ops.py

# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy as np
from numbers import Integral
import math
import six

from paddle import fluid
from paddle.fluid.layer_helper import LayerHelper
from paddle.fluid.initializer import NumpyArrayInitializer
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.regularizer import L2Decay
from ppdet.core.workspace import register, serializable
from ppdet.utils.bbox_utils import bbox_overlaps, box_to_delta

__all__ = [
    'AnchorGenerator', 'AnchorGrid', 'DropBlock', 'RPNTargetAssign',
    'GenerateProposals', 'MultiClassNMS', 'BBoxAssigner', 'MaskAssigner',
    'RoIAlign', 'RoIPool', 'MultiBoxHead', 'SSDLiteMultiBoxHead',
    'SSDOutputDecoder', 'RetinaTargetAssign', 'RetinaOutputDecoder', 'ConvNorm',
    'DeformConvNorm', 'MultiClassSoftNMS', 'LibraBBoxAssigner'
]


def _conv_offset(input, filter_size, stride, padding, act=None, name=None):
    out_channel = filter_size * filter_size * 3
    out = fluid.layers.conv2d(
        input,
        num_filters=out_channel,
        filter_size=filter_size,
        stride=stride,
        padding=padding,
        param_attr=ParamAttr(
            initializer=fluid.initializer.Constant(value=0),
            name=name + ".w_0"),
        bias_attr=ParamAttr(
            initializer=fluid.initializer.Constant(value=0),
            name=name + ".b_0"),
        act=act,
        name=name)
    return out


def DeformConvNorm(input,
                   num_filters,
                   filter_size,
                   stride=1,
                   groups=1,
                   norm_decay=0.,
                   norm_type='affine_channel',
                   norm_groups=32,
                   dilation=1,
                   lr_scale=1,
                   freeze_norm=False,
                   act=None,
                   norm_name=None,
                   initializer=None,
                   bias_attr=False,
                   name=None):
    if bias_attr:
        bias_para = ParamAttr(
            name=name + "_bias",
            initializer=fluid.initializer.Constant(value=0),
            learning_rate=lr_scale * 2)
    else:
        bias_para = False
    offset_mask = _conv_offset(
        input=input,
        filter_size=filter_size,
        stride=stride,
        padding=(filter_size - 1) // 2,
        act=None,
        name=name + "_conv_offset")
    offset_channel = filter_size**2 * 2
    mask_channel = filter_size**2
    offset, mask = fluid.layers.split(
        input=offset_mask,
        num_or_sections=[offset_channel, mask_channel],
        dim=1)
    mask = fluid.layers.sigmoid(mask)
    conv = fluid.layers.deformable_conv(
        input=input,
        offset=offset,
        mask=mask,
        num_filters=num_filters,
        filter_size=filter_size,
        stride=stride,
        padding=(filter_size - 1) // 2 * dilation,
        dilation=dilation,
        groups=groups,
        deformable_groups=1,
        im2col_step=1,
        param_attr=ParamAttr(
            name=name + "_weights",
            initializer=initializer,
            learning_rate=lr_scale),
        bias_attr=bias_para,
        name=name + ".conv2d.output.1")

    norm_lr = 0. if freeze_norm else 1.
    pattr = ParamAttr(
        name=norm_name + '_scale',
        learning_rate=norm_lr * lr_scale,
        regularizer=L2Decay(norm_decay))
    battr = ParamAttr(
        name=norm_name + '_offset',
        learning_rate=norm_lr * lr_scale,
        regularizer=L2Decay(norm_decay))

    if norm_type in ['bn', 'sync_bn']:
        global_stats = True if freeze_norm else False
        out = fluid.layers.batch_norm(
            input=conv,
            act=act,
            name=norm_name + '.output.1',
            param_attr=pattr,
            bias_attr=battr,
            moving_mean_name=norm_name + '_mean',
            moving_variance_name=norm_name + '_variance',
            use_global_stats=global_stats)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
    elif norm_type == 'gn':
        out = fluid.layers.group_norm(
            input=conv,
            act=act,
            name=norm_name + '.output.1',
            groups=norm_groups,
            param_attr=pattr,
            bias_attr=battr)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
    elif norm_type == 'affine_channel':
        scale = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=pattr,
            default_initializer=fluid.initializer.Constant(1.))
        bias = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=battr,
            default_initializer=fluid.initializer.Constant(0.))
        out = fluid.layers.affine_channel(
            x=conv, scale=scale, bias=bias, act=act)

    if freeze_norm:
        scale.stop_gradient = True
        bias.stop_gradient = True
    return out


def ConvNorm(input,
             num_filters,
             filter_size,
             stride=1,
             groups=1,
             norm_decay=0.,
             norm_type='affine_channel',
             norm_groups=32,
             dilation=1,
             lr_scale=1,
             freeze_norm=False,
             act=None,
             norm_name=None,
             initializer=None,
             bias_attr=False,
             name=None):
    fan = num_filters
    if bias_attr:
        bias_para = ParamAttr(
            name=name + "_bias",
            initializer=fluid.initializer.Constant(value=0),
            learning_rate=lr_scale * 2)
    else:
        bias_para = False
    conv = fluid.layers.conv2d(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
        stride=stride,
        padding=((filter_size - 1) // 2) * dilation,
        dilation=dilation,
        groups=groups,
        act=None,
        param_attr=ParamAttr(
            name=name + "_weights",
            initializer=initializer,
            learning_rate=lr_scale),
        bias_attr=bias_para,
        name=name + '.conv2d.output.1')

    norm_lr = 0. if freeze_norm else 1.
    pattr = ParamAttr(
        name=norm_name + '_scale',
        learning_rate=norm_lr * lr_scale,
        regularizer=L2Decay(norm_decay))
    battr = ParamAttr(
        name=norm_name + '_offset',
        learning_rate=norm_lr * lr_scale,
        regularizer=L2Decay(norm_decay))

    if norm_type in ['bn', 'sync_bn']:
        global_stats = True if freeze_norm else False
        out = fluid.layers.batch_norm(
            input=conv,
            act=act,
            name=norm_name + '.output.1',
            param_attr=pattr,
            bias_attr=battr,
            moving_mean_name=norm_name + '_mean',
            moving_variance_name=norm_name + '_variance',
            use_global_stats=global_stats)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
    elif norm_type == 'gn':
        out = fluid.layers.group_norm(
            input=conv,
            act=act,
            name=norm_name + '.output.1',
            groups=norm_groups,
            param_attr=pattr,
            bias_attr=battr)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
    elif norm_type == 'affine_channel':
        scale = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=pattr,
            default_initializer=fluid.initializer.Constant(1.))
        bias = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=battr,
            default_initializer=fluid.initializer.Constant(0.))
        out = fluid.layers.affine_channel(
            x=conv, scale=scale, bias=bias, act=act)
    if freeze_norm:
        scale.stop_gradient = True
        bias.stop_gradient = True
    return out


