WEBVTT 00:00:00.000 --> 00:00:02.410 align:middle line:90% [SQUEAKING] 00:00:02.410 --> 00:00:04.338 align:middle line:90% [RUSTLING] 00:00:04.338 --> 00:00:06.748 align:middle line:90% [CLICKING] 00:00:06.748 --> 00:00:09.262 align:middle line:90% 00:00:09.262 --> 00:00:11.220 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Hello, everybody. 00:00:11.220 --> 00:00:12.803 align:middle line:84% We're going to be looking today at how 00:00:12.803 --> 00:00:17.180 align:middle line:84% we can use machine learning and artificial intelligence in order 00:00:17.180 --> 00:00:22.970 align:middle line:84% to learn more about musical behavior and music theory 00:00:22.970 --> 00:00:25.400 align:middle line:90% and just music in general. 00:00:25.400 --> 00:00:28.190 align:middle line:84% Our first task is going to be how do we 00:00:28.190 --> 00:00:32.610 align:middle line:84% transform music into formats that computers can learn from? 00:00:32.610 --> 00:00:36.680 align:middle line:84% Because a score alone is not something 00:00:36.680 --> 00:00:39.060 align:middle line:84% that can be used for artificial intelligence. 00:00:39.060 --> 00:00:42.710 align:middle line:84% And so our first task is going to be understanding 00:00:42.710 --> 00:00:47.270 align:middle line:84% how feature extraction, the conversion of musical scores 00:00:47.270 --> 00:00:53.930 align:middle line:84% into intelligent numerical data, how feature extraction is 00:00:53.930 --> 00:00:57.050 align:middle line:84% able to make machine learning possible. 00:00:57.050 --> 00:01:00.640 align:middle line:84% So go ahead and load up Jupyter Notebook 00:01:00.640 --> 00:01:04.810 align:middle line:84% and follow along as we do some work with music21, 00:01:04.810 --> 00:01:09.590 align:middle line:84% first separate from the music21 machine learning toolkit, 00:01:09.590 --> 00:01:13.150 align:middle line:84% and then with music21 features in order 00:01:13.150 --> 00:01:19.660 align:middle line:84% to learn more about some genres and other ways of classifying 00:01:19.660 --> 00:01:21.190 align:middle line:90% music. 00:01:21.190 --> 00:01:24.890 align:middle line:84% So go ahead and start Jupyter Notebook. 00:01:24.890 --> 00:01:27.880 align:middle line:84% And we're going to create the first step, which 00:01:27.880 --> 00:01:31.060 align:middle line:84% is to understand what features are 00:01:31.060 --> 00:01:34.900 align:middle line:84% and why we might want to extract them. 00:01:34.900 --> 00:01:37.730 align:middle line:90% So what is a feature? 00:01:37.730 --> 00:01:40.040 align:middle line:84% That's the first thing we're going to work on. 00:01:40.040 --> 00:01:43.340 align:middle line:84% Feature is something really, really, really basic. 00:01:43.340 --> 00:01:45.820 align:middle line:90% Let's say that we have a piece. 00:01:45.820 --> 00:01:49.885 align:middle line:84% We'll load from the music21 corpus, our favorite Bach 00:01:49.885 --> 00:01:50.900 align:middle line:90% chorale. 00:01:50.900 --> 00:01:55.940 align:middle line:84% So you don't even need to know what comes up from this corpus. 00:01:55.940 --> 00:02:01.480 align:middle line:90% Parse bach.bwv66.6. 00:02:01.480 --> 00:02:06.910 align:middle line:84% And just to remind ourselves how it goes, here's the Bach. 00:02:06.910 --> 00:02:07.970 align:middle line:90% So great. 00:02:07.970 --> 00:02:14.840 align:middle line:84% Let's go ahead and choose a feature from this. 00:02:14.840 --> 00:02:16.930 align:middle line:90% Let's do the parts feature. 00:02:16.930 --> 00:02:22.700 align:middle line:84% And we'll call it numParts equals 4. 00:02:22.700 --> 00:02:31.780 align:middle line:84% And let's do the time signature numerator feature equals 4. 00:02:31.780 --> 00:02:32.320 align:middle line:90% There. 00:02:32.320 --> 00:02:41.740 align:middle line:84% We've now extracted manually two features from this chorale. 00:02:41.740 --> 00:02:46.000 align:middle line:84% Feature extraction is the conversion of musical elements, 00:02:46.000 --> 00:02:47.120 align:middle line:90% or any elements. 00:02:47.120 --> 00:02:49.120 align:middle line:84% But musical feature extraction is the conversion 00:02:49.120 --> 00:02:53.410 align:middle line:84% of musical elements into single numbers or lists, arrays, 00:02:53.410 --> 00:02:56.640 align:middle line:90% vectors of numbers. 00:02:56.640 --> 00:03:01.530 align:middle line:84% It is thus the most boring thing in the world, 00:03:01.530 --> 00:03:04.020 align:middle line:90% as you've just seen. 00:03:04.020 --> 00:03:06.700 align:middle line:90% It is just a bunch of numbers. 00:03:06.700 --> 00:03:11.860 align:middle line:84% Feature extraction sucks out the lifeblood of music. 00:03:11.860 --> 00:03:14.370 align:middle line:84% It takes everything that's beautiful in music 00:03:14.370 --> 00:03:16.660 align:middle line:84% and turns it into a bunch of numbers. 00:03:16.660 --> 00:03:18.610 align:middle line:90% It is a life sucker. 00:03:18.610 --> 00:03:23.160 align:middle line:84% It is the zombie of the musical world. 00:03:23.160 --> 00:03:26.580 align:middle line:84% But there are tons of amazing machine learning 00:03:26.580 --> 00:03:28.740 align:middle line:84% and artificial intelligence and deep learning 00:03:28.740 --> 00:03:30.010 align:middle line:90% algorithms in the world. 00:03:30.010 --> 00:03:33.010 align:middle line:84% As far as I know, they all have one thing in common. 00:03:33.010 --> 00:03:36.750 align:middle line:84% They only work on collections of single numbers or vectors 00:03:36.750 --> 00:03:38.670 align:middle line:90% of numbers. 00:03:38.670 --> 00:03:40.720 align:middle line:84% I'm not a machine learning expert, 00:03:40.720 --> 00:03:43.410 align:middle line:84% so maybe there are algorithms that work on two 00:03:43.410 --> 00:03:45.250 align:middle line:84% or multi-dimensional arrays of numbers. 00:03:45.250 --> 00:03:48.930 align:middle line:84% That would be amazing for machine vision or what we do. 00:03:48.930 --> 00:03:52.950 align:middle line:84% But nothing, as far as I know in the artificial intelligence 00:03:52.950 --> 00:03:59.970 align:middle line:84% world, works on things like this notation, or-- 00:03:59.970 --> 00:04:02.240 align:middle line:90% [AUDIO PLAYBACK] 00:04:02.240 --> 00:04:04.010 align:middle line:90% - 00:04:04.010 --> 00:04:06.230 align:middle line:90% 00:04:06.230 --> 00:04:07.340 align:middle line:90% [END PLAYBACK] 00:04:07.340 --> 00:04:08.840 align:middle line:84% MICHAEL SCOTT ASATO CUTHBERT: Sound. 00:04:08.840 --> 00:04:11.570 align:middle line:84% So if we're going to work on doing artificial intelligence 00:04:11.570 --> 00:04:14.780 align:middle line:84% or predictions or anything super amazing on music 00:04:14.780 --> 00:04:17.540 align:middle line:84% notation, at least at present, we 00:04:17.540 --> 00:04:21.769 align:middle line:84% need to convert music to numbers. 00:04:21.769 --> 00:04:24.530 align:middle line:84% So that's what feature extraction is. 00:04:24.530 --> 00:04:29.330 align:middle line:84% Basic feature extraction is pretty simple. 00:04:29.330 --> 00:04:34.160 align:middle line:84% So in this case, let's just extract automatically 00:04:34.160 --> 00:04:36.570 align:middle line:90% the meter from this piece. 00:04:36.570 --> 00:04:40.650 align:middle line:84% So we'll write something that does it. 00:04:40.650 --> 00:04:42.950 align:middle line:84% We'll import from the typing module 00:04:42.950 --> 00:04:45.450 align:middle line:84% because we like to do things like this. 00:04:45.450 --> 00:04:50.120 align:middle line:84% Turn Python, which is a weakly typed language, into slightly 00:04:50.120 --> 00:04:55.930 align:middle line:84% stronger types so that we can understand a little bit better 00:04:55.930 --> 00:04:57.680 align:middle line:90% what's going to go on. 00:04:57.680 --> 00:04:59.980 align:middle line:84% So we'll also import from music21 the meter 00:04:59.980 --> 00:05:01.040 align:middle line:90% and stream modules. 00:05:01.040 --> 00:05:05.320 align:middle line:84% Now let's write something called extract. 00:05:05.320 --> 00:05:10.480 align:middle line:84% Extract meter, which takes in a variable score, which is 00:05:10.480 --> 00:05:13.520 align:middle line:90% a stream.Stream stream object. 00:05:13.520 --> 00:05:15.380 align:middle line:84% Maybe it should be a stream.Score object, 00:05:15.380 --> 00:05:17.860 align:middle line:90% but let's keep it flexible. 00:05:17.860 --> 00:05:21.520 align:middle line:84% And it returns a tuple of two integers. 00:05:21.520 --> 00:05:28.870 align:middle line:84% And we'll say here that returns a 2-tuple of the numerator 00:05:28.870 --> 00:05:43.930 align:middle line:84% and the denominator of the first time signature found in a score. 00:05:43.930 --> 00:05:52.420 align:middle line:84% Returns 0, comma, 0 if the score has no meter. 00:05:52.420 --> 00:05:55.570 align:middle line:84% Why do we return 0, comma, 0 if there's no meter? 00:05:55.570 --> 00:05:57.580 align:middle line:84% In feature extraction, it's generally 00:05:57.580 --> 00:05:59.980 align:middle line:84% assumed that we want to return out 00:05:59.980 --> 00:06:02.450 align:middle line:84% of bound number rather than raising exception 00:06:02.450 --> 00:06:04.060 align:middle line:90% if there's a mistake. 00:06:04.060 --> 00:06:07.960 align:middle line:84% And the reason is you'd hate to be running extract meter 00:06:07.960 --> 00:06:09.920 align:middle line:84% on hundreds or thousands of pieces, 00:06:09.920 --> 00:06:13.150 align:middle line:84% and you're always getting 4, comma, 4, or 3, comma, 4 or 6, 00:06:13.150 --> 00:06:16.760 align:middle line:84% comma 8, or even something wacky like 11, comma, 32. 00:06:16.760 --> 00:06:18.970 align:middle line:84% And it's all working successfully 00:06:18.970 --> 00:06:21.350 align:middle line:84% only to have your entire feature extraction, 00:06:21.350 --> 00:06:23.930 align:middle line:84% artificial intelligence pipeline fail 00:06:23.930 --> 00:06:26.990 align:middle line:84% if there's one piece that doesn't have a time signature. 00:06:26.990 --> 00:06:29.500 align:middle line:84% Like let's say it's Gregorian chant or some kind 00:06:29.500 --> 00:06:31.900 align:middle line:90% of trance meditation. 00:06:31.900 --> 00:06:36.640 align:middle line:84% So that's why we're going to be returning 0, comma, 0. 00:06:36.640 --> 00:06:39.430 align:middle line:84% On the other hand, maybe it's a bad assumption 00:06:39.430 --> 00:06:42.650 align:middle line:84% and we should raise exceptions when something's not right. 00:06:42.650 --> 00:06:45.795 align:middle line:84% And then we should fix our work rather than hiding errors. 00:06:45.795 --> 00:06:48.310 align:middle line:90% 00:06:48.310 --> 00:06:50.650 align:middle line:84% But we'll get to that in just a bit. 00:06:50.650 --> 00:06:52.300 align:middle line:90% First meter. 00:06:52.300 --> 00:06:54.280 align:middle line:84% But first, let's see how we do this. 00:06:54.280 --> 00:06:57.420 align:middle line:84% So let's just remind ourselves we can take from anything 00:06:57.420 --> 00:07:01.140 align:middle line:84% get elements by class, meter.TimeSignature. 00:07:01.140 --> 00:07:03.690 align:middle line:90% 00:07:03.690 --> 00:07:07.950 align:middle line:84% And if we do not have any, the first meter 00:07:07.950 --> 00:07:13.980 align:middle line:84% will return 0, comma, 0, and otherwise 00:07:13.980 --> 00:07:19.710 align:middle line:84% will return the first meter's numerator and the first meter's 00:07:19.710 --> 00:07:23.160 align:middle line:90% denominator. 00:07:23.160 --> 00:07:26.670 align:middle line:84% We're kind of hiding an error in our encoding. 00:07:26.670 --> 00:07:28.120 align:middle line:90% Is that the right thing to do? 00:07:28.120 --> 00:07:30.825 align:middle line:84% And I think this is something that, 00:07:30.825 --> 00:07:32.520 align:middle line:84% when we're talking about it-- and I 00:07:32.520 --> 00:07:36.120 align:middle line:84% would love to talk to anybody directly about ethics 00:07:36.120 --> 00:07:41.160 align:middle line:84% and morality and who artificial intelligence is working for, 00:07:41.160 --> 00:07:44.580 align:middle line:84% because I think there are often important considerations 00:07:44.580 --> 00:07:47.750 align:middle line:84% of artificial intelligence with moral and ethical consequences 00:07:47.750 --> 00:07:50.790 align:middle line:84% when we hide things that it can't do. 00:07:50.790 --> 00:07:52.850 align:middle line:90% And so I want you-- 00:07:52.850 --> 00:07:54.350 align:middle line:84% if you're going to be somebody who's 00:07:54.350 --> 00:07:56.570 align:middle line:84% going to be working with artificial intelligence 00:07:56.570 --> 00:07:58.550 align:middle line:84% and the rest of your life and technology 00:07:58.550 --> 00:08:00.680 align:middle line:84% to think about such things, maybe we 00:08:00.680 --> 00:08:04.250 align:middle line:84% shouldn't be hiding our errors so much. 00:08:04.250 --> 00:08:05.310 align:middle line:90% OK. 00:08:05.310 --> 00:08:08.160 align:middle line:84% So we have this extract meter to get back to this. 00:08:08.160 --> 00:08:10.140 align:middle line:90% Oops. 00:08:10.140 --> 00:08:13.100 align:middle line:84% Doesn't matter if you put a parentheses around a tuple 00:08:13.100 --> 00:08:14.160 align:middle line:90% when you return it. 00:08:14.160 --> 00:08:16.620 align:middle line:84% But if you start with one, you have to do the other. 