如何在卷积网络中找到最优超参数？

Question

我知道scikit-optimize 包，我对贝叶斯优化比较陌生，我想在我当前的卷积神经网络中使用它。但是，我尝试使用Bayesian-optimization 找到卷积神经网络的最佳超参数，但我目前的尝试无法正常工作。

到目前为止，我试图为此目的提出实现，但我的代码无法正常工作，我不知道我的代码的哪一部分仍然存在问题。谁能指出我如何做到这一点？为了找到最佳超参数，在卷积神经网络上使用贝叶斯优化是否有任何有效的实现？有什么可能的想法吗？

更新

我为我的卷积神经网络尝试了GridSearchCV、RandomSearchCV，它的层非常深，使用GridSearchCV 花费了太多时间来完成，即使 2-3 天也无法完成优化。我想使用新的优化框架，如贝叶斯优化（即skopt、optuna）来寻找卷积神经网络的最佳参数和超参数。任何人都可以为my current attempt 1 in colab 和my attempt 2 in colab 提供可能的补救措施和有效方法吗？有什么想法吗？

我目前的尝试：

这是我目前使用scikit-optimize 包进行贝叶斯优化的尝试。这是my attempt in this colab，在那里我进行了所有在卷积神经网络上实施贝叶斯优化的实验，以找到它的最佳超参数：

### function returned to Bayesian Optimization

@use_named_args(dimensions=dimensions)
def bayes_opt(cnn_num_steps, cnn_init_epoch, cnn_max_epoch,
              cnn_learning_rate_decay, cnn_batch_size, cnn_dropout_rate, cnn_init_learning_rate):

    global  iteration, num_steps, init_epoch, max_epoch, learning_rate_decay, dropout_rate, init_learning_rate, batch_size

    num_steps = np.int32(cnn_num_steps)
    batch_size = np.int32(cnn_batch_size)
    learning_rate_decay = np.float32(cnn_learning_rate_decay)
    init_epoch = np.int32(cnn_init_epoch)
    max_epoch = np.int32(cnn_max_epoch)
    dropout_rate = np.float32(cnn_dropout_rate)
    init_learning_rate = np.float32(cnn_init_learning_rate)

    tf.reset_default_graph()
    tf.set_random_seed(randomState)
    sess = tf.Session()

    (train_X, train_y), (test_X, test_y) = cifar10.load_data()
    train_X = train_X.astype('float32') / 255.0
    test_X = test_X.astype('float32') / 255.0

    targets = tf.placeholder(tf.float32, [None, input_size], name="targets")
    
    model_learning_rate = tf.placeholder(tf.float32, None, name="learning_rate")
    model_dropout_rate = tf.placeholder_with_default(0.0, shape=())
    global_step = tf.Variable(0, trainable=False)

    prediction = cnn(model_dropout_rate, model_learning_rate)

    model_learning_rate = tf.train.exponential_decay(learning_rate=model_learning_rate, global_step=global_step, decay_rate=learning_rate_decay,
                                               decay_steps=init_epoch, staircase=False)

    with tf.name_scope('loss'):
        model_loss = tf.losses.mean_squared_error(targets, prediction)

    with tf.name_scope('adam_optimizer'):
        train_step = tf.train.AdamOptimizer(model_learning_rate).minimize(model_loss,global_step=global_step)

    sess.run(tf.global_variables_initializer())

    for epoch_step in range(max_epoch):
        for batch_X, batch_y in generate_batches(train_X, train_y, batch_size):
            train_data_feed = {
                inputs: batch_X,
                targets: batch_y,
                model_learning_rate: init_learning_rate,
                model_dropout_rate: dropout_rate
            }
            sess.run(train_step, train_data_feed)

    ## how to return validation error, any idea?
    ## return validation error
    ## return val_error

my current attempt in colab 仍然存在各种问题，尚未解决。任何人都可以通过使用贝叶斯优化来寻找非常深的卷积神经网络的最佳超参数来提供可能的可行方法吗？有什么想法吗？谢谢！

Answer 1

我会建议你使用Keras Tuner 包为Bayesian Optimization。

以下只是一个小例子，说明如何实现这一目标。

from kerastuner import HyperModel, Objective
import tensorflow as tf
from kerastuner.tuners import BayesianOptimization

# Create the keras tuner model.
class MyHyperModel(HyperModel):
    
    def build(self, hp):
        model = tf.keras.Sequential()
        model.add(tf.keras.layers.Embedding(len(tokenizer.word_index) + 1, embedding_dim))
        for i in range(hp.Int('num_layers', 1, 3)):
            model.add(tf.keras.layers.Conv1D(filters=hp.Choice('num_filters', values=[32, 64], default=64),activation='relu',
                                             kernel_size=3,
                                             bias_initializer='glorot_uniform'))
            model.add(tf.keras.layers.MaxPool1D())
        
        model.add(tf.keras.layers.GlobalMaxPool1D())
        
        for i in range(hp.Int('num_layers_rnn', 1, 3)):
            model.add(tf.keras.layers.Dense(units=hp.Int('units', min_value=32, max_value=512, step=32), activation='relu'))
            model.add(tf.keras.layers.Dropout(0.2))
        
        model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
        
        model.compile(
            optimizer=hp.Choice('optimizer', values= ['Adam', 'Adadelta', 'Adamax']),
            loss='binary_crossentropy',
            metrics=[f1])
        return model