def DropBlock(input, block_size, keep_prob, is_test):
    if is_test:
        return input

    def CalculateGamma(input, block_size, keep_prob):
        input_shape = fluid.layers.shape(input)
        feat_shape_tmp = fluid.layers.slice(input_shape, [0], [3], [4])
        feat_shape_tmp = fluid.layers.cast(feat_shape_tmp, dtype="float32")
        feat_shape_t = fluid.layers.reshape(feat_shape_tmp, [1, 1, 1, 1])
        feat_area = fluid.layers.pow(feat_shape_t, factor=2)

        block_shape_t = fluid.layers.fill_constant(
            shape=[1, 1, 1, 1], value=block_size, dtype='float32')
        block_area = fluid.layers.pow(block_shape_t, factor=2)

        useful_shape_t = feat_shape_t - block_shape_t + 1
        useful_area = fluid.layers.pow(useful_shape_t, factor=2)

        upper_t = feat_area * (1 - keep_prob)
        bottom_t = block_area * useful_area
        output = upper_t / bottom_t
        return output

    gamma = CalculateGamma(input, block_size=block_size, keep_prob=keep_prob)
    input_shape = fluid.layers.shape(input)
    p = fluid.layers.expand_as(gamma, input)

    input_shape_tmp = fluid.layers.cast(input_shape, dtype="int64")
    random_matrix = fluid.layers.uniform_random(
        input_shape_tmp, dtype='float32', min=0.0, max=1.0)
    one_zero_m = fluid.layers.less_than(random_matrix, p)
    one_zero_m.stop_gradient = True
    one_zero_m = fluid.layers.cast(one_zero_m, dtype="float32")

    mask_flag = fluid.layers.pool2d(
        one_zero_m,
        pool_size=block_size,
        pool_type='max',
        pool_stride=1,
        pool_padding=block_size // 2)
    mask = 1.0 - mask_flag

    elem_numel = fluid.layers.reduce_prod(input_shape)
    elem_numel_m = fluid.layers.cast(elem_numel, dtype="float32")
    elem_numel_m.stop_gradient = True

    elem_sum = fluid.layers.reduce_sum(mask)
    elem_sum_m = fluid.layers.cast(elem_sum, dtype="float32")
    elem_sum_m.stop_gradient = True

    output = input * mask * elem_numel_m / elem_sum_m
    return output


@register
@serializable
class AnchorGenerator(object):
    __op__ = fluid.layers.anchor_generator
    __append_doc__ = True

    def __init__(self,
                 stride=[16.0, 16.0],
                 anchor_sizes=[32, 64, 128, 256, 512],
                 aspect_ratios=[0.5, 1., 2.],
                 variance=[1., 1., 1., 1.]):
        super(AnchorGenerator, self).__init__()
        self.anchor_sizes = anchor_sizes
        self.aspect_ratios = aspect_ratios
        self.variance = variance
        self.stride = stride


@register
@serializable
class AnchorGrid(object):
    """Generate anchor grid

    Args:
        image_size (int or list): input image size, may be a single integer or
            list of [h, w]. Default: 512
        min_level (int): min level of the feature pyramid. Default: 3
        max_level (int): max level of the feature pyramid. Default: 7
        anchor_base_scale: base anchor scale. Default: 4
        num_scales: number of anchor scales. Default: 3
        aspect_ratios: aspect ratios. default: [[1, 1], [1.4, 0.7], [0.7, 1.4]]
    """

    def __init__(self,
                 image_size=512,
                 min_level=3,
                 max_level=7,
                 anchor_base_scale=4,
                 num_scales=3,
                 aspect_ratios=[[1, 1], [1.4, 0.7], [0.7, 1.4]]):
        super(AnchorGrid, self).__init__()
        if isinstance(image_size, Integral):
            self.image_size = [image_size, image_size]
        else:
            self.image_size = image_size
        for dim in self.image_size:
            assert dim % 2 ** max_level == 0, \
                "image size should be multiple of the max level stride"
        self.min_level = min_level
        self.max_level = max_level
        self.anchor_base_scale = anchor_base_scale
        self.num_scales = num_scales
        self.aspect_ratios = aspect_ratios

    @property
    def base_cell(self):
        if not hasattr(self, '_base_cell'):
            self._base_cell = self.make_cell()
        return self._base_cell

    def make_cell(self):
        scales = [2**(i / self.num_scales) for i in range(self.num_scales)]
        scales = np.array(scales)
        ratios = np.array(self.aspect_ratios)
        ws = np.outer(scales, ratios[:, 0]).reshape(-1, 1)
        hs = np.outer(scales, ratios[:, 1]).reshape(-1, 1)
        anchors = np.hstack((-0.5 * ws, -0.5 * hs, 0.5 * ws, 0.5 * hs))
        return anchors

    def make_grid(self, stride):
        cell = self.base_cell * stride * self.anchor_base_scale
        x_steps = np.arange(stride // 2, self.image_size[1], stride)
        y_steps = np.arange(stride // 2, self.image_size[0], stride)
        offset_x, offset_y = np.meshgrid(x_steps, y_steps)
        offset_x = offset_x.flatten()
        offset_y = offset_y.flatten()
        offsets = np.stack((offset_x, offset_y, offset_x, offset_y), axis=-1)
        offsets = offsets[:, np.newaxis, :]
        return (cell + offsets).reshape(-1, 4)

    def generate(self):
        return [
            self.make_grid(2**l)
            for l in range(self.min_level, self.max_level + 1)
        ]

    def __call__(self):
        if not hasattr(self, '_anchor_vars'):
            anchor_vars = []
            helper = LayerHelper('anchor_grid')
            for idx, l in enumerate(range(self.min_level, self.max_level + 1)):
                stride = 2**l
                anchors = self.make_grid(stride)
                var = helper.create_parameter(
                    attr=ParamAttr(name='anchors_{}'.format(idx)),
                    shape=anchors.shape,
                    dtype='float32',
                    stop_gradient=True,
                    default_initializer=NumpyArrayInitializer(anchors))
                anchor_vars.append(var)
                var.persistable = True
            self._anchor_vars = anchor_vars

        return self._anchor_vars


@register
@serializable
class RPNTargetAssign(object):
    __op__ = fluid.layers.rpn_target_assign
    __append_doc__ = True

    def __init__(self,
                 rpn_batch_size_per_im=256,
                 rpn_straddle_thresh=0.,
                 rpn_fg_fraction=0.5,
                 rpn_positive_overlap=0.7,
                 rpn_negative_overlap=0.3,
                 use_random=True):
        super(RPNTargetAssign, self).__init__()
        self.rpn_batch_size_per_im = rpn_batch_size_per_im
        self.rpn_straddle_thresh = rpn_straddle_thresh
        self.rpn_fg_fraction = rpn_fg_fraction
        self.rpn_positive_overlap = rpn_positive_overlap
        self.rpn_negative_overlap = rpn_negative_overlap
        self.use_random = use_random


