def max_length(tensor):
    return max(len(t) for t in tensor)

# Helper function to pad the query tensor and slot (target) tensor to the same length.
# Also creates a tensor for teacher forcing.
def create_tensors(input_tensor, target_tensor, nb_sample=9999999, max_len=0):
    len_input, len_target  = max_length(input_tensor), max_length(target_tensor)
    len_input = max(len_input,max_len)
    len_target = max(len_target,max_len)


    # Padding the input and output tensor to the maximum length
    input_data = tf.keras.preprocessing.sequence.pad_sequences(input_tensor,
                                                                 maxlen=len_input,
                                                                 padding='post')

    teacher_data = tf.keras.preprocessing.sequence.pad_sequences(target_tensor,
                                                                  maxlen=len_target ,
                                                                  padding='post')

    target_data = [[teacher_data[n][i+1] for i in range(len(teacher_data[n])-1)] for n in range(len(teacher_data))]
    target_data = tf.keras.preprocessing.sequence.pad_sequences(target_data, maxlen=len_target, padding="post")
    target_data = target_data.reshape((target_data.shape[0], target_data.shape[1], 1))

    nb = len(input_data)
    p = np.random.permutation(nb)
    input_data = input_data[p]
    teacher_data = teacher_data[p]
    target_data = target_data[p]

    return input_data[:min(nb_sample, nb)], teacher_data[:min(nb_sample, nb)], target_data[:min(nb_sample, nb)],
           len_input, len_target

input_data_train, teacher_data_train, target_data_train, \
                  len_input_train, len_target_train  = create_tensors(input_tensor_train, target_tensor_train)
input_data_test, teacher_data_test, target_data_test, \
                 len_input_test, len_target_test  = create_tensors(input_tensor_test, target_tensor_test, max_len=len_input_train)