# 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 paddle import paddle.nn as nn import paddle.nn.functional as F from paddleseg.models import layers from paddleseg.cvlibs import manager from paddleseg.utils import utils @manager.MODELS.add_component class ISANet(nn.Layer): """Interlaced Sparse Self-Attention for Semantic Segmentation. The original article refers to Lang Huang, et al. "Interlaced Sparse Self-Attention for Semantic Segmentation" (https://arxiv.org/abs/1907.12273). Args: num_classes (int): The unique number of target classes. backbone (Paddle.nn.Layer): A backbone network. backbone_indices (tuple): The values in the tuple indicate the indices of output of backbone. isa_channels (int): The channels of ISA Module. down_factor (tuple): Divide the height and width dimension to (Ph, PW) groups. enable_auxiliary_loss (bool, optional): A bool value indicates whether adding auxiliary loss. Default: True. align_corners (bool): An argument of F.interpolate. It should be set to False when the output size of feature is even, e.g. 1024x512, otherwise it is True, e.g. 769x769. Default: False. pretrained (str, optional): The path or url of pretrained model. Default: None. """ def __init__(self, num_classes, backbone, backbone_indices=(2, 3), isa_channels=256, down_factor=(8, 8), enable_auxiliary_loss=True, align_corners=False, pretrained=None): super().__init__() self.backbone = backbone self.backbone_indices = backbone_indices in_channels = [self.backbone.feat_channels[i] for i in backbone_indices] self.head = ISAHead(num_classes, in_channels, isa_channels, down_factor, enable_auxiliary_loss) self.align_corners = align_corners self.pretrained = pretrained self.init_weight() def forward(self, x): feats = self.backbone(x) feats = [feats[i] for i in self.backbone_indices] logit_list = self.head(feats) logit_list = [ F.interpolate( logit, paddle.shape(x)[2:], mode='bilinear', align_corners=self.align_corners, align_mode=1) for logit in logit_list ] return logit_list def init_weight(self): if self.pretrained is not None: utils.load_entire_model(self, self.pretrained) class ISAHead(nn.Layer): """ The ISAHead. Args: num_classes (int): The unique number of target classes. in_channels (tuple): The number of input channels. isa_channels (int): The channels of ISA Module. down_factor (tuple): Divide the height and width dimension to (Ph, PW) groups. enable_auxiliary_loss (bool, optional): A bool value indicates whether adding auxiliary loss. Default: True. """ def __init__(self, num_classes, in_channels, isa_channels, down_factor, enable_auxiliary_loss): super(ISAHead, self).__init__() self.in_channels = in_channels[-1] inter_channels = self.in_channels // 4 self.inter_channels = inter_channels self.down_factor = down_factor self.enable_auxiliary_loss = enable_auxiliary_loss self.in_conv = layers.ConvBNReLU( self.in_channels, inter_channels, 3, bias_attr=False) self.global_relation = SelfAttentionBlock(inter_channels, isa_channels) self.local_relation = SelfAttentionBlock(inter_channels, isa_channels) self.out_conv = layers.ConvBNReLU( inter_channels * 2, inter_channels, 1, bias_attr=False) self.cls = nn.Sequential( nn.Dropout2D(p=0.1), nn.Conv2D(inter_channels, num_classes, 1)) self.aux = nn.Sequential( layers.ConvBNReLU( in_channels=1024, out_channels=256, kernel_size=3, bias_attr=False), nn.Dropout2D(p=0.1), nn.Conv2D(256, num_classes, 1)) def forward(self, feat_list): C3, C4 = feat_list x = self.in_conv(C4) x_shape = paddle.shape(x) P_h, P_w = self.down_factor Q_h, Q_w = paddle.ceil(x_shape[2] / P_h).astype('int32'), paddle.ceil( x_shape[3] / P_w).astype('int32') pad_h, pad_w = (Q_h * P_h - x_shape[2]).astype('int32'), ( Q_w * P_w - x_shape[3]).astype('int32') if pad_h > 0 or pad_w > 0: padding = paddle.concat( [ pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2 ], axis=0) feat = F.pad(x, padding) else: feat = x feat = feat.reshape([0, x_shape[1], Q_h, P_h, Q_w, P_w]) feat = feat.transpose([0, 3, 5, 1, 2, 4]).reshape([-1, self.inter_channels, Q_h, Q_w]) feat = self.global_relation(feat) feat = feat.reshape([x_shape[0], P_h, P_w, x_shape[1], Q_h, Q_w]) feat = feat.transpose([0, 4, 5, 3, 1, 2]).reshape([-1, self.inter_channels, P_h, P_w]) feat = self.local_relation(feat) feat = feat.reshape([x_shape[0], Q_h, Q_w, x_shape[1], P_h, P_w]) feat = feat.transpose([0, 3, 1, 4, 2, 5]).reshape( [0, self.inter_channels, P_h * Q_h, P_w * Q_w]) if pad_h > 0 or pad_w > 0: feat = paddle.slice( feat, axes=[2, 3], starts=[pad_h // 2, pad_w // 2], ends=[pad_h // 2 + x_shape[2], pad_w // 2 + x_shape[3]]) feat = self.out_conv(paddle.concat([feat, x], axis=1)) output = self.cls(feat) if self.enable_auxiliary_loss: auxout = self.aux(C3) return [output, auxout] else: return [output] class SelfAttentionBlock(layers.AttentionBlock): """General self-attention block/non-local block. Args: in_channels (int): Input channels of key/query feature. channels (int): Output channels of key/query transform. """ def __init__(self, in_channels, channels): super(SelfAttentionBlock, self).__init__( key_in_channels=in_channels, query_in_channels=in_channels, channels=channels, out_channels=in_channels, share_key_query=False, query_downsample=None, key_downsample=None, key_query_num_convs=2, key_query_norm=True, value_out_num_convs=1, value_out_norm=False, matmul_norm=True, with_out=False) self.output_project = self.build_project( in_channels, in_channels, num_convs=1, use_conv_module=True) def forward(self, x): context = super(SelfAttentionBlock, self).forward(x, x) return self.output_project(context)