Python mxnet.nd.zeros_like() Examples

def hybrid_forward(self, F, samples):
        """Encode class prediction labels for SigmoidCrossEntropy Loss.

        Parameters
        ----------
        samples : mxnet.nd.NDArray or mxnet.sym.Symbol
            Sampling results with shape (B, N), 1:pos, 0:ignore, -1:negative

        Returns
        -------
        (mxnet.nd.NDArray, mxnet.nd.NDArray)
            (target, mask)
            target is the output label with shape (B, N), 1: pos, 0: negative, -1: ignore
            mask is the mask for label, -1(ignore) labels have mask 0, otherwise mask is 1.

        """
        # notation from samples, 1:pos, 0:ignore, -1:negative
        target = (samples + 1) / 2.
        target = F.where(F.abs(samples) < 1e-5, F.ones_like(target) * -1, target)
        # output: 1: pos, 0: negative, -1: ignore
        mask = F.where(F.abs(samples) > 1e-5, F.ones_like(samples), F.zeros_like(samples))
        return target, mask 

def heatmap_to_coord(heatmaps, bbox_list):
    heatmap_height = heatmaps.shape[2]
    heatmap_width = heatmaps.shape[3]
    coords, maxvals = get_max_pred(heatmaps)
    preds = nd.zeros_like(coords)

    for i, bbox in enumerate(bbox_list):
        x0 = bbox[0]
        y0 = bbox[1]
        x1 = bbox[2]
        y1 = bbox[3]
        w = (x1 - x0) / 2
        h = (y1 - y0) / 2
        center = np.array([x0 + w, y0 + h])
        scale = np.array([w, h])

        w_ratio = coords[i][:, 0] / heatmap_width
        h_ratio = coords[i][:, 1] / heatmap_height
        preds[i][:, 0] = scale[0] * 2 * w_ratio + center[0] - scale[0]
        preds[i][:, 1] = scale[1] * 2 * h_ratio + center[1] - scale[1]
    return preds, maxvals 

def forward(self, samples, matches, anchors, refs):
        """Forward"""
        F = nd
        # TODO(zhreshold): batch_pick, take multiple elements?
        ref_boxes = nd.repeat(refs.reshape((0, 1, -1, 4)), axis=1, repeats=matches.shape[1])
        ref_boxes = nd.split(ref_boxes, axis=-1, num_outputs=4, squeeze_axis=True)
        ref_boxes = nd.concat(*[F.pick(ref_boxes[i], matches, axis=2).reshape((0, -1, 1)) \
            for i in range(4)], dim=2)
        g = self.corner_to_center(ref_boxes)
        a = self.corner_to_center(anchors)
        t0 = ((g[0] - a[0]) / a[2] - self._means[0]) / self._stds[0]
        t1 = ((g[1] - a[1]) / a[3] - self._means[1]) / self._stds[1]
        t2 = (F.log(g[2] / a[2]) - self._means[2]) / self._stds[2]
        t3 = (F.log(g[3] / a[3]) - self._means[3]) / self._stds[3]
        codecs = F.concat(t0, t1, t2, t3, dim=2)
        temp = F.tile(samples.reshape((0, -1, 1)), reps=(1, 1, 4)) > 0.5
        targets = F.where(temp, codecs, F.zeros_like(codecs))
        masks = F.where(temp, F.ones_like(temp), F.zeros_like(temp))
        return targets, masks 

def hybrid_forward(self, F, samples, matches, refs):
        """HybridBlock, handle multi batch correctly

        Parameters
        ----------
        samples: (B, N), value +1 (positive), -1 (negative), 0 (ignore)
        matches: (B, N), value range [0, M)
        refs: (B, M), value range [0, num_fg_class), excluding background

        Returns
        -------
        targets: (B, N), value range [0, num_fg_class + 1), including background

        """
        # samples (B, N) (+1, -1, 0: ignore), matches (B, N) [0, M), refs (B, M)
        # reshape refs (B, M) -> (B, 1, M) -> (B, N, M)
        refs = F.repeat(refs.reshape((0, 1, -1)), axis=1, repeats=matches.shape[1])
        # ids (B, N, M) -> (B, N), value [0, M + 1), 0 reserved for background class
        target_ids = F.pick(refs, matches, axis=2) + 1
        # samples 0: set ignore samples to ignore_label
        targets = F.where(samples > 0.5, target_ids, nd.ones_like(target_ids) * self._ignore_label)
        # samples -1: set negative samples to 0
        targets = F.where(samples < -0.5, nd.zeros_like(targets), targets)
        return targets 

def hybrid_forward(self, F, x):
        pos_x = x.slice_axis(axis=self._axis, begin=1, end=None)
        cls_id = F.argmax(pos_x, self._axis)
        scores = F.pick(pos_x, cls_id, axis=-1)
        mask = scores > self._thresh
        cls_id = F.where(mask, cls_id, F.ones_like(cls_id) * -1)
        scores = F.where(mask, scores, F.zeros_like(scores))
        return cls_id, scores 

