分别来用keras实现通道注意力模块和空间注意力模块。

#通道注意力机制
def channel_attention(input_feature, ratio=8):channel_axis = 1 if K.image_data_format() == "channels_first" else -1channel = input_feature._keras_shape[channel_axis]shared_layer_one = Dense(channel//ratio,activation='relu',kernel_initializer='he_normal',use_bias=True,bias_initializer='zeros')shared_layer_two = Dense(channel,kernel_initializer='he_normal',use_bias=True,bias_initializer='zeros')avg_pool = GlobalAveragePooling2D()(input_feature)    avg_pool = Reshape((1,1,channel))(avg_pool)  # Reshape: width,height,depth#assert avg_pool._keras_shape[1:] == (1,1,channel)avg_pool = shared_layer_one(avg_pool)#assert avg_pool._keras_shape[1:] == (1,1,channel//ratio)avg_pool = shared_layer_two(avg_pool)#assert avg_pool._keras_shape[1:] == (1,1,channel)max_pool = GlobalMaxPooling2D()(input_feature)max_pool = Reshape((1,1,channel))(max_pool)#assert max_pool._keras_shape[1:] == (1,1,channel)max_pool = shared_layer_one(max_pool)#assert max_pool._keras_shape[1:] == (1,1,channel//ratio)max_pool = shared_layer_two(max_pool)#assert max_pool._keras_shape[1:] == (1,1,channel)cbam_feature = Add()([avg_pool,max_pool])   # 处理后的结果相加cbam_feature = Activation('sigmoid')(cbam_feature)  # 获得各通道的权重图if K.image_data_format() == "channels_first":cbam_feature = Permute((3, 1, 2))(cbam_feature)return multiply([input_feature, cbam_feature])

通道注意力：将输入的featuremap，分别经过基于width和height的global max pooling 和global average pooling。

目的：保持通道数不变

"""
我们先分别进行一个通道维度的平均池化和最大池化得到两个 H×W×1 的通道描述，并将这两个描述按照通道拼接在一起；
然后，经过一个 7×7 的卷积层，激活函数为 Sigmoid，得到权重系数 Ms；
最后，拿权重系数和特征 F’ 相乘即可得到缩放后的新特征。
"""
def spatial_attention(input_feature):kernel_size = 7if K.image_data_format() == "channels_first":channel = input_feature._keras_shape[1]cbam_feature = Permute((2,3,1))(input_feature)else:channel = input_feature._keras_shape[-1]cbam_feature = input_featureavg_pool = Lambda(lambda x: K.mean(x, axis=3, keepdims=True))(cbam_feature)  # 对张量求平均值，改变第三维坐标，并保持原本维度#assert avg_pool._keras_shape[-1] == 1max_pool = Lambda(lambda x: K.max(x, axis=3, keepdims=True))(cbam_feature)#assert max_pool._keras_shape[-1] == 1concat = Concatenate(axis=3)([avg_pool, max_pool])  # 拼接#assert concat._keras_shape[-1] == 2cbam_feature = Conv2D(filters = 1,kernel_size=kernel_size,strides=1,padding='same',activation='sigmoid',kernel_initializer='he_normal',use_bias=False)(concat)    #assert cbam_feature._keras_shape[-1] == 1if K.image_data_format() == "channels_first":cbam_feature = Permute((3, 1, 2))(cbam_feature)return multiply([input_feature, cbam_feature])

空间注意力：将输入的featuremap，做一个基于channel的global max pooling 和global average pooling。

目的：保持特征图大小不变

CBAM

将Channel attention模块输出的特征图作为Spatial attention模块的输入特征图

def cbam_block(cbam_feature, ratio=8):cbam_feature = channel_attention(cbam_feature, ratio)cbam_feature = spatial_attention(cbam_feature)return cbam_feature

到底如何做到“随插随用”

核心思想：需要理解输入的是特征图，输出的也是注意力权重与原图相乘后的特征图。

#简单举例：输入特征图，经过卷积，BN层，最后输出的是三者的和，并输入到下一层在相应的位置添加CBAM
inputs = x
residual = layers.Conv2D(filter, kernel_size = (1, 1), strides = strides, padding = 'same')(inputs)
residual = layers.BatchNormalization(axis = bn_axis)(residual)
cbam = cbam_block(residual)
x = layers.add([x, residual, cbam])