00:08:16.620 --> 00:08:19.190 align:middle line:84% OK, so let's take this Bach chorale and let's 00:08:19.190 --> 00:08:26.090 align:middle line:90% run extract meter on the bach. 00:08:26.090 --> 00:08:27.540 align:middle line:90% 0, comma, 0. 00:08:27.540 --> 00:08:29.250 align:middle line:90% Oh, I know why. 00:08:29.250 --> 00:08:33.845 align:middle line:90% Because we forgot to do recurse. 00:08:33.845 --> 00:08:37.490 align:middle line:84% Flat would do just as well, but recurse is a little bit faster. 00:08:37.490 --> 00:08:40.070 align:middle line:84% Oh, by the way, something I didn't mention. 00:08:40.070 --> 00:08:43.640 align:middle line:84% You might have seen that we're sending 00:08:43.640 --> 00:08:46.640 align:middle line:84% in the class, actual class meter.TimeSignature. 00:08:46.640 --> 00:08:50.380 align:middle line:84% We could also put in the string time signature. 00:08:50.380 --> 00:08:53.680 align:middle line:90% Doesn't matter either way there. 00:08:53.680 --> 00:08:57.260 align:middle line:84% But this actually doesn't get us the first meter. 00:08:57.260 --> 00:08:59.630 align:middle line:90% This gets us all the meters. 00:08:59.630 --> 00:09:03.640 align:middle line:84% And so if there are no meters that 00:09:03.640 --> 00:09:05.180 align:middle line:84% says if there's not a single one, 00:09:05.180 --> 00:09:12.290 align:middle line:84% return that, we'll call first meter the first of all meters. 00:09:12.290 --> 00:09:15.460 align:middle line:90% So of course, it's number zero. 00:09:15.460 --> 00:09:17.540 align:middle line:90% Now that should be there. 00:09:17.540 --> 00:09:20.060 align:middle line:90% Now extract_meter bach. 00:09:20.060 --> 00:09:20.620 align:middle line:90% Fantastic. 00:09:20.620 --> 00:09:23.770 align:middle line:90% Returns 4, comma, 4. 00:09:23.770 --> 00:09:29.260 align:middle line:84% So why would we want to extract the meter as two numbers as part 00:09:29.260 --> 00:09:31.000 align:middle line:90% of feature extraction? 00:09:31.000 --> 00:09:35.270 align:middle line:84% We might do it as part of a series of features 00:09:35.270 --> 00:09:37.820 align:middle line:84% in order to make a prediction about a piece. 00:09:37.820 --> 00:09:41.230 align:middle line:84% Let's say we wanted to predict if a piece of music is a waltz 00:09:41.230 --> 00:09:42.430 align:middle line:90% or not. 00:09:42.430 --> 00:09:46.359 align:middle line:84% So we might first think, is this a waltz? 00:09:46.359 --> 00:09:48.854 align:middle line:90% [MUSIC PLAYING] 00:09:48.854 --> 00:09:53.350 align:middle line:90% 00:09:53.350 --> 00:09:55.833 align:middle line:90% Or is this a waltz? 00:09:55.833 --> 00:10:06.190 align:middle line:84% [JOHN-PHILIP SOUSA, "STARS AND STRIPES FOREVER"] 00:10:06.190 --> 00:10:08.180 align:middle line:90% Is this a waltz? 00:10:08.180 --> 00:10:10.620 align:middle line:90% [MUSIC PLAYING] 00:10:10.620 --> 00:10:21.850 align:middle line:90% 00:10:21.850 --> 00:10:23.942 align:middle line:90% So what makes something a waltz? 00:10:23.942 --> 00:10:26.150 align:middle line:84% Well, it probably has something to do with the meter. 00:10:26.150 --> 00:10:29.170 align:middle line:84% We fortunately already have a meter extraction 00:10:29.170 --> 00:10:31.310 align:middle line:84% from our time signature for the piece, 00:10:31.310 --> 00:10:35.680 align:middle line:84% and now we can use this to try to figure out 00:10:35.680 --> 00:10:38.150 align:middle line:90% if something is a waltz or not. 00:10:38.150 --> 00:10:41.650 align:middle line:84% So we would assume that the prevailing time 00:10:41.650 --> 00:10:44.050 align:middle line:84% signature in a piece and the opening time 00:10:44.050 --> 00:10:47.710 align:middle line:84% signature in a piece, if it's a waltz, would be 3, 4. 00:10:47.710 --> 00:10:52.270 align:middle line:84% And therefore, our extraction should be 3, comma, 4. 00:10:52.270 --> 00:10:55.730 align:middle line:84% That is the worst assumption ever. 00:10:55.730 --> 00:10:57.980 align:middle line:84% The most famous waltz of all time, 00:10:57.980 --> 00:11:02.230 align:middle line:84% Johann Strauss' "Blue Danube" waltz, which you just heard, 00:11:02.230 --> 00:11:03.560 align:middle line:90% doesn't start in 3, 4. 00:11:03.560 --> 00:11:07.163 align:middle line:84% It starts with a slow introduction in 6, 8 time. 00:11:07.163 --> 00:11:10.600 align:middle line:90% 00:11:10.600 --> 00:11:13.730 align:middle line:84% I happen to have a random waltz on my computer. 00:11:13.730 --> 00:11:17.770 align:middle line:84% It's Johann Strauss' waltz called on the "Blue Danube." 00:11:17.770 --> 00:11:20.590 align:middle line:84% It's not in the music21 corpus, but it's instead 00:11:20.590 --> 00:11:23.000 align:middle line:90% on my hard drive on my desktop. 00:11:23.000 --> 00:11:28.270 align:middle line:84% So I'll need to import the converter module from music21. 00:11:28.270 --> 00:11:32.060 align:middle line:84% I mean, it's just a wrapper around loading things and stuff. 00:11:32.060 --> 00:11:35.600 align:middle line:84% And now instead of using corpus.parse, 00:11:35.600 --> 00:11:37.430 align:middle line:90% I'll use converter.parse. 00:11:37.430 --> 00:11:41.270 align:middle line:90% I'll give the full filename to-- 00:11:41.270 --> 00:11:46.010 align:middle line:90% file path to my waltz. 00:11:46.010 --> 00:11:48.343 align:middle line:90% It's a music XML file. 00:11:48.343 --> 00:11:52.070 align:middle line:90% 00:11:52.070 --> 00:11:52.890 align:middle line:90% Oh, great. 00:11:52.890 --> 00:11:57.210 align:middle line:84% So the guy who's making this recording just goes online, 00:11:57.210 --> 00:12:00.770 align:middle line:84% grabs a random music XML copy of the "Blue Danube" 00:12:00.770 --> 00:12:05.450 align:middle line:84% from the internet, and this version skips the introduction 00:12:05.450 --> 00:12:06.450 align:middle line:90% completely. 00:12:06.450 --> 00:12:10.740 align:middle line:84% So it has 3, 4 when it should have gotten 6, 8. 00:12:10.740 --> 00:12:12.870 align:middle line:84% So it's going to get the right answer, 00:12:12.870 --> 00:12:15.020 align:middle line:90% but for all the wrong reasons. 00:12:15.020 --> 00:12:17.840 align:middle line:84% Have you ever heard the expression GIGO? 00:12:17.840 --> 00:12:22.070 align:middle line:90% Garbage in equals garbage out. 00:12:22.070 --> 00:12:25.730 align:middle line:84% Let's not do that in our own work, what we're 00:12:25.730 --> 00:12:28.440 align:middle line:90% learning from in this tutorial. 00:12:28.440 --> 00:12:31.670 align:middle line:84% Let's instead make the world of artificial intelligence a better 00:12:31.670 --> 00:12:36.420 align:middle line:84% place and always check whether our training data is correct, 00:12:36.420 --> 00:12:39.090 align:middle line:84% whether this is actually a score of the "Blue Danube," 00:12:39.090 --> 00:12:41.430 align:middle line:84% rather than just the main theme of it. 00:12:41.430 --> 00:12:47.000 align:middle line:84% Let's check things like this before we use it. 00:12:47.000 --> 00:12:48.260 align:middle line:90% Great. 00:12:48.260 --> 00:12:51.210 align:middle line:90% So it's in 3, 4. 00:12:51.210 --> 00:12:54.840 align:middle line:84% So our extract_meter function is going to return 3, 4. 00:12:54.840 --> 00:12:55.600 align:middle line:90% Let's try that. 00:12:55.600 --> 00:13:01.070 align:middle line:90% 00:13:01.070 --> 00:13:01.880 align:middle line:90% Sorry. 00:13:01.880 --> 00:13:03.060 align:middle line:90% Don't get the answer. 00:13:03.060 --> 00:13:04.560 align:middle line:90% Give the question. 00:13:04.560 --> 00:13:05.550 align:middle line:90% There we go. 00:13:05.550 --> 00:13:07.130 align:middle line:84% So now we have our first feature that 00:13:07.130 --> 00:13:09.000 align:middle line:90% might help identify a waltz. 00:13:09.000 --> 00:13:11.480 align:middle line:90% Its time signature. 00:13:11.480 --> 00:13:14.255 align:middle line:84% Now let's do something a little bit more advanced. 00:13:14.255 --> 00:13:16.250 align:middle line:84% Let's look at this piece and see, 00:13:16.250 --> 00:13:19.530 align:middle line:84% hm, is there anything that seems to happen pretty often? 00:13:19.530 --> 00:13:24.170 align:middle line:84% Well, it looks like the last beat, beat three of the piece 00:13:24.170 --> 00:13:29.610 align:middle line:84% or offset two is more often higher than offset one. 00:13:29.610 --> 00:13:31.640 align:middle line:84% I don't know if that's true, but that 00:13:31.640 --> 00:13:33.870 align:middle line:84% looks like something that's a possibility. 00:13:33.870 --> 00:13:39.080 align:middle line:84% So we'll create a more sophisticated feature extractor 00:13:39.080 --> 00:13:45.220 align:middle line:84% which returns the proportion of all the measures where 00:13:45.220 --> 00:13:48.350 align:middle line:90% beat three is above beat one. 00:13:48.350 --> 00:13:50.500 align:middle line:84% First, we'll make the assumption that 00:13:50.500 --> 00:13:52.400 align:middle line:84% beat one is always at offset zero 00:13:52.400 --> 00:13:53.960 align:middle line:90% and beat three is at offset two. 00:13:53.960 --> 00:13:56.510 align:middle line:84% These aren't very good assumptions, 00:13:56.510 --> 00:13:58.760 align:middle line:84% but maybe together with searching for meter, 00:13:58.760 --> 00:14:00.790 align:middle line:84% they'll two will make a good case for 00:14:00.790 --> 00:14:02.540 align:middle line:84% whether something is a waltz or not. 00:14:02.540 --> 00:14:04.750 align:middle line:90% So here we go. 00:14:04.750 --> 00:14:09.700 align:middle line:84% We're going to eventually write something-- proportion 00:14:09.700 --> 00:14:19.300 align:middle line:84% proportion_beat3_above_beat1, which is going to take a score. 00:14:19.300 --> 00:14:21.020 align:middle line:90% Stream.Stream. 00:14:21.020 --> 00:14:24.320 align:middle line:84% It's going to return a float because it's a proportion. 00:14:24.320 --> 00:14:28.000 align:middle line:84% And I'll just say pass now, because in order 00:14:28.000 --> 00:14:29.860 align:middle line:84% to do write this, of course, we're 00:14:29.860 --> 00:14:32.380 align:middle line:84% going to need something that figures out for any given 00:14:32.380 --> 00:14:42.270 align:middle line:84% measure if beat three is above beat one for a single measure, 00:14:42.270 --> 00:14:43.630 align:middle line:90% for one measure. 00:14:43.630 --> 00:14:47.200 align:middle line:84% So we'll take an m, which will be a stream.Measure, 00:14:47.200 --> 00:14:50.040 align:middle line:84% and this is going to return true or false. 00:14:50.040 --> 00:14:50.910 align:middle line:90% Oops. 00:14:50.910 --> 00:14:54.100 align:middle line:84% It's going to return true or false. 00:14:54.100 --> 00:14:56.280 align:middle line:84% Someday we have an incredible production crew 00:14:56.280 --> 00:14:58.180 align:middle line:84% that's going to edit out all these mistakes, 00:14:58.180 --> 00:14:59.320 align:middle line:90% but I don't have that yet. 00:14:59.320 --> 00:15:02.550 align:middle line:84% So we'll just say that measures notes 00:15:02.550 --> 00:15:06.600 align:middle line:84% get all the elements by class note, 00:15:06.600 --> 00:15:09.310 align:middle line:84% and that returns this stream iterator. 00:15:09.310 --> 00:15:12.420 align:middle line:84% For now, I'm going to say recast that stream 00:15:12.420 --> 00:15:14.430 align:middle line:90% iterator as another stream. 00:15:14.430 --> 00:15:17.040 align:middle line:84% You can do this with any of the get something 00:15:17.040 --> 00:15:19.930 align:middle line:90% by something classes in music21. 00:15:19.930 --> 00:15:20.430 align:middle line:90% Great. 00:15:20.430 --> 00:15:24.820 align:middle line:84% So within, this is now a stream of just nothing but notes. 00:15:24.820 --> 00:15:30.270 align:middle line:84% And so I can say the first notes are those notes that 00:15:30.270 --> 00:15:35.150 align:middle line:90% appear on offset zero. 00:15:35.150 --> 00:15:41.180 align:middle line:84% And if there's no note there, return false. 00:15:41.180 --> 00:15:41.840 align:middle line:90% Great. 00:15:41.840 --> 00:15:49.970 align:middle line:84% Now we'll check beat_3_notes are those notes that 00:15:49.970 --> 00:15:56.250 align:middle line:84% are at offset two and that began at offset two. 00:15:56.250 --> 00:15:57.930 align:middle line:90% So that should work pretty well. 00:15:57.930 --> 00:16:05.090 align:middle line:84% And if there's no note that satisfies that condition, 00:16:05.090 --> 00:16:08.270 align:middle line:84% we're going to return false again. 