bbox_head.py

# Copyright (c) 2020 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

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn.initializer import Normal, XavierUniform, KaimingNormal
from paddle.regularizer import L2Decay

from ppdet.core.workspace import register, create
from ppdet.modeling import ops

from .roi_extractor import RoIAlign
from ..shape_spec import ShapeSpec
from ..bbox_utils import bbox2delta
from ppdet.modeling.layers import ConvNormLayer

__all__ = ['TwoFCHead', 'XConvNormHead', 'BBoxHead']


@register
class TwoFCHead(nn.Layer):
    """
    RCNN bbox head with Two fc layers to extract feature

    Args:
        in_channel (int): Input channel which can be derived by from_config
        out_channel (int): Output channel
        resolution (int): Resolution of input feature map, default 7
    """

    def __init__(self, in_channel=256, out_channel=1024, resolution=7):
        super(TwoFCHead, self).__init__()
        self.in_channel = in_channel
        self.out_channel = out_channel
        fan = in_channel * resolution * resolution
        self.fc6 = nn.Linear(
            in_channel * resolution * resolution,
            out_channel,
            weight_attr=paddle.ParamAttr(
                initializer=XavierUniform(fan_out=fan)))
        self.fc6.skip_quant = True

        self.fc7 = nn.Linear(
            out_channel,
            out_channel,
            weight_attr=paddle.ParamAttr(initializer=XavierUniform()))
        self.fc7.skip_quant = True

    @classmethod
    def from_config(cls, cfg, input_shape):
        s = input_shape
        s = s[0] if isinstance(s, (list, tuple)) else s
        return {'in_channel': s.channels}

    @property
    def out_shape(self):
        return [ShapeSpec(channels=self.out_channel, )]

    def forward(self, rois_feat):
        rois_feat = paddle.flatten(rois_feat, start_axis=1, stop_axis=-1)
        fc6 = self.fc6(rois_feat)
        fc6 = F.relu(fc6)
        fc7 = self.fc7(fc6)
        fc7 = F.relu(fc7)
        return fc7


@register
class XConvNormHead(nn.Layer):
    __shared__ = ['norm_type', 'freeze_norm']
    """
    RCNN bbox head with serveral convolution layers

    Args:
        in_channel (int): Input channels which can be derived by from_config
        num_convs (int): The number of conv layers
        conv_dim (int): The number of channels for the conv layers
        out_channel (int): Output channels
        resolution (int): Resolution of input feature map
        norm_type (string): Norm type, bn, gn, sync_bn are available, 
            default `gn`
        freeze_norm (bool): Whether to freeze the norm
        stage_name (string): Prefix name for conv layer,  '' by default
    """

    def __init__(self,
                 in_channel=256,
                 num_convs=4,
                 conv_dim=256,
                 out_channel=1024,
                 resolution=7,
                 norm_type='gn',
                 freeze_norm=False,
                 stage_name=''):
        super(XConvNormHead, self).__init__()
        self.in_channel = in_channel
        self.num_convs = num_convs
        self.conv_dim = conv_dim
        self.out_channel = out_channel
        self.norm_type = norm_type
        self.freeze_norm = freeze_norm

        self.bbox_head_convs = []
        fan = conv_dim * 3 * 3
        initializer = KaimingNormal(fan_in=fan)
        for i in range(self.num_convs):
            in_c = in_channel if i == 0 else conv_dim
            head_conv_name = stage_name + 'bbox_head_conv{}'.format(i)
            head_conv = self.add_sublayer(
                head_conv_name,
                ConvNormLayer(
                    ch_in=in_c,
                    ch_out=conv_dim,
                    filter_size=3,
                    stride=1,
                    norm_type=self.norm_type,
                    freeze_norm=self.freeze_norm,
                    initializer=initializer))
            self.bbox_head_convs.append(head_conv)

        fan = conv_dim * resolution * resolution
        self.fc6 = nn.Linear(
            conv_dim * resolution * resolution,
            out_channel,
            weight_attr=paddle.ParamAttr(
                initializer=XavierUniform(fan_out=fan)),
            bias_attr=paddle.ParamAttr(
                learning_rate=2., regularizer=L2Decay(0.)))