def hybrid_forward(self, F, x):
        scores = x.slice_axis(axis=self._axis, begin=1, end=None)  # b x N x fg_class
        template = F.zeros_like(x.slice_axis(axis=-1, begin=0, end=1))
        cls_ids = []
        for i in range(self._fg_class):
            cls_ids.append(template + i)  # b x N x 1
        cls_id = F.concat(*cls_ids, dim=-1)  # b x N x fg_class
        mask = scores > self._thresh
        cls_id = F.where(mask, cls_id, F.ones_like(cls_id) * -1)
        scores = F.where(mask, scores, F.zeros_like(scores))
        return cls_id, scores 

def hybrid_forward(self, F, x):
        pos_x = x.slice_axis(axis=self._axis, begin=1, end=None)
        cls_id = F.argmax(pos_x, self._axis)
        scores = F.pick(pos_x, cls_id, axis=-1)
        mask = scores > self._thresh
        cls_id = F.where(mask, cls_id, F.ones_like(cls_id) * -1)
        scores = F.where(mask, scores, F.zeros_like(scores))
        return cls_id, scores 

def forward(self, samples, matches, anchors, labels, refs):
        """Encode BBox One entry per category"""
        F = nd
        ref_boxes = F.repeat(refs.reshape((0, 1, -1, 4)), axis=1, repeats=matches.shape[1])
        ref_boxes = F.split(ref_boxes, axis=-1, num_outputs=4, squeeze_axis=True)
        ref_boxes = F.concat(*[F.pick(ref_boxes[i], matches, axis=2).reshape((0, -1, 1)) \
            for i in range(4)], dim=2)
        ref_labels = F.repeat(labels.reshape((0, 1, -1)), axis=1, repeats=matches.shape[1])
        ref_labels = F.pick(ref_labels, matches, axis=2).reshape((0, -1, 1))
        g = self.corner_to_center(ref_boxes)
        a = self.corner_to_center(anchors)
        t0 = ((g[0] - a[0]) / a[2] - self._means[0]) / self._stds[0]
        t1 = ((g[1] - a[1]) / a[3] - self._means[1]) / self._stds[1]
        t2 = (F.log(g[2] / a[2]) - self._means[2]) / self._stds[2]
        t3 = (F.log(g[3] / a[3]) - self._means[3]) / self._stds[3]
        codecs = F.concat(t0, t1, t2, t3, dim=2)
        temp = F.tile(samples.reshape((0, -1, 1)), reps=(1, 1, 4)) > 0.5
        targets = F.where(temp, codecs, F.zeros_like(codecs))
        masks = F.where(temp, F.ones_like(temp), F.zeros_like(temp))
        out_targets = []
        out_masks = []
        # for cid in range(self._num_class):
        out_targets.append(targets)
        # but mask out the one not belong to this class
        out_masks.append(masks )
        all_targets = F.stack(*out_targets, axis=0)
        all_masks = F.stack(*out_masks, axis=0)
        return all_targets, all_masks 

def forward(self, samples, matches, anchors, labels, refs):
        """Encode BBox One entry per category"""
        F = nd
        ref_boxes = F.repeat(refs.reshape((0, 1, -1, 4)), axis=1, repeats=matches.shape[1])
        ref_boxes = F.split(ref_boxes, axis=-1, num_outputs=4, squeeze_axis=True)
        ref_boxes = F.concat(*[F.pick(ref_boxes[i], matches, axis=2).reshape((0, -1, 1)) \
            for i in range(4)], dim=2)
        ref_labels = F.repeat(labels.reshape((0, 1, -1)), axis=1, repeats=matches.shape[1])
        ref_labels = F.pick(ref_labels, matches, axis=2).reshape((0, -1, 1))
        g = self.corner_to_center(ref_boxes)
        a = self.corner_to_center(anchors)
        t0 = ((g[0] - a[0]) / a[2] - self._means[0]) / self._stds[0]
        t1 = ((g[1] - a[1]) / a[3] - self._means[1]) / self._stds[1]
        t2 = (F.log(g[2] / a[2]) - self._means[2]) / self._stds[2]
        t3 = (F.log(g[3] / a[3]) - self._means[3]) / self._stds[3]
        codecs = F.concat(t0, t1, t2, t3, dim=2)
        temp = F.tile(samples.reshape((0, -1, 1)), reps=(1, 1, 4)) > 0.5
        targets = F.where(temp, codecs, F.zeros_like(codecs))
        masks = F.where(temp, F.ones_like(temp), F.zeros_like(temp))
        out_targets = []
        out_masks = []
        for cid in range(self._num_class):
            same_cid = ref_labels == cid
            # keep orig targets
            out_targets.append(targets)
            # but mask out the one not belong to this class
            out_masks.append(masks * same_cid.repeat(axis=-1, repeats=4))
        all_targets = F.stack(*out_targets, axis=0)
        all_masks = F.stack(*out_masks, axis=0)
        return all_targets, all_masks 

def forward(self, cls_pred, box_pred, cls_target, box_target):
        """Compute loss in entire batch across devices."""
        # require results across different devices at this time
        cls_pred, box_pred, cls_target, box_target = [_as_list(x) \
            for x in (cls_pred, box_pred, cls_target, box_target)]
        # cross device reduction to obtain positive samples in entire batch
        num_pos = []
        for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
            pos_samples = (ct > 0)
            num_pos.append(pos_samples.sum())
        num_pos_all = sum([p.asscalar() for p in num_pos])
        if num_pos_all < 1:
            # no positive samples found, return dummy losses
            return nd.zeros((1,)), nd.zeros((1,)), nd.zeros((1,))

