# Number of channels in the input image
CHANNELS = 3

# Dimensions of resized images (input to the neural net)
HEIGHT = 200
WIDTH = 200

# A placeholder for a batch of images
images_placeholder = tf.placeholder(dtype=tf.string, shape=(None,))

# The CloudML Prediction API always "feeds" the Tensorflow graph with
# dynamic batch sizes e.g. (?,).  decode_jpeg only processes scalar
# strings because it cannot guarantee a batch of images would have
# the same output size.  We use tf.map_fn to give decode_jpeg a scalar
# string from dynamic batches.
def decode_and_resize(image_str_tensor):
  """Decodes jpeg string, resizes it and returns a uint8 tensor."""

  image = tf.image.decode_jpeg(image_str_tensor, channels=CHANNELS)

  # Note resize expects a batch_size, but tf_map supresses that index,
  # thus we have to expand then squeeze.  Resize returns float32 in the
  # range [0, uint8_max]
  image = tf.expand_dims(image, 0)
  image = tf.image.resize_bilinear(
      image, [HEIGHT, WIDTH], align_corners=False)
  image = tf.squeeze(image, squeeze_dims=[0])
  image = tf.cast(image, dtype=tf.uint8)
  return image

decoded_images = tf.map_fn(
    decode_and_resize, images_placeholder, back_prop=False, dtype=tf.uint8)

# convert_image_dtype, also scales [0, uint8_max] -> [0, 1).
images = tf.image.convert_image_dtype(decoded_images, dtype=tf.float32)

# Then shift images to [-1, 1) (useful for some models such as Inception)
images = tf.sub(images, 0.5)
images = tf.mul(images, 2.0)

# ...

inputs = {"image_bytes": images_placeholder.name}
tf.add_to_collection("inputs", json.dumps(inputs))

{"image_bytes": {"b64": "dGVzdAo="}}

echo "{"image_bytes": {"b64": "`base64 image.jpg`"}}" > instances

gcloud beta ml predict --instances=instances --model=my_model

{"instances" : [{"image_bytes": {"b64": "dGVzdAo="}}]}