# Copyright (c) 2021 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 paddle from paddle import ParamAttr import paddle.nn as nn import paddle.nn.functional as F import numpy as np class ConvBNLayer(nn.Layer): def __init__(self, input_channels, output_channels, filter_size, stride, padding, name=None): super(ConvBNLayer, self).__init__() self._conv = nn.Conv2D( in_channels=input_channels, out_channels=output_channels, kernel_size=filter_size, stride=stride, padding=padding, weight_attr=ParamAttr(name=name + ".conv.weights"), bias_attr=False) bn_name = name + ".bn" self._bn = nn.BatchNorm( num_channels=output_channels, act="leaky_relu", param_attr=ParamAttr(name=bn_name + ".scale"), bias_attr=ParamAttr(name=bn_name + ".offset"), moving_mean_name=bn_name + ".mean", moving_variance_name=bn_name + ".var") def forward(self, inputs): x = self._conv(inputs) x = self._bn(x) return x class BasicBlock(nn.Layer): def __init__(self, input_channels, output_channels, name=None): super(BasicBlock, self).__init__() self._conv1 = ConvBNLayer(input_channels=input_channels, output_channels=output_channels, filter_size=[ 3, 3], stride=1, padding=1, name=name+'.0') self._max_pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0) self._conv2 = ConvBNLayer(input_channels=output_channels, output_channels=output_channels * 2, filter_size=[3, 3], stride=1, padding=1, name=name+'.1') self._conv3 = ConvBNLayer(input_channels=output_channels*2, output_channels=output_channels, filter_size=[1, 1], stride=1, padding=0, name=name+'.2') def forward(self, x): x = self._conv1(x) x = self._max_pool(x) x = self._conv2(x) x = self._conv3(x) return x class Reorg(nn.Layer): def __init__(self, stride=2): super(Reorg, self).__init__() self.stride = stride def forward(self, x): stride = self.stride assert (x.dim() == 4) B = x.shape[0] C = x.shape[1] H = x.shape[2] W = x.shape[3] assert (H % stride == 0) assert (W % stride == 0) ws = stride hs = stride x = x.reshape([B, C, H // hs, hs, W // ws, ws] ).transpose([0, 1, 2, 4, 3, 5]) x = x.reshape([B, C, H // hs * W // ws, hs * ws] ).transpose([0, 1, 3, 2]) x = x.reshape([B, C, hs * ws, H // hs, W // ws] ).transpose([0, 2, 1, 3, 4]) x = x.reshape([B, hs * ws * C, H // hs, W // ws]) return x class Darknet(nn.Layer): def __init__(self, pretrained=None): super(Darknet, self).__init__() self.pretrained = pretrained self._conv1 = ConvBNLayer( input_channels=3, output_channels=32, filter_size=3, stride=1, padding=1, name='input') self._max_pool1 = nn.MaxPool2D(kernel_size=2, stride=2, padding=0) self._basic_block_11 = BasicBlock( input_channels=32, output_channels=64, name='1.1') self._basic_block_12 = BasicBlock( input_channels=64, output_channels=128, name='1.2') self._basic_block_13 = BasicBlock( input_channels=128, output_channels=256, name='1.3') self._conv2 = ConvBNLayer( input_channels=256, output_channels=512, filter_size=3, stride=1, padding=1, name='up1') self._conv3 = ConvBNLayer( input_channels=512, output_channels=256, filter_size=1, stride=1, padding=0, name='down1') self._conv4 = ConvBNLayer( input_channels=256, output_channels=512, filter_size=3, stride=1, padding=1, name='2.1') self._max_pool2 = nn.MaxPool2D(kernel_size=2, stride=2, padding=0) self._conv5 = ConvBNLayer( input_channels=512, output_channels=1024, filter_size=3, stride=1, padding=1, name='2.2') self._conv6 = ConvBNLayer(input_channels=1024, output_channels=512, filter_size=1, stride=1, padding=0, name='2.3') # ori self._conv7 = ConvBNLayer( input_channels=512, output_channels=1024, filter_size=3, stride=1, padding=1, name='up2') self._conv8 = ConvBNLayer(input_channels=1024, output_channels=512, filter_size=1, stride=1, padding=0, name='down2') self._conv9 = ConvBNLayer( input_channels=512, output_channels=1024, filter_size=3, stride=1, padding=1, name='3.1') self._conv10 = ConvBNLayer( input_channels=1024, output_channels=1024, filter_size=3, stride=1, padding=1, name='3.2') self._conv11 = ConvBNLayer( input_channels=1024, output_channels=1024, filter_size=3, stride=1, padding=1, name='3.3') self._conv12 = ConvBNLayer( input_channels=512, output_channels=64, filter_size=1, stride=1, padding=0, name='4.1') self._reorg = Reorg() self._conv13 = ConvBNLayer( input_channels=1280, output_channels=1024, filter_size=3, stride=1, padding=1, name='5.1') self._conv14 = nn.Conv2D(1024, 425, kernel_size=1) def forward(self, inputs): x = self._conv1(inputs) x = self._max_pool1(x) x = self._basic_block_11(x) x = self._basic_block_12(x) x = self._basic_block_13(x) x = self._conv2(x) x = self._conv3(x) ori = self._conv4(x) x = self._max_pool2(ori) x = self._conv5(x) x = self._conv6(x) x = self._conv7(x) x = self._conv8(x) x = self._conv9(x) x = self._conv10(x) x1 = self._conv11(x) x2 = self._conv12(ori) x2 = self._reorg(x2) x = paddle.concat([x2, x1], 1) x = self._conv13(x) x = self._conv14(x) return x