1 00:00:11,690 --> 00:00:12,080 All right. 2 00:00:12,080 --> 00:00:17,720 So if you saw the past few lectures and you had no idea what was going on, you might be a bit alarmed. 3 00:00:17,750 --> 00:00:22,640 You might be wondering, what the heck is numpy and pandas and where can I learn about them? 4 00:00:22,940 --> 00:00:24,860 Luckily, I've got you covered. 5 00:00:24,980 --> 00:00:31,400 It just so happens that I teach all of this for free in my course called the numpy stack in Python. 6 00:00:31,850 --> 00:00:35,240 As usual, you can find this course by searching for its name. 7 00:00:35,240 --> 00:00:38,930 Or you can click my profile and go to my list of courses. 8 00:00:38,930 --> 00:00:45,620 If you still can't find it, you can go to my website Lazy Programmers Me where I list out all my courses 9 00:00:45,620 --> 00:00:46,730 and link to them there. 10 00:00:51,580 --> 00:00:57,070 Now, because this course is free, I'm going to assume that you've taken the time to do your best to 11 00:00:57,070 --> 00:00:58,550 prepare for this course. 12 00:00:58,570 --> 00:01:05,500 That means knowing the basics of numpy, map, plot, lib pandas, psi, pi and psychic learn. 13 00:01:05,710 --> 00:01:09,330 Of course, when I say basics, that might be a little ambiguous. 14 00:01:09,340 --> 00:01:12,540 What is basic to me may not be basic to you. 15 00:01:12,550 --> 00:01:18,520 So let's enumerate and review what you should have learned so that you can make sure you are well equipped 16 00:01:18,520 --> 00:01:20,080 to handle this course. 17 00:01:24,970 --> 00:01:25,750 For numpy. 18 00:01:25,780 --> 00:01:29,860 I essentially want you to know matrix arithmetic in deep learning. 19 00:01:29,860 --> 00:01:35,830 We work with TensorFlow, which might sound scary, but we actually just mean array like a regular old 20 00:01:35,830 --> 00:01:36,970 programming array. 21 00:01:37,180 --> 00:01:42,730 And if you've been paying attention, the name TensorFlow actually includes the word tensor in it, 22 00:01:42,730 --> 00:01:44,770 so you know that it's going to come in handy. 23 00:01:46,170 --> 00:01:50,960 Sensors or if you prefer arrays, can have any number of dimensions. 24 00:01:50,970 --> 00:01:56,490 We have special names for one dimensional sensors and two dimensional sensors, which we call vectors 25 00:01:56,490 --> 00:01:58,290 and matrices, respectively. 26 00:01:58,380 --> 00:02:03,030 But you should not be afraid if you see a3d tensor or a 40 tensor. 27 00:02:05,350 --> 00:02:07,750 Okay, so what do we mean by arithmetic? 28 00:02:08,020 --> 00:02:14,050 That means your basic arithmetic, operations, addition, subtraction, multiplication and division. 29 00:02:14,260 --> 00:02:19,360 Division doesn't really apply when we're talking about matrices, but there is an analogous concept 30 00:02:19,360 --> 00:02:22,900 of matrix inversion, which you should be familiar with. 31 00:02:23,110 --> 00:02:27,250 You should know when addition, subtraction and multiplication are allowed. 32 00:02:28,140 --> 00:02:32,040 For example, you should know that you can add matrices of different sizes. 33 00:02:33,620 --> 00:02:35,750 Multiplication is the key operation. 34 00:02:35,750 --> 00:02:41,640 However, you should know that matrix multiplication is really just the dot product in disguise. 35 00:02:41,660 --> 00:02:48,050 In fact, whether you're doing a dot product between vectors or matrix multiplication between matrices, 36 00:02:48,050 --> 00:02:49,910 these are the same function in numpy. 