None_local_block in TensorFlow

最新推荐文章于 2022-04-30 17:50:35 发布
最新推荐文章于 2022-04-30 17:50:35 发布
阅读量920 收藏 6
点赞数
分类专栏： cv-dl 文章标签： Deeplearning
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.youkuaiyun.com/xiaoxiaodongshige/article/details/79571435
版权

None_local_block in TensorFlow

this block are not offcially implement of non_local_block in paper

Non-local Neural Networks https://arxiv.org/pdf/1711.07971.pdf

this code just suit for 2D images data, you can change this code easilly to suit for 3D or 4D data

Whole code you can visit https://github.com/jack-Dong/None_local_block

code


def non_local_block(input_tensor, computation_compression=2, mode='dot'):
    if mode not in ['gaussian', 'embedded', 'dot', 'concatenate']:
        raise ValueError('`mode` must be one of `gaussian`, `embedded`, `dot`,`concatenate`')

    input_shape = input_tensor.get_shape().as_list()
    print(input_shape)
    batchsize, dim1, dim2, channels = input_shape

    if mode == 'gaussian':  # Gaussian instantiation
        x1 = tf.reshape(input_tensor, shape=[-1, dim1 * dim2, channels])
        x2 = tf.reshape(input_tensor, shape=[-1, dim1 * dim2, channels])

        f = tf.matmul(x1, x2, transpose_b=True)

        f = tf.reshape(f, shape=[-1, dim1 * dim2 * dim1 * dim2])

        f = tf.nn.softmax(f, axis=-1)

        f = tf.reshape(f, shape=[-1, dim1 * dim2, dim1 * dim2])

        print("gaussian=", f)
    elif mode == 'dot':
        theta = conv2d(input_tensor, channels, channels // 2, 1)  # add BN relu layer before conv will speed up training
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])

        f = tf.matmul(theta, phi, transpose_b=True)

        # scale the values to make it size invarian t
        f = f / (dim1 * dim2 * channels)

        print("dot f=", f)

    elif mode == 'concatenate':  # this operation cost too much memory, so make sure you input a small resolution  feature map like(14X14 7X7)

        theta = conv2d(input_tensor, channels, channels // 2, 1)
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])

        theta_splits = tf.split(theta, dim1 * dim2, 1)
        phi_splits = tf.split(phi, dim1 * dim2, 1)

        theta_split_shape = tf.shape(theta[0])
        print("theta_split_shape", theta_split_shape)

        initial = tf.constant(1.0 / channels, shape=[channels, 1])

        print('initial', initial)
        W_concat = tf.Variable(initial)

        print("W_concat", W_concat)

        f_matrix = []
        for i in range(dim1 * dim2):
            for j in range(dim1 * dim2):
                print(i, '  ', j)
                tmp = tf.concat([theta_splits[i], phi_splits[j]], 2)
                tmp = tf.reshape(tmp, shape=[-1, channels])
                # print(tmp)
                tmp = tf.matmul(tmp, W_concat)
                print(tmp)
                f_matrix.append(tmp)

        f_matrix_tensor = tf.stack(f_matrix, axis=2)
        print('f_matrix_tensor', f_matrix_tensor)

        f = tf.reshape(f_matrix_tensor, shape=[-1, dim1 * dim2, dim1 * dim2])

        f = f / (dim1 * dim2 * channels)

        print("concatenate f=", f)


    else:  # Embedded Gaussian instantiation
        theta = conv2d(input_tensor, channels, channels // 2, 1)
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])<p>

        if computation_compression > 1:
            phi = tf.layers.max_pooling1d(phi, pool_size=2, strides=computation_compression, padding='SAME')
            print('phi', phi)

        f = tf.matmul(theta, phi, transpose_b=True)

        phi_shape = phi.get_shape().as_list()
        f = tf.reshape(f, shape=[-1, dim1 * dim2 * phi_shape[1]])

