政安晨：【Keras机器学习示例演绎】（五十四）—— 使用神经决策森林进行分类

import keras
from keras import layers
from keras.layers import StringLookup
from keras import ops


from tensorflow import data as tf_data
import numpy as np
import pandas as pd

import math

准备数据

CSV_HEADER = [
    "age",
    "workclass",
    "fnlwgt",
    "education",
    "education_num",
    "marital_status",
    "occupation",
    "relationship",
    "race",
    "gender",
    "capital_gain",
    "capital_loss",
    "hours_per_week",
    "native_country",
    "income_bracket",
]

train_data_url = (
    "https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data"
)
train_data = pd.read_csv(train_data_url, header=None, names=CSV_HEADER)

test_data_url = (
    "https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.test"
)
test_data = pd.read_csv(test_data_url, header=None, names=CSV_HEADER)

print(f"Train dataset shape: {train_data.shape}")
print(f"Test dataset shape: {test_data.shape}")

Train dataset shape: (32561, 15)
Test dataset shape: (16282, 15)

删除第一条记录（因为它不是一个有效的数据示例）和类标签中的尾部 "点"。

test_data = test_data[1:]
test_data.income_bracket = test_data.income_bracket.apply(
    lambda value: value.replace(".", "")
)

我们将训练数据和测试数据分割成 CSV 文件存储在本地。

train_data_file = "train_data.csv"
test_data_file = "test_data.csv"

train_data.to_csv(train_data_file, index=False, header=False)
test_data.to_csv(test_data_file, index=False, header=False)

定义数据集元数据

在此，我们定义了数据集的元数据，这些元数据将有助于读取、解析和编码输入特征。

# A list of the numerical feature names.
NUMERIC_FEATURE_NAMES = [
    "age",
    "education_num",
    "capital_gain",
    "capital_loss",
    "hours_per_week",
]
# A dictionary of the categorical features and their vocabulary.
CATEGORICAL_FEATURES_WITH_VOCABULARY = {
    "workclass": sorted(list(train_data["workclass"].unique())),
    "education": sorted(list(train_data["education"].unique())),
    "marital_status": sorted(list(train_data["marital_status"].unique())),
    "occupation": sorted(list(train_data["occupation"].unique())),
    "relationship": sorted(list(train_data["relationship"].unique())),
    "race": sorted(list(train_data["race"].unique())),
    "gender": sorted(list(train_data["gender"].unique())),
    "native_country": sorted(list(train_data["native_country"].unique())),
}
# A list of the columns to ignore from the dataset.
IGNORE_COLUMN_NAMES = ["fnlwgt"]
# A list of the categorical feature names.
CATEGORICAL_FEATURE_NAMES = list(CATEGORICAL_FEATURES_WITH_VOCABULARY.keys())
# A list of all the input features.
FEATURE_NAMES = NUMERIC_FEATURE_NAMES + CATEGORICAL_FEATURE_NAMES
# A list of column default values for each feature.
COLUMN_DEFAULTS = [
    [0.0] if feature_name in NUMERIC_FEATURE_NAMES + IGNORE_COLUMN_NAMES else ["NA"]
    for feature_name in CSV_HEADER
]
# The name of the target feature.
TARGET_FEATURE_NAME = "income_bracket"
# A list of the labels of the target features.
TARGET_LABELS = [" <=50K", " >50K"]