37 00:02:49,940 --> 00:02:51,230 It's just called dot. 38 00:02:52,960 --> 00:02:57,760 Element wise multiplication is also an important operation in deep learning, so you should know how 39 00:02:57,760 --> 00:02:58,870 to do that as well. 40 00:03:03,690 --> 00:03:08,220 In terms of map plot lib, there isn't really anything fancy you need to know how to do. 41 00:03:08,250 --> 00:03:13,890 Just know about line charts and scatter plots since these are what we'll be using 99% of the time. 42 00:03:18,800 --> 00:03:21,610 In terms of pandas, also pretty simple. 43 00:03:21,620 --> 00:03:24,710 We basically use pandas to load in tabular data. 44 00:03:24,980 --> 00:03:30,140 Pandas does some other useful things, but we won't be making full use of its capabilities. 45 00:03:30,260 --> 00:03:33,590 Most of the time it'll just be to load in a CSV file. 46 00:03:38,380 --> 00:03:39,970 Next, we have sci fi. 47 00:03:40,120 --> 00:03:43,190 Now, I don't anticipate really using sci fi at all. 48 00:03:43,210 --> 00:03:44,950 However, that could change. 49 00:03:45,040 --> 00:03:51,640 Sci fi is kind of like a power version of NumPy, so it has more useful statistics, functions more 50 00:03:51,640 --> 00:03:56,560 useful optimization, functions, more useful linear algebra functions, and so forth. 51 00:03:56,710 --> 00:04:01,270 NumPy is more for those lower level operations, like adding and multiplying. 52 00:04:06,110 --> 00:04:07,970 Finally, we have psychic learn. 53 00:04:08,180 --> 00:04:13,100 Now the point of knowing about psychic learn isn't to know about psychic learning per se. 54 00:04:13,670 --> 00:04:17,150 It's to know about the concepts behind using psychic learn. 55 00:04:17,390 --> 00:04:22,310 In other words, you're going to have to learn about the basics of machine learning in order to make 56 00:04:22,310 --> 00:04:24,170 proper use of psychic learning. 57 00:04:24,620 --> 00:04:30,560 Again, I reiterate I teach this for free, so I don't anticipate any problems catching up. 58 00:04:31,100 --> 00:04:33,050 So here are the things you want to remember. 59 00:04:33,050 --> 00:04:34,160 And keep in mind. 60 00:04:34,580 --> 00:04:38,060 First are the basics of classification and regression. 61 00:04:38,270 --> 00:04:44,210 You want to recall how to think of these as geometry problems rather than magic machine learning voodoo. 62 00:04:44,570 --> 00:04:47,720 Second, you want to be familiar with the shape of data. 63 00:04:49,360 --> 00:04:55,300 Typically for a structured data problem, meaning it not working with images, sound or text, but more 64 00:04:55,300 --> 00:04:57,100 traditional business type data. 65 00:04:57,100 --> 00:05:00,490 We're going to have two objects which we call X and Y. 66 00:05:00,670 --> 00:05:05,380 X represents our input samples and Y represents the corresponding targets. 67 00:05:10,220 --> 00:05:11,810 To give you a simple example. 68 00:05:11,810 --> 00:05:18,080 Suppose we want to predict whether or not a student will pass a math exam in order to build a model. 69 00:05:18,080 --> 00:05:23,090 We've collected a bunch of data points for students who took this math exam in the past. 70 00:05:23,330 --> 00:05:29,210 We record several variables, such as how many hours they spent studying and how many hours they spent 71 00:05:29,210 --> 00:05:30,620 playing video games. 72 00:05:30,710 --> 00:05:33,290 We would treat these as our X matrix. 73 00:05:33,530 --> 00:05:39,260 Our Y matrix would be the target or label, which is what we want to predict in the future for future 74 00:05:39,260 --> 00:05:40,190 students. 75 00:05:40,430 --> 00:05:44,660 So in the future we could ask a student, Hey, how many hours did you study? 76 00:05:44,660 --> 00:05:46,850 How many hours did you play video games? 