        # compute element-wise cross entropy loss and sort, then perform negative mining
        cls_losses = []
        box_losses = []
        sum_losses = []
        for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
            pred = nd.log_softmax(cp, axis=-1)
            pos = ct > 0
            cls_loss = -nd.pick(pred, ct, axis=-1, keepdims=False)
            rank = (cls_loss * (pos - 1)).argsort(axis=1).argsort(axis=1)
            hard_negative = rank < (pos.sum(axis=1) * self._negative_mining_ratio).expand_dims(-1)
            # mask out if not positive or negative
            cls_loss = nd.where((pos + hard_negative) > 0, cls_loss, nd.zeros_like(cls_loss))
            cls_losses.append(nd.sum(cls_loss, axis=0, exclude=True) / num_pos_all)

            bp = _reshape_like(nd, bp, bt)
            box_loss = nd.abs(bp - bt)
            box_loss = nd.where(box_loss > self._rho, box_loss - 0.5 * self._rho,
                                (0.5 / self._rho) * nd.square(box_loss))
            # box loss only apply to positive samples
            box_loss = box_loss * pos.expand_dims(axis=-1)
            box_losses.append(nd.sum(box_loss, axis=0, exclude=True) / num_pos_all)
            sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1])

        return sum_losses, cls_losses, box_losses 

def hybrid_forward(self, F, box_preds, gt_boxes):
        """Short summary.

        Parameters
        ----------
        F : mxnet.nd or mxnet.sym
            `F` is mxnet.sym if hybridized or mxnet.nd if not.
        box_preds : mxnet.nd.NDArray
            Predicted bounding boxes.
        gt_boxes : mxnet.nd.NDArray
            Ground-truth bounding boxes.

        Returns
        -------
        (tuple of) mxnet.nd.NDArray
            objectness: 0 for negative, 1 for positive, -1 for ignore.
            center_targets: regression target for center x and y.
            scale_targets: regression target for scale x and y.
            weights: element-wise gradient weights for center_targets and scale_targets.
            class_targets: a one-hot vector for classification.

        """
        with autograd.pause():
            box_preds = box_preds.reshape((0, -1, 4))
            objness_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=1))
            center_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            scale_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            weight_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            class_t = F.ones_like(objness_t.tile(reps=(self._num_class))) * -1
            batch_ious = self._batch_iou(box_preds, gt_boxes)  # (B, N, M)
            ious_max = batch_ious.max(axis=-1, keepdims=True)  # (B, N, 1)
            objness_t = (ious_max > self._ignore_iou_thresh) * -1  # use -1 for ignored
        return objness_t, center_t, scale_t, weight_t, class_t 

def compute_rot_loss(output, target_bin, target_res, mask):
    # output: (B, 128, 8) [bin1_cls[0], bin1_cls[1], bin1_sin, bin1_cos,
    #                 bin2_cls[0], bin2_cls[1], bin2_sin, bin2_cos]
    # target_bin: (B, 128, 2) [bin1_cls, bin2_cls]
    # target_res: (B, 128, 2) [bin1_res, bin2_res]
    # mask: (B, 128, 1)
    output = nd.reshape(output, (-1, 8))
    target_bin = nd.reshape(target_bin, (-1, 2))
    target_res = nd.reshape(target_res, (-1, 2))
    mask = nd.reshape(mask, (-1, 1))
    loss_bin1 = compute_bin_loss(output[:, 0:2], target_bin[:, 0], mask)
    loss_bin2 = compute_bin_loss(output[:, 4:6], target_bin[:, 1], mask)
    loss_res = nd.zeros_like(loss_bin1)

    mask1 = (target_bin[:, 0] > 0).astype('float32')
    if mask1.sum() > 0:
        valid_output1 = output
        valid_target_res1 = target_res

        loss_sin1 = compute_res_loss(valid_output1[:, 2], nd.sin(valid_target_res1[:, 0]), mask1)
        loss_cos1 = compute_res_loss(valid_output1[:, 3], nd.cos(valid_target_res1[:, 0]), mask1)
        loss_res = loss_res + loss_sin1 + loss_cos1

    mask2 = (target_bin[:, 1] > 0).astype('float32')
    if mask2.sum() > 0:
        valid_output2 = output
        valid_target_res2 = target_res