为训练和验证创建 tf_data.Dataset 对象

我们创建了一个输入函数来读取和解析文件，并将特征和标签转换为 tf_data.Dataset 用于训练和验证。我们还通过将目标标签映射到索引对输入进行预处理。

target_label_lookup = StringLookup(
    vocabulary=TARGET_LABELS, mask_token=None, num_oov_indices=0
)


lookup_dict = {}
for feature_name in CATEGORICAL_FEATURE_NAMES:
    vocabulary = CATEGORICAL_FEATURES_WITH_VOCABULARY[feature_name]
    # Create a lookup to convert a string values to an integer indices.
    # Since we are not using a mask token, nor expecting any out of vocabulary
    # (oov) token, we set mask_token to None and num_oov_indices to 0.
    lookup = StringLookup(vocabulary=vocabulary, mask_token=None, num_oov_indices=0)
    lookup_dict[feature_name] = lookup


def encode_categorical(batch_x, batch_y):
    for feature_name in CATEGORICAL_FEATURE_NAMES:
        batch_x[feature_name] = lookup_dict[feature_name](batch_x[feature_name])

    return batch_x, batch_y


def get_dataset_from_csv(csv_file_path, shuffle=False, batch_size=128):
    dataset = (
        tf_data.experimental.make_csv_dataset(
            csv_file_path,
            batch_size=batch_size,
            column_names=CSV_HEADER,
            column_defaults=COLUMN_DEFAULTS,
            label_name=TARGET_FEATURE_NAME,
            num_epochs=1,
            header=False,
            na_value="?",
            shuffle=shuffle,
        )
        .map(lambda features, target: (features, target_label_lookup(target)))
        .map(encode_categorical)
    )

    return dataset.cache()

创建模型输入

def create_model_inputs():
    inputs = {}
    for feature_name in FEATURE_NAMES:
        if feature_name in NUMERIC_FEATURE_NAMES:
            inputs[feature_name] = layers.Input(
                name=feature_name, shape=(), dtype="float32"
            )
        else:
            inputs[feature_name] = layers.Input(
                name=feature_name, shape=(), dtype="int32"
            )
    return inputs

输入特征编码

def encode_inputs(inputs):
    encoded_features = []
    for feature_name in inputs:
        if feature_name in CATEGORICAL_FEATURE_NAMES:
            vocabulary = CATEGORICAL_FEATURES_WITH_VOCABULARY[feature_name]
            # Create a lookup to convert a string values to an integer indices.
            # Since we are not using a mask token, nor expecting any out of vocabulary
            # (oov) token, we set mask_token to None and num_oov_indices to 0.
            value_index = inputs[feature_name]
            embedding_dims = int(math.sqrt(lookup.vocabulary_size()))
            # Create an embedding layer with the specified dimensions.
            embedding = layers.Embedding(
                input_dim=lookup.vocabulary_size(), output_dim=embedding_dims
            )
            # Convert the index values to embedding representations.
            encoded_feature = embedding(value_index)
        else:
            # Use the numerical features as-is.
            encoded_feature = inputs[feature_name]
            if inputs[feature_name].shape[-1] is None:
                encoded_feature = keras.ops.expand_dims(encoded_feature, -1)

        encoded_features.append(encoded_feature)

    encoded_features = layers.concatenate(encoded_features)
    return encoded_features

深度神经决策树

神经决策树模型有两组权重需要学习。第一组是 pi，代表树叶中类别的概率分布。第二组是路由层 decision_fn 的权重，代表前往每个树叶的概率。该模型的前向传递工作原理如下：

该模型希望将输入特征作为一个单一的向量，对批次中某个实例的所有特征进行编码。该向量可以由应用于图像的卷积神经网络（CNN）生成，也可以由应用于结构化数据特征的密集变换生成。

模型首先应用已用特征掩码随机选择要使用的输入特征子集。

然后，模型通过在树的各个层级迭代执行随机路由，计算输入实例到达树叶的概率（mu）。

最后，将到达树叶的概率与树叶上的类概率相结合，生成最终输出。

class NeuralDecisionTree(keras.Model):
    def __init__(self, depth, num_features, used_features_rate, num_classes):
        super().__init__()
        self.depth = depth
        self.num_leaves = 2**depth
        self.num_classes = num_classes

        # Create a mask for the randomly selected features.
        num_used_features = int(num_features * used_features_rate)
        one_hot = np.eye(num_features)
        sampled_feature_indices = np.random.choice(
            np.arange(num_features), num_used_features, replace=False
        )
        self.used_features_mask = ops.convert_to_tensor(
            one_hot[sampled_feature_indices], dtype="float32"
        )

