fasterrcnn 에서 동작하도록 하기 위해서는 keras_cv 에 다음 패치 필요: keras_cv/layers/object_detection/anchor_generator.py

keras-cv--fasterrcnn--patch.txt

"""
fasterrcnn 에서 동작하도록 하기 위해서는 keras_cv 에 다음 패치 필요:

keras_cv/layers/object_detection/anchor_generator.py

line 258:

        stride = self.stride
        # make sure range of `cx` is within limit of `image_width` with
        # `stride`, also for sizes where `image_width % stride != 0`.
        # [W]
        ### XXX math.ceil ==> ops.ceil --shr 2023-08-09
        cx = ops.cast(
            ops.arange(
               # 0.5 * stride, math.ceil(image_width / stride) * stride, stride
               0.5 * stride, ops.ceil(image_width / stride) * stride, stride
            ),
            "float32",
        )
        # make sure range of `cy` is within limit of `image_height` with
        # `stride`, also for sizes where `image_height % stride != 0`.
        # [H]
        ### XXX math.ceil ==> ops.ceil --shr 2023-08-09
        cy = ops.cast(
            ops.arange(
                # 0.5 * stride, math.ceil(image_height / stride) * stride, stride
                0.5 * stride, ops.ceil(image_height / stride) * stride, stride
            ),
            "float32",
        )
""";

if MODEL_TYPE == "fasterrcnn":
    """
    model.predict(image) 호출시에는 다음과 같은 res 반환:
    
    {'boxes': <tf.RaggedTensor [[]]>,
     'confidence': <tf.RaggedTensor [[]]>,
     'classes': <tf.RaggedTensor [[]]>,
     'num_detections': array([0], dtype=int32)}
    """
    
    res = model.predict(images, batch_size=1)

    print(
        "model.predict(images) output:", {k: (v.shape, v.dtype) for k, v in res.items()}
    )

    num_detections = res["num_detections"]
    det_boxes = res["boxes"]
    det_confidence = res["confidence"]
    det_clsses = res["classes"]

    print("images:", type(images), images.shape, images.dtype)
    print(
        "num_detections:",
        type(num_detections),
        num_detections.shape,
        num_detections.dtype,
    )
    print("det_boxes:", type(det_boxes), det_boxes.shape, det_boxes.dtype)
    print(
        "det_confidence:",
        type(det_confidence),
        det_confidence.shape,
        det_confidence.dtype,
    )
    print("det_clsses:", type(det_clsses), det_clsses.shape, det_clsses.dtype)

    # num_detections = num_detections.numpy() # num_detections 는 tensor 아님
    det_boxes = det_boxes.numpy()
    det_confidence = det_confidence.numpy()
    det_clsses = det_clsses.numpy()

    print(
        "num_predictions:",
        num_detections[0],
        "det_clsses:",
        det_clsses[0],
        "det_boxes:",
        det_boxes[0],
        "det_confidence:",
        det_confidence[0],
    )