SE

from keras.layers import GlobalAveragePooling2D, Reshape, Dense, multiply
from keras import backend as Kdef se_block(input_feature, ratio=8):channel_axis = 1 if K.image_data_format() == "channels_first" else -1channel = input_feature._keras_shape[channel_axis]se_feature = GlobalAveragePooling2D()(input_feature)se_feature = Reshape((1, 1, channel))(se_feature)  # 第一步：压缩(Squeeze), reshape成1✖️1✖️C# assert se_feature._keras_shape[1:] == (1,1,channel)# 第二步：激励(Excitation),# 由两个全连接层组成，其中SERatio是一个缩放参数，这个参数的目的是为了减少通道个数从而降低计算量。# 第一个全连接层有(C/radio)个神经元，输入为1×1×C，输出1×1×(C/radio)。# 第二个全连接层有C个神经元，输入为1×1×(C/radio)，输出为1×1×C。se_feature = Dense(channel // ratio,activation='relu',kernel_initializer='he_normal',use_bias=True,bias_initializer='zeros')(se_feature)#assert se_feature._keras_shape[1:] == (1, 1, channel // ratio)se_feature = Dense(channel,activation='sigmoid',kernel_initializer='he_normal',use_bias=True,bias_initializer='zeros')(se_feature)#assert se_feature._keras_shape[1:] == (1, 1, channel)"""# 因为keras默认为channels_last,没修改不需要加这段if K.image_data_format() == 'channels_first':se_feature = Permute((3, 1, 2))(se_feature)"""se_feature = multiply([input_feature, se_feature])return se_feature

ECA

import math
from keras.layers import *
from keras.layers import Activation
from keras.layers import GlobalAveragePooling2D
import keras.backend as K
import tensorflow as tf
def eca_block(input_feature, b=1, gamma=2, name=""):channel_axis = 1 if K.image_data_format() == "channels_first" else -1channel = input_feature.shape[channel_axis]kernel_size = int(abs((math.log(channel, 2) + b) / gamma))kernel_size = kernel_size if kernel_size % 2 else kernel_size + 1avg_pool = GlobalAveragePooling2D()(input_feature)x = Reshape((-1, 1))(avg_pool)x = Conv1D(1, kernel_size=kernel_size, padding="same", name="eca_layer_" + str(name), use_bias=False, )(x)x = Activation('sigmoid')(x)x = Reshape((1, 1, -1))(x)output = multiply([input_feature, x])return output

keras实现注意力机制相关推荐

keras cnn注意力机制_TensorFlow、PyTorch、Keras：NLP框架哪家强
全文共3412字,预计学习时长7分钟在对TensorFlow.PyTorch和Keras做功能对比之前,先来了解一些它们各自的非竞争性柔性特点吧. 非竞争性特点下文介绍了TensorFlow.Py ...
keras cnn注意力机制_从发展历史视角解析Transformer：从全连接CNN到Transformer
编译 | bluemin校对 | 陈彩娴Transformer架构在机器学习领域(尤其是NLP里)是一项热门研究,为我们带来了许多重要成果,比如:GPT-2.GPT-3等写稿机器人:第一代GPT及其性 ...
keras cnn注意力机制_2019 SSA-CNN（自注意力机制）目标检测算法论文阅读笔记
背景 <SSA-CNN Semantic Self-Attention CNN for Pedestrian Detection>是2019 的工作,其作者来自于南洋理工.这篇文章主要是做 ...
tensorflow环境下的识别食物_在TensorFlow+Keras环境下使用RoI池化一步步实现注意力机制...
在本文中,作者解释了感兴趣区域池化(RoI 池化)的基本概念和一般用法,以及如何使用它来实现注意力机制.他一步步给出了在 Keras 和 TensorFlow 环境下使用 RoI 池化的实现. 选自 ...
通道注意力机制keras_在TensorFlow+Keras环境下使用RoI池化一步步实现注意力机制
项目地址:https://gist.github.com/Jsevillamol/0daac5a6001843942f91f2a3daea27a7 理解 RoI 池化 RoI 池化的概念由 Ross ...
浙大 | FcaNet：频域通道注意力机制（keras实现）
FcaNet: Frequency Channel Attention Networks paper:http://arxiv.org/abs/2012.11879 code:https://git ...
WACV 2020 | ULSAM: 超轻量级子空间注意力机制(keras实现)
ULSAM: Ultra-Lightweight Subspace Attention Module for Compact Convolutional Neural Networks paper:h ...
【深度学习】(8) CNN中的通道注意力机制（SEnet、ECAnet），附Tensorflow完整代码
各位同学好,今天和大家分享一下attention注意力机制在CNN卷积神经网络中的应用,重点介绍三种注意力机制,及其代码复现. 在我之前的神经网络专栏的文章中也使用到过注意力机制,比如在MobileN ...
独家 | 感悟注意力机制
作者:Greg Mehdiyev, Ray Hong, Jinghan Yu, Brendan Artley翻译:陈之炎校对:ZRX本文约2800字,建议阅读12分钟本文由Simon Fraser大 ...

keras实现注意力机制

CBAM

到底如何做到“随插随用”

SE

ECA

keras实现注意力机制相关推荐

最新文章

热门文章