        f = tf.nn.def non_local_block(input_tensor, computation_compression=2, mode='dot'):
    if mode not in ['gaussian', 'embedded', 'dot', 'concatenate']:
        raise ValueError('`mode` must be one of `gaussian`, `embedded`, `dot`,`concatenate`')

    input_shape = input_tensor.get_shape().as_list()
    print(input_shape)
    batchsize, dim1, dim2, channels = input_shape

    if mode == 'gaussian':  # Gaussian instantiation
        x1 = tf.reshape(input_tensor, shape=[-1, dim1 * dim2, channels])
        x2 = tf.reshape(input_tensor, shape=[-1, dim1 * dim2, channels])

        f = tf.matmul(x1, x2, transpose_b=True)

        f = tf.reshape(f, shape=[-1, dim1 * dim2 * dim1 * dim2])

        f = tf.nn.softmax(f, axis=-1)

        f = tf.reshape(f, shape=[-1, dim1 * dim2, dim1 * dim2])

        print("gaussian=", f)
    elif mode == 'dot':
        theta = conv2d(input_tensor, channels, channels // 2, 1)  # add BN relu layer before conv will speed up training
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])

        f = tf.matmul(theta, phi, transpose_b=True)

        # scale the values to make it size invarian t
        f = f / (dim1 * dim2 * channels)

        print("dot f=", f)

    elif mode == 'concatenate':  # this operation cost too much memory, so make sure you input a small resolution  feature map like(14X14 7X7)

        theta = conv2d(input_tensor, channels, channels // 2, 1)
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])
<p>
        theta_splits = tf.split(theta, dim1 * dim2, 1)
        phi_splits = tf.split(phi, dim1 * dim2, 1)

        theta_split_shape = tf.shape(theta[0])
        print("theta_split_shape", theta_split_shape)

        initial = tf.constant(1.0 / channels, shape=[channels, 1])

        print('initial', initial)
        W_concat = tf.Variable(initial)

        print("W_concat", W_concat)

        f_matrix = []
        for i in range(dim1 * dim2):
            for j in range(dim1 * dim2):
                print(i, '  ', j)
                tmp = tf.concat([theta_splits[i], phi_splits[j]], 2)
                tmp = tf.reshape(tmp, shape=[-1, channels])
                # print(tmp)
                tmp = tf.matmul(tmp, W_concat)
                print(tmp)
                f_matrix.append(tmp)

        f_matrix_tensor = tf.stack(f_matrix, axis=2)
        print('f_matrix_tensor', f_matrix_tensor)

        f = tf.reshape(f_matrix_tensor, shape=[-1, dim1 * dim2, dim1 * dim2])

        f = f / (dim1 * dim2 * channels)

        print("concatenate f=", f)<p>


    else:  # Embedded Gaussian instantiation
        theta = conv2d(input_tensor, channels, channels // 2, 1)
        theta = tf.reshape(theta, shape=[-1, dim1 * dim2, channels // 2])

        phi = conv2d(input_tensor, channels, channels // 2, 1)
        phi = tf.reshape(phi, shape=[-1, dim1 * dim2, channels // 2])

        if computation_compression > 1:
            phi = tf.layers.max_pooling1d(phi, pool_size=2, strides=computation_compression, padding='SAME')
            print('phi', phi)

        f = tf.matmul(theta, phi, transpose_b=True)

        phi_shape = phi.get_shape().as_list()
        f = tf.reshape(f, shape=[-1, dim1 * dim2 * phi_shape[1]])

        f = tf.nn.softmax(f, axis=-1)

        f = tf.reshape(f, shape=[-1, dim1 * dim2, phi_shape[1]])

        print("Embedded f=", f)

    g = conv2d(input_tensor, channels, channels // 2, 1)
    g = tf.reshape(g, shape=[-1, dim1 * dim2, channels // 2])

