|
| 1 | +# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import paddle |
| 16 | +import paddle.nn as nn |
| 17 | +import paddle.nn.functional as F |
| 18 | + |
| 19 | +from paddleseg.cvlibs import manager |
| 20 | +from paddleseg.models import layers |
| 21 | +from paddleseg.utils import utils |
| 22 | + |
| 23 | + |
| 24 | +@manager.MODELS.add_component |
| 25 | +class GloRe(nn.Layer): |
| 26 | + """ |
| 27 | + The GloRe implementation based on PaddlePaddle. |
| 28 | +
|
| 29 | + The original article refers to: |
| 30 | + Chen, Yunpeng, et al. "Graph-Based Global Reasoning Networks" |
| 31 | + (https://arxiv.org/pdf/1811.12814.pdf) |
| 32 | + |
| 33 | + Args: |
| 34 | + num_classes (int): The unique number of target classes. |
| 35 | + backbone (Paddle.nn.Layer): Backbone network, currently support Resnet50/101. |
| 36 | + backbone_indices (tuple, optional): Two values in the tuple indicate the indices of output of backbone. |
| 37 | + gru_channels (int, optional): The number of input channels in GloRe Unit. Default: 512. |
| 38 | + gru_num_state (int, optional): The number of states in GloRe Unit. Default: 128. |
| 39 | + gru_num_node (tuple, optional): The number of nodes in GloRe Unit. Default: Default: 128. |
| 40 | + enable_auxiliary_loss (bool, optional): A bool value indicates whether adding auxiliary loss. Default: True. |
| 41 | + align_corners (bool, optional): An argument of F.interpolate. It should be set to False when the feature size is even, |
| 42 | + e.g. 1024x512, otherwise it is True, e.g. 769x769. Default: False. |
| 43 | + pretrained (str, optional): The path or url of pretrained model. Default: None. |
| 44 | + """ |
| 45 | + |
| 46 | + def __init__(self, |
| 47 | + num_classes, |
| 48 | + backbone, |
| 49 | + backbone_indices=(2, 3), |
| 50 | + gru_channels=512, |
| 51 | + gru_num_state=128, |
| 52 | + gru_num_node=64, |
| 53 | + enable_auxiliary_loss=True, |
| 54 | + align_corners=False, |
| 55 | + pretrained=None): |
| 56 | + super().__init__() |
| 57 | + |
| 58 | + self.backbone = backbone |
| 59 | + backbone_channels = [ |
| 60 | + backbone.feat_channels[i] for i in backbone_indices |
| 61 | + ] |
| 62 | + |
| 63 | + self.head = GloReHead(num_classes, backbone_indices, backbone_channels, |
| 64 | + gru_channels, gru_num_state, gru_num_node, |
| 65 | + enable_auxiliary_loss) |
| 66 | + self.align_corners = align_corners |
| 67 | + self.pretrained = pretrained |
| 68 | + self.init_weight() |
| 69 | + |
| 70 | + def forward(self, x): |
| 71 | + feat_list = self.backbone(x) |
| 72 | + logit_list = self.head(feat_list) |
| 73 | + return [ |
| 74 | + F.interpolate( |
| 75 | + logit, |
| 76 | + x.shape[2:], |
| 77 | + mode='bilinear', |
| 78 | + align_corners=self.align_corners) for logit in logit_list |
| 79 | + ] |
| 80 | + |
| 81 | + def init_weight(self): |
| 82 | + if self.pretrained is not None: |
| 83 | + utils.load_entire_model(self, self.pretrained) |
| 84 | + |
| 85 | + |
| 86 | +class GloReHead(nn.Layer): |
| 87 | + |
| 88 | + def __init__(self, |
| 89 | + num_classes, |
| 90 | + backbone_indices, |
| 91 | + backbone_channels, |
| 92 | + gru_channels=512, |
| 93 | + gru_num_state=128, |
| 94 | + gru_num_node=64, |
| 95 | + enable_auxiliary_loss=True): |
| 96 | + super().__init__() |
| 97 | + |
| 98 | + in_channels = backbone_channels[1] |
| 99 | + self.conv_bn_relu = layers.ConvBNReLU( |
| 100 | + in_channels, gru_channels, 1, bias_attr=False) |
| 101 | + self.gru_module = GruModule( |
| 102 | + num_input=gru_channels, |
| 103 | + num_state=gru_num_state, |
| 104 | + num_node=gru_num_node) |
| 105 | + |
| 106 | + self.dropout = nn.Dropout(0.1) |