@register
@serializable
class GenerateProposals(object):
    __op__ = fluid.layers.generate_proposals
    __append_doc__ = True

    def __init__(self,
                 pre_nms_top_n=6000,
                 post_nms_top_n=1000,
                 nms_thresh=.5,
                 min_size=.1,
                 eta=1.):
        super(GenerateProposals, self).__init__()
        self.pre_nms_top_n = pre_nms_top_n
        self.post_nms_top_n = post_nms_top_n
        self.nms_thresh = nms_thresh
        self.min_size = min_size
        self.eta = eta


@register
class MaskAssigner(object):
    __op__ = fluid.layers.generate_mask_labels
    __append_doc__ = True
    __shared__ = ['num_classes']

    def __init__(self, num_classes=81, resolution=14):
        super(MaskAssigner, self).__init__()
        self.num_classes = num_classes
        self.resolution = resolution


@register
@serializable
class MultiClassNMS(object):
    __op__ = fluid.layers.multiclass_nms
    __append_doc__ = True

    def __init__(self,
                 score_threshold=.05,
                 nms_top_k=-1,
                 keep_top_k=100,
                 nms_threshold=.5,
                 normalized=False,
                 nms_eta=1.0,
                 background_label=0):
        super(MultiClassNMS, self).__init__()
        self.score_threshold = score_threshold
        self.nms_top_k = nms_top_k
        self.keep_top_k = keep_top_k
        self.nms_threshold = nms_threshold
        self.normalized = normalized
        self.nms_eta = nms_eta
        self.background_label = background_label


@register
@serializable
class MultiClassSoftNMS(object):
    def __init__(
            self,
            score_threshold=0.01,
            keep_top_k=300,
            softnms_sigma=0.5,
            normalized=False,
            background_label=0, ):
        super(MultiClassSoftNMS, self).__init__()
        self.score_threshold = score_threshold
        self.keep_top_k = keep_top_k
        self.softnms_sigma = softnms_sigma
        self.normalized = normalized
        self.background_label = background_label

    def __call__(self, bboxes, scores):
        def create_tmp_var(program, name, dtype, shape, lod_level):
            return program.current_block().create_var(
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)

        def _soft_nms_for_cls(dets, sigma, thres):
            """soft_nms_for_cls"""
            dets_final = []
            while len(dets) > 0:
                maxpos = np.argmax(dets[:, 0])
                dets_final.append(dets[maxpos].copy())
                ts, tx1, ty1, tx2, ty2 = dets[maxpos]
                scores = dets[:, 0]
                # force remove bbox at maxpos
                scores[maxpos] = -1
                x1 = dets[:, 1]
                y1 = dets[:, 2]
                x2 = dets[:, 3]
                y2 = dets[:, 4]
                eta = 0 if self.normalized else 1
                areas = (x2 - x1 + eta) * (y2 - y1 + eta)
                xx1 = np.maximum(tx1, x1)
                yy1 = np.maximum(ty1, y1)
                xx2 = np.minimum(tx2, x2)
                yy2 = np.minimum(ty2, y2)
                w = np.maximum(0.0, xx2 - xx1 + eta)
                h = np.maximum(0.0, yy2 - yy1 + eta)
                inter = w * h
                ovr = inter / (areas + areas[maxpos] - inter)
                weight = np.exp(-(ovr * ovr) / sigma)
                scores = scores * weight
                idx_keep = np.where(scores >= thres)
                dets[:, 0] = scores
                dets = dets[idx_keep]
            dets_final = np.array(dets_final).reshape(-1, 5)
            return dets_final

        def _soft_nms(bboxes, scores):
            class_nums = scores.shape[-1]

            softnms_thres = self.score_threshold
            softnms_sigma = self.softnms_sigma
            keep_top_k = self.keep_top_k

            cls_boxes = [[] for _ in range(class_nums)]
            cls_ids = [[] for _ in range(class_nums)]

            start_idx = 1 if self.background_label == 0 else 0
            for j in range(start_idx, class_nums):
                inds = np.where(scores[:, j] >= softnms_thres)[0]
                scores_j = scores[inds, j]
                rois_j = bboxes[inds, j, :] if len(
                    bboxes.shape) > 2 else bboxes[inds, :]
                dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
                    np.float32, copy=False)
                cls_rank = np.argsort(-dets_j[:, 0])
                dets_j = dets_j[cls_rank]

                cls_boxes[j] = _soft_nms_for_cls(
                    dets_j, sigma=softnms_sigma, thres=softnms_thres)
                cls_ids[j] = np.array([j] * cls_boxes[j].shape[0]).reshape(-1,
                                                                           1)

            cls_boxes = np.vstack(cls_boxes[start_idx:])
            cls_ids = np.vstack(cls_ids[start_idx:])
            pred_result = np.hstack([cls_ids, cls_boxes])

            # Limit to max_per_image detections **over all classes**
            image_scores = cls_boxes[:, 0]
            if len(image_scores) > keep_top_k:
                image_thresh = np.sort(image_scores)[-keep_top_k]
                keep = np.where(cls_boxes[:, 0] >= image_thresh)[0]
                pred_result = pred_result[keep, :]

            return pred_result

        def _batch_softnms(bboxes, scores):
            batch_offsets = bboxes.lod()
            bboxes = np.array(bboxes)
            scores = np.array(scores)
            out_offsets = [0]
            pred_res = []
            if len(batch_offsets) > 0:
                batch_offset = batch_offsets[0]
                for i in range(len(batch_offset) - 1):
                    s, e = batch_offset[i], batch_offset[i + 1]
                    pred = _soft_nms(bboxes[s:e], scores[s:e])
                    out_offsets.append(pred.shape[0] + out_offsets[-1])
                    pred_res.append(pred)
            else:
                assert len(bboxes.shape) == 3
                assert len(scores.shape) == 3
                for i in range(bboxes.shape[0]):
                    pred = _soft_nms(bboxes[i], scores[i])
                    out_offsets.append(pred.shape[0] + out_offsets[-1])
                    pred_res.append(pred)

            res = fluid.LoDTensor()
            res.set_lod([out_offsets])
            if len(pred_res) == 0:
                pred_res = np.array([[1]], dtype=np.float32)
            res.set(np.vstack(pred_res).astype(np.float32), fluid.CPUPlace())
            return res

        pred_result = create_tmp_var(
            fluid.default_main_program(),
            name='softnms_pred_result',
            dtype='float32',
            shape=[-1, 6],
            lod_level=1)
        fluid.layers.py_func(
            func=_batch_softnms, x=[bboxes, scores], out=pred_result)
        return pred_result