00:16:08.270 --> 00:16:12.920 align:middle line:84% But what if there is more than one note 00:16:12.920 --> 00:16:20.240 align:middle line:84% at the beginning of the measure or at beat three? 00:16:20.240 --> 00:16:23.690 align:middle line:84% So we have to make sure that the first note is 00:16:23.690 --> 00:16:27.290 align:middle line:84% the first of the beat three notes 00:16:27.290 --> 00:16:33.800 align:middle line:84% and the beat_3_note is the first of the beat_3_notes. 00:16:33.800 --> 00:16:35.250 align:middle line:90% The third-- the first. 00:16:35.250 --> 00:16:39.860 align:middle line:84% Now we'll go ahead and say if the first note's pitch is lower, 00:16:39.860 --> 00:16:42.830 align:middle line:84% then we're going to measure the pitch by pitch space or midi 00:16:42.830 --> 00:16:47.600 align:middle line:84% number is less than the beat three notes, 00:16:47.600 --> 00:16:50.550 align:middle line:90% pitch space will return false. 00:16:50.550 --> 00:16:54.680 align:middle line:84% Otherwise, we're going to return true. 00:16:54.680 --> 00:16:56.090 align:middle line:90% Great. 00:16:56.090 --> 00:16:58.170 align:middle line:90% Well, let's be good. 00:16:58.170 --> 00:17:00.090 align:middle line:90% We'll make this a second thing. 00:17:00.090 --> 00:17:04.079 align:middle line:90% We'll cut and paste that there. 00:17:04.079 --> 00:17:06.349 align:middle line:90% Meanwhile we're going to-- 00:17:06.349 --> 00:17:09.000 align:middle line:84% let's write a little test on this. 00:17:09.000 --> 00:17:11.450 align:middle line:90% So we have blue. 00:17:11.450 --> 00:17:14.099 align:middle line:90% Let's get measure zero. 00:17:14.099 --> 00:17:17.390 align:middle line:90% So m0. 00:17:17.390 --> 00:17:19.020 align:middle line:90% Actually, let's measure one. 00:17:19.020 --> 00:17:22.280 align:middle line:90% So blue.measure 1. 00:17:22.280 --> 00:17:29.660 align:middle line:84% And we'll see, is_beat_3_above_beat_1. 00:17:29.660 --> 00:17:30.160 align:middle line:90% Oops. 00:17:30.160 --> 00:17:34.390 align:middle line:90% 00:17:34.390 --> 00:17:36.080 align:middle line:90% [AUDIO SPED UP] 00:17:36.080 --> 00:17:43.720 align:middle line:90% 00:17:43.720 --> 00:17:45.260 align:middle line:90% This is why we test things. 00:17:45.260 --> 00:17:46.180 align:middle line:90% Great. 00:17:46.180 --> 00:17:51.400 align:middle line:84% Is_beat_3_above_beat_1 in measure one? 00:17:51.400 --> 00:17:52.850 align:middle line:90% Actually, it is. 00:17:52.850 --> 00:17:54.820 align:middle line:90% Ah. 00:17:54.820 --> 00:17:56.180 align:middle line:90% Oh, I know what it is. 00:17:56.180 --> 00:17:58.880 align:middle line:90% We have a reverse of a sign. 00:17:58.880 --> 00:18:03.980 align:middle line:84% If this pitch is greater than or equal, then we return false. 00:18:03.980 --> 00:18:06.400 align:middle line:90% Actually, if it's-- yeah. 00:18:06.400 --> 00:18:07.810 align:middle line:90% The equal sign there. 00:18:07.810 --> 00:18:10.280 align:middle line:90% And now that's working. 00:18:10.280 --> 00:18:12.560 align:middle line:90% Measure two should be true. 00:18:12.560 --> 00:18:14.960 align:middle line:84% Also measure three should be false. 00:18:14.960 --> 00:18:17.350 align:middle line:90% Let's make sure that works. 00:18:17.350 --> 00:18:18.490 align:middle line:90% Yep. 00:18:18.490 --> 00:18:19.670 align:middle line:90% And false. 00:18:19.670 --> 00:18:20.170 align:middle line:90% Great. 00:18:20.170 --> 00:18:22.840 align:middle line:90% So that seems to be working. 00:18:22.840 --> 00:18:25.550 align:middle line:90% Here's the irony, actually. 00:18:25.550 --> 00:18:29.190 align:middle line:84% Except for the fact that our label would be completely wrong, 00:18:29.190 --> 00:18:32.220 align:middle line:84% having every single thing that should be false be true and true 00:18:32.220 --> 00:18:35.710 align:middle line:84% be false probably wouldn't change our answers very much. 00:18:35.710 --> 00:18:36.760 align:middle line:90% But who knows. 00:18:36.760 --> 00:18:39.240 align:middle line:84% We call this, is_beat_3_above_beat_1 00:18:39.240 --> 00:18:41.070 align:middle line:90% in one measure? 00:18:41.070 --> 00:18:42.470 align:middle line:90% It better do what it says. 00:18:42.470 --> 00:18:42.970 align:middle line:90% OK, great. 00:18:42.970 --> 00:18:45.150 align:middle line:84% Now we're going to get the proportion that 00:18:45.150 --> 00:18:48.150 align:middle line:84% is the proportion of measures in this piece where 00:18:48.150 --> 00:18:49.510 align:middle line:90% beat three is above beat one. 00:18:49.510 --> 00:18:53.760 align:middle line:84% So we'll keep track of the total number of measures 00:18:53.760 --> 00:18:55.470 align:middle line:90% equals zero, the total-- 00:18:55.470 --> 00:18:58.480 align:middle line:84% we'll call them waltz-like measures. 00:18:58.480 --> 00:19:00.910 align:middle line:84% Those where beat three is above beat one. 00:19:00.910 --> 00:19:02.190 align:middle line:90% We'll call it zero. 00:19:02.190 --> 00:19:07.560 align:middle line:84% For every measure in the score, recurse 00:19:07.560 --> 00:19:10.820 align:middle line:90% getElementsByClass measure. 00:19:10.820 --> 00:19:13.650 align:middle line:90% 00:19:13.650 --> 00:19:16.620 align:middle line:84% There's a reason why we'll use getElementsByClass measure 00:19:16.620 --> 00:19:19.950 align:middle line:84% instead of measure zero, one, two, three. 00:19:19.950 --> 00:19:23.430 align:middle line:84% Mostly to do with the fact that sometimes there's two measure 00:19:23.430 --> 00:19:26.660 align:middle line:84% threes or they're missing measure 11 or something like 00:19:26.660 --> 00:19:30.680 align:middle line:84% that, where there's measure 5A and 5B if there's a first ending 00:19:30.680 --> 00:19:31.710 align:middle line:90% or something like that. 00:19:31.710 --> 00:19:33.290 align:middle line:90% Anyhow. 00:19:33.290 --> 00:19:35.960 align:middle line:84% For every measure, we increment the total number of measures 00:19:35.960 --> 00:19:36.660 align:middle line:90% by one. 00:19:36.660 --> 00:19:41.750 align:middle line:90% And now we look at one measure. 00:19:41.750 --> 00:19:48.230 align:middle line:84% And we say if that's true, then the total waltz-like measures 00:19:48.230 --> 00:19:50.450 align:middle line:90% plus equals one. 00:19:50.450 --> 00:19:56.540 align:middle line:84% And we're going to return the total waltz-like measures 00:19:56.540 --> 00:19:59.930 align:middle line:90% divided by the total measures. 00:19:59.930 --> 00:20:05.030 align:middle line:84% We're coding pretty fast, so we'll just put as a to-do later. 00:20:05.030 --> 00:20:09.200 align:middle line:84% What if there are no total measures? 00:20:09.200 --> 00:20:13.000 align:middle line:84% I always make sure that we don't do a division by zero measure 00:20:13.000 --> 00:20:13.500 align:middle line:90% error. 00:20:13.500 --> 00:20:16.170 align:middle line:90% But that can be a later thing. 00:20:16.170 --> 00:20:24.920 align:middle line:84% So let's check this proportion of beat three above beat one 00:20:24.920 --> 00:20:25.820 align:middle line:90% on the "Blue Danube." 00:20:25.820 --> 00:20:30.230 align:middle line:90% 00:20:30.230 --> 00:20:32.870 align:middle line:90% 26%. 00:20:32.870 --> 00:20:34.327 align:middle line:90% Is that really true? 00:20:34.327 --> 00:20:35.410 align:middle line:90% It seems a little bit low. 00:20:35.410 --> 00:20:38.270 align:middle line:90% 00:20:38.270 --> 00:20:41.270 align:middle line:90% One, two down. 00:20:41.270 --> 00:20:43.080 align:middle line:90% Well, we'll check that later. 00:20:43.080 --> 00:20:44.970 align:middle line:90% Turns out it's not that bad. 00:20:44.970 --> 00:20:47.760 align:middle line:90% And bwv-- sorry. 00:20:47.760 --> 00:20:48.830 align:middle line:90% Bach. 00:20:48.830 --> 00:20:50.300 align:middle line:90% Call it this time. 00:20:50.300 --> 00:20:52.040 align:middle line:84% It's actually even a little bit more. 00:20:52.040 --> 00:20:53.130 align:middle line:90% Interesting. 00:20:53.130 --> 00:20:53.990 align:middle line:90% OK. 00:20:53.990 --> 00:20:56.790 align:middle line:84% So in any case, this is a feature we've just defined. 00:20:56.790 --> 00:20:59.360 align:middle line:84% This particular feature figures out how often 00:20:59.360 --> 00:21:01.790 align:middle line:90% beat three is above beat one. 00:21:01.790 --> 00:21:02.990 align:middle line:90% Great. 00:21:02.990 --> 00:21:08.730 align:middle line:84% What we can do now is these are individual features. 00:21:08.730 --> 00:21:12.920 align:middle line:84% Let's create something that does change a whole bunch of features 00:21:12.920 --> 00:21:13.560 align:middle line:90% together. 00:21:13.560 --> 00:21:14.487 align:middle line:90% Well, not this one. 00:21:14.487 --> 00:21:15.570 align:middle line:90% A whole bunch of features. 00:21:15.570 --> 00:21:18.870 align:middle line:90% We'll get four features. 00:21:18.870 --> 00:21:28.030 align:middle line:84% So we'll say our feature extractor will take in a score, 00:21:28.030 --> 00:21:33.520 align:middle line:84% and it's going to return a tuple of a string, an int, an int, 00:21:33.520 --> 00:21:37.900 align:middle line:90% a float, and a Boolean. 00:21:37.900 --> 00:21:38.950 align:middle line:90% OK. 00:21:38.950 --> 00:21:39.763 align:middle line:90% What the heck? 00:21:39.763 --> 00:21:41.180 align:middle line:84% What are these things going to be? 00:21:41.180 --> 00:21:45.040 align:middle line:84% So first, what we're going to do is we're going to take in score 00:21:45.040 --> 00:21:48.460 align:middle line:90% and we're going to-- 00:21:48.460 --> 00:21:52.600 align:middle line:84% well, we're going to say the filename of the score 00:21:52.600 --> 00:21:54.550 align:middle line:84% is going to be that first string. 00:21:54.550 --> 00:21:59.080 align:middle line:84% And you can get the filename from any music21 score 00:21:59.080 --> 00:22:01.820 align:middle line:84% by getting this filePath, which is a path lib, 00:22:01.820 --> 00:22:03.730 align:middle line:84% and getting the name, which gets rid of all 00:22:03.730 --> 00:22:04.880 align:middle line:90% the directories and stuff. 00:22:04.880 --> 00:22:05.380 align:middle line:90% Great. 00:22:05.380 --> 00:22:09.310 align:middle line:84% So we'll get the filename of something and we'll get the-- 00:22:09.310 --> 00:22:12.797 align:middle line:84% next thing we're going to get these two ints, which 00:22:12.797 --> 00:22:13.880 align:middle line:90% are going to be the meter. 00:22:13.880 --> 00:22:18.040 align:middle line:84% So we'll say the meter feature is 00:22:18.040 --> 00:22:20.730 align:middle line:90% where I get that from extract. 00:22:20.730 --> 00:22:24.720 align:middle line:90% Extract_meter on that score. 00:22:24.720 --> 00:22:27.570 align:middle line:84% Next we're going to get this float, which is 00:22:27.570 --> 00:22:30.780 align:middle line:90% going to be the beat3_feature. 00:22:30.780 --> 00:22:37.590 align:middle line:84% So we'll get the beat3_feature is the proportion-- 00:22:37.590 --> 00:22:39.960 align:middle line:90% why do I keep doing this? 00:22:39.960 --> 00:22:44.140 align:middle line:84% Beat_3_above_beat_1 for that score. 00:22:44.140 --> 00:22:44.640 align:middle line:90% Great. 00:22:44.640 --> 00:22:45.430 align:middle line:90% Super. 00:22:45.430 --> 00:22:45.960 align:middle line:90% Oops. 00:22:45.960 --> 00:22:48.690 align:middle line:90% I didn't hit Shift-Enter. 00:22:48.690 --> 00:22:53.350 align:middle line:90% Then we want to get the Boolean. 00:22:53.350 --> 00:22:56.290 align:middle line:84% What we're going to do is if we're training a data set, 00:22:56.290 --> 00:23:00.150 align:middle line:84% we're going to know whether or not something is a waltz. 00:23:00.150 --> 00:23:02.520 align:middle line:84% So we'll say that Boolean is going 00:23:02.520 --> 00:23:05.380 align:middle line:84% to be true or false on whether something is a waltz. 00:23:05.380 --> 00:23:12.130 align:middle line:84% And so we could say is_waltz equals if true 00:23:12.130 --> 00:23:18.060 align:middle line:84% if Waltz is in the filename, or if lowercase waltz 00:23:18.060 --> 00:23:21.210 align:middle line:90% is in the filename. 00:23:21.210 --> 00:23:23.250 align:middle line:90% Filename. 00:23:23.250 --> 00:23:31.020 align:middle line:84% So then we know the filename, numerator and denominator, float 00:23:31.020 --> 00:23:33.945 align:middle line:84% of the proportion beat three, and whether or not filename 00:23:33.945 --> 00:23:34.960 align:middle line:90% is waltz. 00:23:34.960 --> 00:23:38.605 align:middle line:84% This last particular thing, we'll sometimes call this, 00:23:38.605 --> 00:23:40.230 align:middle line:84% instead of is_waltz, we're saying we'll 00:23:40.230 --> 00:23:42.940 align:middle line:90% use generic term ground_truth. 