带门控残差和变量选择网络的分类

import math
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

# Column names.
CSV_HEADER = [
    "age",
    "class_of_worker",
    "detailed_industry_recode",
    "detailed_occupation_recode",
    "education",
    "wage_per_hour",
    "enroll_in_edu_inst_last_wk",
    "marital_stat",
    "major_industry_code",
    "major_occupation_code",
    "race",
    "hispanic_origin",
    "sex",
    "member_of_a_labor_union",
    "reason_for_unemployment",
    "full_or_part_time_employment_stat",
    "capital_gains",
    "capital_losses",
    "dividends_from_stocks",
    "tax_filer_stat",
    "region_of_previous_residence",
    "state_of_previous_residence",
    "detailed_household_and_family_stat",
    "detailed_household_summary_in_household",
    "instance_weight",
    "migration_code-change_in_msa",
    "migration_code-change_in_reg",
    "migration_code-move_within_reg",
    "live_in_this_house_1_year_ago",
    "migration_prev_res_in_sunbelt",
    "num_persons_worked_for_employer",
    "family_members_under_18",
    "country_of_birth_father",
    "country_of_birth_mother",
    "country_of_birth_self",
    "citizenship",
    "own_business_or_self_employed",
    "fill_inc_questionnaire_for_veterans_admin",
    "veterans_benefits",
    "weeks_worked_in_year",
    "year",
    "income_level",
]

data_url = "https://archive.ics.uci.edu/ml/machine-learning-databases/census-income-mld/census-income.data.gz"
data = pd.read_csv(data_url, header=None, names=CSV_HEADER)

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

print(f"Data shape: {data.shape}")
print(f"Test data shape: {test_data.shape}")

data["income_level"] = data["income_level"].apply(
    lambda x: 0 if x == " - 50000." else 1
)
test_data["income_level"] = test_data["income_level"].apply(
    lambda x: 0 if x == " - 50000." else 1
)

random_selection = np.random.rand(len(data.index)) <= 0.85
train_data = data[random_selection]
valid_data = data[~random_selection]

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

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

# Target feature name.
TARGET_FEATURE_NAME = "income_level"
# Weight column name.
WEIGHT_COLUMN_NAME = "instance_weight"
# Numeric feature names.
NUMERIC_FEATURE_NAMES = [
    "age",
    "wage_per_hour",
    "capital_gains",
    "capital_losses",
    "dividends_from_stocks",
    "num_persons_worked_for_employer",
    "weeks_worked_in_year",
]
# Categorical features and their vocabulary lists.
# Note that we add 'v=' as a prefix to all categorical feature values to make
# sure that they are treated as strings.
CATEGORICAL_FEATURES_WITH_VOCABULARY = {
    feature_name: sorted([str(value) for value in list(data[feature_name].unique())])
    for feature_name in CSV_HEADER
    if feature_name
    not in list(NUMERIC_FEATURE_NAMES + [WEIGHT_COLUMN_NAME, TARGET_FEATURE_NAME])
}
# All features names.
FEATURE_NAMES = NUMERIC_FEATURE_NAMES + list(
    CATEGORICAL_FEATURES_WITH_VOCABULARY.keys()
)
# Feature default values.
COLUMN_DEFAULTS = [
    [0.0]
    if feature_name in NUMERIC_FEATURE_NAMES + [TARGET_FEATURE_NAME, WEIGHT_COLUMN_NAME]
    else ["NA"]
    for feature_name in CSV_HEADER
]

from tensorflow.keras.layers import StringLookup


def process(features, target):
    for feature_name in features:
        if feature_name in CATEGORICAL_FEATURES_WITH_VOCABULARY:
            # Cast categorical feature values to string.
            features[feature_name] = tf.cast(features[feature_name], tf.dtypes.string)
    # Get the instance weight.
    weight = features.pop(WEIGHT_COLUMN_NAME)
    return features, target, weight


