TF Serving blog post: serving receiver function

TF_servingPost_serving_receiver_fn.py

import os
import os.path as op
import tensorflow as tf

HEIGHT, WIDTH, CHANNELS = 256, 256, 3


def serving_input_receiver_fn():
    """Convert string encoded images (like base64 strings) into preprocessed tensors"""

    def decode_and_resize(image_str_tensor):
        """Decodes a single image string, preprocesses/resizes it, and returns a reshaped uint8 tensor."""
        image = tf.image.decode_image(image_str_tensor, channels=CHANNELS,
                                      dtype=tf.uint8)
        image = tf.reshape(image, [HEIGHT, WIDTH, CHANNELS])

        return image

    # Run preprocessing function on all images in batch
    input_ph = tf.placeholder(tf.string, shape=[None], name='image_binary')
    images_tensor = tf.map_fn(
        decode_and_resize, input_ph, back_prop=False, dtype=tf.uint8)

    # Cast to float32
    images_tensor = tf.cast(images_tensor, dtype=tf.float32)
    #  Run Xception-specific preprocessing to scale images from [0, 255] to [-1, 1]
    images_tensor = tf.subtract(tf.divide(images_tensor, 127.5), 1)

    return tf.estimator.export.ServingInputReceiver(
        {'input_1': images_tensor},  # The key here needs match the name of your model's first layer
        {'image_bytes': input_ph})   # You can specify the key here, but this is a good default