00:23:42.940 --> 00:23:45.970 align:middle line:84% We'll also see it called the class. 00:23:45.970 --> 00:23:49.720 align:middle line:84% So whether it a belongs to the waltz class or not. 00:23:49.720 --> 00:23:52.510 align:middle line:84% So we can call it waltz or non-waltz or true/false. 00:23:52.510 --> 00:23:53.710 align:middle line:90% Something like that. 00:23:53.710 --> 00:23:56.650 align:middle line:84% Now we want to return all these things as a tuple. 00:23:56.650 --> 00:24:01.660 align:middle line:84% So tuples are created with a comma or with a parentheses. 00:24:01.660 --> 00:24:03.220 align:middle line:90% So we'll use the parentheses. 00:24:03.220 --> 00:24:07.810 align:middle line:84% The filename, the meter_feature's first argument, 00:24:07.810 --> 00:24:11.710 align:middle line:84% which is the numerator, meter_feature's second argument, 00:24:11.710 --> 00:24:15.570 align:middle line:84% which is the denominator, the beat3_feature, 00:24:15.570 --> 00:24:19.430 align:middle line:90% and the ground_truth. 00:24:19.430 --> 00:24:21.750 align:middle line:90% So let's see. 00:24:21.750 --> 00:24:25.370 align:middle line:90% We'll run this feature-- 00:24:25.370 --> 00:24:27.440 align:middle line:90% let's call extractor. 00:24:27.440 --> 00:24:31.340 align:middle line:90% Let's call snippet properly. 00:24:31.340 --> 00:24:32.965 align:middle line:84% Feature extractor on the "Blue Danube." 00:24:32.965 --> 00:24:36.410 align:middle line:90% 00:24:36.410 --> 00:24:38.100 align:middle line:90% I still didn't spell it right. 00:24:38.100 --> 00:24:41.510 align:middle line:90% Why do I copy and paste? 00:24:41.510 --> 00:24:43.430 align:middle line:90% There we go. 00:24:43.430 --> 00:24:48.890 align:middle line:84% Oh, we can see "Blue Danube" waltz is in 3, 4. 00:24:48.890 --> 00:24:52.080 align:middle line:90% 26% has this beat proportion. 00:24:52.080 --> 00:24:53.070 align:middle line:90% And true. 00:24:53.070 --> 00:24:54.860 align:middle line:90% It is a waltz. 00:24:54.860 --> 00:25:04.910 align:middle line:84% Let's run our feature_extractor on bach. 00:25:04.910 --> 00:25:08.580 align:middle line:90% We'll see filename 4, 4, 35. 00:25:08.580 --> 00:25:09.080 align:middle line:90% False. 00:25:09.080 --> 00:25:10.820 align:middle line:90% Great. 00:25:10.820 --> 00:25:14.840 align:middle line:84% So eventually we would want to have a computer system that 00:25:14.840 --> 00:25:20.890 align:middle line:84% can distinguish between a waltz and not by learning 00:25:20.890 --> 00:25:22.990 align:middle line:90% all this information. 00:25:22.990 --> 00:25:26.530 align:middle line:84% Unfortunately, I don't have enough waltzes 00:25:26.530 --> 00:25:28.460 align:middle line:90% in my data set to do this. 00:25:28.460 --> 00:25:31.340 align:middle line:84% So we're going to move to the next best thing. 00:25:31.340 --> 00:25:33.920 align:middle line:90% We're going to get a jig. 00:25:33.920 --> 00:25:38.960 align:middle line:84% So let's go ahead and grab a particular jig. 00:25:38.960 --> 00:25:45.010 align:middle line:84% We have all of these jigs in this data set, which is really 00:25:45.010 --> 00:25:46.550 align:middle line:90% useful for people to know. 00:25:46.550 --> 00:25:51.374 align:middle line:84% It's called Ryan's Mammoth Collection. 00:25:51.374 --> 00:25:53.920 align:middle line:90% 00:25:53.920 --> 00:25:55.160 align:middle line:90% It's a 19th century thing. 00:25:55.160 --> 00:25:55.952 align:middle line:90% You can look it up. 00:25:55.952 --> 00:25:57.200 align:middle line:90% It's very famous. 00:25:57.200 --> 00:25:57.700 align:middle line:90% Hi. 00:25:57.700 --> 00:26:01.390 align:middle line:84% I wanted to jump in for a second and say something 00:26:01.390 --> 00:26:02.510 align:middle line:90% about this collection. 00:26:02.510 --> 00:26:06.410 align:middle line:84% I'm planning on not using again soon. 00:26:06.410 --> 00:26:09.520 align:middle line:84% Since I made this video, I've realized 00:26:09.520 --> 00:26:12.020 align:middle line:84% that within Ryan's Mammoth Collection, 00:26:12.020 --> 00:26:14.740 align:middle line:84% there is a lot of what are called minstrel songs. 00:26:14.740 --> 00:26:17.640 align:middle line:84% That is, songs that were performed 00:26:17.640 --> 00:26:24.010 align:middle line:84% in the style of Negro melodies, but by white performers, 00:26:24.010 --> 00:26:27.390 align:middle line:84% often in blackface and with texts 00:26:27.390 --> 00:26:30.280 align:middle line:84% which are not included in this collection, thank goodness, 00:26:30.280 --> 00:26:33.690 align:middle line:84% but with texts that denigrated Black 00:26:33.690 --> 00:26:36.070 align:middle line:90% Americans and Black experience. 00:26:36.070 --> 00:26:39.630 align:middle line:84% So I'm going to try to change this example as soon as I 00:26:39.630 --> 00:26:45.280 align:middle line:84% can, and my apologies for what I didn't know then 00:26:45.280 --> 00:26:50.490 align:middle line:84% and what I know now and what I plan to continue to do to make 00:26:50.490 --> 00:26:55.110 align:middle line:84% computational musicology a more welcoming place. 00:26:55.110 --> 00:26:57.650 align:middle line:84% So if we're going to be looking at the BloomingMeadowsJig. 00:26:57.650 --> 00:27:00.510 align:middle line:90% 00:27:00.510 --> 00:27:09.060 align:middle line:84% We'll load that blooming equals corpus.parse BloomingMeadowJig. 00:27:09.060 --> 00:27:11.040 align:middle line:90% Don't need the abc. 00:27:11.040 --> 00:27:12.960 align:middle line:84% You can leave it in however you want. 00:27:12.960 --> 00:27:15.870 align:middle line:90% Let's look at this. 00:27:15.870 --> 00:27:19.240 align:middle line:84% In case you haven't seen a jig in a while, 00:27:19.240 --> 00:27:23.790 align:middle line:84% it's kind of helpful for understanding the jig. 00:27:23.790 --> 00:27:25.570 align:middle line:90% Generally speaking, 6, 8. 00:27:25.570 --> 00:27:28.360 align:middle line:84% We've got these violin bow marks. 00:27:28.360 --> 00:27:30.490 align:middle line:84% We could use that as a feature extractor. 00:27:30.490 --> 00:27:33.580 align:middle line:84% If it has violin markings, it's more likely to be a jig. 00:27:33.580 --> 00:27:35.380 align:middle line:90% If it doesn't, it's less likely. 00:27:35.380 --> 00:27:38.770 align:middle line:84% Lots of Bach pieces have violin bow marks. 00:27:38.770 --> 00:27:42.840 align:middle line:84% The Bach Violin Partitas, they're not all jigs. 00:27:42.840 --> 00:27:46.120 align:middle line:84% So it's not going to be perfect, but all these things kind help. 00:27:46.120 --> 00:27:48.750 align:middle line:84% But maybe that meter extractor 6, 00:27:48.750 --> 00:27:51.330 align:middle line:84% 8 is going to be a little bit helpful. 00:27:51.330 --> 00:27:54.530 align:middle line:84% So I happen to know that this little piece is a jig. 00:27:54.530 --> 00:27:56.280 align:middle line:84% It's kind of nice short piece, but there's 00:27:56.280 --> 00:27:57.280 align:middle line:90% a whole bunch of things. 00:27:57.280 --> 00:27:59.590 align:middle line:84% So we're going to get from Ryan's Mammoth Collection. 00:27:59.590 --> 00:28:02.640 align:middle line:84% And we'll use a different thing that we've used before. 00:28:02.640 --> 00:28:04.140 align:middle line:84% Instead of corpus-parse, we're going 00:28:04.140 --> 00:28:07.320 align:middle line:90% to search everything for Ryan. 00:28:07.320 --> 00:28:11.270 align:middle line:84% I think everything in there is as part of this Ryan's Mammoth 00:28:11.270 --> 00:28:11.940 align:middle line:90% Collection. 00:28:11.940 --> 00:28:14.210 align:middle line:84% And we'll see that within Ryan, there's 00:28:14.210 --> 00:28:16.920 align:middle line:84% going to be quite a number of pieces. 00:28:16.920 --> 00:28:18.710 align:middle line:90% 1,063. 00:28:18.710 --> 00:28:22.880 align:middle line:84% You'll see the first time you use corpus search, 00:28:22.880 --> 00:28:27.510 align:middle line:84% it takes some time, but you can use-- 00:28:27.510 --> 00:28:31.520 align:middle line:84% after you do it once, it's often quite a bit faster. 00:28:31.520 --> 00:28:33.020 align:middle line:90% Super fast. 00:28:33.020 --> 00:28:33.830 align:middle line:90% Yeah. 00:28:33.830 --> 00:28:34.530 align:middle line:90% There we go. 00:28:34.530 --> 00:28:35.304 align:middle line:90% OK. 00:28:35.304 --> 00:28:40.770 align:middle line:84% So we're going to have used that beat3_feature. 00:28:40.770 --> 00:28:41.990 align:middle line:90% Why not? 00:28:41.990 --> 00:28:43.407 align:middle line:84% I mean, we already have a feature. 00:28:43.407 --> 00:28:45.157 align:middle line:84% And we're going to use the meters feature. 00:28:45.157 --> 00:28:46.470 align:middle line:90% But let's use one more thing. 00:28:46.470 --> 00:28:48.650 align:middle line:84% I've heard that jigs tend to come 00:28:48.650 --> 00:28:54.320 align:middle line:84% in a few keys, a few key signatures. 00:28:54.320 --> 00:28:56.250 align:middle line:90% Maybe one flat, one sharp. 00:28:56.250 --> 00:28:58.260 align:middle line:84% They don't tend to come in five flats. 00:28:58.260 --> 00:29:00.810 align:middle line:84% That's kind of a hard thing for violins. 00:29:00.810 --> 00:29:06.840 align:middle line:84% So let's go ahead and create a sharp extractor. 00:29:06.840 --> 00:29:08.390 align:middle line:90% So we'll call it-- 00:29:08.390 --> 00:29:13.120 align:middle line:84% let's call it get_sharps, because we'll just 00:29:13.120 --> 00:29:16.090 align:middle line:84% remind ourselves that a feature extractor is nothing 00:29:16.090 --> 00:29:21.280 align:middle line:84% more than something that takes in a score or some other piece 00:29:21.280 --> 00:29:23.240 align:middle line:90% of data and returns numbers-- 00:29:23.240 --> 00:29:24.820 align:middle line:90% in this case, an integer. 00:29:24.820 --> 00:29:27.560 align:middle line:90% We'll just put a try thing. 00:29:27.560 --> 00:29:29.210 align:middle line:90% We'll return the score. 00:29:29.210 --> 00:29:37.330 align:middle line:84% We're going to go through it, recurse, get all the elements-- 00:29:37.330 --> 00:29:41.680 align:middle line:84% use plural this time-- by class, key signature. 00:29:41.680 --> 00:29:45.440 align:middle line:84% Get the first one and return the number of sharps. 00:29:45.440 --> 00:29:49.430 align:middle line:84% Remember, if it's two flats, sharps will be negative 2. 00:29:49.430 --> 00:29:55.150 align:middle line:84% And in case we can't find it, it might be an index error 00:29:55.150 --> 00:29:59.140 align:middle line:90% because there's zero. 00:29:59.140 --> 00:30:02.830 align:middle line:84% You can always, in your own code that you don't turn in 00:30:02.830 --> 00:30:04.310 align:middle line:84% and you don't put something there, 00:30:04.310 --> 00:30:07.690 align:middle line:84% you can just say accept all errors. 00:30:07.690 --> 00:30:09.220 align:middle line:90% Return 0. 00:30:09.220 --> 00:30:12.430 align:middle line:90% Let's double check this. 00:30:12.430 --> 00:30:15.250 align:middle line:90% Do we call it-- yeah, blooming. 00:30:15.250 --> 00:30:20.320 align:middle line:90% Sharps blooming. 00:30:20.320 --> 00:30:21.490 align:middle line:90% That's one. 00:30:21.490 --> 00:30:22.720 align:middle line:90% Let's remind ourselves. 00:30:22.720 --> 00:30:24.480 align:middle line:84% "Blue Danube" should also be one. 00:30:24.480 --> 00:30:28.630 align:middle line:90% 00:30:28.630 --> 00:30:29.320 align:middle line:90% Great. 00:30:29.320 --> 00:30:32.850 align:middle line:84% And that Bach piece should be three. 00:30:32.850 --> 00:30:38.080 align:middle line:90% 00:30:38.080 --> 00:30:38.860 align:middle line:90% Super. 00:30:38.860 --> 00:30:39.800 align:middle line:90% OK. 00:30:39.800 --> 00:30:41.480 align:middle line:84% So that seems to work a little bit. 00:30:41.480 --> 00:30:43.620 align:middle line:84% Now let's write our feature extractor. 00:30:43.620 --> 00:30:47.950 align:middle line:90% 00:30:47.950 --> 00:30:50.210 align:middle line:90% Or our multi feature extractor. 00:30:50.210 --> 00:30:53.463 align:middle line:84% Never really sure if feature extractor 00:30:53.463 --> 00:30:54.880 align:middle line:84% just refers to something that gets 00:30:54.880 --> 00:30:57.318 align:middle line:84% one feature, like get_sharps or something that 00:30:57.318 --> 00:30:58.610 align:middle line:90% gets a whole bunch of features. 00:30:58.610 --> 00:31:00.530 align:middle line:84% People aren't really clear on that. 00:31:00.530 --> 00:31:02.480 align:middle line:84% So we're going to return, again, the string. 00:31:02.480 --> 00:31:04.360 align:middle line:84% It's going to be the filename, two 00:31:04.360 --> 00:31:07.090 align:middle line:84% integers for the metered numerator and denominator. 