        loss_sin2 = compute_res_loss(valid_output2[:, 6], nd.sin(valid_target_res2[:, 1]), mask2)
        loss_cos2 = compute_res_loss(valid_output2[:, 7], nd.cos(valid_target_res2[:, 1]), mask2)
        loss_res = loss_res + loss_sin2 + loss_cos2
    #print("loss_bin1: {}, loss_bin2: {}, loss_sin1: {}, loss_sin2: {}, loss_cos1: {}, loss_cos2: {}".format(loss_bin1, loss_bin2, loss_sin1, loss_sin2, loss_cos1, loss_cos2))
    return loss_bin1 + loss_bin2 + loss_res 

def compute_rot_loss(output, target_bin, target_res, mask):
    # output: (B, 128, 8) [bin1_cls[0], bin1_cls[1], bin1_sin, bin1_cos,
    #                 bin2_cls[0], bin2_cls[1], bin2_sin, bin2_cos]
    # target_bin: (B, 128, 2) [bin1_cls, bin2_cls]
    # target_res: (B, 128, 2) [bin1_res, bin2_res]
    # mask: (B, 128, 1)
    output = nd.reshape(output, (-1, 8))
    target_bin = nd.reshape(target_bin, (-1, 2))
    target_res = nd.reshape(target_res, (-1, 2))
    mask = nd.reshape(mask, (-1, 1))
    loss_bin1 = compute_bin_loss(output[:, 0:2], target_bin[:, 0], mask)
    loss_bin2 = compute_bin_loss(output[:, 4:6], target_bin[:, 1], mask)
    loss_res = nd.zeros_like(loss_bin1)

    mask1 = (target_bin[:, 0] > 0).astype('float32')
    if mask1.sum() > 0:
        valid_output1 = output
        valid_target_res1 = target_res

        loss_sin1 = compute_res_loss(valid_output1[:, 2], nd.sin(valid_target_res1[:, 0]), mask1)
        loss_cos1 = compute_res_loss(valid_output1[:, 3], nd.cos(valid_target_res1[:, 0]), mask1)
        loss_res = loss_res + loss_sin1 + loss_cos1

    mask2 = (target_bin[:, 1] > 0).astype('float32')
    if mask2.sum() > 0:
        valid_output2 = output
        valid_target_res2 = target_res

        loss_sin2 = compute_res_loss(valid_output2[:, 6], nd.sin(valid_target_res2[:, 1]), mask2)
        loss_cos2 = compute_res_loss(valid_output2[:, 7], nd.cos(valid_target_res2[:, 1]), mask2)
        loss_res = loss_res + loss_sin2 + loss_cos2
    #print("loss_bin1: {}, loss_bin2: {}, loss_sin1: {}, loss_sin2: {}, loss_cos1: {}, loss_cos2: {}".format(loss_bin1, loss_bin2, loss_sin1, loss_sin2, loss_cos1, loss_cos2))
    return loss_bin1 + loss_bin2 + loss_res 

def hybrid_forward(self, F, x):
        scores = x.slice_axis(axis=self._axis, begin=1, end=None)  # b x N x fg_class
        template = F.zeros_like(x.slice_axis(axis=-1, begin=0, end=1))
        cls_ids = []
        for i in range(self._fg_class):
            cls_ids.append(template + i)  # b x N x 1
        cls_id = F.concat(*cls_ids, dim=-1)  # b x N x fg_class
        mask = scores > self._thresh
        cls_id = F.where(mask, cls_id, F.ones_like(cls_id) * -1)
        scores = F.where(mask, scores, F.zeros_like(scores))
        return cls_id, scores 

def forward(self, samples, matches, anchors, refs):
        """Not HybridBlock due to use of matches.shape

        Parameters
        ----------
        samples: (B, N) value +1 (positive), -1 (negative), 0 (ignore)
        matches: (B, N) value range [0, M)
        anchors: (B, N, 4) encoded in corner
        refs: (B, M, 4) encoded in corner

        Returns
        -------
        targets: (B, N, 4) transform anchors to refs picked according to matches
        masks: (B, N, 4) only positive anchors has targets

