Rude carnie modified

[abhijithrb]

Made changes to handle memory issue.
In the utils.py, there were ops in the make_multi_crop_batch and make_multi_image_batch that were added to the graph with each iteration of a loop such as:

crop = tf.image.resize_images(image, (RESIZE_FINAL, RESIZE_FINAL))
image = standardize_image(crop)
images.append(image)
image_batch = tf.stack(images)

Hence with each loop iteration, there were nodes added to the graph resulting in an incremental increase in memory usage and thus execution time.
In utils.py, I've modified the ImageCoder class such that the Ops are added just once to the graph.

[dpressel]

Thanks for submitting this. I need a bit more time to review the change. Thanks for your patience

[abhijithrb]

No problem at all.
If you find this to be ok, I can make a PR of my master branch as well which runs both age and gender models in the same code, if that's a change you'd like to consider