input_pipeline.py 文件源码

def load_batches(image_filenames=None,
                 label_filenames=None,
                 shape=(64, 64, 3),
                 batch_size=100):

    min_after_dequeue = 1000
    capacity = min_after_dequeue + 3 * batch_size #As recommended on tf website

    with tf.name_scope('input'):
        image_queue, label_queue = tf.train.slice_input_producer(
                                            [image_filenames, label_filenames],
                                            shuffle=True)

        #File reader and decoder for images and labels goes here
        with tf.name_scope('image'): 

            with tf.name_scope('decode'): 
                file_content = tf.read_file(image_queue)
                image_data = tf.image.decode_jpeg(file_content, channels=3)

            #Any resizing or processing (eg. normalising) goes here
            with tf.name_scope('normalize'): 
                image_data = tf.cast(image_data, tf.float32)
                image_data = tf.image.per_image_standardization(image_data)
                image_data.set_shape(shape)


        with tf.name_scope('label'): 

            with tf.name_scope('decode'): 
                label_data = tf.string_to_number(label_queue, out_type=tf.int32)

        image_batch, label_batch = tf.train.shuffle_batch(
                                            [image_data, label_data],
                                            batch_size=batch_size,
                                            capacity=capacity,
                                            min_after_dequeue=min_after_dequeue,
                                            #,num_threads=1
                                            )

    return image_batch, label_batch