然后一旦创建，您就可以通过以下代码开始模型的训练。

hypermodel = MyHyperModel()

tuner = BayesianOptimization(
    hypermodel,
    objective=Objective('val_f1', direction="max"),
    num_initial_points=50,
    max_trials=15,
    directory='./',
    project_name='real_or_not')

tuner.search(train_dataset,
             epochs=10, validation_data=validation_dataset)

您可以查看此link 的文档。 我还附上了一个指向 Kaggle Notebook 的链接，它演示了我自己编写的 Bayesian Optimization。我附上链接，以便您可以实际尝试该示例。如有任何其他问题，请随时提出。

更新：16/08

您评论说您希望使用Bayesian Optimization 调整以下超参数。我将通过以下方式解决问题。

import tensorflow as tf
from kerastuner import HyperModel, Objective
from kerastuner.tuners import BayesianOptimization

class MyHyperModel(HyperModel):

def build(self, hp):
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Conv2D(filters=hp.Choice('num_filters', values=[32, 64], default=64),activation='relu',
                                         kernel_size=(3,3),
                                         bias_initializer='glorot_uniform', input_shape=(32, 32, 3)))
    model.add(tf.keras.layers.MaxPooling2D())
    for i in range(hp.Int('num_layers', 1, 3)):
        model.add(tf.keras.layers.Conv2D(filters=hp.Choice('num_filters', values=[32, 64], default=64),activation='relu',
                                         kernel_size=(3,3),
                                         bias_initializer='glorot_uniform'))
        model.add(tf.keras.layers.MaxPooling2D())
    
    model.add(tf.keras.layers.Flatten())
    
    for i in range(hp.Int('num_layers_rnn', 1, 3)):
        model.add(tf.keras.layers.Dense(units=hp.Int('units', min_value=32, max_value=512, step=32), activation='relu'))
        model.add(tf.keras.layers.Dropout(rate=hp.Choice('droup_out_rate', values=[0.2, 0.4, 0.5], default=0.2)))
    
    model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
    
    model.compile(
        optimizer=tf.keras.optimizers.Adam(
        hp.Choice('learning_rate', [1e-2, 1e-3, 1e-4])),
        loss='binary_crossentropy',
        metrics=['accuracy'])
    return model


class MyTuner(BayesianOptimization):
  def run_trial(self, trial, *args, **kwargs):
    # You can add additional HyperParameters for preprocessing and custom training loops
    # via overriding `run_trial`
    kwargs['batch_size'] = trial.hyperparameters.Int('batch_size', 32, 256, step=32)
    kwargs['epochs'] = trial.hyperparameters.Int('epochs', 10, 30)
    super(MyTuner, self).run_trial(trial, *args, **kwargs)

hypermodel = MyHyperModel()

tuner = MyTuner(
    hypermodel,
    objective=Objective('val_acc', direction="max"),
    num_initial_points=50,
    max_trials=15,
    directory='./',
    project_name='cnn_bayesian_opt')

tuner.search(train_dataset, validation_data=validation_dataset)

您还可以查看 github 问题，解释如何调整 epochs 和 batch_size here。

以上代码将根据您的要求调整以下参数。

1. number_of_convolutional_filter
2. number_of_hidden_layer
3. drop_rate
4. learning_rate
5. batch_size
6. epochs

Answer 2

Ax platform 是在深度神经网络上使用贝叶斯优化的非常强大的工具。这是我使用ax 的方法如下：

构建 CNN 模型

!pip install ax-platform 

from tensorflow.keras import  models
from ax.service.managed_loop import optimize
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.layers import Conv2D, MaxPooling2D
from tensorflow.keras.layers import Activation, Dropout, Flatten, Dense

(X_train, y_train), (X_test, y_test) = cifar10.load_data()
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)

def build_model(opt, dropout):
    model = models.Sequential()
    model.add(Conv2D(32, kernel_size=(3,3), input_shape=(32,32,3)))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Flatten())
    model.add(Dense(n_hidden))
    model.add(Activation('relu'))
    model.add(Dropout(dropout))
    model.add(Dense(10))
    model.add(Activation('softmax'))
    model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
    return model

训练 CNN 模型

下一步是训练 CNN 模型并返回将用于贝叶斯优化的准确率：

def train_evaluate(param):
    acc = 0
    mymodel = build_model(opt=param["opt"], dropout=param["dropout"])
    mymodel.fit(X_train, y_train, epochs=param["epochs"], batch_size=param["batch_size"],verbose=1, validation_data=[X_test, y_test])
    acc = mymodel.evaluate(X_test, y_test)[1]
    print(param, acc)
    del mymodel
    return acc

运行贝叶斯优化

best_parameters, values, experiment, model = optimize(
     parameters=[
                 {"name": "opt", "type": "choice", "values": ['adam', 'rmsprop', 'sgd']},
                 {"name": "dropout", "type": "choice", "values": [0.0, 0.25, 0.50, 0.75, 0.99]},
                 {"name": "epochs", "type": "choice", "values": [10, 50, 100]},
                 {"name": "batch_size", "type": "choice", "values": [32,64, 100, 128]}
                ],
    evaluation_function=train_evaluate,
    objective_name="acc",
    total_trials=10,
    )

返回最佳参数

data = experiment.fetch_data()
df = data.df
best_arm_name = df.arm_name[df["mean"] == df["mean"].max()].values[0]
best_arm = experiment.arms_by_name[best_arm_name]

print(best_parameters)
print(best_arm)

请注意，您可以添加其他要优化的参数，例如learning_rate、num_hidden_layer，其方式与我在上面展示的方式相同。我希望这可以满足您的需要。如果您还有其他问题，请告诉我。祝你好运！