fake_news-bi_gru.py

from tensorflow.keras import optimizers
from keras.callbacks import EarlyStopping
from tensorflow.keras.layers import Dense, BatchNormalization, Reshape, Activation
from tensorflow.keras.layers import Embedding, GRU, Bidirectional
from tensorflow.keras import Sequential
# Model compilation params
compile_hp = dict()
compile_hp["loss"] = "binary_crossentropy"
compile_hp["optimizer"] = optimizers.Adam(learning_rate = 0.001)
compile_hp["metrics"] = ["accuracy"]
compile_hp["maxlen"] = 500

# Model fitting params
fit_hp = dict()
fit_hp["batch_size"] = 64
fit_hp["epochs"] = 100
fit_hp["validation_split"] = 0.3
# Create callback to select the best model
fit_hp["callbacks"] = EarlyStopping(monitor = "val_loss",
                                         mode = "min",
                                         restore_best_weights = True,
                                         patience = 10)

def bi_gru(loss = "binary_crossentropy",
                optimizer = "adam",
               metrics = ["accuracy"], 
               batch_normalize = False,
               embedding = None,
               maxlen = 500,
               hidden_dense_units = 256,
               dense_kernel_initializer = "glorot_uniform",
               rnn_units = 32,
               rnn_kernel_initializer = "glorot_uniform"):

    # Build model
    model = Sequential(name = "GRU")

    # Add embedding if desired
    if embedding:
        # Embedding contains input shape
        model.add(embedding)
    else:
        # Otherwise reshape data to work with GRU
        model.add(Reshape((maxlen, 1), input_shape = (maxlen, ), name = "Reshaping"))

    # Add GRU
    model.add(Bidirectional(GRU(rnn_units,
                                kernel_initializer = rnn_kernel_initializer),
                                name = "Bidirectional_GRU"))

    # Baseline model
    model.add(Dense(hidden_dense_units, name = "Linear_Dense",
                    kernel_initializer = dense_kernel_initializer))

    # Batch normalised model
    if batch_normalize:
        model.add(BatchNormalization(name = "Batch_Norm1"))

    # Apply non-linear activation, specified in this way to be consistent
    # with the original paper
    model.add(Activation("relu", name = "ReLU_Activation"))

    # Output layer
    model.add(Dense(1, activation = "sigmoid", name = "Output",
                    kernel_initializer = dense_kernel_initializer))
    # Compile model
    model.compile(loss = loss, optimizer = optimizer,
                  metrics = metrics)

    return model

# Set embedding
embedding_layer = "keras"
# Toggle batch normalization
batch_normalize = True

# Build and fit model with embedding
model = bi_gru(**compile_hp, batch_normalize=batch_normalize,
                  embedding = embed_dict[embedding_layer])
model.summary()
history = model.fit(X_train, y_train, **fit_hp)