lstm_module.py

## Setup

import numpy as np
from tensorflow import keras
import tensorflow as tf
from tensorflow.keras import models
from tensorflow.keras import layers

# import optuna
# from optuna.integration import KerasPruningCallback
# from optuna.trial import TrialState

import os
import sys

my_lib_path = os.path.abspath("./")
sys.path.append(my_lib_path)

## Text preprocessing
import preprocessing_module

(
    pad_X_train,
    pad_X_val,
    pad_X_test,
    y_train,
    y_val,
    y_test,
    num_unique_words,
    max_length,
) = preprocessing_module.get_data()


def print_data():
    print(pad_X_train.shape)
    print(y_train.shape)
    print(pad_X_val.shape)
    print(y_val.shape)
    print(pad_X_test.shape)
    print(y_test.shape)
    print(num_unique_words)
    print(max_length)


# LSTM model
def LSTM(
    embedding_dim=32,
    num_units=64,
    num_classes=25,
    num_epochs=20,
    batch_size=32,
    verbosity=0,
    loss_function="sparse_categorical_crossentropy",
    optimizer="adam",
    trial=None,
):
    # create initial model
    if os.path.isdir("LSTM"):
        model = models.load_model("LSTM")
    else:
        model = keras.models.Sequential()
        model.add(
            layers.Embedding(num_unique_words, embedding_dim, input_length=max_length)
        )
        model.save("LSTM")
    if trial is not None:
        ## Optuna
        num_units = trial.suggest_int("Units number", 8, 128, log=True)
    model.add(
        layers.LSTM(
            num_units,
            dropout=0.2,
            activation="tanh",
            recurrent_activation="sigmoid",
            recurrent_dropout=0,
            unroll=False,
            use_bias=True,
        )
    )

    model.add(layers.Dense(num_classes, activation="softmax"))
    # compile model
    model.compile(loss=loss_function, optimizer=optimizer, metrics=["accuracy"])
    # fit the model
    if trial is not None:
        ## Optuna
        num_epochs = trial.suggest_int("Epochs number", 10, 50, log=True)
    model.fit(
        pad_X_train,
        y_train,
        epochs=num_epochs,
        batch_size=batch_size,
        verbose=verbosity,
        validation_data=(pad_X_val, y_val),
    )

    return model

# LSTM model
def cnn(
    embedding_dim=32,
    num_units=64,
    num_classes=25,
    num_epochs=20,
    batch_size=32,
    verbosity=0,
    loss_function="sparse_categorical_crossentropy",
    optimizer="adam",
    trial=None,
):
    # create initial model
    if os.path.isdir("CNN"):
        model = models.load_model("CNN")
    else:
        model = keras.models.Sequential()
        model.add(
            layers.Embedding(num_unique_words, embedding_dim, input_length=max_length)
        )
        model.save("CNN")
    # if trial is not None:
    #     ## Optuna
    #     num_units = trial.suggest_int("Units number", 8, 128, log=True)

    model.add(layers.Conv1D(filters=num_units, kernel_size=3, activation="relu"))
    model.add(layers.Flatten())
    model.add(layers.Dense(num_classes, activation="softmax"))
    # compile model
    model.compile(loss=loss_function, optimizer=optimizer, metrics=["accuracy"])
    # fit the model
    # if trial is not None:
    #     ## Optuna
    #     num_epochs = trial.suggest_int("Epochs number", 10, 50, log=True)
    model.fit(
        pad_X_train,
        y_train,
        epochs=num_epochs,
        batch_size=batch_size,
        verbose=verbosity,
        validation_data=(pad_X_val, y_val),
    )

    return model


def evaluate_model(model):
    # evaluate the model on test data
    scores = model.evaluate(pad_X_test, y_test, verbose=0)

    return scores[1]