diff --git a/TODO/deep-learning-1-setting-up-aws-image-recognition.md b/TODO/deep-learning-1-setting-up-aws-image-recognition.md index b2fb2c0a009..5a18e49954a 100644 --- a/TODO/deep-learning-1-setting-up-aws-image-recognition.md +++ b/TODO/deep-learning-1-setting-up-aws-image-recognition.md @@ -3,104 +3,110 @@ > * 原文作者：[Rutger Ruizendaal](https://medium.com/@r.ruizendaal) > * 译文出自：[掘金翻译计划](https://github.com/xitu/gold-miner) > * 本文永久链接：[https://github.com/xitu/gold-miner/blob/master/TODO/deep-learning-1-setting-up-aws-image-recognition.md](https://github.com/xitu/gold-miner/blob/master/TODO/deep-learning-1-setting-up-aws-image-recognition.md) -> * 译者： -> * 校对者： +> * 译者：[lileizhenshuai](https://github.com/lileizhenshuai) +> * 校对者：[Tina92](https://github.com/Tina92) [sqrthree](https://github.com/sqrthree) -# Deep Learning #1: Setting up AWS & Image Recognition +# 深度学习系列1：设置 AWS & 图像识别 -*This post is part of a series on deep learning. Check-out part 2 *[*here*](https://medium.com/@r.ruizendaal/deep-learning-2-f81ebe632d5c)* and part 3 *[*here*](https://medium.com/@r.ruizendaal/deep-learning-3-more-on-cnns-handling-overfitting-2bd5d99abe5d)*.* +**这篇文章是深度学习系列的第一部分。你可以在[这里](https://github.com/xitu/gold-miner/blob/master/TODO/deep-learning-2-convolutional-neural-networks.md)查看第二部分，以及[这里](https://github.com/xitu/gold-miner/blob/master/TODO/deep-learning-3-more-on-cnns-handling-overfitting.md)查看第三部分。** ![](https://cdn-images-1.medium.com/max/1600/1*y3guCmNkYLF2uR09Fslh5g.png) -This week: classifying images of cats and dogs -Welcome to this first entry in this series on practical deep learning. In this entry I will setup the Amazon Web Services (AWS) instance and use a pre-trained model to classify images of cats and dogs. +本周的任务：对猫和狗的图像进行分类 -In this complete series I will be blogging about my process in the first part of the Fast AI deep learning course. This course was first given at the Data Institute at the University of San Francisco and is now available as a MOOC. Recently the authors gave part 2 of the course which will become available online in a couple of months. The main reason for following this course is my extreme interest in deep learning. I have found many online resources regarding machine learning but practical courses on deep learning seem to be a rarity. Deep learning seems to be an exclusive group that is just a little harder to get into. The first thing needed to start on deep learning is a GPU. In this course we use the p2 instance from AWS. Let’s get that set up. +欢迎阅读本系列第一篇关于实战深度学习的文章。在本文中，我将创建 Amazon Web Services（AWS）实例，并使用预先训练的模型对猫和狗的图像进行分类。 -The first week of this course really focused on the setup. Getting your deep learning setup right can take a while and it is important to get everything working correctly. This includes setting up AWS, creating and configuring the GPU instance, setting up the process of ssh-ing into the server and managing your directories. +在这个完整的系列里，我会记录下我在 Fast AI 深度学习课程的第一部分内容的进度。这门课程最初是由旧金山大学数据研究所提供的，并且现在能够在 MOOC 上观看。最近，这门课的作者提供了第二部分的内容，并且在接下来的几个月都可以在网上观看。我上这门课的主要是因为我对深度学习有着强烈的兴趣。我在网上发现了许多关于机器学习的课程，但有关深度学习的实战课程还是比较少见的。深度学习似乎因为进入门槛略高一点，而被单独列出。开始深度学习之前我们首先需要一个 GPU，在这门课程里我们会使用 AWS 的 p2 实例。现在让我们一起来准备它。 -I ran into some issues with permissions on my internship laptop. Let me give you one tip that will save a lot of time in trying to bypass this: Make sure you have full administrator access on your laptop before attempting this. Some lovely engineers offered to setup the GPU instance for me, but they didn’t have time to do it soon. So I decided to take matters into my own hands. +这门课程第一周，我们会把重点放在准备工作上。正确地准备深度学习需要一点时间，但这对一切能正确运行很重要。这包括了设置 AWS，创建和配置 GPU 实例，设置 ssh 连接服务器以及管理你的目录。 -The scrips for setting up the AWS instance are written in bash. If you’re working on a Windows machine you will need a program that can handle this. I’m using Cygwin. I want to share some issues (and their solutions) that I ran into during the install. You can skip this if you’re not following the Fast AI course and just reading along. Some issues that I ran into during the setup process were: +我在实习期用的笔记本电脑上遇到了一些权限问题。我有个建议能够避免这个问题，从而帮你节省大量时间：在尝试操作之前，确保你在你的笔记本电脑上拥有完整的管理员权限。一些热情的工程师提出帮助我设置 GPU 实例，但是他们不能马上帮我搞定，所以我决定自己来。 -- The bash scripts throw an error +用来设置 AWS 的脚本是用 bash 写的，如果你用的是 Windows 操作系统，那么你需要一个能够处理它的程序，我用的是 Cygwin。我想分享一些在设置过程中我遇到的问题（以及对应的解决方案）。如果你没有在上 Fast AI 课程，你可以跳过这部分继续阅读。我在设置过程中所遇到的问题有： -I have read some possible explanations for this, but not a clear solution that worked for me. The setup script of the course on Github is now split in two scripts: setup_p2.sh and setup_instance.sh. In case you cannot get these two scripts to work you can use [this](https://github.com/ericschwarzkopf/courses/blob/dc06ce745a30850e7937858fb26a67df2aff329d/setup/setup_p2.sh) script to setup your p2 instance. If the script does not run be sure to try the raw version as well. +- bash 脚本报错 -I had a similar issue with the aws-alias.sh script. Adding a ‘ at the end of line 7 fixed this issue. Here is a before and after of line 7: + 我看过一些可能的原因，但是没有一个是对我有用的解决方案。Github 上这个课程的设置脚本有两个：setup_p2.sh 和 setup_instance.sh。如果上面那两个脚本不能用，你可以用[这个](https://github.com/ericschwarzkopf/courses/blob/dc06ce745a30850e7937858fb26a67df2aff329d/setup/setup_p2.sh)脚本试试。但如果这个脚本还是不行，请务必再尝试使用原始版本的脚本。 - alias aws-state='aws ec2 describe-instances --instance-ids $instanceId --query "Reservations[0].Instances[0].State.Name" + 我在 aws-alias.sh 这个脚本上也遇到了同样的问题，在第七行的末尾加上 `'` 能够解决这个问题。下面是修改前和修改后的第七行： - alias aws-state='aws ec2 describe-instances --instance-ids $instanceId --query "Reservations[0].Instances[0].State.Name"' + > alias aws-state='aws ec2 describe-instances --instance-ids $instanceId --query "Reservations[0].Instances[0].State.Name" + + > alias aws-state='aws ec2 describe-instances --instance-ids $instanceId --query "Reservations[0].Instances[0].State.Name"' -[Here](https://gist.github.com/LeCoupa/122b12050f5fb267e75f) is a Bash cheat sheet for everyone who is not familiar with Bash. I greatly recommend this since you will need Bash to interact with your instance. + [这里](https://gist.github.com/LeCoupa/122b12050f5fb267e75f)有一个为不熟悉 Bash 的人准备的 Bash 备忘录，因为你需要通过 Bash 来和你的实例进行交互，所以我非常推荐你去看看。 -- The Anaconda install. The video mentions that you should install Anaconda before installing Cygwin. This can be a bit confusing as you need to use the ‘Cygwin python’ to run the pip commands in there and not a local Anaconda distribution. +- Anaconda 的安装。视频中提到你需要在安装 Cygwin 之前先安装 Anaconda。你可能感到有些疑惑，因为你需要用“Cygwin python”来运行 pip 命令而不是一个本地的 Anaconda 分发版。 -Additionally, [this](https://github.com/TomLous/practical-deep-learning) repository has a nice step-by-step guide on getting your instance running. +另外，[这个](https://github.com/TomLous/practical-deep-learning)仓库有一个手把手的教程教你如何让你的实例运行起来。 --- -#### Getting started with Deep Learning +#### 开始深度学习 -After some issues I got my GPU instance running. Time to get started with deep learning! A quick disclaimer: in these blogs I won’t be repeating exactly what is listed in the lesson notes, there is no need for that. I will be highlighting some things that I found really interesting, as well as issues and ideas that I ran into while going through the lesson. +解决了一些问题之后我总算让我的 GPU 实例运行起来了。是时候开始深度学习了！一个简短的免责声明：在这一系列博客中，我不会重复已经在课程笔记中列出的内容，因为没必要。我会强调一些我觉得很有趣的事情，以及我在课程中遇到的问题和一些想法。 -Let’s start with the first question that is probably on your mind: **What is deep learning and why is it experiencing such hype right now?** +让我们从第一个可能已经在你脑海中的问题开始：**什么是深度学习？它现在为什么被炒得这么火？** -Deep learning simply is an artificial neural network with multiple hidden layers, this makes them ‘deep’. A general neural network only has one, maybe two hidden layers. A deep neural network has much more hidden layers. They also have different types of layers than the ‘simple’ ones in the normal neural network. +深度学习只是一个有着多个隐含层的人造神经网络，隐含层让它变得“深度”。一般的神经网络只有一层或者两层的隐含层，而一个深度神经网络有更多的隐含层。它们也具有与一般神经网络中的“简单”层不同类型的层。 ![](https://cdn-images-1.medium.com/max/1600/1*CcQPggEbLgej32mVF2lalg.png) -(Shallow) Neural Network -Currently deep learning is consistently beating performance on well-known datasets. Therefore deep learning has been experiencing a lot of hype. There are three reasons for the popularity of deep learning: +(浅) 神经网络 -- Infinitely flexible function -- All-purpose parameter fitting -- Fast and scalable +目前，深度学习在一些著名的数据集上不断地有着出色的表现，所以深度学习也经历了不少的炒作。深度学习的流行有三个原因： -The Neural Network is modeled after the human brain. According to the universal approximation theorem it can theoretically solve any function. The Neural Network is trained through backward propagation, which allows us to fit the parameters of the model to all these different functions. The last reason is the main one for the recent achievements in deep learning. Because of advancements in the gaming industry and the developments of powerful GPUs it is now possible to train deep neural networks in a fast and scalable way. +- 无限灵活的函数 +- 通用参数拟合 +- 迅速以及可拓展 -In this first lesson the goal is to use a pre-trained model, namely Vgg16 to classify images of cats and dogs. Vgg16 is a lightweight version of the model that won the Imagenet challenge in 2014. This is a yearly challenge and probably the biggest one in computer vision. We can take this pre-trained model and apply it to our dataset of cat and dog images. Our dataset has been edited by the authors of the course to make sure it is in the right format for our model. The original dataset can be found on [Kaggle](https://www.kaggle.com/c/dogs-vs-cats). When this competition was originally run in 2013, the state of the art was 80% accuracy. Our simple model will already achieve 97% accuracy. Mind-blowing right? This is how some of the pictures and their predicted labels look: +神经网络是通过模仿人脑而设计的。根据通用近似定理，它理论上能拟合任何函数。神经网络通过反向传播算法来训练，这使得我们能够调整模型的参数来适应不同的函数。最后一个原因，也是深度学习近期取得众多成就的主要原因。因为游戏行业的进步和 GPU 计算能力的强劲发展，现在我们以非常快速和可扩展的方式来训练深层的神经网络。 + +在第一节课里，我们的目标是使用一个叫做 Vgg16 的预先训练好的模型，来对猫和狗的图片进行分类。Vgg16 是 2014 年赢得 Imagenet 比赛模型的一个轻量级版本。这是一个年度的比赛并且可能是计算机视觉方面最大的一个比赛。我们可以利用这预先训练好的模型，并且把它应用到我们的猫和狗的图片数据集上。