tensorflow.keras preprocessor in tensorflow serving preprocessing?

Question

I'm converting a tensorflow.keras model to an estimator and using estimator.export_saved_mode(serving_input_receiver_fn=serving_input_receiver_fn) to get my model ready for TensorFlow Serving. Here's what my serving_input_receiver_fn looks like:

def image_preprocessing(image):   
    image = tf.expand_dims(image, 0)
    image = tf.image.resize_bilinear(image, [HEIGHT, WIDTH], align_corners=False)
    image = tf.squeeze(image, axis=[0])
    image = tf.cast(image, dtype=tf.uint8)
    return image

def serving_input_receiver_fn():
    def prepare_image(image_str_tensor):
        image = tf.image.decode_jpeg(image_str_tensor, channels=CHANNELS)
        return image_preprocessing(image)

    input_ph = tf.placeholder(tf.string, shape=[None], name='image_binary')
    images_tensor = tf.map_fn(prepare_image, input_ph, back_prop=False, dtype=tf.uint8)
    images_tensor = tf.image.convert_image_dtype(images_tensor, dtype=tf.float32)

    return tf.estimator.export.ServingInputReceiver(
        {model.input_names[0]: images_tensor},
        {'image_bytes': input_ph})

Is there a way to continue taking uint8 as input, but convert it to float32 and then apply a tensorflow.keras preprocessing function such as tensorflow.keras.applications.xception.preprocess_input?

I'm not sure how to normalize this input according to the mean/std of a tensorflow.keras.applications model. Before adding the above my model accepted json serialized lists of numpy arrays and I would keras normalize them client side. Now that it accepts uint8 base64 encoded byte strings I'm not sure how to move the keras normalizing over to this function.

sdcbr sdcbr · Accepted Answer · 2019-01-24T08:13:14

If you'll look up the Keras implementation of Xception on github, you'll notice that it uses the general imagenet preprocessing function that is used by most other pretrained models in that repo. This function should work on numpy arrays as well as on Keras/TensorFlow tensors. Hence you should be able to import it and add it to your image_preprocessing function.

Now it's not quite clear why you want to cast to tf.uint8. I would expect a preprocessing function that looks like this:

def image_preprocessing(im):
    im = tf.expand_dims(im, 0)
    im = tf.image.resize_bilinear(im, [HEIGHT, WIDTH], align_corners=False)
    im = tf.squeeze(im, axis=[0])
    im = preprocess_input(im)
    return im

tensorflow.keras preprocessor in tensorflow serving preprocessing?

1 Answers