帧（frame）

除了 Heap 以外，其他的内存占用叫做 Stack（栈）。简单说，Stack 是由于函数运行而临时占用的内存区域。

这种因为用户主动请求而划分出来的内存区域，叫做 Heap（堆）。

Heap 的一个重要特点就是不会自动消失，必须手动释放，或者由垃圾回收机制来回收。

32位 CPU、64位 CPU 这样的名称，其实指的就是寄存器的大小。

一级缓存和二级缓存

学习汇编语言，首先必须了解两个知识点：寄存器和内存模型。

计算机真正能够理解的是低级语言，它专门用来控制硬件。汇编语言就是低级语言，直接描述/控制 CPU 的运行。如果你想了解 CPU 到底干了些什么，以及代码的运行步骤，就一定要学习汇编语言。

然而，仍然存在由于噪声和颜色变化引起的一些边缘像素。为了解决这些杂散响应，必须用弱梯度值过滤边缘像素，并保留具有高梯度值的边缘像素，可以通过选择高低阈值来实现。

非极大值抑制

卷积

有限差分

通过单元测试，开发人员和测试人员可以快速查找 C#、Visual Basic 和 C++ 项目中各个类的方法中的逻辑错误。

accumulator instructions

存储器泄漏只会在程序运行的时间内持续。

在现代操作系统中，一个应用程序使用的常规内存在程序终止时被释放。这表示一个短暂运行的应用程序中的内存泄漏不会导致严重后果。

内存泄漏并非指内存在物理上的消失，而是应用程序分配某段内存后，由于设计错误，导致在释放该段内存之前就失去了对该段内存的控制，从而造成了内存的浪费。

内存泄漏指由于疏忽或错误造成程序未能释放已经不再使用的内存。

This is very important when designing a Gaussian kernel of fixed length.

Shown graphically, we see the familiar bell shaped Gaussian distribution.

Thedistributionisassumed to have a mean of 0.

standarddeviationofthedistribution

Denoising

stereo reconstruction

基于进程的多任务处理是程序的并发执行。基于线程的多任务处理是同一程序的片段的并发执行。

amend

结果是：sizeof(strcut A)值为8；sizeof(struct B)的值却是12。

tweaked

trim the audio

messy situation

季节

为

collectively referred to as

daunting

escalated

distinct

incantation

另类风格的makefiles

opaque string values

sibling

whopping

This programmers manual describes Mini-XML version 2.12, a small XML parsing library that you can use to read and write XML data files in your C and C++ applications.

XML 被设计用来传输和存储数据。

XML 指可扩展标记语言（eXtensible Markup Language）。

向前移动

现在，Git 仓库中有五个对象：三个表示文件快照内容的 blob 对象；一个记录着目录树内容及其中各个文件对应 blob 对象索引的 tree 对象；以及一个包含指向 tree 对象（根目录）的索引和其他提交信息元数据的 commit 对象。概念上来说，仓库中的各个对象保存的数据和相互关系看起来如图 3-1 所示：

在图像测量过程以及机器视觉应用中，为确定空间物体表面某点的三维几何位置与其在图像中对应点之间的相互关系，必须建立相机成像的几何模型，这些几何模型参数就是相机参数。在大多数条件下这些参数必须通过实验与计算才能得到，这个求解参数的过程就称之为相机标定（或摄像机标定）

一句话就是世界坐标到像素坐标的映射，当然这个世界坐标是我们人为去定义的，标定就是已知标定控制点的世界坐标和像素坐标我们去解算这个映射关系，一旦这个关系解算出来了我们就可以由点的像素坐标去反推它的世界坐标，当然有了这个世界坐标，我们就可以进行测量等其他后续操作了～

相机内参数是与相机自身特性相关的参数，比如相机的焦距、像素大小等；

sophisticated features

terrific

fine grained classification

do more sophisticated stuff

clobbering

得到的复合函数记作 g ∘ f : X → Z，定义为对 X 中的所有 x，(g ∘ f )(x) = g(f(x))。

直观地说，复合两个函数是把两个函数链接在一起的过程，内函数的输出就是外函数的输入。

在数学中，复合函数（英语：function composition）是指逐点地把一个函数作用于另一个函数的结果，所得到的第三个函数。

This characteristic of vector-Jacobian product makes it very convenient to feed external gradients into a model that has non-scalar output.

y.grad_fn

Each tensor has a .grad_fn attribute that references a Function that has created the Tensor (except for Tensors created by the user - their grad_fn is None).