我们的数据集已经被课程的作者编辑过了，以确保它的格式正确。原始的数据集可以在 [Kaggle](https://www.kaggle.com/c/dogs-vs-cats) 上找到。这场比赛最初是在 2013 年进行的，那时的准确率是 80％。而我们的简单模型已经能够达到 97％的准确度。大脑现在还清醒吧？下面是一些照片和他们被预测的标记： ![](https://cdn-images-1.medium.com/max/1600/1*y3guCmNkYLF2uR09Fslh5g.png) -Predicted labels for dogs and cats -The target labels are setup using a process called one-hot-encoding which is often used for categorical variables. [1. 0.] refers to a cat and [0. 1.] refers to a dog. Instead of having one variable named ‘target’ with two levels 0 and 1, we create an array with two values. You can look at these variables as ‘cats’ and ‘dogs’. If the variable is true it gets labeled as a 1 and otherwise as a 0. In a multi-classification problem this can mean that your output vector looks like this: [0 0 0 0 0 0 0 1 0 0 0]. In this case the Vgg16 model outputs the probability of the image belonging to class ‘cat’ and the probability of the image belonging to the class ‘dog’. The next challenge is to tweak this model so we can apply it to another dataset. +狗狗们和猫猫们被预测的标记 + +我们用叫做独热编码的方法来处理目标标记，这是分类问题中常用的方法。[1. 0.] 说明图片中是一只猫， [0. 1.] 则说明是一只狗。我们没有用一个叫做“目标”的有 0 和 1 两种取值的变量，而是创建了一个包含两个值的数组。你可以把这些变量看成“猫猫”和“狗狗”。如果变量为正，那么它就会被标记为 1，否则就是 0。在一个多分类问题中，这意味着你的输出向量可能长成这样：[0 0 0 0 0 0 0 1 0 0 0]。在这个例子中，Vgg16 模型会输出图片属于“猫”这个类别的可能性以及属于“狗”这个类别的可能性。接下来的一个挑战是调整这个模型，以便我们将其应用于另一个数据集。 --- -#### **Dogs vs. Cats Redux** +#### **狗狗还是猫猫 终极版** -Essentially this is the same dataset as the previous one, but it is not pre-processed by the authors of the course. The Kaggle Command Line Interface (CLI) provides a quick way to download the dataset. It can be installed via pip. A dollar sign is often used to show that a command is run in the terminal. +本质上这个数据集和先前的是同一个数据集，但是没有被课程作者预处理过。Kaggle 命令行接口（CLI）提供了一个快捷的方法来下载这个数据集，可以通过 pip 来安装。一个美元标志通常用来表示命令运行在终端中。 $ pip install kaggle-cli -The training set contains 25.000 labeled images of dogs and cats, while the test set contains 12.500 unlabeled images. In order to finetune the parameters we also create a validation set by taking a small part of the training set. It is also useful to set-up a ‘sample’ of the full dataset that you can use to quickly check if your model is working during the building proces. +训练数据集中有 25000 张已经被标记为猫或是的狗的图片，测试数据集中则包含 12500 张未被标记的图片。为了调整参数，我们还通过占用训练集的一小部分来创建验证数据集。设置一个完整数据集的“样本”也很有用，可以用来快速检查你的模型在构建过程中是否正常工作。 -In order to run our model we use the Keras library. This library sits on top of the popular deep learning libraries Theano and TensorFlow. Keras basically makes it more intuitive to code your network. This means that you can focus more on the structure of the network and worry less about the TensorFlow API. In order to know which picture belongs to which class Keras looks at the directory it is stored in. Therefore, it is important to make sure you move the images to the correct directories. The bash commands that are needed to do this can be run directly from the Jupyter Notebook where we do all our coding. [This](https://www.cyberciti.biz/faq/mv-command-howto-move-folder-in-linux-terminal/) link contains additional information on these commands. +我们使用 Keras 库来运行我们的模型，这个库是基于 Thenao 和 TensorFlow 的最流行的深度学习库之一。Keras 能够让你更加直观地来编写神经网络，这意味着你能够更多地关注神经网络的架构而不用担心 TensorFlow API。因为Keras 通过查看图片所属的目录来确定它的类别，所以把图片移动到正确的目录非常的重要。这些操作所需的 bash 命令可以直接在 Jupyter Notebook 中运行，也就是我们写代码的地方。