        # Initialize the weights of the classes in leaves.
        self.pi = self.add_weight(
            initializer="random_normal",
            shape=[self.num_leaves, self.num_classes],
            dtype="float32",
            trainable=True,
        )

        # Initialize the stochastic routing layer.
        self.decision_fn = layers.Dense(
            units=self.num_leaves, activation="sigmoid", name="decision"
        )

    def call(self, features):
        batch_size = ops.shape(features)[0]

        # Apply the feature mask to the input features.
        features = ops.matmul(
            features, ops.transpose(self.used_features_mask)
        )  # [batch_size, num_used_features]
        # Compute the routing probabilities.
        decisions = ops.expand_dims(
            self.decision_fn(features), axis=2
        )  # [batch_size, num_leaves, 1]
        # Concatenate the routing probabilities with their complements.
        decisions = layers.concatenate(
            [decisions, 1 - decisions], axis=2
        )  # [batch_size, num_leaves, 2]

        mu = ops.ones([batch_size, 1, 1])

        begin_idx = 1
        end_idx = 2
        # Traverse the tree in breadth-first order.
        for level in range(self.depth):
            mu = ops.reshape(mu, [batch_size, -1, 1])  # [batch_size, 2 ** level, 1]
            mu = ops.tile(mu, (1, 1, 2))  # [batch_size, 2 ** level, 2]
            level_decisions = decisions[
                :, begin_idx:end_idx, :
            ]  # [batch_size, 2 ** level, 2]
            mu = mu * level_decisions  # [batch_size, 2**level, 2]
            begin_idx = end_idx
            end_idx = begin_idx + 2 ** (level + 1)

        mu = ops.reshape(mu, [batch_size, self.num_leaves])  # [batch_size, num_leaves]
        probabilities = keras.activations.softmax(self.pi)  # [num_leaves, num_classes]
        outputs = ops.matmul(mu, probabilities)  # [batch_size, num_classes]
        return outputs

深度神经决策森林

神经决策森林模型由一组同时训练的神经决策树组成。森林模型的输出是各树的平均输出。

class NeuralDecisionForest(keras.Model):
    def __init__(self, num_trees, depth, num_features, used_features_rate, num_classes):
        super().__init__()
        self.ensemble = []
        # Initialize the ensemble by adding NeuralDecisionTree instances.
        # Each tree will have its own randomly selected input features to use.
        for _ in range(num_trees):
            self.ensemble.append(
                NeuralDecisionTree(depth, num_features, used_features_rate, num_classes)
            )

    def call(self, inputs):
        # Initialize the outputs: a [batch_size, num_classes] matrix of zeros.
        batch_size = ops.shape(inputs)[0]
        outputs = ops.zeros([batch_size, num_classes])

        # Aggregate the outputs of trees in the ensemble.
        for tree in self.ensemble:
            outputs += tree(inputs)
        # Divide the outputs by the ensemble size to get the average.
        outputs /= len(self.ensemble)
        return outputs

最后，让我们来设置训练和评估模型的代码。

learning_rate = 0.01
batch_size = 265
num_epochs = 10


def run_experiment(model):
    model.compile(
        optimizer=keras.optimizers.Adam(learning_rate=learning_rate),
        loss=keras.losses.SparseCategoricalCrossentropy(),
        metrics=[keras.metrics.SparseCategoricalAccuracy()],
    )

    print("Start training the model...")
    train_dataset = get_dataset_from_csv(
        train_data_file, shuffle=True, batch_size=batch_size
    )

    model.fit(train_dataset, epochs=num_epochs)
    print("Model training finished")

    print("Evaluating the model on the test data...")
    test_dataset = get_dataset_from_csv(test_data_file, batch_size=batch_size)

    _, accuracy = model.evaluate(test_dataset)
    print(f"Test accuracy: {round(accuracy * 100, 2)}%")

