pyramidbox.py

import numpy as np

import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import Xavier
from paddle.fluid.initializer import Constant
from paddle.fluid.initializer import Bilinear
from paddle.fluid.regularizer import L2Decay


def conv_bn(input, filter, ksize, stride, padding, act='relu', bias_attr=False):
    conv = fluid.layers.conv2d(
        input=input,
        filter_size=ksize,
        num_filters=filter,
        stride=stride,
        padding=padding,
        act=None,
        bias_attr=bias_attr)
    return fluid.layers.batch_norm(input=conv, act=act)


def conv_block(input, groups, filters, ksizes, strides=None, with_pool=True):
    assert len(filters) == groups
    assert len(ksizes) == groups
    strides = [1] * groups if strides is None else strides
    w_attr = ParamAttr(learning_rate=1., initializer=Xavier())
    b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
    conv = input
    for i in xrange(groups):
        conv = fluid.layers.conv2d(
            input=conv,
            num_filters=filters[i],
            filter_size=ksizes[i],
            stride=strides[i],
            padding=(ksizes[i] - 1) / 2,
            param_attr=w_attr,
            bias_attr=b_attr,
            act='relu')
    if with_pool:
        pool = fluid.layers.pool2d(
            input=conv,
            pool_size=2,
            pool_type='max',
            pool_stride=2,
            ceil_mode=True)
        return conv, pool
    else:
        return conv


class PyramidBox(object):
    def __init__(self,
                 data_shape,
                 num_classes=None,
                 use_transposed_conv2d=True,
                 is_infer=False,
                 sub_network=False):
        """
        TODO(qingqing): add comments.
        """
        self.data_shape = data_shape
        self.min_sizes = [16., 32., 64., 128., 256., 512.]
        self.steps = [4., 8., 16., 32., 64., 128.]
        self.num_classes = num_classes
        self.use_transposed_conv2d = use_transposed_conv2d
        self.is_infer = is_infer
        self.sub_network = sub_network

        # the base network is VGG with atrous layers
        self._input()
        self._vgg()
        if sub_network:
            self._low_level_fpn()
            self._cpm_module()
            self._pyramidbox()
        else:
            self._vgg_ssd()

    def feeds(self):
        if self.is_infer:
            return [self.image]
        else:
            return [self.image, self.face_box, self.head_box, self.gt_label]

    def _input(self):
        self.image = fluid.layers.data(
            name='image', shape=self.data_shape, dtype='float32')
        if not self.is_infer:
            self.face_box = fluid.layers.data(
                name='face_box', shape=[4], dtype='float32', lod_level=1)
            self.head_box = fluid.layers.data(
                name='head_box', shape=[4], dtype='float32', lod_level=1)
            self.gt_label = fluid.layers.data(
                name='gt_label', shape=[1], dtype='int32', lod_level=1)

    def _vgg(self):
        self.conv1, self.pool1 = conv_block(self.image, 2, [64] * 2, [3] * 2)
        self.conv2, self.pool2 = conv_block(self.pool1, 2, [128] * 2, [3] * 2)

        #priorbox min_size is 16
        self.conv3, self.pool3 = conv_block(self.pool2, 3, [256] * 3, [3] * 3)
        #priorbox min_size is 32
        self.conv4, self.pool4 = conv_block(self.pool3, 3, [512] * 3, [3] * 3)
        #priorbox min_size is 64
        self.conv5, self.pool5 = conv_block(self.pool4, 3, [512] * 3, [3] * 3)

