Example of Using a Raw Image Dataset in the Feature Store¶

images_df = featurestore.get_training_dataset("sample_mnist")

import tensorflow as tf
import numpy as np

def build_tf_graph(images_dir):
    """
    A simple computational graph for reading in a jpg into a tensor
    """
    img_filenames = tf.gfile.Glob(images_dir + "/*.jpg")
    num_images = len(img_filenames)
    img_queue = tf.train.string_input_producer(img_filenames)
    img_reader = tf.WholeFileReader()
    # Operation for reading a single file from the queue
    file_name_op, file_contents_op = img_reader.read(img_queue)
    # Operation for decoding JPEG to tensor
    img_to_tensor_op = tf.image.decode_jpeg(file_contents_op)
    return img_to_tensor_op, file_name_op, num_images

def run_graph_for_reading_images(sess, images_dir, num_images, img_op, file_name_op):
    """
    Run the tf-graph for reading all images into tensors
    """
    image_tensors = []
    image_filenames_read = []
    for i in range(num_images):
        # these two must be run in the same call to sess.run() otherwise they become unsynced which messes up labels for validation set..
        img_tensor, file_name_str = sess.run([img_op, file_name_op])
        image_tensors.append(img_tensor)
        image_filenames_read.append(file_name_str)
    return np.array(image_tensors), np.array(image_filenames_read)

def init_graph():
    """ 
    Initialize the graph and variables for Tensorflow engine 
    """
    # get operation for initializing the global variables in the graph
    init_g = tf.global_variables_initializer()
    
    # create a session for encapsulating the environment where 
    # operations can be run and tensors can be evaluated
    sess = tf.Session()
    
    # run the initialization operation
    sess.run(init_g)
    return sess

def setup_tf():
    """
    Setup TF session
    """
    # Initialize TF
    sess = init_graph()

    # Get coordinator for threads to be able to read
    coord = tf.train.Coordinator()

    # Starts all queue runners in the graph and return list of the threads
    threads = tf.train.start_queue_runners(coord=coord, sess=sess)
    
    return sess, coord, threads

images_dir = featurestore.get_training_dataset_path("sample_mnist")
img_to_tensor_op, file_name_op, num_images = build_tf_graph(images_dir)
sess, coord, threads = setup_tf()
image_tensors, image_filenames_read = run_graph_for_reading_images(sess, 
                                                                   images_dir, 
                                                                   num_images, 
                                                                   img_to_tensor_op, 
                                                                   file_name_op
                                                                  )