fpn.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.functional as F
from paddle import ParamAttr
from paddle.nn import Layer
from paddle.nn import Conv2D
from paddle.nn.initializer import XavierUniform
from paddle.regularizer import L2Decay
from ppdet.core.workspace import register, serializable
from ..shape_spec import ShapeSpec


@register
@serializable
class FPN(Layer):
    def __init__(self,
                 in_channels,
                 out_channel,
                 spatial_scales=[0.25, 0.125, 0.0625, 0.03125],
                 has_extra_convs=False,
                 extra_stage=1,
                 use_c5=True,
                 relu_before_extra_convs=True):

        super(FPN, self).__init__()
        self.out_channel = out_channel
        for s in range(extra_stage):
            spatial_scales = spatial_scales + [spatial_scales[-1] / 2.]
        self.spatial_scales = spatial_scales
        self.has_extra_convs = has_extra_convs
        self.extra_stage = extra_stage
        self.use_c5 = use_c5
        self.relu_before_extra_convs = relu_before_extra_convs

        self.lateral_convs = []
        self.fpn_convs = []
        fan = out_channel * 3 * 3

        for i in range(len(in_channels)):
            if i == 3:
                lateral_name = 'fpn_inner_res5_sum'
            else:
                lateral_name = 'fpn_inner_res{}_sum_lateral'.format(i + 2)
            in_c = in_channels[i]
            lateral = self.add_sublayer(
                lateral_name,
                Conv2D(
                    in_channels=in_c,
                    out_channels=out_channel,
                    kernel_size=1,
                    weight_attr=ParamAttr(
                        initializer=XavierUniform(fan_out=in_c))))
            self.lateral_convs.append(lateral)

            fpn_name = 'fpn_res{}_sum'.format(i + 2)
            fpn_conv = self.add_sublayer(
                fpn_name,
                Conv2D(
                    in_channels=out_channel,
                    out_channels=out_channel,
                    kernel_size=3,
                    padding=1,
                    weight_attr=ParamAttr(
                        initializer=XavierUniform(fan_out=fan))))
            self.fpn_convs.append(fpn_conv)

        # add extra conv levels for RetinaNet(use_c5)/FCOS(use_p5)
        if self.has_extra_convs:
            for lvl in range(self.extra_stage):  # P6 P7 ...
                if lvl == 0 and self.use_c5:
                    in_c = in_channels[-1]
                else:
                    in_c = out_channel
                extra_fpn_name = 'fpn_{}'.format(lvl + 2)
                extra_fpn_conv = self.add_sublayer(
                    extra_fpn_name,
                    Conv2D(
                        in_channels=in_c,
                        out_channels=out_channel,
                        kernel_size=3,
                        stride=2,
                        padding=1,
                        weight_attr=ParamAttr(
                            initializer=XavierUniform(fan_out=fan))))
                self.fpn_convs.append(extra_fpn_conv)

    @classmethod
    def from_config(cls, cfg, input_shape):
        return {
            'in_channels': [i.channels for i in input_shape],
            'spatial_scales': [1.0 / i.stride for i in input_shape],
        }

    def forward(self, body_feats):
        laterals = []
        num_levels = len(body_feats)
        for i in range(num_levels):
            laterals.append(self.lateral_convs[i](body_feats[i]))

        for i in range(1, num_levels):
            lvl = num_levels - i
            upsample = F.interpolate(
                laterals[lvl],
                scale_factor=2.,
                mode='nearest', )
            laterals[lvl - 1] += upsample

        fpn_output = []
        for lvl in range(num_levels):
            fpn_output.append(self.fpn_convs[lvl](laterals[lvl]))

        if self.extra_stage > 0:
            # use max pool to get more levels on top of outputs (Faster R-CNN, Mask R-CNN)
            if not self.has_extra_convs:
                assert self.extra_stage == 1, 'extra_stage should be 1 if FPN has not extra convs'
                fpn_output.append(F.max_pool2d(fpn_output[-1], 1, stride=2))
            # add extra conv levels for RetinaNet(use_c5)/FCOS(use_p5)
            else:
                if self.use_c5:
                    extra_source = body_feats[-1]
                else:
                    extra_source = fpn_output[-1]
                fpn_output.append(self.fpn_convs[num_levels](extra_source))

                for i in range(1, self.extra_stage):
                    if self.relu_before_extra_convs:
                        fpn_output.append(self.fpn_convs[num_levels + i](F.relu(
                            fpn_output[-1])))
                    else:
                        fpn_output.append(self.fpn_convs[num_levels + i](
                            fpn_output[-1]))
        return fpn_output

    @property
    def out_shape(self):
        return [
            ShapeSpec(
                channels=self.out_channel, stride=1. / s)
            for s in self.spatial_scales
        ]