        # fc6 and fc7 in paper, priorbox min_size is 128
        self.conv6 = conv_block(
            self.pool5, 2, [1024, 1024], [3, 1], with_pool=False)
        # conv6_1 and conv6_2 in paper, priorbox min_size is 256
        self.conv7 = conv_block(
            self.conv6, 2, [256, 512], [1, 3], [1, 2], with_pool=False)
        # conv7_1 and conv7_2 in paper, priorbox mini_size is 512
        self.conv8 = conv_block(
            self.conv7, 2, [128, 256], [1, 3], [1, 2], with_pool=False)

    def _low_level_fpn(self):
        """
        Low-level feature pyramid network.
        """

        def fpn(up_from, up_to):
            ch = up_to.shape[1]
            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
            conv1 = fluid.layers.conv2d(
                up_from, ch, 1, act='relu', bias_attr=b_attr)
            if self.use_transposed_conv2d:
                w_attr = ParamAttr(
                    learning_rate=0.,
                    regularizer=L2Decay(0.),
                    initializer=Bilinear())
                upsampling = fluid.layers.conv2d_transpose(
                    conv1,
                    ch,
                    output_size=None,
                    filter_size=4,
                    padding=1,
                    stride=2,
                    groups=ch,
                    param_attr=w_attr,
                    bias_attr=False,
                    use_cudnn=True)
            else:
                upsampling = fluid.layers.resize_bilinear(
                    conv1, out_shape=up_to.shape[2:])

            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
            conv2 = fluid.layers.conv2d(
                up_to, ch, 1, act='relu', bias_attr=b_attr)
            if self.is_infer:
                upsampling = fluid.layers.crop(upsampling, shape=conv2)
            # eltwise mul
            conv_fuse = upsampling * conv2
            return conv_fuse

        self.lfpn2_on_conv5 = fpn(self.conv6, self.conv5)
        self.lfpn1_on_conv4 = fpn(self.lfpn2_on_conv5, self.conv4)
        self.lfpn0_on_conv3 = fpn(self.lfpn1_on_conv4, self.conv3)

    def _cpm_module(self):
        """
        Context-sensitive Prediction Module 
        """

        def cpm(input):
            # residual
            branch1 = conv_bn(input, 1024, 1, 1, 0, None)
            branch2a = conv_bn(input, 256, 1, 1, 0, act='relu')
            branch2b = conv_bn(branch2a, 256, 3, 1, 1, act='relu')
            branch2c = conv_bn(branch2b, 1024, 1, 1, 0, None)
            sum = branch1 + branch2c
            rescomb = fluid.layers.relu(x=sum)

            # ssh
            b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
            ssh_1 = fluid.layers.conv2d(rescomb, 256, 3, 1, 1, bias_attr=b_attr)
            ssh_dimred = fluid.layers.conv2d(
                rescomb, 128, 3, 1, 1, act='relu', bias_attr=b_attr)
            ssh_2 = fluid.layers.conv2d(
                ssh_dimred, 128, 3, 1, 1, bias_attr=b_attr)
            ssh_3a = fluid.layers.conv2d(
                ssh_dimred, 128, 3, 1, 1, act='relu', bias_attr=b_attr)
            ssh_3b = fluid.layers.conv2d(ssh_3a, 128, 3, 1, 1, bias_attr=b_attr)

            ssh_concat = fluid.layers.concat([ssh_1, ssh_2, ssh_3b], axis=1)
            ssh_out = fluid.layers.relu(x=ssh_concat)
            return ssh_out

        self.ssh_conv3 = cpm(self.lfpn0_on_conv3)
        self.ssh_conv4 = cpm(self.lfpn1_on_conv4)
        self.ssh_conv5 = cpm(self.lfpn2_on_conv5)
        self.ssh_conv6 = cpm(self.conv6)
        self.ssh_conv7 = cpm(self.conv7)
        self.ssh_conv8 = cpm(self.conv8)

    def _l2_norm_scale(self, input, init_scale=1.0, channel_shared=False):
        from paddle.fluid.layer_helper import LayerHelper
        helper = LayerHelper("Scale")
        l2_norm = fluid.layers.l2_normalize(
            input, axis=1)  # l2 norm along channel
        shape = [1] if channel_shared else [input.shape[1]]
        scale = helper.create_parameter(
            attr=helper.param_attr,
            shape=shape,
            dtype=input.dtype,
            default_initializer=Constant(init_scale))
        out = fluid.layers.elementwise_mul(
            x=l2_norm, y=scale, axis=-1 if channel_shared else 1)
        return out

    def _pyramidbox(self):
        """
        Get prior-boxes and pyramid-box
        """
        self.ssh_conv3_norm = self._l2_norm_scale(
            self.ssh_conv3, init_scale=10.)
        self.ssh_conv4_norm = self._l2_norm_scale(self.ssh_conv4, init_scale=8.)
        self.ssh_conv5_norm = self._l2_norm_scale(self.ssh_conv5, init_scale=5.)