        """
        F = nd
        # TODO(zhreshold): batch_pick, take multiple elements?
        # refs [B, M, 4], anchors [B, N, 4], samples [B, N], matches [B, N]
        # refs [B, M, 4] -> reshape [B, 1, M, 4] -> repeat [B, N, M, 4]
        ref_boxes = F.repeat(refs.reshape((0, 1, -1, 4)), axis=1, repeats=matches.shape[1])
        # refs [B, N, M, 4] -> 4 * [B, N, M]
        ref_boxes = F.split(ref_boxes, axis=-1, num_outputs=4, squeeze_axis=True)
        # refs 4 * [B, N, M] -> pick from matches [B, N, 1] -> concat to [B, N, 4]
        ref_boxes = F.concat(*[F.pick(ref_boxes[i], matches, axis=2).reshape((0, -1, 1)) \
                               for i in range(4)], dim=2)
        # transform based on x, y, w, h
        # g [B, N, 4], a [B, N, 4] -> codecs [B, N, 4]
        g = self.corner_to_center(ref_boxes)
        a = self.corner_to_center(anchors)
        t0 = ((g[0] - a[0]) / a[2] - self._means[0]) / self._stds[0]
        t1 = ((g[1] - a[1]) / a[3] - self._means[1]) / self._stds[1]
        t2 = (F.log(g[2] / a[2]) - self._means[2]) / self._stds[2]
        t3 = (F.log(g[3] / a[3]) - self._means[3]) / self._stds[3]
        codecs = F.concat(t0, t1, t2, t3, dim=2)
        # samples [B, N] -> [B, N, 1] -> [B, N, 4] -> boolean
        temp = F.tile(samples.reshape((0, -1, 1)), reps=(1, 1, 4)) > 0.5
        # fill targets and masks [B, N, 4]
        targets = F.where(temp, codecs, F.zeros_like(codecs))
        masks = F.where(temp, F.ones_like(temp), F.zeros_like(temp))
        return targets, masks 

def get_final_preds(batch_heatmaps, center, scale):
    coords, maxvals = get_max_pred(batch_heatmaps)

    heatmap_height = batch_heatmaps.shape[2]
    heatmap_width = batch_heatmaps.shape[3]

    # post-processing
    for n in range(coords.shape[0]):
        for p in range(coords.shape[1]):
            hm = batch_heatmaps[n][p]
            px = int(nd.floor(coords[n][p][0] + 0.5).asscalar())
            py = int(nd.floor(coords[n][p][1] + 0.5).asscalar())
            if 1 < px < heatmap_width-1 and 1 < py < heatmap_height-1:
                diff = nd.concat(hm[py][px+1] - hm[py][px-1],
                                 hm[py+1][px] - hm[py-1][px],
                                 dim=0)
                coords[n][p] += nd.sign(diff) * .25

    preds = nd.zeros_like(coords)

    # Transform back
    for i in range(coords.shape[0]):
        preds[i] = transform_preds(coords[i], center[i], scale[i],
                                   [heatmap_width, heatmap_height])

    return preds, maxvals 

def hybrid_forward(self, F, pred, label):
        """Compute loss"""
        softmaxout = F.contrib.SoftmaxOHEMOutput(
            pred, label.astype(pred.dtype), ignore_label=self._ignore_label,
            multi_output=self._sparse_label,
            use_ignore=True, normalization='valid' if self._size_average else 'null',
            thresh=0.6, min_keep=256)
        loss = -F.pick(F.log(softmaxout), label, axis=1, keepdims=True)
        loss = F.where(label.expand_dims(axis=1) == self._ignore_label,
                       F.zeros_like(loss), loss)
        return F.mean(loss, axis=self._batch_axis, exclude=True) 

def hybrid_forward(self, F, pred, label):
        """Compute loss"""
        softmaxout = F.SoftmaxOutput(
            pred, label.astype(pred.dtype), ignore_label=self._ignore_label,
            multi_output=self._sparse_label,
            use_ignore=True, normalization='valid' if self._size_average else 'null')
        loss = -F.pick(F.log(softmaxout), label, axis=1, keepdims=True)
        loss = F.where(label.expand_dims(axis=1) == self._ignore_label,
                       F.zeros_like(loss), loss)
        return F.mean(loss, axis=self._batch_axis, exclude=True) 

def hybrid_forward(self, F, box_preds, gt_boxes):
        """Short summary.

        Parameters
        ----------
        F : mxnet.nd or mxnet.sym
            `F` is mxnet.sym if hybridized or mxnet.nd if not.
        box_preds : mxnet.nd.NDArray
            Predicted bounding boxes.
        gt_boxes : mxnet.nd.NDArray
            Ground-truth bounding boxes.

        Returns
        -------
        (tuple of) mxnet.nd.NDArray
            objectness: 0 for negative, 1 for positive, -1 for ignore.
            center_targets: regression target for center x and y.
            scale_targets: regression target for scale x and y.
            weights: element-wise gradient weights for center_targets and scale_targets.
            class_targets: a one-hot vector for classification.

        """
        with autograd.pause():
            box_preds = box_preds.reshape((0, -1, 4))
            objness_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=1))
            center_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            scale_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            weight_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            class_t = F.ones_like(objness_t.tile(reps=(self._num_class))) * -1
            batch_ious = self._batch_iou(box_preds, gt_boxes)  # (B, N, M)
            ious_max = batch_ious.max(axis=-1, keepdims=True)  # (B, N, 1)
            objness_t = (ious_max > self._ignore_iou_thresh) * -1  # use -1 for ignored
        return objness_t, center_t, scale_t, weight_t, class_t 

def hybrid_forward(self, F, box_preds, gt_boxes):
        """Short summary.

        Parameters
        ----------
        F : mxnet.nd or mxnet.sym
            `F` is mxnet.sym if hybridized or mxnet.nd if not.
        box_preds : mxnet.nd.NDArray
            Predicted bounding boxes.
        gt_boxes : mxnet.nd.NDArray
            Ground-truth bounding boxes.

