start = time.time()
dataset_manager = DatasetManager(dataset_name)
data = dataset_manager.read_dataset()
train, test = dataset_manager.split_data(data,
train_ratio,
split="temporal")
train, val = dataset_manager.get_train_val_data(train, sample_size,
val_sample_size)
print("Done: %s" % (time.time() - start))
print('Encoding data...')
start = time.time()
dt_train = dataset_manager.encode_data(train)
dt_val = dataset_manager.encode_data(val)
dt_test = dataset_manager.encode_data(test)
X, y = dataset_manager.generate_3d_data(dt_train, max_len)
X_val, y_val = dataset_manager.generate_3d_data(dt_val, max_len)
print("Done: %s" % (time.time() - start))
print('Building and compiling the model...')
start = time.time()
data_dim = X.shape[2]
model = Sequential()
model.add(
LSTM(lstmsize, input_shape=(max_len, data_dim), return_sequences=True))
model.add(BatchNormalization())
model.add(
LSTM(lstmsize2, input_shape=(max_len, lstmsize),
return_sequences=True))
model.add(BatchNormalization())
model.add(Dropout(dropout))
dataset_manager.calculate_divisors(dt_train)
elif normalize_over == "all":
dt_all = dataset_manager.extract_timestamp_features(data)
dt_all = dataset_manager.extract_duration_features(dt_all)
dataset_manager.calculate_divisors(dt_all)
else:
print("unknown normalization mode")
dt_train = dataset_manager.normalize_data(dt_train)
max_len = dataset_manager.get_max_case_length(dt_train)
activity_cols = [col for col in dt_train.columns if col.startswith("act")]
n_activities = len(activity_cols)
data_dim = n_activities + 5
X, y_a, y_t = dataset_manager.generate_3d_data(dt_train, max_len)
print("Done: %s"%(time.time() - start))
# compile a model with same parameters that was trained, and load the weights of the trained model
print('Training model...')
start = time.time()
main_input = Input(shape=(max_len, data_dim), name='main_input')
# train a 2-layer LSTM with one shared layer
l1 = LSTM(lstmsize, input_shape=(max_len, data_dim), consume_less='gpu', init='glorot_uniform', return_sequences=True, dropout_W=dropout)(main_input) # the shared layer
b1 = BatchNormalization(axis=1)(l1)
l2_1 = LSTM(lstmsize, consume_less='gpu', init='glorot_uniform', return_sequences=False, dropout_W=dropout)(b1) # the layer specialized in activity prediction
b2_1 = BatchNormalization()(l2_1)
l2_2 = LSTM(lstmsize, consume_less='gpu', init='glorot_uniform', return_sequences=False, dropout_W=dropout)(b1) # the layer specialized in time prediction