    if computation_compression > 1 and mode == 'embedded':
        g = tf.layers.max_pooling1d(g, pool_size=2, strides=computation_compression, padding='SAME')
        print('g', g)

    y = tf.matmul(f, g)

    print('y=', y)

    y = tf.reshape(y, shape=[-1, dim1, dim2, channels // 2])

    y = conv2d(y, channels // 2, channels, kernel_size=3)
    print('y=', y)

    residual = input_tensor + y

    return residualsoftmax(f, axis=-1)

        f = tf.reshape(f, shape=[-1, dim1 * dim2, phi_shape[1]])

        print("Embedded f=", f)

    g = conv2d(input_tensor, channels, channels // 2, 1)
    g = tf.reshape(g, shape=[-1, dim1 * dim2, channels // 2])

    if computation_compression > 1 and mode == 'embedded':
        g = tf.layers.max_pooling1d(g, pool_size=2, strides=computation_compression, padding='SAME')
        print('g', g)

    y = tf.matmul(f, g)

    print('y=', y)

    y = tf.reshape(y, shape=[-1, dim1, dim2, channels // 2])

    y = conv2d(y, channels // 2, channels, kernel_size=3)
    print('y=', y)

    residual = input_tensor + y

    return residual
tensorflow 2.0 kerasd代码-收集整理归纳总结一 non-local
苏打水的杯子的博客
01-19 484
keras自带的Reshape层不需要写batch的维度，但是tensorflow的reshape需要完整的维度。
深度学习平台demo(基础知识）- Keras相关知识点整理(tensorflow2.4)
jiugeshao的专栏
03-18 2566
具体的一些知识点还是从官方文档获取, 博主只列举几个常用的（曾经项目中用的一些功能）。 应用 Applications - Keras 中文文档https://keras.io/zh/applications/Overfit and underfit | TensorFlow Corehttps://tensorflow.google.cn/tutorials/keras/overfit_and_underfit?hl=zh-cn 1.plot_model函数可以可视化出所设计出的模型结构 2....
None-local
Jason_mmt的博客
10-23 829
https://zhuanlan.zhihu.com/p/33345791
Non-local blcok (self-attention)
qq_34535410的博客
05-05 1627
# -*- coding: utf-8 -*- """ @Author : zhwzhong @License : (C) Copyright 2013-2018, hit @Contact : zhwzhong.hit@gmail.com @Software: PyCharm @File : nlrn.py @Time : 2019/5/5 16:27 @D...
NONE LOCAL survey 2018-2019
Aries_M的博客
10-16 811
none local neural network non-local Neural Networks for Video Classification CVPR 2018 motivated from `non-local mean operation, instead of adding layers, the authors choose to use none local mechan...
Non-local 代码实现
weixin_42096202的博客
06-21 4083
# Non-local module class Nl(nn.Module): def __init__(self, wn, in_channels, key_channels, value_channels, out_channels=None, scale=1): super(Nl, self).__init__() # self.scale = sca...
自动将代码升级到TensorFlow 2代码——tf_upgrade_v2.py
一摩尔自由的博客
10-09 6658
TensorFlow升级到2.0后给出了代码自动升级的脚本和使用方法，详情可见官网 TensorFlow官网：https://www.tensorflow.org/guide/upgrade 代码链接： tf_upgrade_v2.py # Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed u...
python install tensorflow 国内源-清华源安装
热门推荐
mac_block的博客
04-08 1万+