77 00:05:46,850 --> 00:05:52,550 And then from that information alone, we could make a prediction about whether or not they will pass 78 00:05:52,550 --> 00:05:53,510 the math exam. 79 00:05:55,210 --> 00:06:00,280 Why would be whether or not the student pastor failed and coded as zero or one. 80 00:06:00,910 --> 00:06:07,630 In general, we say that the x matrix is of size nn by dx, where nn is the number of samples and d 81 00:06:07,630 --> 00:06:09,130 is the number of features. 82 00:06:09,340 --> 00:06:14,410 Why is just a one dimensional array of length n which you can also think of as a vector? 83 00:06:19,160 --> 00:06:24,110 You also want to know the steps we take in general when we are writing a machine learning script to 84 00:06:24,110 --> 00:06:26,120 test that a model we've just learned about. 85 00:06:26,390 --> 00:06:28,760 This is going to be pretty much the same every time. 86 00:06:28,760 --> 00:06:32,960 So if you don't know this, you're going to be extremely confused. 87 00:06:33,200 --> 00:06:35,000 The steps are as follows. 88 00:06:35,270 --> 00:06:37,510 Number one, we load in the data. 89 00:06:37,520 --> 00:06:40,010 That's just the X and Y we just talked about. 90 00:06:40,190 --> 00:06:45,740 We will use pandas for this when we have simple data sets, or if we have something more complicated, 91 00:06:45,740 --> 00:06:47,720 we'll write our own python code. 92 00:06:49,100 --> 00:06:55,370 Number two, we can optionally split the data into train and test sets depending on what we are doing. 93 00:06:55,820 --> 00:07:02,240 Note that sometimes the term test set is used interchangeably with validation set, in which case the 94 00:07:02,240 --> 00:07:04,790 true test set would mean something else. 95 00:07:04,940 --> 00:07:06,450 Don't get too fussy over this. 96 00:07:06,470 --> 00:07:07,610 It is what it is. 97 00:07:09,010 --> 00:07:11,230 Number three, we build a model. 98 00:07:11,350 --> 00:07:17,140 Now, in Python, we usually use object oriented programming, which means our model takes the form 99 00:07:17,140 --> 00:07:18,820 of a Python object. 100 00:07:19,300 --> 00:07:22,390 In PyTorch, we still use object oriented programming. 101 00:07:22,390 --> 00:07:26,950 In fact, probably more so than what you're used to if you've only used scikit learn. 102 00:07:27,130 --> 00:07:31,900 Specifically, PyTorch uses subclasses in order to build models. 103 00:07:32,020 --> 00:07:36,700 These models do not work like what you are used to in psych learn, but it's similar. 104 00:07:37,000 --> 00:07:40,120 Ultimately, you should still recognize the following steps. 105 00:07:42,400 --> 00:07:45,970 Number four, we fit the model in psych learn. 106 00:07:46,000 --> 00:07:51,970 This just means calling the fit function on the model and passing in your training data in a deep learning. 107 00:07:51,970 --> 00:07:55,780 This is a little more complicated because we have many more options. 108 00:07:55,930 --> 00:08:01,390 Since training in deep learning involves an iterative training algorithm known as gradient descent, 109 00:08:01,420 --> 00:08:07,270 we typically plot a loss per iteration when we are finished training to confirm that our training process 110 00:08:07,270 --> 00:08:09,130 has converged as expected. 111 00:08:09,490 --> 00:08:12,730 If it didn't, we have to go back and make changes. 112 00:08:13,210 --> 00:08:16,660 And finally, number five, we evaluate the model. 113 00:08:16,780 --> 00:08:21,910 This involves checking the accuracy of our model on both the train and validation sets. 114 00:08:22,450 --> 00:08:26,950 At this time, we can also check to see if the model is overfitting or under fitting. 115 00:08:28,520 --> 00:08:31,970 And once this is all done, we can make predictions with our model. 116 00:08:32,830 --> 00:08:33,700 In PyTorch. 117 00:08:33,700 --> 00:08:35,650 This is a little more complicated than inside. 