00:31:07.090 --> 00:31:08.010 align:middle line:90% Pretty important. 00:31:08.010 --> 00:31:11.980 align:middle line:84% Float for that kind of silly beat three thing. 00:31:11.980 --> 00:31:18.420 align:middle line:84% But we'll add one more integer on whether or not-- 00:31:18.420 --> 00:31:20.380 align:middle line:90% yeah, on the number of sharps. 00:31:20.380 --> 00:31:22.570 align:middle line:84% And instead of returning a Boolean at the end, 00:31:22.570 --> 00:31:25.110 align:middle line:84% we're going to switch this last little thing 00:31:25.110 --> 00:31:27.930 align:middle line:84% to returning an integer for whether or not 00:31:27.930 --> 00:31:29.320 align:middle line:90% something is a jig. 00:31:29.320 --> 00:31:30.370 align:middle line:90% The ground truth. 00:31:30.370 --> 00:31:37.980 align:middle line:84% And it's because most of the machine learning toolkits really 00:31:37.980 --> 00:31:42.600 align:middle line:84% like working with numbers, so might as well make 00:31:42.600 --> 00:31:44.400 align:middle line:90% our ground truth a number. 00:31:44.400 --> 00:31:47.430 align:middle line:84% Make a filename, the filePath.name. 00:31:47.430 --> 00:31:48.430 align:middle line:90% Again, as before. 00:31:48.430 --> 00:31:49.510 align:middle line:84% You don't really need that there, 00:31:49.510 --> 00:31:51.180 align:middle line:84% but it's going to be really helpful for figuring out 00:31:51.180 --> 00:31:53.580 align:middle line:84% why something's working or what particular pieces are 00:31:53.580 --> 00:31:54.630 align:middle line:90% wrong or right. 00:31:54.630 --> 00:31:58.080 align:middle line:84% I'm hoping I still had proportion somewhere 00:31:58.080 --> 00:31:59.987 align:middle line:90% in my clipboard, but OK. 00:31:59.987 --> 00:32:01.320 align:middle line:90% So proportion_beat3_above_beat1. 00:32:01.320 --> 00:32:02.778 align:middle line:84% I always hit Shift-Enter too early. 00:32:02.778 --> 00:32:05.780 align:middle line:84% Sharps feature is going to be get_sharps from score. 00:32:05.780 --> 00:32:07.260 align:middle line:90% And our ground_truth. 00:32:07.260 --> 00:32:09.343 align:middle line:84% In this case, we said it's going to be an integer. 00:32:09.343 --> 00:32:14.420 align:middle line:84% So we'll say int jig in filename or jig in filename. 00:32:14.420 --> 00:32:15.780 align:middle line:90% Two things to note. 00:32:15.780 --> 00:32:19.370 align:middle line:90% One, Python int on true-- 00:32:19.370 --> 00:32:22.770 align:middle line:84% int on true returns one, int on false returns zero. 00:32:22.770 --> 00:32:27.170 align:middle line:84% In fact, in earlier versions of Python, true and false 00:32:27.170 --> 00:32:29.760 align:middle line:84% were just kind of aliases for one and zero. 00:32:29.760 --> 00:32:31.590 align:middle line:84% Thankfully, that's a little bit better. 00:32:31.590 --> 00:32:36.090 align:middle line:84% The other thing is, again, we've learned about garbage in, 00:32:36.090 --> 00:32:36.810 align:middle line:90% garbage out. 00:32:36.810 --> 00:32:42.170 align:middle line:84% I don't actually know if every single thing in this collection 00:32:42.170 --> 00:32:46.140 align:middle line:84% that has the word jig in its title is a jig, 00:32:46.140 --> 00:32:48.860 align:middle line:84% and I definitely don't know that there's 00:32:48.860 --> 00:32:52.520 align:middle line:84% every piece in this collection that isn't a jig 00:32:52.520 --> 00:32:55.903 align:middle line:84% or that is a jig has jig in the title or that-- yeah. 00:32:55.903 --> 00:32:58.320 align:middle line:84% That there might be jigs that don't have jig in the title, 00:32:58.320 --> 00:33:03.400 align:middle line:84% just as most rags in and ragtime and Scott Joplin 00:33:03.400 --> 00:33:07.430 align:middle line:84% have rag in their title, Maple Leaf Rag and so on. 00:33:07.430 --> 00:33:08.900 align:middle line:90% But some of them don't. 00:33:08.900 --> 00:33:10.400 align:middle line:90% The entertainer is the rag. 00:33:10.400 --> 00:33:12.550 align:middle line:84% It doesn't have rag in his title. 00:33:12.550 --> 00:33:16.540 align:middle line:84% So our ground_truth isn't exactly 00:33:16.540 --> 00:33:18.830 align:middle line:90% the world's gospel truth. 00:33:18.830 --> 00:33:22.630 align:middle line:84% This is something that people neglect all the time 00:33:22.630 --> 00:33:23.870 align:middle line:90% in machine learning. 00:33:23.870 --> 00:33:28.900 align:middle line:84% They keep training the computer to replicate truths 00:33:28.900 --> 00:33:31.690 align:middle line:90% that aren't necessarily true. 00:33:31.690 --> 00:33:33.950 align:middle line:84% And you'll see this all over the place. 00:33:33.950 --> 00:33:37.390 align:middle line:84% I was just working with a master's student on a project 00:33:37.390 --> 00:33:39.370 align:middle line:84% where we were trying to train optical music 00:33:39.370 --> 00:33:43.270 align:middle line:84% recognition to produce the exact ground truth that we 00:33:43.270 --> 00:33:45.280 align:middle line:90% had found encoded. 00:33:45.280 --> 00:33:48.620 align:middle line:84% The thing is, the ground truth had lots of errors too. 00:33:48.620 --> 00:33:50.570 align:middle line:90% So that's something to know. 00:33:50.570 --> 00:33:53.480 align:middle line:84% Anyhow, getting back to what are we going to return. 00:33:53.480 --> 00:33:55.550 align:middle line:84% Take a second and see if you can figure this out 00:33:55.550 --> 00:33:58.780 align:middle line:84% based on what we've said the return value is 00:33:58.780 --> 00:34:00.980 align:middle line:90% and what we've defined here. 00:34:00.980 --> 00:34:01.480 align:middle line:90% Great. 00:34:01.480 --> 00:34:08.120 align:middle line:84% So we're going to return the filename, meter_feature 0, 00:34:08.120 --> 00:34:15.820 align:middle line:84% meter_feature 1, beat3_feature, sharps_feature, 00:34:15.820 --> 00:34:18.639 align:middle line:90% and the ground_truth. 00:34:18.639 --> 00:34:20.920 align:middle line:90% Great. 00:34:20.920 --> 00:34:23.690 align:middle line:84% OK, so now we just need to iterate over everything. 00:34:23.690 --> 00:34:25.100 align:middle line:90% I want to show something. 00:34:25.100 --> 00:34:28.280 align:middle line:84% When you iterate over a corpus search object, 00:34:28.280 --> 00:34:30.290 align:middle line:84% there's this thing called a metadata bundle. 00:34:30.290 --> 00:34:34.370 align:middle line:84% That is to say, you haven't actually parsed 1,063 pieces. 00:34:34.370 --> 00:34:38.120 align:middle line:84% It just knows about the existence of 1,063 pieces. 00:34:38.120 --> 00:34:47.080 align:middle line:84% So when we say for bundle in Ryan, we can print the bundle 00:34:47.080 --> 00:34:53.260 align:middle line:84% and we can see that we have all of these different entries 00:34:53.260 --> 00:34:54.639 align:middle line:90% for pieces. 00:34:54.639 --> 00:34:59.610 align:middle line:84% So what we'll go ahead and do is instead of getting-- 00:34:59.610 --> 00:35:00.910 align:middle line:90% but not actual pieces. 00:35:00.910 --> 00:35:07.210 align:middle line:84% So to get the pieces, we can say piece_parsed equals bundle. 00:35:07.210 --> 00:35:11.745 align:middle line:84% Actually, we call it-- instead of bundle, say metadata entry. 00:35:11.745 --> 00:35:12.960 align:middle line:90% Since outputs end there. 00:35:12.960 --> 00:35:17.220 align:middle line:90% Metadata_entry in Ryan entry. 00:35:17.220 --> 00:35:20.620 align:middle line:84% Piece_parsed is the metadata_entry.parse. 00:35:20.620 --> 00:35:23.340 align:middle line:90% That's going to take some time. 00:35:23.340 --> 00:35:27.960 align:middle line:84% Then we're going to print out the feature_extraction 00:35:27.960 --> 00:35:32.190 align:middle line:90% on the piece_parsed. 00:35:32.190 --> 00:35:35.940 align:middle line:84% You're going to learn to use this little button 00:35:35.940 --> 00:35:39.480 align:middle line:84% in a second, the Stop button, because this is going to take 00:35:39.480 --> 00:35:41.500 align:middle line:90% some time if you hit Run. 00:35:41.500 --> 00:35:43.780 align:middle line:84% But you're going to have to do it in a little bit. 00:35:43.780 --> 00:35:44.280 align:middle line:90% Oops. 00:35:44.280 --> 00:35:46.050 align:middle line:90% Feature_extractor. 00:35:46.050 --> 00:35:48.120 align:middle line:84% Used to call it feature_extraction. 00:35:48.120 --> 00:35:53.680 align:middle line:90% That and exact_meter. 00:35:53.680 --> 00:35:54.970 align:middle line:90% Extract_meter. 00:35:54.970 --> 00:35:55.750 align:middle line:90% Here we go. 00:35:55.750 --> 00:35:58.370 align:middle line:84% So we're going to put this back here. 00:35:58.370 --> 00:36:02.360 align:middle line:90% And let's go ahead and run this. 00:36:02.360 --> 00:36:04.460 align:middle line:90% Oops. 00:36:04.460 --> 00:36:05.660 align:middle line:90% Oh my goodness. 00:36:05.660 --> 00:36:08.310 align:middle line:90% I cannot spell anything today. 00:36:08.310 --> 00:36:08.810 align:middle line:90% OK. 00:36:08.810 --> 00:36:09.870 align:middle line:90% Here we go. 00:36:09.870 --> 00:36:11.700 align:middle line:90% Let's go ahead and running this. 00:36:11.700 --> 00:36:15.260 align:middle line:84% It shows that AdmiralsHornpipe is-- 00:36:15.260 --> 00:36:18.440 align:middle line:84% I don't know this piece, but it's in 2/2. 00:36:18.440 --> 00:36:23.930 align:middle line:84% Really seldom beat three was offset 00:36:23.930 --> 00:36:26.490 align:middle line:90% to higher than offset one. 00:36:26.490 --> 00:36:26.990 align:middle line:90% Wow. 00:36:26.990 --> 00:36:29.220 align:middle line:90% That is really, really low. 00:36:29.220 --> 00:36:32.610 align:middle line:90% And it's in one sharp. 00:36:32.610 --> 00:36:33.300 align:middle line:90% Oh, good. 00:36:33.300 --> 00:36:34.430 align:middle line:90% Flats do work. 00:36:34.430 --> 00:36:35.940 align:middle line:90% And it is zero. 00:36:35.940 --> 00:36:37.620 align:middle line:90% It is not a jig. 00:36:37.620 --> 00:36:39.260 align:middle line:90% It is a hornpipe. 00:36:39.260 --> 00:36:42.410 align:middle line:84% I know we all know these differences between reels 00:36:42.410 --> 00:36:45.260 align:middle line:84% and hornpipes and jigs and all these things. 00:36:45.260 --> 00:36:47.670 align:middle line:84% Well, maybe some of us got our pirates in. 00:36:47.670 --> 00:36:51.470 align:middle line:84% OK, we can see here BannocksOBarleyMealJig, 00:36:51.470 --> 00:36:53.780 align:middle line:90% which is in 6/8. 00:36:53.780 --> 00:36:55.430 align:middle line:90% It has this feature. 00:36:55.430 --> 00:36:59.870 align:middle line:84% It's in two sharps and there's a one here, so it is a jig. 00:36:59.870 --> 00:37:01.287 align:middle line:90% So that works pretty well. 00:37:01.287 --> 00:37:02.870 align:middle line:84% And we can scroll all the way through. 00:37:02.870 --> 00:37:04.078 align:middle line:90% Wow, there's a lot of pieces. 00:37:04.078 --> 00:37:05.120 align:middle line:90% OK, great. 00:37:05.120 --> 00:37:06.620 align:middle line:90% And some of them are jigs. 00:37:06.620 --> 00:37:07.545 align:middle line:90% Some of them are not. 00:37:07.545 --> 00:37:10.420 align:middle line:90% 00:37:10.420 --> 00:37:11.304 align:middle line:90% OK. 00:37:11.304 --> 00:37:15.050 align:middle line:84% So now we've extracted all this data. 00:37:15.050 --> 00:37:19.390 align:middle line:84% If we put all of this into some sort of file somewhere, 00:37:19.390 --> 00:37:21.620 align:middle line:90% we will have it all done. 00:37:21.620 --> 00:37:23.500 align:middle line:84% So I'm going to take a two second break 00:37:23.500 --> 00:37:28.210 align:middle line:84% and show how we move from feature extraction 00:37:28.210 --> 00:37:30.700 align:middle line:90% to machine learning. 00:37:30.700 --> 00:37:32.290 align:middle line:84% OK, one of the most important things 00:37:32.290 --> 00:37:35.440 align:middle line:84% we're going to be thinking about when we think about machine 00:37:35.440 --> 00:37:38.620 align:middle line:84% learning is that we're going to be 00:37:38.620 --> 00:37:42.530 align:middle line:84% training the computer on a certain set of data, 00:37:42.530 --> 00:37:45.580 align:middle line:84% training our algorithms, training our neural networks 00:37:45.580 --> 00:37:48.520 align:middle line:84% or whatever, artificial intelligence and machine 00:37:48.520 --> 00:37:52.100 align:middle line:84% learning algorithm we choose to use on some portion of the data. 00:37:52.100 --> 00:37:55.400 align:middle line:84% And then we want to see, how good is the computer doing? 