train_risk_classifier.py

import os
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, Model
import argparse
from datetime import datetime

# Define the function to create the multiple instance learning (MIL) model
def create_simple_model2(instance_shape, max_length):
    inputs = layers.Input(shape=(max_length, instance_shape[-1]), name="bag_input")
    flatten = layers.TimeDistributed(layers.Flatten())(inputs)
    dense_1 = layers.TimeDistributed(layers.Dense(256, activation="relu"))(flatten)
    dropout_1 = layers.TimeDistributed(layers.Dropout(0.5))(dense_1)
    dense_2 = layers.TimeDistributed(layers.Dense(64, activation="relu"))(dropout_1)
    dropout_2 = layers.TimeDistributed(layers.Dropout(0.5))(dense_2)
    aggregated = layers.GlobalAveragePooling1D()(dropout_2)
    norm_1 = layers.LayerNormalization()(aggregated)
    output = layers.Dense(1, activation="sigmoid")(norm_1)
    return Model(inputs, output)

def create_simple_model(instance_shape, max_length, num_heads=4, key_dim=64):
    inputs = layers.Input(shape=(max_length, instance_shape[-1]), name="bag_input")
    flatten = layers.TimeDistributed(layers.Flatten())(inputs)
    dense_1 = layers.TimeDistributed(layers.Dense(256, activation="relu"))(flatten)
    dropout_1 = layers.TimeDistributed(layers.Dropout(0.5))(dense_1)
    dense_2 = layers.TimeDistributed(layers.Dense(64, activation="relu"))(dropout_1)
    dropout_2 = layers.TimeDistributed(layers.Dropout(0.5))(dense_2)
    attention_output, attention_scores = layers.MultiHeadAttention(
        num_heads=num_heads,
        key_dim=key_dim,
        value_dim=64,
        dropout=0.1,
        use_bias=True
    )(query=dropout_2, value=dropout_2, key=dropout_2, return_attention_scores=True)
    aggregated = layers.GlobalAveragePooling1D()(attention_output)
    norm_1 = layers.LayerNormalization()(aggregated)
    output = layers.Dense(1, activation="sigmoid")(norm_1)
    return Model(inputs, output)

# Function to compute class weights
def compute_class_weights(labels):
    negative_count = len(np.where(labels == 0)[0])
    positive_count = len(np.where(labels == 1)[0])
    total_count = negative_count + positive_count
    return {0: (1 / negative_count) * (total_count / 2), 1: (1 / positive_count) * (total_count / 2)}

# Function to generate batches of data
def data_generator(data, labels, batch_size=1):
    class_weights = compute_class_weights(labels)
    while True:
        for i in range(0, len(data), batch_size):
            batch_data = np.array(data[i:i + batch_size], dtype=np.float32)
            batch_labels = np.array(labels[i:i + batch_size], dtype=np.float32)
            batch_weights = np.array([class_weights[int(label)] for label in batch_labels], dtype=np.float32)
            yield batch_data, batch_labels, batch_weights

# Learning rate scheduler
def lr_scheduler(epoch, lr):
    decay_rate = 0.1
    decay_step = 10
    if epoch % decay_step == 0 and epoch:
        return lr * decay_rate
    return lr

# Function to train the model
def train(train_data, train_labels, val_data, val_labels, model, save_dir):
    model_path = os.path.join(save_dir, "risk_classifier_model.h5")
    model_checkpoint = tf.keras.callbacks.ModelCheckpoint(model_path, monitor="val_loss", verbose=1, mode="min", save_best_only=True, save_weights_only=False)
    early_stopping = tf.keras.callbacks.EarlyStopping(monitor="val_loss", patience=10, mode="min")
    lr_callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler)
    model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy", "AUC"])
    train_gen = data_generator(train_data, train_labels)
    val_gen = data_generator(val_data, val_labels)
    model.fit(train_gen, steps_per_epoch=len(train_data), validation_data=val_gen, validation_steps=len(val_data), epochs=50, batch_size=1, callbacks=[early_stopping, model_checkpoint, lr_callback], verbose=1)
    return model

if __name__ == "__main__":
    # Command line arguments
    parser = argparse.ArgumentParser(description='Train a multiple instance learning classifier on risk data.')
    parser.add_argument('--data_file', type=str, required=True, help='Path to the saved .npz file with training and validation data.')
    parser.add_argument('--save_dir', type=str, default='./model_save/', help='Directory to save the model.')
    parser.add_argument('--epochs', type=int, default=50, help='Number of training epochs.')

    args = parser.parse_args()

    if not os.path.exists(args.save_dir):
        os.makedirs(args.save_dir)

    # Load the preprocessed data
    data = np.load(args.data_file)
    train_X, train_Y = data['train_X'], data['train_Y']
    validate_X, validate_Y = data['validate_X'], data['validate_Y']

    # Create the model
    instance_shape = (train_X.shape[-1],)
    max_length = train_X.shape[1]
    model = create_simple_model(instance_shape, max_length)

    # Train the model
    trained_model = train(train_X, train_Y, validate_X, validate_Y, model, args.save_dir)

    # Final message after training and saving the model
    print(f"Model saved successfully to {os.path.join(args.save_dir, 'risk_classifier_model.h5')}")