@register
@serializable
class MultiClassDiouNMS(object):
    def __init__(
            self,
            score_threshold=0.05,
            keep_top_k=100,
            nms_threshold=0.5,
            normalized=False,
            background_label=0, ):
        super(MultiClassDiouNMS, self).__init__()
        self.score_threshold = score_threshold
        self.nms_threshold = nms_threshold
        self.keep_top_k = keep_top_k
        self.normalized = normalized
        self.background_label = background_label

    def __call__(self, bboxes, scores):
        def create_tmp_var(program, name, dtype, shape, lod_level):
            return program.current_block().create_var(
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)

        def _calc_diou_term(dets1, dets2):
            eps = 1.e-10
            eta = 0 if self.normalized else 1

            x1, y1, x2, y2 = dets1[0], dets1[1], dets1[2], dets1[3]
            x1g, y1g, x2g, y2g = dets2[0], dets2[1], dets2[2], dets2[3]

            cx = (x1 + x2) / 2
            cy = (y1 + y2) / 2
            w = x2 - x1 + eta
            h = y2 - y1 + eta

            cxg = (x1g + x2g) / 2
            cyg = (y1g + y2g) / 2
            wg = x2g - x1g + eta
            hg = y2g - y1g + eta

            x2 = np.maximum(x1, x2)
            y2 = np.maximum(y1, y2)

            # A or B
            xc1 = np.minimum(x1, x1g)
            yc1 = np.minimum(y1, y1g)
            xc2 = np.maximum(x2, x2g)
            yc2 = np.maximum(y2, y2g)

            # DIOU term
            dist_intersection = (cx - cxg)**2 + (cy - cyg)**2
            dist_union = (xc2 - xc1)**2 + (yc2 - yc1)**2
            diou_term = (dist_intersection + eps) / (dist_union + eps)
            return diou_term

        def _diou_nms_for_cls(dets, thres):
            """_diou_nms_for_cls"""
            scores = dets[:, 0]
            x1 = dets[:, 1]
            y1 = dets[:, 2]
            x2 = dets[:, 3]
            y2 = dets[:, 4]
            eta = 0 if self.normalized else 1
            areas = (x2 - x1 + eta) * (y2 - y1 + eta)
            dt_num = dets.shape[0]
            order = np.array(range(dt_num))

            keep = []
            while order.size > 0:
                i = order[0]
                keep.append(i)
                xx1 = np.maximum(x1[i], x1[order[1:]])
                yy1 = np.maximum(y1[i], y1[order[1:]])
                xx2 = np.minimum(x2[i], x2[order[1:]])
                yy2 = np.minimum(y2[i], y2[order[1:]])

                w = np.maximum(0.0, xx2 - xx1 + eta)
                h = np.maximum(0.0, yy2 - yy1 + eta)
                inter = w * h
                ovr = inter / (areas[i] + areas[order[1:]] - inter)

                diou_term = _calc_diou_term([x1[i], y1[i], x2[i], y2[i]], [
                    x1[order[1:]], y1[order[1:]], x2[order[1:]], y2[order[1:]]
                ])

                inds = np.where(ovr - diou_term <= thres)[0]

                order = order[inds + 1]

            dets_final = dets[keep]
            return dets_final

        def _diou_nms(bboxes, scores):
            bboxes = np.array(bboxes)
            scores = np.array(scores)
            class_nums = scores.shape[-1]

            score_threshold = self.score_threshold
            nms_threshold = self.nms_threshold
            keep_top_k = self.keep_top_k

            cls_boxes = [[] for _ in range(class_nums)]
            cls_ids = [[] for _ in range(class_nums)]

            start_idx = 1 if self.background_label == 0 else 0
            for j in range(start_idx, class_nums):
                inds = np.where(scores[:, j] >= score_threshold)[0]
                scores_j = scores[inds, j]
                rois_j = bboxes[inds, j, :]
                dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
                    np.float32, copy=False)
                cls_rank = np.argsort(-dets_j[:, 0])
                dets_j = dets_j[cls_rank]

                cls_boxes[j] = _diou_nms_for_cls(dets_j, thres=nms_threshold)
                cls_ids[j] = np.array([j] * cls_boxes[j].shape[0]).reshape(-1,
                                                                           1)

            cls_boxes = np.vstack(cls_boxes[start_idx:])
            cls_ids = np.vstack(cls_ids[start_idx:])
            pred_result = np.hstack([cls_ids, cls_boxes]).astype(np.float32)

            # Limit to max_per_image detections **over all classes**
            image_scores = cls_boxes[:, 0]
            if len(image_scores) > keep_top_k:
                image_thresh = np.sort(image_scores)[-keep_top_k]
                keep = np.where(cls_boxes[:, 0] >= image_thresh)[0]
                pred_result = pred_result[keep, :]

            res = fluid.LoDTensor()
            res.set_lod([[0, pred_result.shape[0]]])
            if pred_result.shape[0] == 0:
                pred_result = np.array([[1]], dtype=np.float32)
            res.set(pred_result, fluid.CPUPlace())

            return res

        pred_result = create_tmp_var(
            fluid.default_main_program(),
            name='diou_nms_pred_result',
            dtype='float32',
            shape=[-1, 6],
            lod_level=0)
        fluid.layers.py_func(
            func=_diou_nms, x=[bboxes, scores], out=pred_result)
        return pred_result


@register
class BBoxAssigner(object):
    __op__ = fluid.layers.generate_proposal_labels
    __append_doc__ = True
    __shared__ = ['num_classes']

    def __init__(self,
                 batch_size_per_im=512,
                 fg_fraction=.25,
                 fg_thresh=.5,
                 bg_thresh_hi=.5,
                 bg_thresh_lo=0.,
                 bbox_reg_weights=[0.1, 0.1, 0.2, 0.2],
                 num_classes=81,
                 shuffle_before_sample=True):
        super(BBoxAssigner, self).__init__()
        self.batch_size_per_im = batch_size_per_im
        self.fg_fraction = fg_fraction
        self.fg_thresh = fg_thresh
        self.bg_thresh_hi = bg_thresh_hi
        self.bg_thresh_lo = bg_thresh_lo
        self.bbox_reg_weights = bbox_reg_weights
        self.class_nums = num_classes
        self.use_random = shuffle_before_sample


