Learning Memory-guided Normality代码学习笔记
Learning Memory-guided Normality代码学习笔记
记忆模块核心
Memory部分的核心在于以下定义Memory类的部分。
class Memory(nn.Module):
def __init__(self, memory_size, feature_dim, key_dim, temp_update, temp_gather):
super(Memory, self).__init__()
# Constants
self.memory_size = memory_size
self.feature_dim = feature_dim
self.key_dim = key_dim
self.temp_update = temp_update
self.temp_gather = temp_gather
def hard_neg_mem(self, mem, i):
similarity = torch.matmul(mem,torch.t(self.keys_var))
similarity[:,i] = -1
_, max_idx = torch.topk(similarity, 1, dim=1)
return self.keys_var[max_idx]
def random_pick_memory(self, mem, max_indices):
m, d = mem.size()
output = []
for i in range(m):
flattened_indices = (max_indices==i).nonzero()
a, _ = flattened_indices.size()
if a != 0:
number = np.random.choice(a, 1)
output.append(flattened_indices[number, 0])
else:
output.append(-1)
return torch.tensor(output)
def get_update_query(self, mem, max_indices, update_indices, score, query, train):
m, d = mem.size()
if train:
query_update = torch.zeros((m,d)).cuda()
# random_update = torch.zeros((m,d)).cuda()
for i in range(m):
idx = torch.nonzero(max_indices.squeeze(1)==i)
a, _ = idx.size()
if a != 0:
query_update[i] = torch.sum(((score[idx,i] / torch.max(score[:,i])) *query[idx].squeeze(1)), dim=0)
else:
query_update[i] = 0
return query_update
else:
query_update = torch.zeros((m,d)).cuda()
for i in range(m):
idx = torch.nonzero(max_indices.squeeze(1)==i)
a, _ = idx.size()
if a != 0:
query_update[i] = torch.sum(((score[idx,i] / torch.max(score[:,i])) *query[idx].squeeze(1)), dim=0)
else:
query_update[i] = 0
return query_update
def get_score(self, mem, query):
bs, h,w,d = query.size()
m, d = mem.size()
score = torch.matmul(query, torch.t(mem))# b X h X w X m
score = score.view(bs*h*w, m)# (b X h X w) X m
score_query = F.softmax(score, dim=0)
score_memory = F.softmax(score,dim=1)
return score_query, score_memory
def forward(self, query, keys, train=True):
batch_size, dims,h,w = query.size() # b X d X h X w
query = F.normalize(query, dim=1)
query = query.permute(0,2,3,1) # b X h X w X d
#train
if train:
#losses
separateness_loss, compactness_loss = self.gather_loss(query,keys, train)
# read
updated_query, softmax_score_query,softmax_score_memory = self.read(query, keys)
#update
updated_memory = self.update(query, keys, train)
return updated_query, updated_memory, softmax_score_query, softmax_score_memory, separateness_loss, compactness_loss
#test
else:
# loss
compactness_loss, query_re, top1_keys, keys_ind = self.gather_loss(query,keys, train)
# read
updated_query, softmax_score_query,softmax_score_memory = self.read(query, keys)
#update
updated_memory = keys
return updated_query, updated_memory, softmax_score_query, softmax_score_memory, query_re, top1_keys,keys_ind, compactness_loss
def update(self, query, keys,train):
batch_size, h,w,dims = query.size() # b X h X w X d
softmax_score_query, softmax_score_memory = self.get_score(keys, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
_, gathering_indices = torch.topk(softmax_score_memory, 1, dim=1)
_, updating_indices = torch.topk(softmax_score_query, 1, dim=0)
if train:
query_update = self.get_update_query(keys, gathering_indices, updating_indices, softmax_score_query, query_reshape,train)