pip install tensorflow -i https://pypi.tuna.tsinghua.edu.cn/simple Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple Collecting tensorflow Downloading https://pypi.tuna.tsinghua.edu.cn/packages/dd/0b/6fd33732d436a6735094b6e44359e13c10901e71.
tensorflow2- Grad-CAM类激活可视化
chenvvei的博客
04-24 5788
文章目录Grad-CAM类激活可视化可配置参数Grad-CAM算法测试一下创建叠加的可视化让我们尝试另一张图片参考 Grad-CAM类激活可视化 如何获取图像分类模型的类激活热图。 import numpy as np import tensorflow as tf from tensorflow import keras # Display from IPython.display import Image, display import matplotlib.pyplot as plt import
Python库 | keras_cv_attention_models-1.1.3-py3-none-any.whl
03-21
Keras CV Attention Models引入了多种先进的注意力机制，如SEBlock（Squeeze-and-Excitation）、CBAM（Channel and Spatial Attention Module）、Non-local Block等，这些机制能够提升模型对关键特征的识别能力。...
Non-local_pytorch:非本地块的实现
05-12
非本地_火炬 。 陈述 您可以在lib /中找到不同种类的非本地块。 您可以按照以下显示的“运行步骤”形象化Non_local Attention Map。 该代码在MNIST数据集上进行了测试。 您可以在lib / network.py中选择非本地块的类型。 如果我的代码有问题，请与我联系，谢谢！ 环境 python 3.7.7 pytorch 1.4.0（> = 1.2.0，1.8.1也可以） opencv 4.2.0.34（其他人也可以） 可视化 在第一个非本地层中。 在第二个非本地层中。 运行步骤 在lib / network.py中选择非本地块的类型。 from lib.non_local_concatenation import NONLocalBlock2D from lib.non_local_gaussian import NONLocalBlock2D f
Non-local总结
judgechen1997的博客
05-10 1107
Non-local能够引入全局的信息，就能很好地解决local操作无法看清全局的情况，为后面的层带去更丰富的信息。（可以和FC以及普通local conv对比下，FC也是non-local连接，而且是global的） 这两篇blog感觉写的很好： https://zhuanlan.zhihu.com/p/33345791 https://blog.youkuaiyun.com/u014380165/article/details/80012001 自己总结下， Non-local module的核心就是如下两个公式：
Non-local的详细解析
weixin_43489950的博客
09-26 1002
https://zhuanlan.zhihu.com/p/33345791
深度学习4：TensorFlow基础结构——Session，Variable，Placeholder，Optimizer
慢慢积累
01-25 675
1.会话Session 参考 http://blog.youkuaiyun.com/hanging_gardens/article/details/72784392 https://morvanzhou.github.io/tutorials/machine-learning/tensorflow/ http://wiki.jikexueyuan.com/project/tensorflow-zh/t
MobileNets(V1)的Tensorflow实现
chenxy_bwave的专栏
04-30 2261
MobileNetV1 Tensorflow实现及在cifar10数据集上的实验
Non-local block 即插即用
Only_ZeRo_IT的博客
04-20 3046
import torch import torch.nn as nn import torchvision class NonLocalBlock(nn.Module): def __init__(self, channel): super(NonLocalBlock, self).__init__() self.inter_channel = channel // 2 self.conv_phi = nn.Con...
如何在Tensorflow实现ResNet网络
Ftwhale的博客
04-16 1080
实现ResNet使用的也是先搭建网络，然后评价迭代耗时的方法，在这里我们把 代码分成了ResNet_struct.py和ResNet_run.py，其中ResNet_struct.py用来搭建网络，ResNet_run.py用来评测网络，今天主要介绍ResNet_struct.py，我们先从搭建网络开始，导入一些需要用到的库： 接下来创建一个Block类，这个类用来配置残差学习模型的大小。初始化...
视觉注意力机制 | Non-local模块与Self-attention的之间的关系与区别？
AI算法修炼营的博客
02-28 5706
点击上方“AI算法修炼营”，选择加星标或“置顶”标题以下，全是干货什么是视觉中的注意力机制？计算机视觉（computer vision）中的注意力机制（attention）的基本思想就是...
Non-local U-Nets
qq_36321330的博客
04-11 1199
code tensorflow Non-local 由三个部分组成： U-Net、Residual Blocks、Global Aggregation Block U-Net 作者将原始u-net的concatenation -> sum的原因： 求和不会增加特征图的数量，因此会减少下一层中可训练参数的数量。 具有求和的skip connection可以被视为远程残差连接，可以促进模型的...