@register
class LibraBBoxAssigner(object):
    __shared__ = ['num_classes']

    def __init__(self,
                 batch_size_per_im=512,
                 fg_fraction=.25,
                 fg_thresh=.5,
                 bg_thresh_hi=.5,
                 bg_thresh_lo=0.,
                 bbox_reg_weights=[0.1, 0.1, 0.2, 0.2],
                 num_classes=81,
                 shuffle_before_sample=True,
                 is_cls_agnostic=False,
                 num_bins=3):
        super(LibraBBoxAssigner, self).__init__()
        self.batch_size_per_im = batch_size_per_im
        self.fg_fraction = fg_fraction
        self.fg_thresh = fg_thresh
        self.bg_thresh_hi = bg_thresh_hi
        self.bg_thresh_lo = bg_thresh_lo
        self.bbox_reg_weights = bbox_reg_weights
        self.class_nums = num_classes
        self.use_random = shuffle_before_sample
        self.is_cls_agnostic = is_cls_agnostic
        self.num_bins = num_bins

    def __call__(
            self,
            rpn_rois,
            gt_classes,
            is_crowd,
            gt_boxes,
            im_info, ):
        return self.generate_proposal_label_libra(
            rpn_rois=rpn_rois,
            gt_classes=gt_classes,
            is_crowd=is_crowd,
            gt_boxes=gt_boxes,
            im_info=im_info,
            batch_size_per_im=self.batch_size_per_im,
            fg_fraction=self.fg_fraction,
            fg_thresh=self.fg_thresh,
            bg_thresh_hi=self.bg_thresh_hi,
            bg_thresh_lo=self.bg_thresh_lo,
            bbox_reg_weights=self.bbox_reg_weights,
            class_nums=self.class_nums,
            use_random=self.use_random,
            is_cls_agnostic=self.is_cls_agnostic,
            is_cascade_rcnn=False)

    def generate_proposal_label_libra(
            self, rpn_rois, gt_classes, is_crowd, gt_boxes, im_info,
            batch_size_per_im, fg_fraction, fg_thresh, bg_thresh_hi,
            bg_thresh_lo, bbox_reg_weights, class_nums, use_random,
            is_cls_agnostic, is_cascade_rcnn):
        num_bins = self.num_bins

        def create_tmp_var(program, name, dtype, shape, lod_level=None):
            return program.current_block().create_var(
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)

        def _sample_pos(max_overlaps, max_classes, pos_inds, num_expected):
            if len(pos_inds) <= num_expected:
                return pos_inds
            else:
                unique_gt_inds = np.unique(max_classes[pos_inds])
                num_gts = len(unique_gt_inds)
                num_per_gt = int(round(num_expected / float(num_gts)) + 1)

                sampled_inds = []
                for i in unique_gt_inds:
                    inds = np.nonzero(max_classes == i)[0]
                    before_len = len(inds)
                    inds = list(set(inds) & set(pos_inds))
                    after_len = len(inds)
                    if len(inds) > num_per_gt:
                        inds = np.random.choice(
                            inds, size=num_per_gt, replace=False)
                    sampled_inds.extend(list(inds))  # combine as a new sampler
                if len(sampled_inds) < num_expected:
                    num_extra = num_expected - len(sampled_inds)
                    extra_inds = np.array(
                        list(set(pos_inds) - set(sampled_inds)))
                    assert len(sampled_inds)+len(extra_inds) == len(pos_inds), \
                        "sum of sampled_inds({}) and extra_inds({}) length must be equal with pos_inds({})!".format(
                            len(sampled_inds), len(extra_inds), len(pos_inds))
                    if len(extra_inds) > num_extra:
                        extra_inds = np.random.choice(
                            extra_inds, size=num_extra, replace=False)
                    sampled_inds.extend(extra_inds.tolist())
                elif len(sampled_inds) > num_expected:
                    sampled_inds = np.random.choice(
                        sampled_inds, size=num_expected, replace=False)
                return sampled_inds

        def sample_via_interval(max_overlaps, full_set, num_expected, floor_thr,
                                num_bins, bg_thresh_hi):
            max_iou = max_overlaps.max()
            iou_interval = (max_iou - floor_thr) / num_bins
            per_num_expected = int(num_expected / num_bins)

            sampled_inds = []
            for i in range(num_bins):
                start_iou = floor_thr + i * iou_interval
                end_iou = floor_thr + (i + 1) * iou_interval

                tmp_set = set(
                    np.where(
                        np.logical_and(max_overlaps >= start_iou, max_overlaps <
                                       end_iou))[0])
                tmp_inds = list(tmp_set & full_set)

                if len(tmp_inds) > per_num_expected:
                    tmp_sampled_set = np.random.choice(
                        tmp_inds, size=per_num_expected, replace=False)
                else:
                    tmp_sampled_set = np.array(tmp_inds, dtype=np.int)
                sampled_inds.append(tmp_sampled_set)

            sampled_inds = np.concatenate(sampled_inds)
            if len(sampled_inds) < num_expected:
                num_extra = num_expected - len(sampled_inds)
                extra_inds = np.array(list(full_set - set(sampled_inds)))
                assert len(sampled_inds)+len(extra_inds) == len(full_set), \
                    "sum of sampled_inds({}) and extra_inds({}) length must be equal with full_set({})!".format(
                            len(sampled_inds), len(extra_inds), len(full_set))

                if len(extra_inds) > num_extra:
                    extra_inds = np.random.choice(
                        extra_inds, num_extra, replace=False)
                sampled_inds = np.concatenate([sampled_inds, extra_inds])

            return sampled_inds

        def _sample_neg(max_overlaps,
                        max_classes,
                        neg_inds,
                        num_expected,
                        floor_thr=-1,
                        floor_fraction=0,
                        num_bins=3,
                        bg_thresh_hi=0.5):
            if len(neg_inds) <= num_expected:
                return neg_inds
            else:
                # balance sampling for negative samples
                neg_set = set(neg_inds)
                if floor_thr > 0:
                    floor_set = set(
                        np.where(
                            np.logical_and(max_overlaps >= 0, max_overlaps <
                                           floor_thr))[0])
                    iou_sampling_set = set(
                        np.where(max_overlaps >= floor_thr)[0])
                elif floor_thr == 0:
                    floor_set = set(np.where(max_overlaps == 0)[0])
                    iou_sampling_set = set(
                        np.where(max_overlaps > floor_thr)[0])
                else:
                    floor_set = set()
                    iou_sampling_set = set(
                        np.where(max_overlaps > floor_thr)[0])
                    floor_thr = 0

                floor_neg_inds = list(floor_set & neg_set)
                iou_sampling_neg_inds = list(iou_sampling_set & neg_set)

                num_expected_iou_sampling = int(num_expected *
                                                (1 - floor_fraction))
                if len(iou_sampling_neg_inds) > num_expected_iou_sampling:
                    if num_bins >= 2:
                        iou_sampled_inds = sample_via_interval(
                            max_overlaps,
                            set(iou_sampling_neg_inds),
                            num_expected_iou_sampling, floor_thr, num_bins,
                            bg_thresh_hi)
                    else:
                        iou_sampled_inds = np.random.choice(
                            iou_sampling_neg_inds,
                            size=num_expected_iou_sampling,
                            replace=False)
                else:
                    iou_sampled_inds = np.array(
                        iou_sampling_neg_inds, dtype=np.int)
                num_expected_floor = num_expected - len(iou_sampled_inds)
                if len(floor_neg_inds) > num_expected_floor:
                    sampled_floor_inds = np.random.choice(
                        floor_neg_inds, size=num_expected_floor, replace=False)
                else:
                    sampled_floor_inds = np.array(floor_neg_inds, dtype=np.int)
                sampled_inds = np.concatenate(
                    (sampled_floor_inds, iou_sampled_inds))
                if len(sampled_inds) < num_expected:
                    num_extra = num_expected - len(sampled_inds)
                    extra_inds = np.array(list(neg_set - set(sampled_inds)))
                    if len(extra_inds) > num_extra:
                        extra_inds = np.random.choice(
                            extra_inds, size=num_extra, replace=False)
                    sampled_inds = np.concatenate((sampled_inds, extra_inds))
                return sampled_inds