updated_memory = F.normalize(query_update + keys, dim=1)
else:
query_update = self.get_update_query(keys, gathering_indices, updating_indices, softmax_score_query, query_reshape, train)
updated_memory = F.normalize(query_update + keys, dim=1)
return updated_memory.detach()
def pointwise_gather_loss(self, query_reshape, keys, gathering_indices, train):
n,dims = query_reshape.size() # (b X h X w) X d
loss_mse = torch.nn.MSELoss(reduction='none')
pointwise_loss = loss_mse(query_reshape, keys[gathering_indices].squeeze(1).detach())
return pointwise_loss
def gather_loss(self,query, keys, train):
batch_size, h,w,dims = query.size() # b X h X w X d
if train:
loss = torch.nn.TripletMarginLoss(margin=1.0)
loss_mse = torch.nn.MSELoss()
softmax_score_query, softmax_score_memory = self.get_score(keys, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
_, gathering_indices = torch.topk(softmax_score_memory, 2, dim=1)
#1st, 2nd closest memories
pos = keys[gathering_indices[:,0]]
neg = keys[gathering_indices[:,1]]
top1_loss = loss_mse(query_reshape, pos.detach())
gathering_loss = loss(query_reshape,pos.detach(), neg.detach())
return gathering_loss, top1_loss
else:
loss_mse = torch.nn.MSELoss()
softmax_score_query, softmax_score_memory = self.get_score(keys, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
_, gathering_indices = torch.topk(softmax_score_memory, 1, dim=1)
gathering_loss = loss_mse(query_reshape, keys[gathering_indices].squeeze(1).detach())
return gathering_loss, query_reshape, keys[gathering_indices].squeeze(1).detach(), gathering_indices[:,0]
def read(self, query, updated_memory):
batch_size, h,w,dims = query.size() # b X h X w X d
softmax_score_query, softmax_score_memory = self.get_score(updated_memory, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
concat_memory = torch.matmul(softmax_score_memory.detach(), updated_memory) # (b X h X w) X d
updated_query = torch.cat((query_reshape, concat_memory), dim = 1) # (b X h X w) X 2d
updated_query = updated_query.view(batch_size, h, w, 2*dims)
updated_query = updated_query.permute(0,3,1,2)
return updated_query, softmax_score_query, softmax_score_memory
Update过程
调用get_update_query(self, mem, max_indices, update_indices, score, query, train)函数计算\(query\_ dpdate= \sum_{k \in U_{t}^M} v_t^{'k,m}q_t^k\)
然后计算\(f(P^m+query_dpdate)\)
文中对f的描述为L2正则。
看一下get_update_query函数的定义:
def get_update_query(self, mem, max_indices, update_indices, score, query, train):
m, d = mem.size()
if train:
query_update = torch.zeros((m,d)).cuda()
# random_update = torch.zeros((m,d)).cuda()
for i in range(m):
idx = torch.nonzero(max_indices.squeeze(1)==i)
a, _ = idx.size()
if a != 0:
query_update[i] = torch.sum(((score[idx,i] / torch.max(score[:,i])) *query[idx].squeeze(1)), dim=0)
else:
query_update[i] = 0
return query_update
else:
query_update = torch.zeros((m,d)).cuda()
for i in range(m):
idx = torch.nonzero(max_indices.squeeze(1)==i)
a, _ = idx.size()
if a != 0:
query_update[i] = torch.sum(((score[idx,i] / torch.max(score[:,i])) *query[idx].squeeze(1)), dim=0)
else:
query_update[i] = 0
return query_update
