preprocess_tensorflow.py

preprocess_tensorflow.py

@tf.function
def preprocess_tf(x):
    """
    Preprocessing for Keras (MobileNetV2, ResNetV2).
    :param x: np.asarray([image, image, ...], dtype="float32") in RGB
    :return: normalized image tf style (RGB)
    """
    batch, height, width, channels = x.shape
    x = tf.cast(x, tf.float32)

    # ! do not use tf.constant as they are not right now serializable when saving model for .h5
    # ! https://stackoverflow.com/questions/47066635/checkpointing-keras-model-typeerror-cant-pickle-thread-lock-objects
    # mean_tensor = tf.constant([127.5, 127.5, 127.5], dtype=tf.float32, shape=[1, 1, 1, 3], name="mean")
    # one_tensor = tf.constant([1.0, 1.0, 1.0], dtype=tf.float32, shape=[1, 1, 1, 3], name="one")

    mean_tensor = np.asarray([[[[127.5, 127.5, 127.5]]]], dtype=np.float32)
    one_tensor = np.asarray([[[[1.0, 1.0, 1.0]]]], dtype=np.float32)

    x = tf.keras.backend.reshape(x, (-1, 3))
    result = (x / mean_tensor) - one_tensor
    return tf.keras.backend.reshape(result, (-1, height, width, channels))