00:37:55.400 --> 00:37:58.990 align:middle line:84% So we're going to hold out some portion of the data 00:37:58.990 --> 00:38:00.070 align:middle line:90% to be our test. 00:38:00.070 --> 00:38:04.390 align:middle line:84% To see, well, if the training data is somewhat like the test 00:38:04.390 --> 00:38:06.700 align:middle line:84% data, but the computer's seen the training data 00:38:06.700 --> 00:38:08.480 align:middle line:84% and it hasn't seen the test data, 00:38:08.480 --> 00:38:13.570 align:middle line:84% how well does it learn from the training data 00:38:13.570 --> 00:38:18.080 align:middle line:84% in order to make identifications in the test data? 00:38:18.080 --> 00:38:21.800 align:middle line:84% And there are lots of different ways of doing this. 00:38:21.800 --> 00:38:24.410 align:middle line:84% Please take a machine learning class. 00:38:24.410 --> 00:38:27.590 align:middle line:84% Take an artificial intelligence class so you can learn about one 00:38:27.590 --> 00:38:31.960 align:middle line:84% hold out cross-validation or 1/10 hold out. 00:38:31.960 --> 00:38:34.433 align:middle line:84% Lots of different ways that we can do this better 00:38:34.433 --> 00:38:35.600 align:middle line:90% than what we're going to do. 00:38:35.600 --> 00:38:40.210 align:middle line:84% But instead, for now, to make things very simple, 00:38:40.210 --> 00:38:42.730 align:middle line:84% for whom this is going to be your first introduction 00:38:42.730 --> 00:38:45.790 align:middle line:84% to artificial intelligence and machine learning, what we're 00:38:45.790 --> 00:38:49.060 align:middle line:84% going to do is we're going to divide all of our training data 00:38:49.060 --> 00:38:52.260 align:middle line:84% into two parts, all of our data into half of it 00:38:52.260 --> 00:38:55.300 align:middle line:84% as the test data and half of it as the training data. 00:38:55.300 --> 00:38:58.800 align:middle line:84% And I'm just going to go through and almost always do everything 00:38:58.800 --> 00:39:02.130 align:middle line:84% in, this one to test, this one training, this one 00:39:02.130 --> 00:39:03.610 align:middle line:90% test, this one training. 00:39:03.610 --> 00:39:06.150 align:middle line:84% It's just going to make things easier. 00:39:06.150 --> 00:39:09.600 align:middle line:84% But there are better ways to do it. 00:39:09.600 --> 00:39:10.450 align:middle line:90% Enough for now. 00:39:10.450 --> 00:39:14.070 align:middle line:84% Let's get to some artificial intelligence and machine 00:39:14.070 --> 00:39:16.440 align:middle line:90% learning. 00:39:16.440 --> 00:39:17.010 align:middle line:90% OK. 00:39:17.010 --> 00:39:20.950 align:middle line:84% Let's go ahead and open up two files for writing. 00:39:20.950 --> 00:39:24.210 align:middle line:90% I'm going to call them-- 00:39:24.210 --> 00:39:25.240 align:middle line:90% they're on my desktop. 00:39:25.240 --> 00:39:29.130 align:middle line:84% Oh, now you know my secret, secret path. 00:39:29.130 --> 00:39:32.358 align:middle line:90% We'll call it train_data.tab. 00:39:32.358 --> 00:39:33.900 align:middle line:84% We're going to open this for writing, 00:39:33.900 --> 00:39:37.320 align:middle line:90% and we'll call it as training. 00:39:37.320 --> 00:39:40.710 align:middle line:84% If you're not familiar with Python file 00:39:40.710 --> 00:39:44.610 align:middle line:84% handles, please take a second to look this up 00:39:44.610 --> 00:39:47.950 align:middle line:84% because otherwise this won't make much sense. 00:39:47.950 --> 00:39:53.210 align:middle line:84% But essentially we're going to have two files so we're 00:39:53.210 --> 00:39:56.690 align:middle line:84% going to be able to write to, going back and forth, 00:39:56.690 --> 00:39:57.990 align:middle line:90% one versus the other. 00:39:57.990 --> 00:40:00.930 align:middle line:84% There's going to be better ways of doing this, 00:40:00.930 --> 00:40:03.630 align:middle line:84% but this is how we're going to do it first. 00:40:03.630 --> 00:40:04.170 align:middle line:90% OK. 00:40:04.170 --> 00:40:06.470 align:middle line:90% We're going to need to say-- 00:40:06.470 --> 00:40:09.650 align:middle line:84% for a lot of various algorithms, we're 00:40:09.650 --> 00:40:13.110 align:middle line:84% going to need to be able to say what our column names are. 00:40:13.110 --> 00:40:15.590 align:middle line:84% So we'll say our columns are going 00:40:15.590 --> 00:40:18.900 align:middle line:84% to be a series of tab separated values. 00:40:18.900 --> 00:40:21.680 align:middle line:84% That's what these dot tab files are going to be. 00:40:21.680 --> 00:40:27.240 align:middle line:84% A series of tab separated values that join these titles, 00:40:27.240 --> 00:40:32.600 align:middle line:84% filename, numerator-- doesn't matter what these are called-- 00:40:32.600 --> 00:40:37.220 align:middle line:90% denominator beat3 proportion. 00:40:37.220 --> 00:40:40.080 align:middle line:84% I'm not going to make myself type that again. 00:40:40.080 --> 00:40:44.940 align:middle line:84% And we'll say is_jig at the end for a ground truth. 00:40:44.940 --> 00:40:46.040 align:middle line:90% Fantastic. 00:40:46.040 --> 00:40:46.950 align:middle line:90% OK. 00:40:46.950 --> 00:40:48.480 align:middle line:90% Those are columns. 00:40:48.480 --> 00:40:50.270 align:middle line:84% And first I'm going to do is write 00:40:50.270 --> 00:40:57.110 align:middle line:84% to each of our different files, write our columns with a line 00:40:57.110 --> 00:41:00.020 align:middle line:90% break at the end. 00:41:00.020 --> 00:41:02.720 align:middle line:84% Our testing is going to write columns 00:41:02.720 --> 00:41:05.390 align:middle line:90% with a line break at the end. 00:41:05.390 --> 00:41:06.120 align:middle line:90% Great. 00:41:06.120 --> 00:41:08.720 align:middle line:84% Now we want to describe what kind of value, what kind 00:41:08.720 --> 00:41:11.660 align:middle line:84% of value will a machine learning toolkit 00:41:11.660 --> 00:41:16.340 align:middle line:84% expect to get from filename, string, numerator, integer, 00:41:16.340 --> 00:41:18.570 align:middle line:90% beat3, float. 00:41:18.570 --> 00:41:19.830 align:middle line:90% Things like that. 00:41:19.830 --> 00:41:24.290 align:middle line:84% So we'll say our descriptions are going to be-- 00:41:24.290 --> 00:41:28.200 align:middle line:84% and these are pretty standard formats but they're not Python. 00:41:28.200 --> 00:41:29.610 align:middle line:90% So look at this. 00:41:29.610 --> 00:41:30.500 align:middle line:90% It's not str. 00:41:30.500 --> 00:41:31.860 align:middle line:90% It's string. 00:41:31.860 --> 00:41:36.750 align:middle line:84% And we'll say that our integers are discrete data with an i. 00:41:36.750 --> 00:41:40.670 align:middle line:90% Not discreet like indiscreet. 00:41:40.670 --> 00:41:41.910 align:middle line:90% OK, I won't make a joke. 00:41:41.910 --> 00:41:44.930 align:middle line:84% But discrete like integer values. 00:41:44.930 --> 00:41:48.440 align:middle line:84% But beat3 data can be continuous. 00:41:48.440 --> 00:41:51.590 align:middle line:90% It can be 1.2, 1.21, so on. 00:41:51.590 --> 00:41:55.250 align:middle line:84% And then another discrete data for the number of sharps. 00:41:55.250 --> 00:41:59.060 align:middle line:84% And then for is_jib, it's also at 0, 1, 00:41:59.060 --> 00:42:00.950 align:middle line:84% so it's going to be discrete data. 00:42:00.950 --> 00:42:02.320 align:middle line:90% Fantastic. 00:42:02.320 --> 00:42:07.750 align:middle line:84% We're going to write thees to these files, the descriptions 00:42:07.750 --> 00:42:08.650 align:middle line:90% also. 00:42:08.650 --> 00:42:09.250 align:middle line:90% Boom. 00:42:09.250 --> 00:42:10.510 align:middle line:90% OK. 00:42:10.510 --> 00:42:11.010 align:middle line:90% Great. 00:42:11.010 --> 00:42:14.290 align:middle line:90% 00:42:14.290 --> 00:42:18.110 align:middle line:84% We also tend to put these meta variables in, 00:42:18.110 --> 00:42:22.010 align:middle line:84% and they say, if anything is special, 00:42:22.010 --> 00:42:23.660 align:middle line:90% if anything is unlike the rest. 00:42:23.660 --> 00:42:27.380 align:middle line:84% Well, the ones that are special is the first thing. 00:42:27.380 --> 00:42:29.960 align:middle line:90% String is just a metadata. 00:42:29.960 --> 00:42:34.210 align:middle line:84% That is to say, it just means something 00:42:34.210 --> 00:42:37.900 align:middle line:84% so that we can more easily diagnose problems and look 00:42:37.900 --> 00:42:38.600 align:middle line:90% things up later. 00:42:38.600 --> 00:42:39.950 align:middle line:90% We don't actually need it. 00:42:39.950 --> 00:42:43.180 align:middle line:84% We're not going to study the filename. 00:42:43.180 --> 00:42:46.750 align:middle line:84% And then these are just normal things. 00:42:46.750 --> 00:42:50.260 align:middle line:84% Numerator, denominator, beat3, number of sharps. 00:42:50.260 --> 00:42:54.370 align:middle line:84% And then at the very end, the final thing is our class. 00:42:54.370 --> 00:42:56.640 align:middle line:84% The thing that we're trying to learn. 00:42:56.640 --> 00:43:04.320 align:middle line:84% We're trying to learn, does this information, 98.11112, 00:43:04.320 --> 00:43:07.350 align:middle line:90% describe a jig or not? 00:43:07.350 --> 00:43:10.180 align:middle line:84% The class of objects that is a jig. 00:43:10.180 --> 00:43:10.680 align:middle line:90% Great. 00:43:10.680 --> 00:43:12.520 align:middle line:90% So these are our meta variables. 00:43:12.520 --> 00:43:18.905 align:middle line:84% And we're going to write them to, again, our-- 00:43:18.905 --> 00:43:19.780 align:middle line:90% do I still have that? 00:43:19.780 --> 00:43:21.900 align:middle line:90% No. 00:43:21.900 --> 00:43:27.990 align:middle line:84% I'm going to write that to our training and to our testing. 00:43:27.990 --> 00:43:30.490 align:middle line:90% Great. 00:43:30.490 --> 00:43:30.990 align:middle line:90% Here. 00:43:30.990 --> 00:43:33.520 align:middle line:90% I can do this anyhow. 00:43:33.520 --> 00:43:38.881 align:middle line:90% We can cat that file. 00:43:38.881 --> 00:43:42.090 align:middle line:84% These magic things are stuff you can do in Jupyter that you 00:43:42.090 --> 00:43:44.180 align:middle line:90% can't do in normal Python. 00:43:44.180 --> 00:43:49.770 align:middle line:90% But it'll see train_data tab. 00:43:49.770 --> 00:43:50.270 align:middle line:90% Yep. 00:43:50.270 --> 00:43:52.920 align:middle line:84% We can see what we've written out there. 00:43:52.920 --> 00:43:54.000 align:middle line:90% OK, great. 00:43:54.000 --> 00:43:55.730 align:middle line:84% Now we're going to take this and we're 00:43:55.730 --> 00:43:57.540 align:middle line:84% going to write more things to it. 00:43:57.540 --> 00:44:04.520 align:middle line:84% We're going to write into these two files information 00:44:04.520 --> 00:44:07.020 align:middle line:84% about from our feature extraction. 00:44:07.020 --> 00:44:13.640 align:middle line:84% So we're going to go back for each piece in Ryan. 00:44:13.640 --> 00:44:17.090 align:middle line:90% Enumerate Ryan. 00:44:17.090 --> 00:44:20.930 align:middle line:84% And enumerate says our first thing goes into piece 00:44:20.930 --> 00:44:21.953 align:middle line:90% and we'll call it i. 00:44:21.953 --> 00:44:23.370 align:middle line:84% Actually, We'll not call it piece. 00:44:23.370 --> 00:44:26.310 align:middle line:90% We'll call it metadata_entry. 00:44:26.310 --> 00:44:27.860 align:middle line:90% Great. 00:44:27.860 --> 00:44:32.540 align:middle line:84% And we'll say piece_parsed is our metadata_entry.parse. 00:44:32.540 --> 00:44:35.840 align:middle line:90% 00:44:35.840 --> 00:44:36.500 align:middle line:90% Great. 00:44:36.500 --> 00:44:43.230 align:middle line:84% And our features are for feature_extractor 00:44:43.230 --> 00:44:47.480 align:middle line:90% on piece_parsed. 00:44:47.480 --> 00:44:48.510 align:middle line:90% Piece_parsed. 00:44:48.510 --> 00:44:53.430 align:middle line:84% And now we're going to make a string out of that feature. 00:44:53.430 --> 00:44:57.140 align:middle line:84% String equals tab separated version 00:44:57.140 --> 00:45:02.240 align:middle line:84% of the string value of each element 00:45:02.240 --> 00:45:05.280 align:middle line:90% for the element in features. 00:45:05.280 --> 00:45:06.052 align:middle line:90% OK. 00:45:06.052 --> 00:45:09.080 align:middle line:90% And we should call it f. 00:45:09.080 --> 00:45:11.510 align:middle line:90% Fantastic. 00:45:11.510 --> 00:45:15.860 align:middle line:90% And if i is even. 00:45:15.860 --> 00:45:19.310 align:middle line:84% No, if i is odd, we'll write it to the training 00:45:19.310 --> 00:45:25.910 align:middle line:90% data with a new line. 00:45:25.910 --> 00:45:29.540 align:middle line:84% And otherwise we'll write it to the testing 00:45:29.540 --> 00:45:35.180 align:middle line:90% data with a new line. 00:45:35.180 --> 00:45:36.020 align:middle line:90% Great. 00:45:36.020 --> 00:45:38.015 align:middle line:84% Hopefully I didn't type anything wrong. 00:45:38.015 --> 00:45:39.140 align:middle line:90% We'll find out in a second. 00:45:39.140 --> 00:45:42.220 align:middle line:90% 00:45:42.220 --> 00:45:43.090 align:middle line:90% Great. 