神经网络参数调优遇RuntimeError：layers参数无法克隆问题求助

Hey there, let's break down why this error is popping up and how to fix it step by step.

The Root Cause #

GridSearchCV works by cloning your model instance for every hyperparameter combination in your param_grid. For this cloning to succeed, your model class must store all the hyperparameters you're tuning as instance attributes (like self.layers, self.activation), and these attributes must match exactly what's passed to the constructor. If your model doesn't save layers as an instance attribute, or modifies it without updating the instance, GridSearchCV can't recreate the model and throws this error.

Step-by-Step Fixes #

1. Ensure Your Model Saves All Tuned Parameters as Instance Attributes

First, check your custom model's __init__ method. You need to explicitly assign every hyperparameter you're tuning to a self. attribute. For example:

Bad (missing instance attribute):

class MyNeuralNet(Model):
    def __init__(self, layers, activation='relu'):
        super().__init__()
        # Oops! We didn't save layers as self.layers
        self._build_network(layers, activation)

Good (saves parameters properly):

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Input

class MyNeuralNet(Model):
    def __init__(self, layers=[20], activation='relu', batch_size=128, epochs=30):
        super().__init__()
        # Save ALL tuned parameters as instance attributes
        self.layers = layers
        self.activation = activation
        self.batch_size = batch_size
        self.epochs = epochs
        self.trained_model = None

    def _build_network(self):
        # Use the saved instance attributes to build the network
        input_layer = Input(shape=(X_train.shape[1],))
        x = input_layer
        for units in self.layers:
            x = Dense(units, activation=self.activation)(x)
        output_layer = Dense(1, activation='sigmoid')(x)  # Adjust based on your task
        self.trained_model = Model(inputs=input_layer, outputs=output_layer)
        self.trained_model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

    def fit(self, X, y):
        if self.trained_model is None:
            self._build_network()
        self.trained_model.fit(X, y, batch_size=self.batch_size, epochs=self.epochs, verbose=0)
        return self

    def predict(self, X):
        return self.trained_model.predict(X)

2. Inherit from Scikit-Learn's BaseEstimator (Optional but Recommended)

If you're using a custom model, inheriting from sklearn.base.BaseEstimator and ClassifierMixin/RegressorMixin will handle most of the parameter cloning logic automatically. This ensures GridSearchCV can properly access and set your model's parameters:

from sklearn.base import BaseEstimator, ClassifierMixin

class MyNeuralNet(BaseEstimator, ClassifierMixin, Model):
    def __init__(self, layers=[20], activation='relu', batch_size=128, epochs=30):
        super().__init__()
        self.layers = layers
        self.activation = activation
        self.batch_size = batch_size
        self.epochs = epochs
        self.trained_model = None

    # ... keep the _build_network, fit, predict methods from above ...

3. Verify Parameter Names Match Between param_grid and Model Constructor

Double-check that every key in your param_grid exactly matches the parameter name in your model's __init__ method. From your code, param_grid uses layers which looks correct, but typos (like layer instead of layers) are easy to make—so confirm this matches perfectly.

4. Avoid Modifying Parameters Without Updating Instance Attributes

If your model modifies a parameter (e.g., converting a list to a tuple) in __init__, make sure you save the original parameter value (not the modified one) as the instance attribute. GridSearchCV uses the original parameter values to clone the model, so modifying them can break the cloning process.

Test the Fix #

Once you've updated your model class, re-run your GridSearchCV code:

layers = [[20], [40, 20], [45, 30, 15]]
activations = ['sigmoid', 'relu']
param_grid = dict(layers=layers, activation=activations, batch_size=[128, 256], epochs=[30])

# Initialize your corrected model
model = MyNeuralNet()

grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=1, verbose=2)
grid_result = grid.fit(X_train, y_train)

This should resolve the cloning error, and GridSearchCV will be able to iterate through your hyperparameter combinations properly.

内容的提问来源于stack exchange，提问作者Madushan