Function fbresnet_mapper

examples/ImageNetModels/imagenet_utils.py:177–261

Note: compared to fbresnet_augmentor, it lacks some photometric augmentation that may have a small effect (0.1~0.2%) on accuracy.

(isTrain)

def fbresnet_mapper(isTrain):
    """
    Note: compared to fbresnet_augmentor, it
    lacks some photometric augmentation that may have a small effect (0.1~0.2%) on accuracy.
    """
    JPEG_OPT = {'fancy_upscaling': True, 'dct_method': 'INTEGER_ACCURATE'}

    def uint8_resize_bicubic(image, shape):
        ret = tf.image.resize_bicubic([image], shape)
        return tf.cast(tf.clip_by_value(ret, 0, 255), tf.uint8)[0]

    def resize_shortest_edge(image, image_shape, size):
        shape = tf.cast(image_shape, tf.float32)
        w_greater = tf.greater(image_shape[0], image_shape[1])
        shape = tf.cond(w_greater,
                        lambda: tf.cast([shape[0] / shape[1] * size, size], tf.int32),
                        lambda: tf.cast([size, shape[1] / shape[0] * size], tf.int32))

        return uint8_resize_bicubic(image, shape)

    def center_crop(image, size):
        image_height = tf.shape(image)[0]
        image_width = tf.shape(image)[1]

        offset_height = (image_height - size) // 2
        offset_width = (image_width - size) // 2
        image = tf.slice(image, [offset_height, offset_width, 0], [size, size, -1])
        return image

    def lighting(image, std, eigval, eigvec):
        v = tf.random_normal(shape=[3], stddev=std) * eigval
        inc = tf.matmul(eigvec, tf.reshape(v, [3, 1]))
        image = tf.cast(tf.cast(image, tf.float32) + tf.reshape(inc, [3]), image.dtype)
        return image

    def validation_mapper(byte):
        image = tf.image.decode_jpeg(
            tf.reshape(byte, shape=[]), 3, **JPEG_OPT)
        image = resize_shortest_edge(image, tf.shape(image), 256)
        image = center_crop(image, 224)
        image = tf.reverse(image, axis=[2])  # to BGR
        return image

    def training_mapper(byte):
        jpeg_shape = tf.image.extract_jpeg_shape(byte)  # hwc
        bbox_begin, bbox_size, distort_bbox = tf.image.sample_distorted_bounding_box(
            jpeg_shape,
            bounding_boxes=tf.zeros(shape=[0, 0, 4]),
            min_object_covered=0,
            aspect_ratio_range=[0.75, 1.33],
            area_range=[0.08, 1.0],
            max_attempts=10,
            use_image_if_no_bounding_boxes=True)

        is_bad = tf.reduce_sum(tf.cast(tf.equal(bbox_size, jpeg_shape), tf.int32)) >= 2

        def good():
            offset_y, offset_x, _ = tf.unstack(bbox_begin)