00:45:43.090 --> 00:45:44.660 align:middle line:90% It seems to be working. 00:45:44.660 --> 00:45:46.922 align:middle line:90% It's slow. 00:45:46.922 --> 00:45:48.880 align:middle line:84% One of the things I'm going to eventually learn 00:45:48.880 --> 00:45:52.340 align:middle line:84% is that everything's slower when I'm screen recording. 00:45:52.340 --> 00:45:56.740 align:middle line:84% So we could see that, yep, there's 00:45:56.740 --> 00:45:58.670 align:middle line:90% a bunch of things in here. 00:45:58.670 --> 00:46:00.680 align:middle line:84% Most important one, can we get a jig? 00:46:00.680 --> 00:46:04.630 align:middle line:84% It has a one at the end if it's a jig, zero at the end 00:46:04.630 --> 00:46:07.750 align:middle line:90% if it's something else. 00:46:07.750 --> 00:46:08.990 align:middle line:90% These jigs are 6/8. 00:46:08.990 --> 00:46:10.660 align:middle line:90% Oh, look, there's a jig. 00:46:10.660 --> 00:46:12.140 align:middle line:90% It begins at 9/8. 00:46:12.140 --> 00:46:14.448 align:middle line:84% So maybe it's 8 that's more important. 00:46:14.448 --> 00:46:14.990 align:middle line:90% I don't know. 00:46:14.990 --> 00:46:17.180 align:middle line:84% We'll see if the computer can figure this out. 00:46:17.180 --> 00:46:19.880 align:middle line:84% We don't need to look at all 1,000 entries. 00:46:19.880 --> 00:46:24.590 align:middle line:84% Jupyter does a nice way of not putting that all together. 00:46:24.590 --> 00:46:25.910 align:middle line:90% OK, great. 00:46:25.910 --> 00:46:28.960 align:middle line:84% So now we have something that we can 00:46:28.960 --> 00:46:32.410 align:middle line:84% feed into an artificial intelligence classification 00:46:32.410 --> 00:46:33.230 align:middle line:90% program. 00:46:33.230 --> 00:46:36.350 align:middle line:84% So we don't have one so we're going to write one of our own. 00:46:36.350 --> 00:46:39.930 align:middle line:84% So we're going to say neural network. 00:46:39.930 --> 00:46:41.020 align:middle line:90% Just kidding. 00:46:41.020 --> 00:46:42.910 align:middle line:84% No, that's not what this class is about. 00:46:42.910 --> 00:46:44.110 align:middle line:90% That's other classes. 00:46:44.110 --> 00:46:45.370 align:middle line:90% We're going to install one. 00:46:45.370 --> 00:46:51.930 align:middle line:84% So go ahead and install orange3, which is a really great package. 00:46:51.930 --> 00:46:53.850 align:middle line:84% Oh, hey, I already have it installed. 00:46:53.850 --> 00:46:56.110 align:middle line:84% Yours probably takes a little bit longer than that. 00:46:56.110 --> 00:46:56.730 align:middle line:90% Great. 00:46:56.730 --> 00:46:58.710 align:middle line:90% Super. 00:46:58.710 --> 00:47:00.810 align:middle line:90% So now I have orange3. 00:47:00.810 --> 00:47:05.490 align:middle line:84% Orange3 names its stuff kind of funny. 00:47:05.490 --> 00:47:09.100 align:middle line:84% So I always-- because I've been using it for a long time, 00:47:09.100 --> 00:47:11.560 align:middle line:90% I just import it. 00:47:11.560 --> 00:47:12.970 align:middle line:90% It's not orange3. 00:47:12.970 --> 00:47:14.050 align:middle line:90% I just call it orange. 00:47:14.050 --> 00:47:14.682 align:middle line:90% Great. 00:47:14.682 --> 00:47:16.140 align:middle line:84% Now what we're going to do is we're 00:47:16.140 --> 00:47:20.490 align:middle line:84% going to load in our training data 00:47:20.490 --> 00:47:26.046 align:middle line:84% from an orange data table based on that information. 00:47:26.046 --> 00:47:28.830 align:middle line:90% 00:47:28.830 --> 00:47:31.232 align:middle line:90% So a train_data.tab. 00:47:31.232 --> 00:47:35.540 align:middle line:90% 00:47:35.540 --> 00:47:39.560 align:middle line:84% I reverted myself to using camelcase things. 00:47:39.560 --> 00:47:41.270 align:middle line:84% I probably shouldn't, but I'm not 00:47:41.270 --> 00:47:44.900 align:middle line:90% going to rewrite everything now. 00:47:44.900 --> 00:47:48.330 align:middle line:84% So it doesn't look like much, but I have a lot in my sleeve. 00:47:48.330 --> 00:47:48.900 align:middle line:90% OK. 00:47:48.900 --> 00:47:51.390 align:middle line:84% So we're going to load those two. 00:47:51.390 --> 00:47:52.310 align:middle line:90% Great. 00:47:52.310 --> 00:47:55.110 align:middle line:84% We can look at train_data little bit. 00:47:55.110 --> 00:47:55.610 align:middle line:90% Yep. 00:47:55.610 --> 00:47:57.380 align:middle line:90% Whole bunch of stuff. 00:47:57.380 --> 00:48:02.790 align:middle line:84% These are, I think, matplotlib-ish type things. 00:48:02.790 --> 00:48:03.690 align:middle line:90% OK, great. 00:48:03.690 --> 00:48:08.000 align:middle line:84% So what we want to do is create a bunch 00:48:08.000 --> 00:48:13.190 align:middle line:84% of learners, things that can classify and understand 00:48:13.190 --> 00:48:15.810 align:middle line:84% data and try to make guesses about things. 00:48:15.810 --> 00:48:20.400 align:middle line:84% Our first learner is kind of like the placebo of learners. 00:48:20.400 --> 00:48:23.120 align:middle line:84% It's called the majority learner. 00:48:23.120 --> 00:48:27.060 align:middle line:84% And there's other ways of doing this, but we'll use this one. 00:48:27.060 --> 00:48:33.570 align:middle line:84% And the majority learner is pretty cool. 00:48:33.570 --> 00:48:35.820 align:middle line:90% It just looks at the data. 00:48:35.820 --> 00:48:41.840 align:middle line:84% It looks at what we put in the training data and it says, 00:48:41.840 --> 00:48:43.050 align:middle line:90% you know what? 00:48:43.050 --> 00:48:47.390 align:middle line:84% There's about three times as many things that are not jigs 00:48:47.390 --> 00:48:49.110 align:middle line:90% as there are jigs. 00:48:49.110 --> 00:48:52.760 align:middle line:84% So I can get 3 out of 4 right if I just 00:48:52.760 --> 00:48:54.930 align:middle line:84% learn to always vote with the majority. 00:48:54.930 --> 00:48:58.655 align:middle line:84% I don't know anything about music, but I'll just say-- 00:48:58.655 --> 00:49:00.510 align:middle line:84% I'll always say that's not a jig. 00:49:00.510 --> 00:49:01.260 align:middle line:90% And you know what? 00:49:01.260 --> 00:49:03.530 align:middle line:90% I'll usually be right. 00:49:03.530 --> 00:49:06.650 align:middle line:84% So never classify things according to, 00:49:06.650 --> 00:49:09.350 align:middle line:90% hey, is this right or wrong? 00:49:09.350 --> 00:49:11.190 align:middle line:84% Did it get it right more than half the time? 00:49:11.190 --> 00:49:11.690 align:middle line:90% No. 00:49:11.690 --> 00:49:15.090 align:middle line:84% Classify how you're doing against, 00:49:15.090 --> 00:49:17.100 align:middle line:84% did it do better than the majority learner? 00:49:17.100 --> 00:49:19.920 align:middle line:90% 00:49:19.920 --> 00:49:22.070 align:middle line:84% You're going to now, if you don't 00:49:22.070 --> 00:49:24.350 align:middle line:84% know what this is, we're going to learn 00:49:24.350 --> 00:49:31.210 align:middle line:84% what a k nearest neighbors is, knnLearner. 00:49:31.210 --> 00:49:33.280 align:middle line:90% Classification. 00:49:33.280 --> 00:49:34.840 align:middle line:90% Classification. 00:49:34.840 --> 00:49:37.870 align:middle line:90% KnnLearner. 00:49:37.870 --> 00:49:41.320 align:middle line:84% And this is a pretty standard artificial intelligence 00:49:41.320 --> 00:49:44.380 align:middle line:84% learning algorithm and so I'm not 00:49:44.380 --> 00:49:45.980 align:middle line:84% going to say too much about that. 00:49:45.980 --> 00:49:48.700 align:middle line:84% So we create a kind of learner thing, 00:49:48.700 --> 00:49:52.030 align:middle line:84% and our classifier is a learner that's 00:49:52.030 --> 00:49:58.330 align:middle line:84% going to work on a particular set of data, our training data. 00:49:58.330 --> 00:49:59.240 align:middle line:90% OK, good. 00:49:59.240 --> 00:50:01.970 align:middle line:84% So we create a learner, create a classifier. 00:50:01.970 --> 00:50:03.520 align:middle line:84% Hasn't actually learned anything yet, 00:50:03.520 --> 00:50:07.250 align:middle line:84% but it's going to be ready for doing that. 00:50:07.250 --> 00:50:09.836 align:middle line:90% OK. 00:50:09.836 --> 00:50:12.190 align:middle line:84% Now we're going to check how often does 00:50:12.190 --> 00:50:14.240 align:middle line:84% the majority learner get things-- 00:50:14.240 --> 00:50:16.870 align:middle line:84% how many times does it get it correct 00:50:16.870 --> 00:50:22.840 align:middle line:84% and how many times does the k nearest neighbor's learner 00:50:22.840 --> 00:50:24.730 align:middle line:90% get things correct? 00:50:24.730 --> 00:50:30.870 align:middle line:84% And how many total things are there? 00:50:30.870 --> 00:50:34.150 align:middle line:90% We get that as a float. 00:50:34.150 --> 00:50:39.300 align:middle line:84% Still thinking too much how to think like Python too. 00:50:39.300 --> 00:50:43.590 align:middle line:90% So how many test data there are. 00:50:43.590 --> 00:50:44.610 align:middle line:90% Go and see. 00:50:44.610 --> 00:50:47.330 align:middle line:90% OK, 532 in the test data. 00:50:47.330 --> 00:50:50.790 align:middle line:90% 00:50:50.790 --> 00:50:52.980 align:middle line:84% The training data does not need to be 00:50:52.980 --> 00:50:55.170 align:middle line:84% exactly the same size, or even close 00:50:55.170 --> 00:50:57.300 align:middle line:84% to the same size as the test data. 00:50:57.300 --> 00:50:58.470 align:middle line:90% But in this case, they are. 00:50:58.470 --> 00:50:59.050 align:middle line:90% Great. 00:50:59.050 --> 00:51:01.650 align:middle line:84% So now we're going to iterate through each 00:51:01.650 --> 00:51:12.600 align:middle line:84% of the rows of our test data not our training data. 00:51:12.600 --> 00:51:16.140 align:middle line:84% So we're going to know what row number we are on, i, 00:51:16.140 --> 00:51:17.530 align:middle line:90% and then test row. 00:51:17.530 --> 00:51:19.620 align:middle line:84% And then the majority learner is going 00:51:19.620 --> 00:51:24.490 align:middle line:84% to make a guess based on the data in the test row. 00:51:24.490 --> 00:51:26.740 align:middle line:84% However, the majority learner is really stupid, 00:51:26.740 --> 00:51:31.350 align:middle line:84% so it's guess is always going to be, it's not a jig. 00:51:31.350 --> 00:51:34.270 align:middle line:84% Then k nearest neighbor is going to make a guess. 00:51:34.270 --> 00:51:36.450 align:middle line:84% We hope that k nearest neighbor is 00:51:36.450 --> 00:51:40.480 align:middle line:84% making a guess based on some intelligence that it has, 00:51:40.480 --> 00:51:42.790 align:middle line:90% or fake intelligence. 00:51:42.790 --> 00:51:45.340 align:middle line:84% I mean, it's called artificial intelligence for a reason. 00:51:45.340 --> 00:51:51.880 align:middle line:84% And then we get to see in this case what the actual answer is, 00:51:51.880 --> 00:51:52.720 align:middle line:90% what the class-- 00:51:52.720 --> 00:51:53.920 align:middle line:90% whether it's a jig or not. 00:51:53.920 --> 00:51:56.160 align:middle line:90% So that's our real answer. 00:51:56.160 --> 00:51:58.470 align:middle line:84% So we know this but k nearest neighbor 00:51:58.470 --> 00:52:00.660 align:middle line:90% is not allowed to look at this. 00:52:00.660 --> 00:52:03.730 align:middle line:84% We really hope the computer's not peeking, 00:52:03.730 --> 00:52:07.390 align:middle line:84% otherwise all artificial intelligence is kind of screwed. 00:52:07.390 --> 00:52:07.950 align:middle line:90% OK. 00:52:07.950 --> 00:52:10.440 align:middle line:84% We'll just print out some test data 00:52:10.440 --> 00:52:22.330 align:middle line:84% that the row i is a little bit of information about our set. 00:52:22.330 --> 00:52:23.770 align:middle line:90% That should be good. 00:52:23.770 --> 00:52:24.270 align:middle line:90% Great. 00:52:24.270 --> 00:52:38.293 align:middle line:84% We'll print out that the answer to row i is jig, if realAnswer. 00:52:38.293 --> 00:52:40.850 align:middle line:90% 00:52:40.850 --> 00:52:42.770 align:middle line:90% I think that works. 00:52:42.770 --> 00:52:44.320 align:middle line:90% Else not jig. 00:52:44.320 --> 00:52:50.510 align:middle line:90% 00:52:50.510 --> 00:52:52.410 align:middle line:90% And then we'll put a tab. 00:52:52.410 --> 00:52:56.450 align:middle line:84% We'll say k nearest neighbor got it right or wrong. 00:52:56.450 --> 00:53:00.140 align:middle line:84% Right if knn's-- if the k nearest neighbors guess is 00:53:00.140 --> 00:53:05.480 align:middle line:84% the same as the realAnswer, else wrong. 00:53:05.480 --> 00:53:06.530 align:middle line:90% Whew. 00:53:06.530 --> 00:53:09.200 align:middle line:90% That's some pretty tough stuff. 