        Returns
        -------
        (tuple of) mxnet.nd.NDArray
            objectness: 0 for negative, 1 for positive, -1 for ignore.
            center_targets: regression target for center x and y.
            scale_targets: regression target for scale x and y.
            weights: element-wise gradient weights for center_targets and scale_targets.
            class_targets: a one-hot vector for classification.

        """
        with autograd.pause():
            box_preds = box_preds.reshape((0, -1, 4))
            objness_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=1))
            center_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            scale_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            weight_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            class_t = F.ones_like(objness_t.tile(reps=(self._num_class))) * -1
            batch_ious = self._batch_iou(box_preds, gt_boxes)  # (B, N, M)
            ious_max = batch_ious.max(axis=-1, keepdims=True)  # (B, N, 1)
            objness_t = (ious_max > self._ignore_iou_thresh) * -1  # use -1 for ignored
        return objness_t, center_t, scale_t, weight_t, class_t 

def get_final_preds(batch_heatmaps, center, scale):
    coords, maxvals = get_max_pred(batch_heatmaps)

    heatmap_height = batch_heatmaps.shape[2]
    heatmap_width = batch_heatmaps.shape[3]

    # post-processing
    for n in range(coords.shape[0]):
        for p in range(coords.shape[1]):
            hm = batch_heatmaps[n][p]
            px = int(nd.floor(coords[n][p][0] + 0.5).asscalar())
            py = int(nd.floor(coords[n][p][1] + 0.5).asscalar())
            if 1 < px < heatmap_width-1 and 1 < py < heatmap_height-1:
                diff = nd.concat(hm[py][px+1] - hm[py][px-1],
                                 hm[py+1][px] - hm[py-1][px],
                                 dim=0)
                coords[n][p] += nd.sign(diff) * .25

    preds = nd.zeros_like(coords)

    # Transform back
    for i in range(coords.shape[0]):
        preds[i] = transform_preds(coords[i], center[i], scale[i],
                                   [heatmap_width, heatmap_height])

    return preds, maxvals 

def hybrid_forward(self, F, pred, label):
        """Compute loss"""
        softmaxout = F.contrib.SoftmaxOHEMOutput(
            pred, label.astype(pred.dtype), ignore_label=self._ignore_label,
            multi_output=self._sparse_label,
            use_ignore=True, normalization='valid' if self._size_average else 'null',
            thresh=0.6, min_keep=256)
        loss = -F.pick(F.log(softmaxout), label, axis=1, keepdims=True)
        loss = F.where(label.expand_dims(axis=1) == self._ignore_label,
                       F.zeros_like(loss), loss)
        return F.mean(loss, axis=self._batch_axis, exclude=True) 

def hybrid_forward(self, F, pred, label):
        """Compute loss"""
        softmaxout = F.SoftmaxOutput(
            pred, label.astype(pred.dtype), ignore_label=self._ignore_label,
            multi_output=self._sparse_label,
            use_ignore=True, normalization='valid' if self._size_average else 'null')
        if self._sparse_label:
            loss = -F.pick(F.log(softmaxout), label, axis=1, keepdims=True)
        else:
            label = _reshape_like(F, label, pred)
            loss = -F.sum(F.log(softmaxout) * label, axis=-1, keepdims=True)
        loss = F.where(label.expand_dims(axis=1) == self._ignore_label,
                       F.zeros_like(loss), loss)
        return F.mean(loss, axis=self._batch_axis, exclude=True) 

def hybrid_forward(self, F, box_preds, gt_boxes):
        """Short summary.

        Parameters
        ----------
        F : mxnet.nd or mxnet.sym
            `F` is mxnet.sym if hybridized or mxnet.nd if not.
        box_preds : mxnet.nd.NDArray
            Predicted bounding boxes.
        gt_boxes : mxnet.nd.NDArray
            Ground-truth bounding boxes.

        Returns
        -------
        (tuple of) mxnet.nd.NDArray
            objectness: 0 for negative, 1 for positive, -1 for ignore.
            center_targets: regression target for center x and y.
            scale_targets: regression target for scale x and y.
            weights: element-wise gradient weights for center_targets and scale_targets.
            class_targets: a one-hot vector for classification.