        def _sample_rois(rpn_rois, gt_classes, is_crowd, gt_boxes, im_info,
                         batch_size_per_im, fg_fraction, fg_thresh,
                         bg_thresh_hi, bg_thresh_lo, bbox_reg_weights,
                         class_nums, use_random, is_cls_agnostic,
                         is_cascade_rcnn):
            rois_per_image = int(batch_size_per_im)
            fg_rois_per_im = int(np.round(fg_fraction * rois_per_image))

            # Roidb
            im_scale = im_info[2]
            inv_im_scale = 1. / im_scale
            rpn_rois = rpn_rois * inv_im_scale
            if is_cascade_rcnn:
                rpn_rois = rpn_rois[gt_boxes.shape[0]:, :]
            boxes = np.vstack([gt_boxes, rpn_rois])
            gt_overlaps = np.zeros((boxes.shape[0], class_nums))
            box_to_gt_ind_map = np.zeros((boxes.shape[0]), dtype=np.int32)
            if len(gt_boxes) > 0:
                proposal_to_gt_overlaps = bbox_overlaps(boxes, gt_boxes)

                overlaps_argmax = proposal_to_gt_overlaps.argmax(axis=1)
                overlaps_max = proposal_to_gt_overlaps.max(axis=1)
                # Boxes which with non-zero overlap with gt boxes
                overlapped_boxes_ind = np.where(overlaps_max > 0)[0]

                overlapped_boxes_gt_classes = gt_classes[overlaps_argmax[
                    overlapped_boxes_ind]]

                for idx in range(len(overlapped_boxes_ind)):
                    gt_overlaps[overlapped_boxes_ind[
                        idx], overlapped_boxes_gt_classes[idx]] = overlaps_max[
                            overlapped_boxes_ind[idx]]
                    box_to_gt_ind_map[overlapped_boxes_ind[
                        idx]] = overlaps_argmax[overlapped_boxes_ind[idx]]

            crowd_ind = np.where(is_crowd)[0]
            gt_overlaps[crowd_ind] = -1

            max_overlaps = gt_overlaps.max(axis=1)
            max_classes = gt_overlaps.argmax(axis=1)

            # Cascade RCNN Decode Filter
            if is_cascade_rcnn:
                ws = boxes[:, 2] - boxes[:, 0] + 1
                hs = boxes[:, 3] - boxes[:, 1] + 1
                keep = np.where((ws > 0) & (hs > 0))[0]
                boxes = boxes[keep]
                max_overlaps = max_overlaps[keep]
                fg_inds = np.where(max_overlaps >= fg_thresh)[0]
                bg_inds = np.where((max_overlaps < bg_thresh_hi) & (
                    max_overlaps >= bg_thresh_lo))[0]
                fg_rois_per_this_image = fg_inds.shape[0]
                bg_rois_per_this_image = bg_inds.shape[0]
            else:
                # Foreground
                fg_inds = np.where(max_overlaps >= fg_thresh)[0]
                fg_rois_per_this_image = np.minimum(fg_rois_per_im,
                                                    fg_inds.shape[0])
                # Sample foreground if there are too many
                if fg_inds.shape[0] > fg_rois_per_this_image:
                    if use_random:
                        fg_inds = _sample_pos(max_overlaps, max_classes,
                                              fg_inds, fg_rois_per_this_image)
                fg_inds = fg_inds[:fg_rois_per_this_image]

                # Background
                bg_inds = np.where((max_overlaps < bg_thresh_hi) & (
                    max_overlaps >= bg_thresh_lo))[0]
                bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
                bg_rois_per_this_image = np.minimum(bg_rois_per_this_image,
                                                    bg_inds.shape[0])
                assert bg_rois_per_this_image >= 0, "bg_rois_per_this_image must be >= 0 but got {}".format(
                    bg_rois_per_this_image)

                # Sample background if there are too many
                if bg_inds.shape[0] > bg_rois_per_this_image:
                    if use_random:
                        # libra neg sample
                        bg_inds = _sample_neg(
                            max_overlaps,
                            max_classes,
                            bg_inds,
                            bg_rois_per_this_image,
                            num_bins=num_bins,
                            bg_thresh_hi=bg_thresh_hi)
                bg_inds = bg_inds[:bg_rois_per_this_image]

            keep_inds = np.append(fg_inds, bg_inds)
            sampled_labels = max_classes[keep_inds]  # N x 1
            sampled_labels[fg_rois_per_this_image:] = 0
            sampled_boxes = boxes[keep_inds]  # N x 324
            sampled_gts = gt_boxes[box_to_gt_ind_map[keep_inds]]
            sampled_gts[fg_rois_per_this_image:, :] = gt_boxes[0]
            bbox_label_targets = _compute_targets(
                sampled_boxes, sampled_gts, sampled_labels, bbox_reg_weights)
            bbox_targets, bbox_inside_weights = _expand_bbox_targets(
                bbox_label_targets, class_nums, is_cls_agnostic)
            bbox_outside_weights = np.array(
                bbox_inside_weights > 0, dtype=bbox_inside_weights.dtype)
            # Scale rois
            sampled_rois = sampled_boxes * im_scale

