def optimize_model(args):
print(args)
datasets_helper = Dataset_Helper(False)
datasets_helper.set_wanted_datasets([args['dataset_num']])
datasets_helper.next_dataset()
tokenizer = Tokenizer(num_words=args['num_of_words'])
generator = datasets_helper.text_generator()
tokenizer.fit_on_texts(generator)
optimizer = create_optimizer(args['optimizer'], args['learning_rate'])
model = resolve_network_type(args['network_type'])
model.set_params(args)
model.optimizer = optimizer
if args['network_type'] == 'embedding':
model.tokenizer = tokenizer
model.compile_model()
model.fit(datasets_helper=datasets_helper,
tokenizer=tokenizer,
validation_count=500)
results = model.evaluate(datasets_helper=datasets_helper,
tokenizer=tokenizer)
print(results)
args['results_saver'].write_any(
'logs', [get_important_params_from_args(results[1], args)], 'a')
del model
del tokenizer
del generator
del datasets_helper
tf.compat.v2.keras.backend.clear_session()
return -np.amax(results[1])
preprocess = True
datasets_helper = Dataset_Helper(preprocess)
results_saver = LogWriter(log_file_desc=simpledialog.askstring(
title="Test Name", prompt="Insert test name:", initialvalue='CONV_GRU_'))
results = []
num_of_words = 10000
while datasets_helper.next_dataset():
results_saver.add_log("Starting testing dataset {}".format(
datasets_helper.get_dataset_name()))
validation_count = datasets_helper.get_num_of_train_texts() // 10
tokenizer = Tokenizer(num_words=num_of_words,
filters='#$%&()*+-<=>@[\\]^_`{|}~\t\n',
lower=False,
split=' ')
generator = datasets_helper.text_generator()
results_saver.add_log("Starting preprocessing and tokenization.")
tokenizer.fit_on_texts(generator)
results_saver.add_log("Done. Building model now.")
batch_size = 256
gauss_noise = 0.5
epochs = 1
val_split = 0.2
val_data_count = int(datasets_helper.get_num_of_train_texts() * val_split)
model = Sequential()
model.add(Conv1D(40, 1, activation='relu', input_shape=(1, num_of_words)))
model.add(MaxPooling1D(1))
model.add(Conv1D(40, 1, activation='relu'))
model.add(GRU(40, dropout=0.1, recurrent_dropout=0.5))
model.add(Dense(datasets_helper.get_num_of_topics(), activation='softmax'))