在定义中,我们需要看到\(v_t^{'k,m}\)的计算。代码是通过(score[idx,i] / torch.max(score[:,i])实现的,进一步,我们需要查看\(v_t^{k,m}\)的计算过程。这个参数与\(w\)一样是权重,文中通过get_score函数计算权重,如下为此函数的定义:
def get_score(self, mem, query):
#计算权重$w_t^{k,m}$
bs, h,w,d = query.size()
m, d = mem.size()
score = torch.matmul(query, torch.t(mem))# b X h X w X m
score = score.view(bs*h*w, m)# (b X h X w) X m
score_query = F.softmax(score, dim=0)
score_memory = F.softmax(score,dim=1)
return score_query, score_memory
实现了文献中的权重计算
Read过程
def read(self, query, updated_memory):
#Read部分
batch_size, h,w,dims = query.size() # b X h X w X d
softmax_score_query, softmax_score_memory = self.get_score(updated_memory, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
concat_memory = torch.matmul(softmax_score_memory.detach(), updated_memory) # (b X h X w) X d
# 权重和memory获得加权均值
updated_query = torch.cat((query_reshape, concat_memory), dim = 1) # (b X h X w) X 2d
# 进行拼接
updated_query = updated_query.view(batch_size, h, w, 2*dims)
updated_query = updated_query.permute(0,3,1,2)
return updated_query, softmax_score_query, softmax_score_memory
核心部分在代码中给出了注释。
forward过程
separateness_loss, compactness_loss = self.gather_loss(query,keys, train)
# read
updated_query, softmax_score_query,softmax_score_memory = self.read(query, keys)
#update
updated_memory = self.update(query, keys, train)
return updated_query, updated_memory, softmax_score_query, softmax_score_memory, separateness_loss, compactness_loss
分别调用update函数和read函数
需要说明损失函数的定义,\(L = L_{rec} + \lambda _cL_{compact}+ \lambda _sL_{separate}\)
代码中通过gather_loss函数实现。
def gather_loss(self,query, keys, train):
batch_size, h,w,dims = query.size() # b X h X w X d
if train:
loss = torch.nn.TripletMarginLoss(margin=1.0)
# 计算Feature separateness loss的主要函数
loss_mse = torch.nn.MSELoss()
# 计算均方差损失
softmax_score_query, softmax_score_memory = self.get_score(keys, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
_, gathering_indices = torch.topk(softmax_score_memory, 2, dim=1)
#1st, 2nd closest memories
pos = keys[gathering_indices[:,0]]
neg = keys[gathering_indices[:,1]]
top1_loss = loss_mse(query_reshape, pos.detach())
gathering_loss = loss(query_reshape,pos.detach(), neg.detach())
return gathering_loss, top1_loss
else:
loss_mse = torch.nn.MSELoss()
softmax_score_query, softmax_score_memory = self.get_score(keys, query)
query_reshape = query.contiguous().view(batch_size*h*w, dims)
_, gathering_indices = torch.topk(softmax_score_memory, 1, dim=1)
gathering_loss = loss_mse(query_reshape, keys[gathering_indices].squeeze(1).detach())
return gathering_loss, query_reshape, keys[gathering_indices].squeeze(1).detach(), gathering_indices[:,0]
Learning Memory-guided Normality代码学习笔记的更多相关文章
- DeepLearnToolbox-master代码学习笔记
卷积神经网络(CNN)博大精深,网上资料浩如烟海,让初学者无从下手.笔者以为,学习编程还是从代码实例入们最好.目前,学习CNN最好的代码实例就是,DeepLearnToolbox-master,不用装 ...
- 《Learning Play! Framework 2》学习笔记——案例研究1(Templating System)
注解: 这是对<Learning Play! Framework 2>第三章的学习 本章是一个显示聊天记录的项目,只有一个页面,可以自动对聊天记录进行排序.分组和显示,并整合使用了less ...
- Machine Learning In Action 第二章学习笔记: kNN算法
本文主要记录<Machine Learning In Action>中第二章的内容.书中以两个具体实例来介绍kNN(k nearest neighbors),分别是: 约会对象预测 手写数 ...
- C# 好代码学习笔记(1):文件操作、读取文件、Debug/Trace 类、Conditional条件编译、CLS
目录 1,文件操作 2,读取文件 3,Debug .Trace类 4,条件编译 5,MethodImpl 特性 5,CLSCompliantAttribute 6,必要时自定义类型别名 目录: 1,文 ...