            # Faster RCNN blobs
            frcn_blobs = dict(
                rois=sampled_rois,
                labels_int32=sampled_labels,
                bbox_targets=bbox_targets,
                bbox_inside_weights=bbox_inside_weights,
                bbox_outside_weights=bbox_outside_weights)
            return frcn_blobs

        def _compute_targets(roi_boxes, gt_boxes, labels, bbox_reg_weights):
            assert roi_boxes.shape[0] == gt_boxes.shape[0]
            assert roi_boxes.shape[1] == 4
            assert gt_boxes.shape[1] == 4

            targets = np.zeros(roi_boxes.shape)
            bbox_reg_weights = np.asarray(bbox_reg_weights)
            targets = box_to_delta(
                ex_boxes=roi_boxes, gt_boxes=gt_boxes, weights=bbox_reg_weights)

            return np.hstack([labels[:, np.newaxis], targets]).astype(
                np.float32, copy=False)

        def _expand_bbox_targets(bbox_targets_input, class_nums,
                                 is_cls_agnostic):
            class_labels = bbox_targets_input[:, 0]
            fg_inds = np.where(class_labels > 0)[0]
            bbox_targets = np.zeros((class_labels.shape[0], 4 * class_nums
                                     if not is_cls_agnostic else 4 * 2))
            bbox_inside_weights = np.zeros(bbox_targets.shape)
            for ind in fg_inds:
                class_label = int(class_labels[
                    ind]) if not is_cls_agnostic else 1
                start_ind = class_label * 4
                end_ind = class_label * 4 + 4
                bbox_targets[ind, start_ind:end_ind] = bbox_targets_input[ind,
                                                                          1:]
                bbox_inside_weights[ind, start_ind:end_ind] = (1.0, 1.0, 1.0,
                                                               1.0)
            return bbox_targets, bbox_inside_weights

        def generate_func(
                rpn_rois,
                gt_classes,
                is_crowd,
                gt_boxes,
                im_info, ):
            rpn_rois_lod = rpn_rois.lod()[0]
            gt_classes_lod = gt_classes.lod()[0]

            # convert
            rpn_rois = np.array(rpn_rois)
            gt_classes = np.array(gt_classes)
            is_crowd = np.array(is_crowd)
            gt_boxes = np.array(gt_boxes)
            im_info = np.array(im_info)

            rois = []
            labels_int32 = []
            bbox_targets = []
            bbox_inside_weights = []
            bbox_outside_weights = []
            lod = [0]

            for idx in range(len(rpn_rois_lod) - 1):
                rois_si = rpn_rois_lod[idx]
                rois_ei = rpn_rois_lod[idx + 1]

                gt_si = gt_classes_lod[idx]
                gt_ei = gt_classes_lod[idx + 1]
                frcn_blobs = _sample_rois(
                    rpn_rois[rois_si:rois_ei], gt_classes[gt_si:gt_ei],
                    is_crowd[gt_si:gt_ei], gt_boxes[gt_si:gt_ei], im_info[idx],
                    batch_size_per_im, fg_fraction, fg_thresh, bg_thresh_hi,
                    bg_thresh_lo, bbox_reg_weights, class_nums, use_random,
                    is_cls_agnostic, is_cascade_rcnn)
                lod.append(frcn_blobs['rois'].shape[0] + lod[-1])
                rois.append(frcn_blobs['rois'])
                labels_int32.append(frcn_blobs['labels_int32'].reshape(-1, 1))
                bbox_targets.append(frcn_blobs['bbox_targets'])
                bbox_inside_weights.append(frcn_blobs['bbox_inside_weights'])
                bbox_outside_weights.append(frcn_blobs['bbox_outside_weights'])

            rois = np.vstack(rois)
            labels_int32 = np.vstack(labels_int32)
            bbox_targets = np.vstack(bbox_targets)
            bbox_inside_weights = np.vstack(bbox_inside_weights)
            bbox_outside_weights = np.vstack(bbox_outside_weights)

            # create lod-tensor for return
            # notice that the func create_lod_tensor does not work well here
            ret_rois = fluid.LoDTensor()
            ret_rois.set_lod([lod])
            ret_rois.set(rois.astype("float32"), fluid.CPUPlace())

            ret_labels_int32 = fluid.LoDTensor()
            ret_labels_int32.set_lod([lod])
            ret_labels_int32.set(labels_int32.astype("int32"), fluid.CPUPlace())

            ret_bbox_targets = fluid.LoDTensor()
            ret_bbox_targets.set_lod([lod])
            ret_bbox_targets.set(
                bbox_targets.astype("float32"), fluid.CPUPlace())

            ret_bbox_inside_weights = fluid.LoDTensor()
            ret_bbox_inside_weights.set_lod([lod])
            ret_bbox_inside_weights.set(
                bbox_inside_weights.astype("float32"), fluid.CPUPlace())

            ret_bbox_outside_weights = fluid.LoDTensor()
            ret_bbox_outside_weights.set_lod([lod])
            ret_bbox_outside_weights.set(
                bbox_outside_weights.astype("float32"), fluid.CPUPlace())

            return ret_rois, ret_labels_int32, ret_bbox_targets, ret_bbox_inside_weights, ret_bbox_outside_weights

        rois = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'rois', 
            dtype='float32',
            shape=[-1, 4], )
        bbox_inside_weights = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_inside_weights', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        bbox_outside_weights = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_outside_weights', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        bbox_targets = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_targets', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        labels_int32 = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'labels_int32', 
            dtype='int32',
            shape=[-1, 1], )

        outs = [
            rois, labels_int32, bbox_targets, bbox_inside_weights,
            bbox_outside_weights
        ]

        fluid.layers.py_func(
            func=generate_func,
            x=[rpn_rois, gt_classes, is_crowd, gt_boxes, im_info],
            out=outs)
        return outs


@register
class RoIAlign(object):
    __op__ = fluid.layers.roi_align
    __append_doc__ = True

    def __init__(self, resolution=7, spatial_scale=1. / 16, sampling_ratio=0):
        super(RoIAlign, self).__init__()
        if isinstance(resolution, Integral):
            resolution = [resolution, resolution]
        self.pooled_height = resolution[0]
        self.pooled_width = resolution[1]
        self.spatial_scale = spatial_scale
        self.sampling_ratio = sampling_ratio


@register
class RoIPool(object):
    __op__ = fluid.layers.roi_pool
    __append_doc__ = True

    def __init__(self, resolution=7, spatial_scale=1. / 16):
        super(RoIPool, self).__init__()
        if isinstance(resolution, Integral):
            resolution = [resolution, resolution]
        self.pooled_height = resolution[0]
        self.pooled_width = resolution[1]
        self.spatial_scale = spatial_scale


@register
class MultiBoxHead(object):
    __op__ = fluid.layers.multi_box_head
    __append_doc__ = True

    def __init__(self,
                 min_ratio=20,
                 max_ratio=90,
                 base_size=300,
                 min_sizes=[60.0, 105.0, 150.0, 195.0, 240.0, 285.0],
                 max_sizes=[[], 150.0, 195.0, 240.0, 285.0, 300.0],
                 aspect_ratios=[[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.],
                                [2., 3.]],
                 steps=None,
                 offset=0.5,
                 flip=True,
                 min_max_aspect_ratios_order=False,
                 kernel_size=1,
                 pad=0):
        super(MultiBoxHead, self).__init__()
        self.min_ratio = min_ratio
        self.max_ratio = max_ratio
        self.base_size = base_size
        self.min_sizes = min_sizes
        self.max_sizes = max_sizes
        self.aspect_ratios = aspect_ratios
        self.steps = steps
        self.offset = offset
        self.flip = flip
        self.min_max_aspect_ratios_order = min_max_aspect_ratios_order
        self.kernel_size = kernel_size
        self.pad = pad


