IssueTracker

Issue summary:
It would be a nice addition to Cloud ML Engine to be able to call a code such as the one below ([1]), being "GCSuri" the Cloud Storage URI to an accessible image in GCS, in order to perform predictions in ML Engine.
As explained in the documentation [2], the current supported prediction input data formats are:
 - BATCH: Text file with JSON strings / TFRecords file
 - ONLINE: Text file with JSON instance / CSV file

By supporting reading images directly from GCS, it would be possible to simply upload the original images to GCS and the model would handle the necessary transformations.

Current workarounds:
 - Use the supported and recommended approach to provide input data (as in [2]).


[1]
def read_img(GCSuri):
    img_str = tf.read_file(GCSuri)
    img_tensor = tf.image.decode_image(image_string)
    return image_tensor
[2] https://cloud.google.com/ml-engine/docs/tensorflow/prediction-overview#prediction_input_data