用Keras写一个non local block
05-10
非局部块（Non-local Block）是一种用于处理图像、视频等数据的神经网络模块，它可以捕捉全局上下文信息，从而提高模型的性能。在本文中，我们将使用 Keras 实现非局部块。 首先，我们需要导入 Keras 和相关的库： ```python import tensorflow as tf from tensorflow.keras import layers ``` 接下来，我们定义一个名为 NonLocalBlock 的 Keras 层： ```python class NonLocalBlock(layers.Layer): def __init__(self, inter_channels=None, sub_sample=True): super(NonLocalBlock, self).__init__() self.inter_channels = inter_channels self.sub_sample = sub_sample def build(self, input_shape): self.input_channels = input_shape[-1] if self.inter_channels is None: self.inter_channels = self.input_channels // 2 self.theta = layers.Conv2D(filters=self.inter_channels, kernel_size=(1, 1), strides=(1, 1), padding='same') self.phi = layers.Conv2D(filters=self.inter_channels, kernel_size=(1, 1), strides=(1, 1), padding='same') self.g = layers.Conv2D(filters=self.inter_channels, kernel_size=(1, 1), strides=(1, 1), padding='same') self.out_conv = layers.Conv2D(filters=self.input_channels, kernel_size=(1, 1), strides=(1, 1), padding='same') if self.sub_sample: self.pool = layers.MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='same') def call(self, inputs): if self.sub_sample: x = self.pool(inputs) else: x = inputs batch_size = tf.shape(x)[0] spatial_size = tf.shape(x)[1:3] theta = self.theta(x) phi = self.phi(x) g = self.g(x) theta = tf.reshape(theta, [batch_size, -1, self.inter_channels]) phi = tf.reshape(phi, [batch_size, -1, self.inter_channels]) g = tf.reshape(g, [batch_size, -1, self.inter_channels]) theta_phi = tf.matmul(theta, phi, transpose_b=True) theta_phi = tf.nn.softmax(theta_phi, axis=-1) g = tf.matmul(theta_phi, g) g = tf.reshape(g, [batch_size, *spatial_size, self.inter_channels]) out = self.out_conv(g) if self.sub_sample: out = tf.image.resize(out, spatial_size, tf.image.ResizeMethod.NEAREST_NEIGHBOR) return out + inputs ``` 这个 NonLocalBlock 层有两个参数：inter_channels 和 sub_sample。其中，inter_channels 表示非局部块中使用的卷积核数量，sub_sample 表示是否对输入进行下采样。在 __init__ 函数中，我们将这两个参数保存为类属性。在 build 函数中，我们定义了四个卷积层和一个最终输出的卷积层。其中，theta、phi 和 g 分别表示输入经过三个不同卷积层后的特征图，out_conv 表示最终输出的特征图。在 call 函数中，我们首先对输入进行下采样（如果需要），然后将 theta、phi 和 g 分别进行 reshape 操作，以便进行矩阵乘法。接下来，我们计算 theta 和 phi 的点积，并通过 softmax 函数计算权重。最后，我们将权重与 g 进行矩阵乘法，并将结果 reshape 成与输入相同的形状。最后，我们将这个结果与输入相加，并返回结果。 最后，我们可以测试一下这个 NonLocalBlock 层： ```python inputs = layers.Input(shape=(None, None, 64)) x = NonLocalBlock(inter_channels=32, sub_sample=True)(inputs) model = tf.keras.Model(inputs, x) ``` 这个模型的输入是一个形状为 (None, None, 64) 的特征图，输出与输入相同的形状。我们可以使用 model.summary() 来查看模型的结构。 这就是用 Keras 实现非局部块的方法。
评论 1
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值