        def permute_and_reshape(input, last_dim):
            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
            new_shape = [
                trans.shape[0], np.prod(trans.shape[1:]) / last_dim, last_dim
            ]
            return fluid.layers.reshape(trans, shape=new_shape)

        face_locs, face_confs = [], []
        head_locs, head_confs = [], []
        boxes, vars = [], []
        inputs = [
            self.ssh_conv3_norm, self.ssh_conv4_norm, self.ssh_conv5_norm,
            self.ssh_conv6, self.ssh_conv7, self.ssh_conv8
        ]
        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
        for i, input in enumerate(inputs):
            mbox_loc = fluid.layers.conv2d(input, 8, 3, 1, 1, bias_attr=b_attr)
            face_loc, head_loc = fluid.layers.split(
                mbox_loc, num_or_sections=2, dim=1)
            face_loc = permute_and_reshape(face_loc, 4)
            head_loc = permute_and_reshape(head_loc, 4)

            mbox_conf = fluid.layers.conv2d(input, 6, 3, 1, 1, bias_attr=b_attr)
            face_conf1, face_conf3, head_conf = fluid.layers.split(
                mbox_conf, num_or_sections=[1, 3, 2], dim=1)
            face_conf3_maxin = fluid.layers.reduce_max(
                face_conf3, dim=1, keep_dim=True)
            face_conf = fluid.layers.concat(
                [face_conf1, face_conf3_maxin], axis=1)

            face_conf = permute_and_reshape(face_conf, 2)
            head_conf = permute_and_reshape(head_conf, 2)

            face_locs.append(face_loc)
            face_confs.append(face_conf)

            head_locs.append(head_loc)
            head_confs.append(head_conf)

            box, var = fluid.layers.prior_box(
                input,
                self.image,
                min_sizes=[self.min_sizes[i]],
                steps=[self.steps[i]] * 2,
                aspect_ratios=[1.],
                clip=False,
                flip=True,
                offset=0.5)
            box = fluid.layers.reshape(box, shape=[-1, 4])
            var = fluid.layers.reshape(var, shape=[-1, 4])

            boxes.append(box)
            vars.append(var)

        self.face_mbox_loc = fluid.layers.concat(face_locs, axis=1)
        self.face_mbox_conf = fluid.layers.concat(face_confs, axis=1)

        self.head_mbox_loc = fluid.layers.concat(head_locs, axis=1)
        self.head_mbox_conf = fluid.layers.concat(head_confs, axis=1)

        self.prior_boxes = fluid.layers.concat(boxes)
        self.box_vars = fluid.layers.concat(vars)

    def _vgg_ssd(self):
        self.conv3_norm = self._l2_norm_scale(self.conv3, init_scale=10.)
        self.conv4_norm = self._l2_norm_scale(self.conv4, init_scale=8.)
        self.conv5_norm = self._l2_norm_scale(self.conv5, init_scale=5.)

        def permute_and_reshape(input, last_dim):
            trans = fluid.layers.transpose(input, perm=[0, 2, 3, 1])
            new_shape = [
                trans.shape[0], np.prod(trans.shape[1:]) / last_dim, last_dim
            ]
            return fluid.layers.reshape(trans, shape=new_shape)

        locs, confs = [], []
        boxes, vars = [], []
        b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))