    @classmethod
    def from_config(cls, cfg, input_shape):
        s = input_shape
        s = s[0] if isinstance(s, (list, tuple)) else s
        return {'in_channel': s.channels}

    @property
    def out_shape(self):
        return [ShapeSpec(channels=self.out_channel, )]

    def forward(self, rois_feat):
        for i in range(self.num_convs):
            rois_feat = F.relu(self.bbox_head_convs[i](rois_feat))
        rois_feat = paddle.flatten(rois_feat, start_axis=1, stop_axis=-1)
        fc6 = F.relu(self.fc6(rois_feat))
        return fc6


@register
class BBoxHead(nn.Layer):
    __shared__ = ['num_classes']
    __inject__ = ['bbox_assigner', 'bbox_loss']
    """
    RCNN bbox head

    Args:
        head (nn.Layer): Extract feature in bbox head
        in_channel (int): Input channel after RoI extractor
        roi_extractor (object): The module of RoI Extractor
        bbox_assigner (object): The module of Box Assigner, label and sample the 
            box.
        with_pool (bool): Whether to use pooling for the RoI feature.
        num_classes (int): The number of classes
        bbox_weight (List[float]): The weight to get the decode box 
    """

    def __init__(self,
                 head,
                 in_channel,
                 roi_extractor=RoIAlign().__dict__,
                 bbox_assigner='BboxAssigner',
                 with_pool=False,
                 num_classes=80,
                 bbox_weight=[10., 10., 5., 5.],
                 bbox_loss=None):
        super(BBoxHead, self).__init__()
        self.head = head
        self.roi_extractor = roi_extractor
        if isinstance(roi_extractor, dict):
            self.roi_extractor = RoIAlign(**roi_extractor)
        self.bbox_assigner = bbox_assigner

        self.with_pool = with_pool
        self.num_classes = num_classes
        self.bbox_weight = bbox_weight
        self.bbox_loss = bbox_loss

        self.bbox_score = nn.Linear(
            in_channel,
            self.num_classes + 1,
            weight_attr=paddle.ParamAttr(initializer=Normal(
                mean=0.0, std=0.01)))
        self.bbox_score.skip_quant = True

        self.bbox_delta = nn.Linear(
            in_channel,
            4 * self.num_classes,
            weight_attr=paddle.ParamAttr(initializer=Normal(
                mean=0.0, std=0.001)))
        self.bbox_delta.skip_quant = True
        self.assigned_label = None
        self.assigned_rois = None

    @classmethod
    def from_config(cls, cfg, input_shape):
        roi_pooler = cfg['roi_extractor']
        assert isinstance(roi_pooler, dict)
        kwargs = RoIAlign.from_config(cfg, input_shape)
        roi_pooler.update(kwargs)
        kwargs = {'input_shape': input_shape}
        head = create(cfg['head'], **kwargs)
        return {
            'roi_extractor': roi_pooler,
            'head': head,
            'in_channel': head.out_shape[0].channels
        }

    def forward(self, body_feats=None, rois=None, rois_num=None, inputs=None):
        """
        body_feats (list[Tensor]): Feature maps from backbone
        rois (list[Tensor]): RoIs generated from RPN module
        rois_num (Tensor): The number of RoIs in each image
        inputs (dict{Tensor}): The ground-truth of image
        """
        if self.training:
            rois, rois_num, targets = self.bbox_assigner(rois, rois_num, inputs)
            self.assigned_rois = (rois, rois_num)
            self.assigned_targets = targets

        rois_feat = self.roi_extractor(body_feats, rois, rois_num)
        bbox_feat = self.head(rois_feat)
        if self.with_pool:
            feat = F.adaptive_avg_pool2d(bbox_feat, output_size=1)
            feat = paddle.squeeze(feat, axis=[2, 3])
        else:
            feat = bbox_feat
        scores = self.bbox_score(feat)
        deltas = self.bbox_delta(feat)

        if self.training:
            loss = self.get_loss(scores, deltas, targets, rois,
                                 self.bbox_weight)
            return loss, bbox_feat
        else:
            pred = self.get_prediction(scores, deltas)
            return pred, self.head

    def get_loss(self, scores, deltas, targets, rois, bbox_weight):
        """
        scores (Tensor): scores from bbox head outputs
        deltas (Tensor): deltas from bbox head outputs
        targets (list[List[Tensor]]): bbox targets containing tgt_labels, tgt_bboxes and tgt_gt_inds
        rois (List[Tensor]): RoIs generated in each batch
        """
        # TODO: better pass args
        tgt_labels, tgt_bboxes, tgt_gt_inds = targets
        tgt_labels = paddle.concat(tgt_labels) if len(
            tgt_labels) > 1 else tgt_labels[0]
        tgt_labels = tgt_labels.cast('int64')
        tgt_labels.stop_gradient = True
        loss_bbox_cls = F.cross_entropy(
            input=scores, label=tgt_labels, reduction='mean')
        # bbox reg

