def vdcnn_model(X_train, y_train, X_test, y_test, sequence_max_length):
model = VDCNN(num_classes=y_train.shape[1],
sequence_length=sequence_max_length)
if y_train.shape[1] == 1:
model.compile(loss='binary_crossentropy',
optimizer=SGD(lr=0.0001, momentum=0.9),
metrics=['accuracy'])
else:
model.compile(loss='categorical_crossentropy',
optimizer=SGD(lr=0.0001, momentum=0.9),
metrics=['accuracy'])
model.fit(X_train,
y_train,
validation_data=(X_test, y_test),
epochs=50,
batch_size=128)
scores = model.evaluate(X_test, y_test)
accuracy = scores[1] * 100
y_pred = model.predict(X_test)
return accuracy, y_pred
def train(x_train, y_train, x_test, y_test):
session_ts = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
log_dir = str(pathlib.Path(FLAGS.train_log_dir) / session_ts)
if FLAGS.dataset_type == "embeddings":
embedding_input = True
embedding_dim = x_train.shape[-1]
else:
embedding_input = False
embedding_dim = 16
# Build model
model = VDCNN(
num_classes=y_train.shape[1],
depth=FLAGS.depth,
sequence_length=FLAGS.sequence_length,
shortcut=FLAGS.shortcut,
pool_type=FLAGS.pool_type,
sort=FLAGS.sort,
use_bias=FLAGS.use_bias,
embedding_input=embedding_input,
embedding_dim=embedding_dim,
)
model.compile(
optimizer=tf.keras.optimizers.SGD(lr=FLAGS.lr, momentum=0.9),
loss="categorical_crossentropy",
metrics=["acc"],
)
# Save model architecture
model_json = model.to_json()
with open("vdcnn_model.json", "w") as json_file:
json_file.write(model_json)
time_str = datetime.datetime.now().isoformat()
print("{}: Model saved as json.".format(time_str))
print("")
# Trainer
# Tensorboard and extra callback to support steps history
tensorboard = tf.keras.callbacks.TensorBoard(log_dir=log_dir,
histogram_freq=50,
write_graph=True,
write_images=True)
checkpointer = tf.keras.callbacks.ModelCheckpoint(
filepath="./checkpoints/vdcnn_weights_val_acc_{val_acc:.4f}.h5",
save_freq="epoch",
verbose=1,
save_best_only=True,
mode="max",
monitor="val_acc",
)
loss_history = custom_callbacks.LossHistory(model,
tensorboard,
logdir=log_dir)
evaluate_step = custom_callbacks.EvaluateStep(
model,
checkpointer,
tensorboard,
FLAGS.evaluate_every,
FLAGS.batch_size,
x_test,
y_test,
log_dir,
)
# Fit model
model.fit(
x_train,
y_train,
batch_size=FLAGS.batch_size,
epochs=FLAGS.num_epochs,
validation_data=(x_test, y_test),
verbose=1,
callbacks=[
checkpointer,
tensorboard,
# loss_history,
evaluate_step,
],
)
print("-" * 30)
time_str = datetime.datetime.now().isoformat()
print("{}: Done training.".format(time_str))
tf.keras.backend.clear_session()
print("-" * 30)
print()