Tensor and Function are interconnected and build up an acyclic graph, that encodes a complete history of computation.

There’s one more class which is very important for autograd implementation - a Function.

torch.Tensor is the central class of the package. If you set its attribute .requires_grad as True, it starts to track all operations on it. When you finish your computation you can call .backward() and have all the gradients computed automatically. The gradient for this tensor will be accumulated into .grad attribute.

terms

an acyclic graph

[2018 IJCV] Depth-Based Hand Pose Estimation: Methods, Data, and Challenges. [PDF] [Project Page] [Code]

but 64-bit compiler can only run on 64-bit Windows

Finding the right compiler for your operating platform (Windows, Mac OS X, Ubuntu), for your JDK (32-bit, 64-bit), and figuring out the correct compiler options is the hardest part to get the JNI working!!!

MinGW

能够在 Android 应用中使用原生代码对于想执行以下一项或多项操作的开发者特别有用： 在平台之间移植其应用。 重复使用现有库，或者提供其自己的库供重复使用。 在某些情况下提高性能，特别是像游戏这种计算密集型应用。

Android NDK 是一组允许您将 C 或 C++（“原生代码”）嵌入到 Android 应用中的工具。

np.tile

在这三个点中，动作片出现的频率为三分之二，爱情片出现的频率为三分之一，所以该红色圆点标记的电影为动作片。这个判别过程就是k-近邻算法。

给

Remember that we padded all our sentences to have the same length (59 for our data set).

To allow various hyperparameter configurations we put our code into a TextCNN class, generating the model graph in the init function.

The first layers embeds words into low-dimensional vectors.

semantic clustering

Channels

Pooling Layers

invariance to translation

Now consider a cyclic list and imagine the slow and fast pointers are two runners racing around a circle track. The fast runner will eventually meet the slow runner. Why? Consider this case (we name it case A) - The fast runner is just one step behind the slow runner. In the next iteration, they both increment one and two steps respectively and meet each other.

Partition the shuffled (X, Y) into mini-batches of size mini_batch_size (here 64).

The shuffling step ensures that examples will be split randomly into different mini-batches.

synchronously

What you should remember:

In Stochastic Gradient Descent, you use only 1 training example before updating the gradients. When the training set is large, SGD can be faster. But the parameters will "oscillate" toward the minimum rather than converge smoothly.

A variant of this is Stochastic Gradient Descent (SGD), which is equivalent to mini-batch gradient descent where each mini-batch has just 1 example.

omit

We encourage you to experiment with different algorithms to learn first-hand what works well and how techniques compare.

pre-extracted

You are using a 3 layer neural network, and will add dropout to the first and second hidden layers. We will not apply dropout to the input layer or output layer.

diagonal

Brute Force

tuning

In five courses, you will learn the foundations of Deep Learning, understand how to build neural networks, and learn how to lead successful machine learning projects. You will learn about Convolutional networks, RNNs, LSTM, Adam, Dropout, BatchNorm, Xavier/He initialization, and more. You will work on case studies from healthcare, autonomous driving, sign language reading, music generation, and natural language processing. You will master not only the theory, but also see how it is applied in industry. You will practice all these ideas in Python and in TensorFlow, which we will teach.

So after completing it, you will be able to apply deep learning to a your own applications.

When you finish this, you will have finished the last programming assignment of Week 4, and also the last programming assignment of this course!

General methodology

0.5

standardize the images

Linear Forward

that for every forward function, there is a corresponding backward function. That is why at every step of your forward module you will be storing some values in a cache.

