)。正如您在示例中所看到的，估计器参数是一个回归变量。在你的例子中，它应该是你的模型'SkorchNNR‘。