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,
        shuffle=shuffle,
    ).map(process)

    return dataset

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=tf.float32
            )
        else:
            inputs[feature_name] = layers.Input(
                name=feature_name, shape=(), dtype=tf.string
            )
    return inputs

def encode_inputs(inputs, encoding_size):
    encoded_features = []
    for feature_name in inputs:
        if feature_name in CATEGORICAL_FEATURES_WITH_VOCABULARY:
            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.
            index = StringLookup(
                vocabulary=vocabulary, mask_token=None, num_oov_indices=0
            )
            # Convert the string input values into integer indices.
            value_index = index(inputs[feature_name])
            # Create an embedding layer with the specified dimensions
            embedding_ecoder = layers.Embedding(
                input_dim=len(vocabulary), output_dim=encoding_size
            )
            # Convert the index values to embedding representations.
            encoded_feature = embedding_ecoder(value_index)
        else:
            # Project the numeric feature to encoding_size using linear transformation.
            encoded_feature = tf.expand_dims(inputs[feature_name], -1)
            encoded_feature = layers.Dense(units=encoding_size)(encoded_feature)
        encoded_features.append(encoded_feature)
    return encoded_features

class GatedLinearUnit(layers.Layer):
    def __init__(self, units):
        super().__init__()
        self.linear = layers.Dense(units)
        self.sigmoid = layers.Dense(units, activation="sigmoid")

    def call(self, inputs):
        return self.linear(inputs) * self.sigmoid(inputs)

class GatedResidualNetwork(layers.Layer):
    def __init__(self, units, dropout_rate):
        super().__init__()
        self.units = units
        self.elu_dense = layers.Dense(units, activation="elu")
        self.linear_dense = layers.Dense(units)
        self.dropout = layers.Dropout(dropout_rate)
        self.gated_linear_unit = GatedLinearUnit(units)
        self.layer_norm = layers.LayerNormalization()
        self.project = layers.Dense(units)

    def call(self, inputs):
        x = self.elu_dense(inputs)
        x = self.linear_dense(x)
        x = self.dropout(x)
        if inputs.shape[-1] != self.units:
            inputs = self.project(inputs)
        x = inputs + self.gated_linear_unit(x)
        x = self.layer_norm(x)
        return x

class VariableSelection(layers.Layer):
    def __init__(self, num_features, units, dropout_rate):
        super().__init__()
        self.grns = list()
        # Create a GRN for each feature independently
        for idx in range(num_features):
            grn = GatedResidualNetwork(units, dropout_rate)
            self.grns.append(grn)
        # Create a GRN for the concatenation of all the features
        self.grn_concat = GatedResidualNetwork(units, dropout_rate)
        self.softmax = layers.Dense(units=num_features, activation="softmax")

    def call(self, inputs):
        v = layers.concatenate(inputs)
        v = self.grn_concat(v)
        v = tf.expand_dims(self.softmax(v), axis=-1)

        x = []
        for idx, input in enumerate(inputs):
            x.append(self.grns[idx](input))
        x = tf.stack(x, axis=1)

        outputs = tf.squeeze(tf.matmul(v, x, transpose_a=True), axis=1)
        return outputs

def create_model(encoding_size):
    inputs = create_model_inputs()
    feature_list = encode_inputs(inputs, encoding_size)
    num_features = len(feature_list)

    features = VariableSelection(num_features, encoding_size, dropout_rate)(
        feature_list
    )

    outputs = layers.Dense(units=1, activation="sigmoid")(features)
    model = keras.Model(inputs=inputs, outputs=outputs)
    return model

learning_rate = 0.001
dropout_rate = 0.15
batch_size = 265
num_epochs = 20
encoding_size = 16

model = create_model(encoding_size)
model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=learning_rate),
    loss=keras.losses.BinaryCrossentropy(),
    metrics=[keras.metrics.BinaryAccuracy(name="accuracy")],
)


# Create an early stopping callback.
early_stopping = tf.keras.callbacks.EarlyStopping(
    monitor="val_loss", patience=5, restore_best_weights=True
)

print("Start training the model...")
train_dataset = get_dataset_from_csv(
    train_data_file, shuffle=True, batch_size=batch_size
)
valid_dataset = get_dataset_from_csv(valid_data_file, batch_size=batch_size)
model.fit(
    train_dataset,
    epochs=num_epochs,
    validation_data=valid_dataset,
    callbacks=[early_stopping],
)
print("Model training finished.")

print("Evaluating model performance...")
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)}%")