Function _multibox_ref_numpy

tests/python/relax/test_op_vision.py:1492–1534

Numpy reference aligned with ``topi.vision.multibox_transform_loc``.

(
    cls_pred, loc_pred, anchor, variances, clip=False, threshold=0.0, keep_background=True
)

def _multibox_ref_numpy(
    cls_pred, loc_pred, anchor, variances, clip=False, threshold=0.0, keep_background=True
):
    """Numpy reference aligned with ``topi.vision.multibox_transform_loc``."""

    def _softmax(x, axis):
        x_max = np.max(x, axis=axis, keepdims=True)
        exp = np.exp(x - x_max)
        return exp / np.sum(exp, axis=axis, keepdims=True)

    B, C, N = cls_pred.shape
    loc = loc_pred.reshape(B, N, 4)
    scores = _softmax(cls_pred.astype("float64"), axis=1).astype(np.float32)
    if threshold > 0.0:
        scores = np.where(scores >= threshold, scores, 0.0).astype(np.float32)
    if not keep_background:
        scores = scores.copy()
        scores[:, 0, :] = 0.0
    vx, vy, vw, vh = variances
    boxes = np.zeros((B, N, 4), dtype=np.float32)
    for b in range(B):
        for a in range(N):
            left, top, right, bottom = anchor[0, a, :]
            ay = (top + bottom) * 0.5
            ax = (left + right) * 0.5
            ah = bottom - top
            aw = right - left
            ex, ey, ew, eh = loc[b, a, :]
            ycenter = ey * vy * ah + ay
            xcenter = ex * vx * aw + ax
            half_h = 0.5 * np.exp(eh * vh) * ah
            half_w = 0.5 * np.exp(ew * vw) * aw
            ymin = ycenter - half_h
            xmin = xcenter - half_w
            ymax = ycenter + half_h
            xmax = xcenter + half_w
            if clip:
                ymin = np.clip(ymin, 0.0, 1.0)
                xmin = np.clip(xmin, 0.0, 1.0)
                ymax = np.clip(ymax, 0.0, 1.0)
                xmax = np.clip(xmax, 0.0, 1.0)
            boxes[b, a, :] = (ymin, xmin, ymax, xmax)
    return boxes, scores


@pytest.mark.skipif(not env.has_llvm(), reason="need llvm")