- 1.JAVA中使用JNI调用C++代码学习笔记
Java 之JNI编程1.什么是JNI? JNI:(Java Natibe Inetrface)缩写. 2.为什么要学习JNI? Java 是跨平台的语言,但是在有些时候仍然是有需要调用本地代码 ( ...
- APM代码学习笔记1
libraries目录 传感器 AP_InertialSensor 惯性导航传感器 就是陀螺仪加速计 AP_Baro 气压计 居然支持BMP085 在我印象中APM一直用高端的MS5611 AP_Co ...
- boost timer代码学习笔记
socket连接中需要判断超时 所以这几天看了看boost中计时器的文档和示例 一共有五个例子 从简单的同步等待到异步调用超时处理 先看第一个例子 // timer1.cpp: 定义控制台应用程序的入 ...
- Hands on Machine Learning with Sklearn and TensorFlow学习笔记——机器学习概览
一.什么是机器学习? 计算机程序利用经验E(训练数据)学习任务T(要做什么,即目标),性能是P(性能指标),如果针对任务T的性能P随着经验E不断增长,成为机器学习.[这是汤姆米切尔在1997年定义] ...
- cc代码学习笔记1
#define #define INT32 int #define INT8 char #define CHAR char #define SSHORT signed short #define IN ...
随机推荐
- 话说CAS
一.前言 cas 一般认为是compare and swap 也可以认为是compare and set cas涉及三个值 1)P 变量内存地址 2)E 期望值 ,CPU做计算之前拿出来的旧值 3) ...
- C# 基础 - string 和 Datetime
1. string 1. 格式化填充 string str = "this {0} a {1}"; Console.WriteLine(string.Format(str, &qu ...
- 超详细Linux新手快速入门(一)——Linux的介绍安装以及虚拟机的介绍安装
一.Linux的介绍 1.Linux和Windows的比较 Linux是一款操作系统,其性能稳定,因其防火墙组件高效安全.简单易配置,所以获得了追求速度和安全的一些企业和人群的青睐.与我们日常所熟知 ...
- 【Azure 微服务】PowerShell中,用Connect-ServiceFabricCluster命令无法连接到sf-test.chinaeast2.cloudapp.chinacloudapi.cn:19000 问题分析
问题描述 Azure Service Fabric提供了PowerShell的指令来进行创建,管理资源,如Get-ServiceFabricClusterHealth 获取当前集群的健康状态,但这些命 ...
- [LOJ 572] Misaka Network 与求和
一.题目 点此看题 二.解法 直接推柿子吧: \[\sum_{i=1}^n\sum_{j=1}^nf(\gcd(i,j))^k \] \[\sum_{d=1}^nf(d)^k\sum_{i=1}^{n ...
- springboot系列四:springboot整合mybatis jsp
一.用IDEA 创建maven项目 项目目录结构 1.添加pom jar依赖 <?xml version="1.0" encoding="UTF-8"?& ...
- 如何在 ASP.Net Core 中实现 健康检查
健康检查 常用于判断一个应用程序能否对 request 请求进行响应,ASP.Net Core 2.2 中引入了 健康检查 中间件用于报告应用程序的健康状态. ASP.Net Core 中的 健康检查 ...
- Apache Pulsar 在能源互联网领域的落地实践
关于 Apache Pulsar Apache Pulsar 是 Apache 软件基金会顶级项目,是下一代云原生分布式消息流平台,集消息.存储.轻量化函数式计算为一体,采用计算与存储分离架构设计,支 ...
- 归并排序(JAVA语言)
public class merge { public static void main(String[] args) { // TODO Auto-generated method stub int ...
- HarmonyOS三方件开发指南(15)-LoadingView功能介绍
目录: 1. LoadingView组件功能介绍2. Lottie使用方法3. Lottie开发实现4.<HarmonyOS三方件开发指南>系列文章合集 1. LoadingView组件功 ...