import onnx
from onnxruntime.quantization import quantize_dynamic, QuantType

model_fp32 = 'yolov4.onnx'
model_quant = 'yolov4_uint8.onnx'
quantized_model = quantize_dynamic(
    model_fp32,
    model_quant,
    weight_type=QuantType.QUInt8
)

onnxsim ./yolov4_uint8.onnx ./yolov4_uint8.onnx

def image_preprocess(image: np.ndarray, target_size: tuple, gt_boxes: np.ndarray = None) -> np.ndarray:  # type: ignore
    """
    Preprocess the image by resizing and padding.

    Args:
        image (np.ndarray): Input image array of shape (height, width, channels).
        target_size (tuple): Target size for resizing (height, width).
        gt_boxes (np.ndarray, optional): Ground truth boxes. Defaults to None.

    Returns:
        np.ndarray: Resized and padded image.
    """

    ih, iw = target_size
    h, w, _ = image.shape

    # Calculate the scale and new width/height
    scale = min(iw / w, ih / h)
    nw, nh = int(scale * w), int(scale * h)

    # Resize the image
    image_resized = cv2.resize(image, (nw, nh))

    # Create a padded image with the new dimensions
    image_padded = np.full(shape=[ih, iw, 3], fill_value=128.0)
    dw, dh = (iw - nw) // 2, (ih - nh) // 2
    image_padded[dh:nh + dh, dw:nw + dw, :] = image_resized
    image_padded = image_padded / 255.

    # If ground truth boxes are provided, adjust their coordinates
    if gt_boxes is not None:
        gt_boxes[:, [0, 2]] = gt_boxes[:, [0, 2]] * scale + dw
        gt_boxes[:, [1, 3]] = gt_boxes[:, [1, 3]] * scale + dh
        return image_padded, gt_boxes

    return image_padded

def postprocess_bbbox(pred_bbox: np.ndarray, ANCHORS: np.ndarray, STRIDES: np.ndarray, XYSCALE: list[float] = [1, 1, 1]) -> np.ndarray:
    """
    Postprocess bounding box predictions to get final predictions.

    Args:
        pred_bbox (np.ndarray): Predicted bounding boxes.
        ANCHORS (np.ndarray): Anchor values for bounding boxes.
        STRIDES (np.ndarray): Stride values for bounding boxes.
        XYSCALE (list[float], optional): Scaling factors for bounding boxes. Defaults to [1, 1, 1].

    Returns:
        np.ndarray: Postprocessed bounding boxes.
    """

    for i, pred in enumerate(pred_bbox):
        conv_shape = pred.shape
        output_size = conv_shape[1]

        # Extract dx, dy, dw, dh
        conv_raw_dxdy = pred[:, :, :, :, 0:2]
        conv_raw_dwdh = pred[:, :, :, :, 2:4]

        # Generate the grid
        xy_grid = np.meshgrid(np.arange(output_size), np.arange(output_size))
        xy_grid = np.expand_dims(np.stack(xy_grid, axis=-1), axis=2)
        xy_grid = np.tile(np.expand_dims(xy_grid, axis=0), [1, 1, 1, 3, 1])
        xy_grid = xy_grid.astype(np.float32)

        # Calculate pred_xy and pred_wh
        pred_xy = ((special.expit(conv_raw_dxdy) *
                   XYSCALE[i]) - 0.5 * (XYSCALE[i] - 1) + xy_grid) * STRIDES[i]
        pred_wh = (np.exp(conv_raw_dwdh) * ANCHORS[i])
        pred[:, :, :, :, 0:4] = np.concatenate([pred_xy, pred_wh], axis=-1)

    # Reshape and concatenate the bounding boxes
    pred_bbox = [np.reshape(x, (-1, np.shape(x)[-1])) for x in pred_bbox]
    pred_bbox = np.concatenate(pred_bbox, axis=0)
    return pred_bbox

def postprocess_boxes(pred_bbox: np.ndarray, org_img_shape: tuple, input_size: int, score_threshold: float) -> np.ndarray:
    """
    Postprocess bounding boxes by resizing and removing invalid ones.

    Args:
        pred_bbox (np.ndarray): Predicted bounding boxes.
        org_img_shape (tuple): Original image shape (height, width).
        input_size (int): Size of the input image after preprocessing.
        score_threshold (float): Threshold for valid bounding boxes.

    Returns:
        np.ndarray: Resized and filtered bounding boxes.
    """

    valid_scale = [0, np.inf]
    pred_bbox = np.array(pred_bbox)
    pred_xywh = pred_bbox[:, 0:4]
    pred_conf = pred_bbox[:, 4]
    pred_prob = pred_bbox[:, 5:]

    # (x, y, w, h) --> (xmin, ymin, xmax, ymax)
    pred_coor = np.concatenate([pred_xywh[:, :2] - pred_xywh[:, 2:] * 0.5,
                                pred_xywh[:, :2] + pred_xywh[:, 2:] * 0.5], axis=-1)

    # (xmin, ymin, xmax, ymax) -> (xmin_org, ymin_org, xmax_org, ymax_org)
    org_h, org_w = org_img_shape
    resize_ratio = min(input_size / org_w, input_size / org_h)
    dw = (input_size - resize_ratio * org_w) / 2
    dh = (input_size - resize_ratio * org_h) / 2

    # Resize coordinates
    pred_coor[:, 0::2] = 1.0 * (pred_coor[:, 0::2] - dw) / resize_ratio
    pred_coor[:, 1::2] = 1.0 * (pred_coor[:, 1::2] - dh) / resize_ratio

    # clip some boxes that are out of range
    pred_coor = np.concatenate([np.maximum(pred_coor[:, :2], [0, 0]),
                                np.minimum(pred_coor[:, 2:], [org_w - 1, org_h - 1])], axis=-1)
    invalid_mask = np.logical_or(
        (pred_coor[:, 0] > pred_coor[:, 2]), (pred_coor[:, 1] > pred_coor[:, 3]))
    pred_coor[invalid_mask] = 0

    # discard some invalid boxes
    bboxes_scale = np.sqrt(np.multiply.reduce(
        pred_coor[:, 2:4] - pred_coor[:, 0:2], axis=-1))
    scale_mask = np.logical_and(
        (valid_scale[0] < bboxes_scale), (bboxes_scale < valid_scale[1]))

    # discard some boxes with low scores
    classes = np.argmax(pred_prob, axis=-1)
    scores = pred_conf * pred_prob[np.arange(len(pred_coor)), classes]
    score_mask = scores > score_threshold
    mask = np.logical_and(scale_mask, score_mask)
    coors, scores, classes = pred_coor[mask], scores[mask], classes[mask]

    return np.concatenate([coors, scores[:, np.newaxis], classes[:, np.newaxis]], axis=-1)