[这个](https://www.cyberciti.biz/faq/mv-command-howto-move-folder-in-linux-terminal/)链接包含了额外的一些关于这些命令的信息。 -One epoch, which is a full pass through the dataset, takes 10 minutes on my Amazon p2 instance. In this case that dataset is the training set which consists of 23.000 images. The other 2000 images are in the validation set. I decided to use 3 epochs here. The accuracy on the validation set is around 98%. After training the model we can take a look at some of the correctly classified images. In this case we use the probabilities of the image being a cat. 1.0 refers to full confidence that the image is of a cat and 0.0 that the image is of a dog. +一个 epoch，也就是在数据集完整地跑一遍，在我的 Amazon p2 实例上花费了 10 分钟时间。在这个例子里数据集是包含 23000 张图片的训练数据集，另外的 2000 张图片被保留下来作为验证数据集。在这里我决定使用 3 个 epoch。在验证数据集上的准确度在 98% 左右。训练好模型之后，我们可以看一些被正确分类的图片。在这个例子里，我们用图片中是一只猫的概率作为结果。1.0 表示模型非常自信地认为图片中是一只猫，而 0.0 则表示图片中是一只狗。 ![](https://cdn-images-1.medium.com/max/1600/1*fgOX3G_imeRsodKuBBA8Tg.png) -Correctly classified images -Now let’s take a look at some of the wrongly classified images. As we can see most of them are taken from far away and feature multiple animals. The original Vgg model was used for images where one thing of the target class was clearly visible in the picture. Am I the only one who finds the fourth picture slightly terrifying? +被正确分类的图片 + +现在让我们来看一些被错误分类的图片。正如我们所见，这些图片大部分是从远处拍摄的，并且图片里有多种动物。原始的 Vgg 模型是用在图片中只有一种清晰可见目标类别中的。只有我觉得第四张图片有点可怕吗？ ![](https://cdn-images-1.medium.com/max/1600/1*jD6t1ifVrrGq571eh5lqhA.png) -Incorrectly classified images -Finally, these are the images that the model was most uncertain about. This means that the probability was closest to 0.5 (where 1 is a cat and 0 a dog). The fourth picture features a cat where only the face is visible. The first and third picture are rectangular and not square like the the pictures the original model was trained on. +被错误分类的图片 + +最后，这些是模型对其类别最不确定的一些图片。这意味着概率非常接近 0.5（1 代表是一只猫而 0 代表是一只狗）。第四张图片中的猫只有一张脸露出来。第一张和第三张图片是长方形的而不是原模型训练集中的正方形。 ![](https://cdn-images-1.medium.com/max/1600/1*zlSUpvspBf9zYm175uaY1w.png) -Images where the model is most uncertain -That’s it for this week. Personally I can’t wait to get started on lesson 2 and learn more about the internals of the model. Hopefully we will also start on building a model from scratch with Keras! +模型最不确定的图片 -Also, thanks to everyone who is updating the Github scripts. It helped a lot! Another thank you to everyone on the Fast AI forums, you’re awesome. +这就是这周的内容。就我个人而言，我已经迫不及待地想要开始第二周的课程并且学习更多关于这个模型的内部细节。希望我们也能开始利用 Keras 从头构建一个模型。 -If you liked this posts be sure to recommend it so others can see it. You can also follow this profile to keep up with my process in the Fast AI course. See you there! +同时，感谢所有更新 GitHub 脚本的人，这可帮了大忙！另外也要感谢所有参与 Fast AI 论坛的人，你们太棒了。 +如果你喜欢这篇文章，请把它推荐给你的朋友们，让更多人的看到它。你也可以按照这篇文章，跟上我在 Fast AI 课程中的进度。到时候那里见！ --- > [掘金翻译计划](https://github.com/xitu/gold-miner) 是一个翻译优质互联网技术文章的社区，文章来源为 [掘金](https://juejin.im) 上的英文分享文章。内容覆盖 [Android](https://github.com/xitu/gold-miner#android)、[iOS](https://github.com/xitu/gold-miner#ios)、[React](https://github.com/xitu/gold-miner#react)、[前端](https://github.com/xitu/gold-miner#前端)、[后端](https://github.com/xitu/gold-miner#后端)、[产品](https://github.com/xitu/gold-miner#产品)、[设计](https://github.com/xitu/gold-miner#设计) 等领域，想要查看更多优质译文请持续关注 [掘金翻译计划](https://github.com/xitu/gold-miner)。 +