Function _parse_eval_data

axlearn/vision/input_detection.py:245–307

Parses single data for evaluation.

(
    data: dict[str, Any],
    *,
    output_size: tuple[int, int],
    output_stride: int,
    max_num_instances: int = 100,
    **kwargs,
)

def _parse_eval_data(
    data: dict[str, Any],
    *,
    output_size: tuple[int, int],
    output_stride: int,
    max_num_instances: int = 100,
    **kwargs,
):
    """Parses single data for evaluation."""
    del kwargs  # Delete train-specific kwargs, e.g. kwargs for data augmentation.

    classes = data["groundtruth_classes"]
    boxes = data["groundtruth_boxes"]

    # Gets original image and its size.
    image = data["image"]
    image_shape = tf.shape(input=image)[0:2]

    # Normalizes image with mean and std pixel values.
    image = whiten(image)

    # Converts boxes from normalized coordinates to pixel coordinates.
    boxes = utils_detection.denormalize_boxes_tf(boxes, image_shape)

    # Resizes and crops image.
    image, image_info = resize_and_crop_image(
        image,
        desired_size=output_size,
        padded_size=utils_detection.compute_padded_size(output_size, output_stride),
        aug_scale_min=1.0,  # No scale jitter for evaluation.
        aug_scale_max=1.0,  # No scale jitter for evaluation.
    )
    image_data = {"image": image, "image_info": image_info}
    image_height, image_width, _ = image.get_shape().as_list()

    # Resizes and crops boxes.
    image_scale = image_info[2, :]
    offset = image_info[3, :]
    boxes = utils_detection.resize_and_crop_boxes(boxes, image_scale, image_info[1, :], offset)
    # Filters out ground truth boxes that are all zeros.
    indices = utils_detection.get_non_empty_box_indices(boxes)
    boxes = tf.gather(boxes, indices)
    classes = tf.gather(classes, indices)

    # Sets up groundtruth data for evaluation.
    groundtruths = {
        "source_id": data["source_id"],
        "height": data["height"],
        "width": data["width"],
        "num_detections": tf.shape(data["groundtruth_classes"]),
        "image_info": image_info,
        "boxes": utils_detection.denormalize_boxes_tf(data["groundtruth_boxes"], image_shape),
        "classes": data["groundtruth_classes"],
        "areas": data["groundtruth_area"],
        "is_crowds": tf.cast(data["groundtruth_is_crowd"], tf.int32),
    }
    groundtruths["source_id"] = utils_detection.process_source_id(groundtruths["source_id"])
    labels = {