Here are the formulas you need

Remember that back propagation is used to calculate the gradient of the loss function with respect to the parameters.

outputs a row vector A[L]A[L]A^{[L]} containing your predictions

Now you will implement forward and backward propagation. You need to compute the cost, because you want to check if your model is actually learning.

Linear-Activation Forward

Hence, you will implement a function that does the LINEAR forward step followed by an ACTIVATION forward step.

W1 = [[ 0.01788628 0.0043651 0.00096497 -0.01863493 -0.00277388] [-0.00354759 -0.00082741 -0.00627001 -0.00043818 -0.00477218] [-0.01313865 0.00884622 0.00881318 0.01709573 0.00050034] [-0.00404677 -0.0054536 -0.01546477 0.00982367 -0.01101068]]

"W" + str(L)

We will store n[l]n[l]n^{[l]}, the number of units in different layers, in a variable layer_dims.

Create and initialize the parameters of the 2-layer neural network.

The model's structure is: LINEAR -> RELU -> LINEAR -> SIGMOID.

You will write two helper functions that will initialize the parameters for your model.

The cached values are useful for computing gradients.

Figure 1

Outline of the Assignment

The model's structure is [LINEAR -> RELU] ×× \times (L-1) -> LINEAR -> SIGMOID. I.e., it has L−1L−1L-1 layers using a ReLU activation function followed by an output layer with a sigmoid activation function.

And we defined a friend circle is a group of students who are direct or indirect friends.

transitive

It is the first of many examples where a good algorithm—in this case, weighted quick union—makes the difference between being able to efficiently solve a problem and not being able to address it at all.

all its elements

decision boundaries

Integrate parts 4.1, 4.2 and 4.3 in nn_model()

−∑i=0my(i)log(a[2](i))

Simple Logistic Regression

The size of the input layer is: n_x = 5 The size of the hidden layer is: n_h = 2 The size of the output layer is: n_y = 3

The general methodology to build a Neural Network is to

0.5

Keep track of the carry using a variable

the next node

While you are traversing the list, change the current node's next pointer to point to its previous element.

Our original goal for this book was to cover the 50 algorithms that every programmer should know.

String

ADT

Which of the following are true? (Check all that apply.)

建立

欧几里得空间也称为三维座标空间(coordinate space)

称为向量加法与纯量乘法

粗浅地说，空间是一个赋予某种数学结构的集合，该数学结构决定空间的名称。

最终，数学家以抽象的公理化方式定义出一个数学结构，称为向量空间。任何一个数学实体只要满足这些规范都可归类为向量空间。

对任一实矩阵，是半正定矩阵。

设为一个阶实矩阵，阶方阵称为Gramian或Gram矩阵，也有人称之为交互乘积(cross-product)矩阵。

在几何向量空间，向量和的点积(dot product)，或称内积(inner product)，定义为 。

任意二次型都可以转换为等价的，是实对称矩阵。

正定矩阵的概念建立于二次型之上。若是一实对称矩阵且任一满足，我们称是正定的，

矩阵

推得

反对称矩阵(anti-symmetric)

axes

Each row vector Xi{\bf X}_i

The set of 5 observations, measuring 3 variables,

multi_class

Features

Dimensionality

boolean

a classic and very easy binary classification dataset

but now with 32 channels holding the output from each of the filters

To specify that the output tensor should have the same width and height values as the input tensor, we set padding=same here, which instructs TensorFlow to add 0 values to the edges of the output tensor to preserve width and height of 28.

a hyperparameter that we can tune

monochrome images

expect

3,920

skeleton

volatile

overfitting

symmetry breaking

moderately sophisticated

max pool with 2x2 filters and stride 2

shrinks volumes spatially

3

(N - F) / stride + 1

5

convolve

nterspersed

preserve spatial structure

3

First without the brain stuff

Where yy is our predicted probability distribution, and y′y' is the true distribution (the one-hot vector with the digit labels).

cross-entropy

tweaking

flipped from

symbolic variables

interacting operations

overhead

tallies

parser

enqueues

dequeues

auxiliary

exhaust

protocol

versatile

axes