| 107 | + self.classifier = nn.Conv2D(512, num_classes, kernel_size=1) |
| 108 | + self.auxlayer = layers.AuxLayer( |
| 109 | + in_channels=backbone_channels[0], |
| 110 | + inter_channels=backbone_channels[0] // 4, |
| 111 | + out_channels=num_classes) |
| 112 | + |
| 113 | + self.backbone_indices = backbone_indices |
| 114 | + self.enable_auxiliary_loss = enable_auxiliary_loss |
| 115 | + |
| 116 | + def forward(self, feat_list): |
| 117 | + |
| 118 | + logit_list = [] |
| 119 | + x = feat_list[self.backbone_indices[1]] |
| 120 | + |
| 121 | + feature = self.conv_bn_relu(x) |
| 122 | + gru_output = self.gru_module(feature) |
| 123 | + output = self.dropout(gru_output) |
| 124 | + logit = self.classifier(output) |
| 125 | + logit_list.append(logit) |
| 126 | + |
| 127 | + if self.enable_auxiliary_loss: |
| 128 | + low_level_feat = feat_list[self.backbone_indices[0]] |
| 129 | + auxiliary_logit = self.auxlayer(low_level_feat) |
| 130 | + logit_list.append(auxiliary_logit) |
| 131 | + |
| 132 | + return logit_list |
| 133 | + |
| 134 | + |
| 135 | +class GCN(nn.Layer): |
| 136 | + def __init__(self, num_state, num_node, bias=False): |
| 137 | + super(GCN, self).__init__() |
| 138 | + self.conv1 = nn.Conv1D(num_node, num_node, kernel_size=1) |
| 139 | + self.relu = nn.ReLU() |
| 140 | + self.conv2 = nn.Conv1D( |
| 141 | + num_state, num_state, kernel_size=1, bias_attr=bias) |
| 142 | + |
| 143 | + def forward(self, x): |
| 144 | + h = self.conv1(paddle.transpose(x, perm=(0, 2, 1))) |
| 145 | + h = paddle.transpose(h, perm=(0, 2, 1)) |
| 146 | + h = h + x |
| 147 | + h = self.relu(self.conv2(h)) |
| 148 | + return h |
| 149 | + |
| 150 | + |
| 151 | +class GruModule(nn.Layer): |
| 152 | + def __init__(self, |
| 153 | + num_input=512, |
| 154 | + num_state=128, |
| 155 | + num_node=64, |
| 156 | + normalize=False): |
| 157 | + super(GruModule, self).__init__() |
| 158 | + self.normalize = normalize |
| 159 | + self.num_state = num_state |
| 160 | + self.num_node = num_node |
| 161 | + self.reduction_dim = nn.Conv2D(num_input, num_state, kernel_size=1) |
| 162 | + self.projection_mat = nn.Conv2D(num_input, num_node, kernel_size=1) |
| 163 | + self.gcn = GCN(num_state=self.num_state, num_node=self.num_node) |
| 164 | + self.extend_dim = nn.Conv2D( |
| 165 | + self.num_state, num_input, kernel_size=1, bias_attr=False) |
| 166 | + self.extend_bn = nn.SyncBatchNorm(num_input, epsilon=1e-4) |
| 167 | + |
| 168 | + def forward(self, input): |
| 169 | + n, c, h, w = input.shape |
| 170 | + # B, C, H, W |
| 171 | + reduction_dim = self.reduction_dim(input) |
| 172 | + # B, N, H, W |
| 173 | + mat_B = self.projection_mat(input) |
| 174 | + # B, C, H*W |
| 175 | + reshaped_reduction = paddle.reshape( |
| 176 | + reduction_dim, shape=[n, self.num_state, h * w]) |
| 177 | + # B, N, H*W |
| 178 | + reshaped_B = paddle.reshape(mat_B, shape=[n, self.num_node, h * w]) |
| 179 | + # B, N, H*W |
| 180 | + reproject = reshaped_B |
| 181 | + # B, C, N |
| 182 | + node_state_V = paddle.matmul( |
| 183 | + reshaped_reduction, paddle.transpose( |
| 184 | + reshaped_B, perm=[0, 2, 1])) |
| 185 | + |
| 186 | + if self.normalize: |
| 187 | + node_state_V = node_state_V * (1. / reshaped_reduction.shape[2]) |
| 188 | + |
| 189 | + # B, C, N |
| 190 | + gcn_out = self.gcn(node_state_V) |
| 191 | + # B, C, H*W |
| 192 | + Y = paddle.matmul(gcn_out, reproject) |
| 193 | + # B, C, H, W |
| 194 | + Y = paddle.reshape(Y, shape=[n, self.num_state, h, w]) |
| 195 | + Y_extend = self.extend_dim(Y) |
| 196 | + Y_extend = self.extend_bn(Y_extend) |
| 197 | + |
| 198 | + out = input + Y_extend |
| 199 | + return out |
0 commit comments