        cls_agnostic_bbox_reg = deltas.shape[1] == 4

        fg_inds = paddle.nonzero(
            paddle.logical_and(tgt_labels >= 0, tgt_labels <
                               self.num_classes)).flatten()

        cls_name = 'loss_bbox_cls'
        reg_name = 'loss_bbox_reg'
        loss_bbox = {}

        loss_weight = 1.
        if fg_inds.numel() == 0:
            fg_inds = paddle.zeros([1], dtype='int32')
            loss_weight = 0.

        if cls_agnostic_bbox_reg:
            reg_delta = paddle.gather(deltas, fg_inds)
        else:
            fg_gt_classes = paddle.gather(tgt_labels, fg_inds)

            reg_row_inds = paddle.arange(fg_gt_classes.shape[0]).unsqueeze(1)
            reg_row_inds = paddle.tile(reg_row_inds, [1, 4]).reshape([-1, 1])

            reg_col_inds = 4 * fg_gt_classes.unsqueeze(1) + paddle.arange(4)

            reg_col_inds = reg_col_inds.reshape([-1, 1])
            reg_inds = paddle.concat([reg_row_inds, reg_col_inds], axis=1)

            reg_delta = paddle.gather(deltas, fg_inds)
            reg_delta = paddle.gather_nd(reg_delta, reg_inds).reshape([-1, 4])
        rois = paddle.concat(rois) if len(rois) > 1 else rois[0]
        tgt_bboxes = paddle.concat(tgt_bboxes) if len(
            tgt_bboxes) > 1 else tgt_bboxes[0]

        reg_target = bbox2delta(rois, tgt_bboxes, bbox_weight)
        reg_target = paddle.gather(reg_target, fg_inds)
        reg_target.stop_gradient = True

        if self.bbox_loss is not None:
            reg_delta = self.bbox_transform(reg_delta)
            reg_target = self.bbox_transform(reg_target)
            loss_bbox_reg = self.bbox_loss(
                reg_delta, reg_target).sum() / tgt_labels.shape[0]
            loss_bbox_reg *= self.num_classes
        else:
            loss_bbox_reg = paddle.abs(reg_delta - reg_target).sum(
            ) / tgt_labels.shape[0]

        loss_bbox[cls_name] = loss_bbox_cls * loss_weight
        loss_bbox[reg_name] = loss_bbox_reg * loss_weight

        return loss_bbox

    def bbox_transform(self, deltas, weights=[0.1, 0.1, 0.2, 0.2]):
        wx, wy, ww, wh = weights

        deltas = paddle.reshape(deltas, shape=(0, -1, 4))

        dx = paddle.slice(deltas, axes=[2], starts=[0], ends=[1]) * wx
        dy = paddle.slice(deltas, axes=[2], starts=[1], ends=[2]) * wy
        dw = paddle.slice(deltas, axes=[2], starts=[2], ends=[3]) * ww
        dh = paddle.slice(deltas, axes=[2], starts=[3], ends=[4]) * wh

        dw = paddle.clip(dw, -1.e10, np.log(1000. / 16))
        dh = paddle.clip(dh, -1.e10, np.log(1000. / 16))

        pred_ctr_x = dx
        pred_ctr_y = dy
        pred_w = paddle.exp(dw)
        pred_h = paddle.exp(dh)

        x1 = pred_ctr_x - 0.5 * pred_w
        y1 = pred_ctr_y - 0.5 * pred_h
        x2 = pred_ctr_x + 0.5 * pred_w
        y2 = pred_ctr_y + 0.5 * pred_h

        x1 = paddle.reshape(x1, shape=(-1, ))
        y1 = paddle.reshape(y1, shape=(-1, ))
        x2 = paddle.reshape(x2, shape=(-1, ))
        y2 = paddle.reshape(y2, shape=(-1, ))

        return paddle.concat([x1, y1, x2, y2])

    def get_prediction(self, score, delta):
        bbox_prob = F.softmax(score)
        return delta, bbox_prob

    def get_head(self, ):
        return self.head

    def get_assigned_targets(self, ):
        return self.assigned_targets

    def get_assigned_rois(self, ):
        return self.assigned_rois