        """
        with autograd.pause():
            box_preds = box_preds.reshape((0, -1, 4))
            objness_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=1))
            center_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            scale_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            weight_t = F.zeros_like(box_preds.slice_axis(axis=-1, begin=0, end=2))
            class_t = F.ones_like(objness_t.tile(reps=(self._num_class))) * -1
            batch_ious = self._batch_iou(box_preds, gt_boxes)  # (B, N, M)
            ious_max = batch_ious.max(axis=-1, keepdims=True)  # (B, N, 1)
            objness_t = (ious_max > self._ignore_iou_thresh) * -1  # use -1 for ignored
        return objness_t, center_t, scale_t, weight_t, class_t 

def build_graph_validate_gt_obj(img, gt_ids, bbox, spatial_feat,
                                bbox_improvement=True, overlap=False):
    '''given ground truth bbox and label, build graph for validation'''
    n_batch = img.shape[0]
    img_size = img.shape[2:4]
    bbox[:, :, 0] /= img_size[1]
    bbox[:, :, 1] /= img_size[0]
    bbox[:, :, 2] /= img_size[1]
    bbox[:, :, 3] /= img_size[0]
    ctx = img.context

    g_batch = []
    for btc in range(n_batch):
        inds = np.where(bbox[btc].sum(1).asnumpy() > 0)[0].tolist()
        if len(inds) == 0:
            continue
        n_nodes = len(inds)
        g_pred = dgl.DGLGraph()
        g_pred.add_nodes(n_nodes, {'pred_bbox': bbox[btc, inds],
                                   'node_feat': spatial_feat[btc, inds],
                                   'node_class_pred': gt_ids[btc, inds, 0],
                                   'node_class_logit': nd.zeros_like(gt_ids[btc, inds, 0], ctx=ctx)})

        edge_list = []
        for i in range(n_nodes - 1):
            for j in range(i + 1, n_nodes):
                edge_list.append((i, j))
        src, dst = tuple(zip(*edge_list))
        g_pred.add_edges(src, dst)
        g_pred.add_edges(dst, src)

        n_nodes = g_pred.number_of_nodes()
        n_edges = g_pred.number_of_edges()
        if bbox_improvement:
            g_pred.ndata['pred_bbox'] = bbox_improve(g_pred.ndata['pred_bbox'])
        g_pred.edata['rel_bbox'] = extract_edge_bbox(g_pred)
        g_pred.edata['batch_id'] = nd.zeros((n_edges, 1), ctx = ctx) + btc

        g_batch.append(g_pred)

    if len(g_batch) == 0:
        return None 
    if len(g_batch) > 1:
        return dgl.batch(g_batch)
    return g_batch[0] 

def refine_bbox_nd(bbox, bbox_delta, im_info=None, means=None, stds=None):

    xmin, ymin, xmax, ymax = nd.split(data=bbox, num_outputs=4, axis=1)
    bbox_width = xmax - xmin + 1.
    bbox_height = ymax - ymin + 1.
    center_x = 0.5 * (xmin + xmax)
    center_y = 0.5 * (ymin + ymax)

    bbox_delta_reshape = nd.Reshape(data=bbox_delta, shape=(0, -1, 4))
    dx, dy, dw, dh = nd.split(data=bbox_delta_reshape, 
        num_outputs=4, axis=2, squeeze_axis=1)
    if (means is not None) and (stds is not None):
        dx = dx * stds[0] + means[0]
        dy = dy * stds[1] + means[1]
        dw = dw * stds[2] + means[2]
        dh = dh * stds[3] + means[3]

    refine_center_x = nd.broadcast_add(lhs=center_x,
        rhs=nd.broadcast_mul(lhs=bbox_width, rhs=dx))
    refine_center_y = nd.broadcast_add(lhs=center_y,
        rhs=nd.broadcast_mul(lhs=bbox_height, rhs=dy))
    refined_width = nd.broadcast_mul(lhs=bbox_width, rhs=nd.exp(dw))
    refined_height = nd.broadcast_mul(lhs=bbox_height, rhs=nd.exp(dh))
    w_offset = 0.5 * (refined_width - 1.)
    h_offset = 0.5 * (refined_height - 1.)
    refined_xmin = nd.expand_dims(refine_center_x - w_offset, axis=1)
    refined_ymin = nd.expand_dims(refine_center_y - h_offset, axis=1)
    refined_xmax = nd.expand_dims(refine_center_x + w_offset, axis=1)
    refined_ymax = nd.expand_dims(refine_center_y + h_offset, axis=1)

    refined_bbox = nd.concat(refined_xmin, refined_ymin, refined_xmax, refined_ymax, dim=1)
    if im_info is not None:
        # assume im_info [[height, width, scale]] with shape (1,3)
        im_hw = nd.slice_axis(im_info, axis=1, begin=0, end=2)
        im_wh = nd.reverse(im_hw, axis=1)
        im_wh = im_wh - 1.
        im_wh = nd.tile(data=im_wh, reps=(1, 2))
        im_wh = nd.Reshape(im_wh, shape=(1, 4, 1))
        refined_bbox = nd.broadcast_minimum(lhs=refined_bbox, rhs=im_wh)
        refined_bbox = nd.broadcast_maximum(lhs=refined_bbox,
            rhs=nd.zeros_like(refined_bbox))
    # print refined_bbox.debug_str()
    return refined_bbox 

def refine_bbox_nd(bbox, bbox_delta, im_info=None, means=None, stds=None):

    xmin, ymin, xmax, ymax = nd.split(data=bbox, num_outputs=4, axis=1)
    bbox_width = xmax - xmin + 1.
    bbox_height = ymax - ymin + 1.
    center_x = 0.5 * (xmin + xmax)
    center_y = 0.5 * (ymin + ymax)