118 00:08:35,650 --> 00:08:39,490 You learn because PyTorch doesn't really work with numpy arrays. 119 00:08:39,610 --> 00:08:42,670 Instead, PyTorch works with torch sensors. 120 00:08:42,670 --> 00:08:47,940 So there will be a lot of times where we have to convert back and forth between numpy array and torch 121 00:08:47,950 --> 00:08:50,920 tensor depending on what we need at the time. 122 00:08:51,040 --> 00:08:55,270 This might seem strange at first, but you'll get used to it as you get more practice. 123 00:08:56,630 --> 00:09:01,340 This process encapsulates the typical life cycle of a machine learning model. 124 00:09:01,370 --> 00:09:07,460 You load in your data, train a model, evaluate it, and then use it to make future predictions. 125 00:09:11,600 --> 00:09:11,920 Okay. 126 00:09:11,960 --> 00:09:14,210 So just a quick update to this lecture. 127 00:09:14,360 --> 00:09:20,870 I won't name any names, but on certain platforms, meaning certain platforms I personally don't own. 128 00:09:20,900 --> 00:09:26,300 They've made it very hard for students such as yourselves to find free courses. 129 00:09:26,510 --> 00:09:28,550 This doesn't mean that they don't exist. 130 00:09:28,550 --> 00:09:34,370 It simply means that the people behind these platforms have made them very hard for you to find. 131 00:09:34,520 --> 00:09:38,030 Obviously, this is in their best interests, but not yours. 132 00:09:38,060 --> 00:09:41,270 That being said, there is an easy way around this. 133 00:09:41,270 --> 00:09:47,390 Specifically, it is better to simply go to my website, which is deep learning courses dot com, where 134 00:09:47,390 --> 00:09:50,270 it should be very easy to find my free course. 135 00:09:54,580 --> 00:09:59,740 Now you might be wondering what's the difference between this website and other websites where you can 136 00:09:59,740 --> 00:10:00,970 find my courses? 137 00:10:01,240 --> 00:10:07,060 The answer is that the courses on my website are always a superset of courses which have the same name 138 00:10:07,060 --> 00:10:08,320 on other sites. 139 00:10:08,530 --> 00:10:13,870 In other words, if you use deep learning courses dot com, you are getting the best value for each 140 00:10:13,870 --> 00:10:15,250 dollar you spend. 141 00:10:19,600 --> 00:10:24,700 Now, just as a quick side note, what if you do really want to get the course on this website? 142 00:10:24,730 --> 00:10:26,320 Well, you still can. 143 00:10:26,470 --> 00:10:32,140 As mentioned, you simply need to go to my profile, which has a list of all the courses I teach and 144 00:10:32,140 --> 00:10:34,660 scroll through them to find the relevant course. 145 00:10:34,960 --> 00:10:39,460 I'm not sure that they show up in any particular order, so you just have to scroll through them all. 146 00:10:44,090 --> 00:10:48,170 Finally, just a brief note on why there are two versions of this course. 147 00:10:48,200 --> 00:10:51,020 You see, there's a free version and the paid version. 148 00:10:51,440 --> 00:10:56,960 The reason there's two is because, again, this platform, which by the way, I don't work for, made 149 00:10:56,960 --> 00:11:00,320 some changes a few years ago to how free courses work. 150 00:11:00,410 --> 00:11:04,460 Therefore, it was by necessity that two versions were created. 151 00:11:04,880 --> 00:11:08,950 Remember that instructors such as myself are not employees of this company. 152 00:11:08,960 --> 00:11:14,630 We are simply users of this website, just like how you are a user of Gmail, Hotmail and what have 153 00:11:14,630 --> 00:11:15,020 you. 154 00:11:15,680 --> 00:11:17,300 So I'm a user of this website. 155 00:11:17,300 --> 00:11:19,670 I don't have any control over how it works.