        # conv3
        mbox_loc = fluid.layers.conv2d(
            self.conv3_norm, 4, 3, 1, 1, bias_attr=b_attr)
        loc = permute_and_reshape(mbox_loc, 4)
        mbox_conf = fluid.layers.conv2d(
            self.conv3_norm, 4, 3, 1, 1, bias_attr=b_attr)
        conf1, conf3 = fluid.layers.split(
            mbox_conf, num_or_sections=[1, 3], dim=1)
        conf3_maxin = fluid.layers.reduce_max(conf3, dim=1, keep_dim=True)
        conf = fluid.layers.concat([conf1, conf3_maxin], axis=1)
        conf = permute_and_reshape(conf, 2)
        box, var = fluid.layers.prior_box(
            self.conv3_norm,
            self.image,
            min_sizes=[16.],
            steps=[4, 4],
            aspect_ratios=[1.],
            clip=False,
            flip=True,
            offset=0.5)
        box = fluid.layers.reshape(box, shape=[-1, 4])
        var = fluid.layers.reshape(var, shape=[-1, 4])

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

        min_sizes = [32., 64., 128., 256., 512.]
        steps = [8., 16., 32., 64., 128.]
        inputs = [
            self.conv4_norm, self.conv5_norm, self.conv6, self.conv7, self.conv8
        ]
        for i, input in enumerate(inputs):
            mbox_loc = fluid.layers.conv2d(input, 4, 3, 1, 1, bias_attr=b_attr)
            loc = permute_and_reshape(mbox_loc, 4)

            mbox_conf = fluid.layers.conv2d(input, 2, 3, 1, 1, bias_attr=b_attr)
            conf = permute_and_reshape(mbox_conf, 2)
            box, var = fluid.layers.prior_box(
                input,
                self.image,
                min_sizes=[min_sizes[i]],
                steps=[steps[i]] * 2,
                aspect_ratios=[1.],
                clip=False,
                flip=True,
                offset=0.5)
            box = fluid.layers.reshape(box, shape=[-1, 4])
            var = fluid.layers.reshape(var, shape=[-1, 4])

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

        self.face_mbox_loc = fluid.layers.concat(locs, axis=1)
        self.face_mbox_conf = fluid.layers.concat(confs, axis=1)
        self.prior_boxes = fluid.layers.concat(boxes)
        self.box_vars = fluid.layers.concat(vars)

    def vgg_ssd_loss(self):
        loss = fluid.layers.ssd_loss(
            self.face_mbox_loc,
            self.face_mbox_conf,
            self.face_box,
            self.gt_label,
            self.prior_boxes,
            self.box_vars,
            overlap_threshold=0.35,
            neg_overlap=0.35)
        loss = fluid.layers.reduce_sum(loss)
        return loss

    def train(self):
        face_loss = fluid.layers.ssd_loss(
            self.face_mbox_loc,
            self.face_mbox_conf,
            self.face_box,
            self.gt_label,
            self.prior_boxes,
            self.box_vars,
            overlap_threshold=0.35,
            neg_overlap=0.35)
        face_loss.persistable = True
        head_loss = fluid.layers.ssd_loss(
            self.head_mbox_loc,
            self.head_mbox_conf,
            self.head_box,
            self.gt_label,
            self.prior_boxes,
            self.box_vars,
            overlap_threshold=0.35,
            neg_overlap=0.35)
        head_loss.persistable = True
        face_loss = fluid.layers.reduce_sum(face_loss)
        face_loss.persistable = True
        head_loss = fluid.layers.reduce_sum(head_loss)
        head_loss.persistable = True
        total_loss = face_loss + head_loss
        total_loss.persistable = True
        return face_loss, head_loss, total_loss

    def infer(self, main_program=None):
        if main_program is None:
            test_program = fluid.default_main_program().clone(for_test=True)
        else:
            test_program = main_program.clone(for_test=True)
        with fluid.program_guard(test_program):
            face_nmsed_out = fluid.layers.detection_output(
                self.face_mbox_loc,
                self.face_mbox_conf,
                self.prior_boxes,
                self.box_vars,
                nms_threshold=0.3,
                nms_top_k=5000,
                keep_top_k=750,
                score_threshold=0.01)
        return test_program, face_nmsed_out