    bbox_delta_reshape = nd.Reshape(data=bbox_delta, shape=(0, -1, 4))
    dx, dy, dw, dh = nd.split(data=bbox_delta_reshape, 
        num_outputs=4, axis=2, squeeze_axis=1)
    if (means is not None) and (stds is not None):
        dx = dx * stds[0] + means[0]
        dy = dy * stds[1] + means[1]
        dw = dw * stds[2] + means[2]
        dh = dh * stds[3] + means[3]

    refine_center_x = nd.broadcast_add(lhs=center_x,
        rhs=nd.broadcast_mul(lhs=bbox_width, rhs=dx))
    refine_center_y = nd.broadcast_add(lhs=center_y,
        rhs=nd.broadcast_mul(lhs=bbox_height, rhs=dy))
    refined_width = nd.broadcast_mul(lhs=bbox_width, rhs=nd.exp(dw))
    refined_height = nd.broadcast_mul(lhs=bbox_height, rhs=nd.exp(dh))
    w_offset = 0.5 * (refined_width - 1.)
    h_offset = 0.5 * (refined_height - 1.)
    refined_xmin = nd.expand_dims(refine_center_x - w_offset, axis=1)
    refined_ymin = nd.expand_dims(refine_center_y - h_offset, axis=1)
    refined_xmax = nd.expand_dims(refine_center_x + w_offset, axis=1)
    refined_ymax = nd.expand_dims(refine_center_y + h_offset, axis=1)

    refined_bbox = nd.concat(refined_xmin, refined_ymin, refined_xmax, refined_ymax, dim=1)
    if im_info is not None:
        # assume im_info [[height, width, scale]] with shape (1,3)
        im_hw = nd.slice_axis(im_info, axis=1, begin=0, end=2)
        im_wh = nd.reverse(im_hw, axis=1)
        im_wh = im_wh - 1.
        im_wh = nd.tile(data=im_wh, reps=(1, 2))
        im_wh = nd.Reshape(im_wh, shape=(1, 4, 1))
        refined_bbox = nd.broadcast_minimum(lhs=refined_bbox, rhs=im_wh)
        refined_bbox = nd.broadcast_maximum(lhs=refined_bbox,
            rhs=nd.zeros_like(refined_bbox))
    # print refined_bbox.debug_str()
    return refined_bbox 

def forward(self, cls_pred, box_pred, cls_target, box_target):
        """Compute loss in entire batch across devices."""
        # require results across different devices at this time
        cls_pred, box_pred, cls_target, box_target = [_as_list(x) \
            for x in (cls_pred, box_pred, cls_target, box_target)]
        # cross device reduction to obtain positive samples in entire batch
        num_pos = []
        for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
            pos_samples = (ct > 0)
            num_pos.append(pos_samples.sum())
        num_pos_all = sum([p.asscalar() for p in num_pos])
        if num_pos_all < 1 and self._min_hard_negatives < 1:
            # no positive samples and no hard negatives, return dummy losses
            cls_losses = [nd.sum(cp * 0) for cp in cls_pred]
            box_losses = [nd.sum(bp * 0) for bp in box_pred]
            sum_losses = [nd.sum(cp * 0) + nd.sum(bp * 0) for cp, bp in zip(cls_pred, box_pred)]
            return sum_losses, cls_losses, box_losses


        # compute element-wise cross entropy loss and sort, then perform negative mining
        cls_losses = []
        box_losses = []
        sum_losses = []
        for cp, bp, ct, bt in zip(*[cls_pred, box_pred, cls_target, box_target]):
            pred = nd.log_softmax(cp, axis=-1)
            pos = ct > 0
            cls_loss = -nd.pick(pred, ct, axis=-1, keepdims=False)
            rank = (cls_loss * (pos - 1)).argsort(axis=1).argsort(axis=1)
            hard_negative = rank < nd.maximum(self._min_hard_negatives, pos.sum(axis=1)
                                              * self._negative_mining_ratio).expand_dims(-1)
            # mask out if not positive or negative
            cls_loss = nd.where((pos + hard_negative) > 0, cls_loss, nd.zeros_like(cls_loss))
            cls_losses.append(nd.sum(cls_loss, axis=0, exclude=True) / max(1., num_pos_all))

            bp = _reshape_like(nd, bp, bt)
            box_loss = nd.abs(bp - bt)
            box_loss = nd.where(box_loss > self._rho, box_loss - 0.5 * self._rho,
                                (0.5 / self._rho) * nd.square(box_loss))
            # box loss only apply to positive samples
            box_loss = box_loss * pos.expand_dims(axis=-1)
            box_losses.append(nd.sum(box_loss, axis=0, exclude=True) / max(1., num_pos_all))
            sum_losses.append(cls_losses[-1] + self._lambd * box_losses[-1])

        return sum_losses, cls_losses, box_losses