@register
@serializable
class SSDLiteMultiBoxHead(object):
    def __init__(self,
                 min_ratio=20,
                 max_ratio=90,
                 base_size=300,
                 min_sizes=None,
                 max_sizes=None,
                 aspect_ratios=[[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.],
                                [2., 3.]],
                 steps=None,
                 offset=0.5,
                 flip=True,
                 clip=False,
                 pad=0,
                 conv_decay=0.0):
        super(SSDLiteMultiBoxHead, self).__init__()
        self.min_ratio = min_ratio
        self.max_ratio = max_ratio
        self.base_size = base_size
        self.min_sizes = min_sizes
        self.max_sizes = max_sizes
        self.aspect_ratios = aspect_ratios
        self.steps = steps
        self.offset = offset
        self.flip = flip
        self.pad = pad
        self.clip = clip
        self.conv_decay = conv_decay

    def _separable_conv(self, input, num_filters, name):
        dwconv_param_attr = ParamAttr(
            name=name + 'dw_weights', regularizer=L2Decay(self.conv_decay))
        num_filter1 = input.shape[1]
        depthwise_conv = fluid.layers.conv2d(
            input=input,
            num_filters=num_filter1,
            filter_size=3,
            stride=1,
            padding="SAME",
            groups=int(num_filter1),
            act=None,
            use_cudnn=False,
            param_attr=dwconv_param_attr,
            bias_attr=False)
        bn_name = name + '_bn'
        bn_param_attr = ParamAttr(
            name=bn_name + "_scale", regularizer=L2Decay(0.0))
        bn_bias_attr = ParamAttr(
            name=bn_name + "_offset", regularizer=L2Decay(0.0))
        bn = fluid.layers.batch_norm(
            input=depthwise_conv,
            param_attr=bn_param_attr,
            bias_attr=bn_bias_attr,
            moving_mean_name=bn_name + '_mean',
            moving_variance_name=bn_name + '_variance')
        bn = fluid.layers.relu6(bn)
        pwconv_param_attr = ParamAttr(
            name=name + 'pw_weights', regularizer=L2Decay(self.conv_decay))
        pointwise_conv = fluid.layers.conv2d(
            input=bn,
            num_filters=num_filters,
            filter_size=1,
            stride=1,
            act=None,
            use_cudnn=True,
            param_attr=pwconv_param_attr,
            bias_attr=False)
        return pointwise_conv

    def __call__(self, inputs, image, num_classes):
        def _permute_and_reshape(input, last_dim):
            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
            compile_shape = [0, -1, last_dim]
            return fluid.layers.reshape(trans, shape=compile_shape)

        def _is_list_or_tuple_(data):
            return (isinstance(data, list) or isinstance(data, tuple))

        if self.min_sizes is None and self.max_sizes is None:
            num_layer = len(inputs)
            self.min_sizes = []
            self.max_sizes = []
            step = int(
                math.floor(((self.max_ratio - self.min_ratio)) / (num_layer - 2
                                                                  )))
            for ratio in six.moves.range(self.min_ratio, self.max_ratio + 1,
                                         step):
                self.min_sizes.append(self.base_size * ratio / 100.)
                self.max_sizes.append(self.base_size * (ratio + step) / 100.)
            self.min_sizes = [self.base_size * .10] + self.min_sizes
            self.max_sizes = [self.base_size * .20] + self.max_sizes

        locs, confs = [], []
        boxes, mvars = [], []

        for i, input in enumerate(inputs):
            min_size = self.min_sizes[i]
            max_size = self.max_sizes[i]
            if not _is_list_or_tuple_(min_size):
                min_size = [min_size]
            if not _is_list_or_tuple_(max_size):
                max_size = [max_size]
            step = [
                self.steps[i] if self.steps else 0.0, self.steps[i]
                if self.steps else 0.0
            ]
            box, var = fluid.layers.prior_box(
                input,
                image,
                min_sizes=min_size,
                max_sizes=max_size,
                steps=step,
                aspect_ratios=self.aspect_ratios[i],
                variance=[0.1, 0.1, 0.2, 0.2],
                clip=self.clip,
                flip=self.flip,
                offset=0.5)

            num_boxes = box.shape[2]
            box = fluid.layers.reshape(box, shape=[-1, 4])
            var = fluid.layers.reshape(var, shape=[-1, 4])
            num_loc_output = num_boxes * 4
            num_conf_output = num_boxes * num_classes
            # get loc
            mbox_loc = self._separable_conv(input, num_loc_output,
                                            "loc_{}".format(i + 1))
            loc = _permute_and_reshape(mbox_loc, 4)
            # get conf
            mbox_conf = self._separable_conv(input, num_conf_output,
                                             "conf_{}".format(i + 1))
            conf = _permute_and_reshape(mbox_conf, num_classes)

            locs.append(loc)
            confs.append(conf)
            boxes.append(box)
            mvars.append(var)

        ssd_mbox_loc = fluid.layers.concat(locs, axis=1)
        ssd_mbox_conf = fluid.layers.concat(confs, axis=1)
        prior_boxes = fluid.layers.concat(boxes)
        box_vars = fluid.layers.concat(mvars)

        prior_boxes.stop_gradient = True
        box_vars.stop_gradient = True
        return ssd_mbox_loc, ssd_mbox_conf, prior_boxes, box_vars


@register
@serializable
class SSDOutputDecoder(object):
    __op__ = fluid.layers.detection_output
    __append_doc__ = True

    def __init__(self,
                 nms_threshold=0.45,
                 nms_top_k=400,
                 keep_top_k=200,
                 score_threshold=0.01,
                 nms_eta=1.0,
                 background_label=0):
        super(SSDOutputDecoder, self).__init__()
        self.nms_threshold = nms_threshold
        self.background_label = background_label
        self.nms_top_k = nms_top_k
        self.keep_top_k = keep_top_k
        self.score_threshold = score_threshold
        self.nms_eta = nms_eta


@register
@serializable
class RetinaTargetAssign(object):
    __op__ = fluid.layers.retinanet_target_assign
    __append_doc__ = True

    def __init__(self, positive_overlap=0.5, negative_overlap=0.4):
        super(RetinaTargetAssign, self).__init__()
        self.positive_overlap = positive_overlap
        self.negative_overlap = negative_overlap


@register
@serializable
class RetinaOutputDecoder(object):
    __op__ = fluid.layers.retinanet_detection_output
    __append_doc__ = True

    def __init__(self,
                 score_thresh=0.05,
                 nms_thresh=0.3,
                 pre_nms_top_n=1000,
                 detections_per_im=100,
                 nms_eta=1.0):
        super(RetinaOutputDecoder, self).__init__()
        self.score_threshold = score_thresh
        self.nms_threshold = nms_thresh
        self.nms_top_k = pre_nms_top_n
        self.keep_top_k = detections_per_im
        self.nms_eta = nms_eta