实验 1：训练决策树模型

在本实验中，我们使用所有输入特征训练一个神经决策树模型。

num_trees = 10
depth = 10
used_features_rate = 1.0
num_classes = len(TARGET_LABELS)


def create_tree_model():
    inputs = create_model_inputs()
    features = encode_inputs(inputs)
    features = layers.BatchNormalization()(features)
    num_features = features.shape[1]

    tree = NeuralDecisionTree(depth, num_features, used_features_rate, num_classes)

    outputs = tree(features)
    model = keras.Model(inputs=inputs, outputs=outputs)
    return model


tree_model = create_tree_model()
run_experiment(tree_model)

Start training the model...
Epoch 1/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 5s 26ms/step - loss: 0.5308 - sparse_categorical_accuracy: 0.8150
Epoch 2/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3476 - sparse_categorical_accuracy: 0.8429
Epoch 3/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3312 - sparse_categorical_accuracy: 0.8478
Epoch 4/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3247 - sparse_categorical_accuracy: 0.8495
Epoch 5/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3202 - sparse_categorical_accuracy: 0.8512
Epoch 6/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3158 - sparse_categorical_accuracy: 0.8536
Epoch 7/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3116 - sparse_categorical_accuracy: 0.8572
Epoch 8/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3071 - sparse_categorical_accuracy: 0.8608
Epoch 9/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 11ms/step - loss: 0.3026 - sparse_categorical_accuracy: 0.8630
Epoch 10/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.2975 - sparse_categorical_accuracy: 0.8653
Model training finished
Evaluating the model on the test data...
 62/62 ━━━━━━━━━━━━━━━━━━━━ 1s 13ms/step - loss: 0.3279 - sparse_categorical_accuracy: 0.8463
Test accuracy: 85.08%

实验 2：训练森林模型

在本实验中，我们使用 num_trees 树训练神经决策森林，每棵树随机使用 50%的输入特征。通过设置 used_features_rate 变量，可以控制每棵树使用的特征数量。此外，与之前的实验相比，我们将深度设置为 5，而不是 10。

num_trees = 25
depth = 5
used_features_rate = 0.5


def create_forest_model():
    inputs = create_model_inputs()
    features = encode_inputs(inputs)
    features = layers.BatchNormalization()(features)
    num_features = features.shape[1]

    forest_model = NeuralDecisionForest(
        num_trees, depth, num_features, used_features_rate, num_classes
    )

    outputs = forest_model(features)
    model = keras.Model(inputs=inputs, outputs=outputs)
    return model


forest_model = create_forest_model()

run_experiment(forest_model)

Start training the model...
Epoch 1/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 47s 202ms/step - loss: 0.5469 - sparse_categorical_accuracy: 0.7915
Epoch 2/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3459 - sparse_categorical_accuracy: 0.8494
Epoch 3/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3268 - sparse_categorical_accuracy: 0.8523
Epoch 4/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3195 - sparse_categorical_accuracy: 0.8524
Epoch 5/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3149 - sparse_categorical_accuracy: 0.8539
Epoch 6/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3112 - sparse_categorical_accuracy: 0.8556
Epoch 7/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 10ms/step - loss: 0.3079 - sparse_categorical_accuracy: 0.8566
Epoch 8/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 9ms/step - loss: 0.3050 - sparse_categorical_accuracy: 0.8582
Epoch 9/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 9ms/step - loss: 0.3021 - sparse_categorical_accuracy: 0.8595
Epoch 10/10
 123/123 ━━━━━━━━━━━━━━━━━━━━ 1s 9ms/step - loss: 0.2992 - sparse_categorical_accuracy: 0.8617
Model training finished
Evaluating the model on the test data...
 62/62 ━━━━━━━━━━━━━━━━━━━━ 5s 39ms/step - loss: 0.3145 - sparse_categorical_accuracy: 0.8503
Test accuracy: 85.55%