00:53:09.200 --> 00:53:11.310 align:middle line:84% Our thing can't even get that right. 00:53:11.310 --> 00:53:11.853 align:middle line:90% Good. 00:53:11.853 --> 00:53:12.353 align:middle line:90% OK. 00:53:12.353 --> 00:53:14.960 align:middle line:90% 00:53:14.960 --> 00:53:16.945 align:middle line:84% We'll print a blank line after that. 00:53:16.945 --> 00:53:19.730 align:middle line:90% 00:53:19.730 --> 00:53:24.550 align:middle line:84% If the majority guessed right, then the majority 00:53:24.550 --> 00:53:29.140 align:middle line:90% correct increment. 00:53:29.140 --> 00:53:35.920 align:middle line:84% If k nearest neighbors guess is right-- 00:53:35.920 --> 00:53:40.180 align:middle line:84% so we're just checking at this point how good is this. 00:53:40.180 --> 00:53:44.260 align:middle line:84% k nearest neighbor's correct, it is going to be one bigger. 00:53:44.260 --> 00:53:45.940 align:middle line:90% Great. 00:53:45.940 --> 00:53:49.150 align:middle line:84% So we'll go ahead and let that run. 00:53:49.150 --> 00:53:51.430 align:middle line:90% Oops. 00:53:51.430 --> 00:53:53.450 align:middle line:90% Colons are very nice. 00:53:53.450 --> 00:53:54.490 align:middle line:90% Oops. 00:53:54.490 --> 00:53:55.530 align:middle line:90% I screwed something up. 00:53:55.530 --> 00:53:59.440 align:middle line:90% 00:53:59.440 --> 00:54:00.580 align:middle line:90% Oops. 00:54:00.580 --> 00:54:02.350 align:middle line:90% There we go. 00:54:02.350 --> 00:54:06.010 align:middle line:84% I have to cast this as a list because-- 00:54:06.010 --> 00:54:06.860 align:middle line:90% where is it? 00:54:06.860 --> 00:54:10.150 align:middle line:84% This data can't be sliced, but lists can. 00:54:10.150 --> 00:54:10.760 align:middle line:90% OK. 00:54:10.760 --> 00:54:12.450 align:middle line:90% Is that going on? 00:54:12.450 --> 00:54:14.930 align:middle line:84% I probably need to reset my data, 00:54:14.930 --> 00:54:16.820 align:middle line:90% so let's go ahead and do that. 00:54:16.820 --> 00:54:18.350 align:middle line:90% It's going to be printing. 00:54:18.350 --> 00:54:19.180 align:middle line:90% Cool. 00:54:19.180 --> 00:54:20.500 align:middle line:90% Not jig. 00:54:20.500 --> 00:54:22.740 align:middle line:84% k nearest neighbors got it wrong. 00:54:22.740 --> 00:54:23.895 align:middle line:90% Not jig. 00:54:23.895 --> 00:54:25.270 align:middle line:84% k nearest neighbors got it right. 00:54:25.270 --> 00:54:26.250 align:middle line:90% Good. 00:54:26.250 --> 00:54:28.840 align:middle line:84% So, hey, we see a lot of got it right. 00:54:28.840 --> 00:54:29.410 align:middle line:90% I like that. 00:54:29.410 --> 00:54:30.250 align:middle line:90% I like that. 00:54:30.250 --> 00:54:31.000 align:middle line:90% OK. 00:54:31.000 --> 00:54:33.810 align:middle line:90% Here's the moment of truth. 00:54:33.810 --> 00:54:39.790 align:middle line:84% Without machine learning, there's a majority learner. 00:54:39.790 --> 00:54:42.030 align:middle line:90% Majority correct. 00:54:42.030 --> 00:54:48.430 align:middle line:84% Always put things into print percentages. 00:54:48.430 --> 00:54:52.000 align:middle line:84% Always round your percentages to something reasonable. 00:54:52.000 --> 00:54:55.170 align:middle line:84% Nobody wants to see all those dangling digits. 00:54:55.170 --> 00:54:55.870 align:middle line:90% OK. 00:54:55.870 --> 00:54:58.060 align:middle line:90% So it got that number right. 00:54:58.060 --> 00:55:06.430 align:middle line:84% And we'll say with knn machine learning, 00:55:06.430 --> 00:55:10.240 align:middle line:90% we got the knn correct. 00:55:10.240 --> 00:55:12.450 align:middle line:90% How well do we do? 00:55:12.450 --> 00:55:15.660 align:middle line:84% Without machine learning, got it 77% right. 00:55:15.660 --> 00:55:20.560 align:middle line:84% k nearest neighbors got it 95% right. 00:55:20.560 --> 00:55:22.200 align:middle line:90% Wow. 00:55:22.200 --> 00:55:23.560 align:middle line:90% OK, that's pretty awesome. 00:55:23.560 --> 00:55:25.530 align:middle line:84% That's a lot better than I thought 00:55:25.530 --> 00:55:30.030 align:middle line:90% when I thought I would do this. 00:55:30.030 --> 00:55:31.995 align:middle line:84% By the way, one of the things we can do. 00:55:31.995 --> 00:55:38.830 align:middle line:84% You probably want to keep track of your data and don't do this, 00:55:38.830 --> 00:55:40.630 align:middle line:84% but we'll just go ahead and check. 00:55:40.630 --> 00:55:43.920 align:middle line:84% Does it make a difference if we swap the training and the test 00:55:43.920 --> 00:55:44.420 align:middle line:90% data? 00:55:44.420 --> 00:55:47.880 align:middle line:90% 00:55:47.880 --> 00:55:48.930 align:middle line:90% Oops. 00:55:48.930 --> 00:55:50.020 align:middle line:90% Not training data. 00:55:50.020 --> 00:55:51.190 align:middle line:90% Train_data. 00:55:51.190 --> 00:55:55.200 align:middle line:90% 00:55:55.200 --> 00:55:57.810 align:middle line:90% Do we get a fluke? 00:55:57.810 --> 00:56:01.230 align:middle line:90% And we can go ahead and run. 00:56:01.230 --> 00:56:02.710 align:middle line:90% 72 and 94. 00:56:02.710 --> 00:56:04.810 align:middle line:90% So yeah, still pretty good. 00:56:04.810 --> 00:56:06.430 align:middle line:90% Doesn't matter which one we do. 00:56:06.430 --> 00:56:11.010 align:middle line:84% I'm going to switch that back because I will totally 00:56:11.010 --> 00:56:12.130 align:middle line:90% make that mistake. 00:56:12.130 --> 00:56:13.240 align:middle line:90% But yeah. 00:56:13.240 --> 00:56:16.210 align:middle line:84% So we're doing really well with this. 00:56:16.210 --> 00:56:18.810 align:middle line:84% So this is how you can do artificial intelligence 00:56:18.810 --> 00:56:19.885 align:middle line:90% with machine learning. 00:56:19.885 --> 00:56:23.300 align:middle line:90% 00:56:23.300 --> 00:56:25.740 align:middle line:84% But that's been a little bit hard. 00:56:25.740 --> 00:56:30.770 align:middle line:84% So fortunately, music21 has some tools 00:56:30.770 --> 00:56:34.230 align:middle line:84% to make a lot of these things even easier here. 00:56:34.230 --> 00:56:37.460 align:middle line:84% So music21 feature extraction will be the last topic 00:56:37.460 --> 00:56:40.340 align:middle line:84% that I'm going to talk about today. 00:56:40.340 --> 00:56:43.700 align:middle line:84% So we can go ahead and import from music21 00:56:43.700 --> 00:56:45.063 align:middle line:90% the features module. 00:56:45.063 --> 00:56:49.560 align:middle line:90% 00:56:49.560 --> 00:56:50.060 align:middle line:90% Great. 00:56:50.060 --> 00:56:52.820 align:middle line:84% So within the features module, there 00:56:52.820 --> 00:56:55.050 align:middle line:90% are things called DataSets. 00:56:55.050 --> 00:56:59.420 align:middle line:84% So let's create training_set, which 00:56:59.420 --> 00:57:05.990 align:middle line:84% will be a DataSet with a class label of is_jig. 00:57:05.990 --> 00:57:07.620 align:middle line:84% We call it ground truth or whatever. 00:57:07.620 --> 00:57:13.190 align:middle line:84% And we'll create an identical one for our test set. 00:57:13.190 --> 00:57:14.490 align:middle line:90% Testing_set. 00:57:14.490 --> 00:57:15.050 align:middle line:90% Great. 00:57:15.050 --> 00:57:17.650 align:middle line:84% So these are just things that are 00:57:17.650 --> 00:57:22.060 align:middle line:84% ready to add two important things. 00:57:22.060 --> 00:57:26.710 align:middle line:84% A set of data that is streamed, so a whole bunch of streams 00:57:26.710 --> 00:57:29.810 align:middle line:84% will be added to both the training and the test set, 00:57:29.810 --> 00:57:35.110 align:middle line:84% and also a set of feature extractors. 00:57:35.110 --> 00:57:39.070 align:middle line:84% I'm not going to explain today how to write new feature 00:57:39.070 --> 00:57:42.560 align:middle line:84% extractors in the way music 21 expects them, 00:57:42.560 --> 00:57:45.220 align:middle line:84% but they're basically pretty similar to what we've just 00:57:45.220 --> 00:57:47.230 align:middle line:84% been writing with functions, except they're 00:57:47.230 --> 00:57:50.380 align:middle line:90% kind of expected to be classes. 00:57:50.380 --> 00:57:54.130 align:middle line:84% But we don't really have to because music21 00:57:54.130 --> 00:57:57.730 align:middle line:84% has a whole bunch of feature extractors. 00:57:57.730 --> 00:58:01.150 align:middle line:84% Some of them come from this amazing toolkit called 00:58:01.150 --> 00:58:03.850 align:middle line:84% jSymbolic, or they're re-interpretations 00:58:03.850 --> 00:58:08.810 align:middle line:84% of Corey McKay's jSymbolic toolkit, which is in Java, 00:58:08.810 --> 00:58:14.230 align:middle line:84% and they have numbers like D1 is the overall dynamic range 00:58:14.230 --> 00:58:14.840 align:middle line:90% features. 00:58:14.840 --> 00:58:17.460 align:middle line:84% Not all of them have been implemented 00:58:17.460 --> 00:58:20.410 align:middle line:90% I's are instrument features. 00:58:20.410 --> 00:58:24.360 align:middle line:84% You can probably see that the percentage of all notes being 00:58:24.360 --> 00:58:28.480 align:middle line:84% played by electric guitars, if you know something about jigs, 00:58:28.480 --> 00:58:34.200 align:middle line:84% is probably not going to be very helpful to classifying 00:58:34.200 --> 00:58:36.400 align:middle line:84% whether or not something is a jig. 00:58:36.400 --> 00:58:39.600 align:middle line:84% But maybe the average melodic interval 00:58:39.600 --> 00:58:42.330 align:middle line:84% will tell you something like that, or how often there 00:58:42.330 --> 00:58:44.940 align:middle line:84% are melodic thirds or melodic fifths, 00:58:44.940 --> 00:58:47.620 align:middle line:84% or maybe other things might tell signs. 00:58:47.620 --> 00:58:52.530 align:middle line:84% So these are melody features, pitch features, and so on. 00:58:52.530 --> 00:58:53.806 align:middle line:90% There's rhythm features. 00:58:53.806 --> 00:58:55.306 align:middle line:84% That's going to be pretty important. 00:58:55.306 --> 00:58:58.380 align:middle line:90% 00:58:58.380 --> 00:59:01.330 align:middle line:84% R31, initial time signature feature. 00:59:01.330 --> 00:59:07.030 align:middle line:84% R32, compound or simple meter features, and so on. 00:59:07.030 --> 00:59:10.800 align:middle line:84% So we could add a bunch of these features in, 00:59:10.800 --> 00:59:13.140 align:middle line:90% and we remember the names. 00:59:13.140 --> 00:59:15.670 align:middle line:84% Within the music21 documentation, 00:59:15.670 --> 00:59:18.030 align:middle line:84% there will be certain things that 00:59:18.030 --> 00:59:21.030 align:middle line:84% will tell you TripleMeterFeature set 00:59:21.030 --> 00:59:23.790 align:middle line:84% to 1 if the numerator of the initial time signature 00:59:23.790 --> 00:59:26.200 align:middle line:90% is 3 or 0 otherwise. 00:59:26.200 --> 00:59:30.420 align:middle line:84% And this shows you how it can work. 00:59:30.420 --> 00:59:34.710 align:middle line:84% So we can give a particular piece and it will tell us that 00:59:34.710 --> 00:59:39.460 align:middle line:84% bwv66.6 is 0, not in triple meter. 00:59:39.460 --> 00:59:41.160 align:middle line:90% OK. 00:59:41.160 --> 00:59:46.020 align:middle line:84% We've already done something really, really great. 00:59:46.020 --> 00:59:49.620 align:middle line:84% And so I will tell you that if we use this very similar feature 00:59:49.620 --> 00:59:51.580 align:middle line:84% extractors to the ones we did before, 00:59:51.580 --> 00:59:55.390 align:middle line:84% we're going to get the same results. 00:59:55.390 --> 00:59:57.030 align:middle line:84% But one of the other things we can do 00:59:57.030 --> 01:00:00.910 align:middle line:84% is we can just use a whole bunch of pretty random ones. 01:00:00.910 --> 01:00:07.170 align:middle line:84% So we'll get these features by ID. 01:00:07.170 --> 01:00:10.320 align:middle line:84% And these are a bunch of features that I used 01:00:10.320 --> 01:00:15.350 align:middle line:90% to classify various pieces of-- 01:00:15.350 --> 01:00:17.460 align:middle line:90% I think it was 16th-- 01:00:17.460 --> 01:00:18.590 align:middle line:90% 15th century music. 01:00:18.590 --> 01:00:20.340 align:middle line:90% No, 16th century music. 01:00:20.340 --> 01:00:23.540 align:middle line:84% So they're not exactly super tuned 01:00:23.540 --> 01:00:26.420 align:middle line:90% for what we're going to do. 01:00:26.420 --> 01:00:30.980 align:middle line:84% A lot of them are melodic features and so on. 01:00:30.980 --> 01:00:31.820 align:middle line:90% Let me see. 01:00:31.820 --> 01:00:36.810 align:middle line:84% Scroll over, see which ones haven't done pitch features. 01:00:36.810 --> 01:00:44.940 align:middle line:84% Pitch feature 8, 10, rhythm feature 15. 01:00:44.940 --> 01:00:45.440 align:middle line:90% Great. 01:00:45.440 --> 01:00:48.470 align:middle line:90% And we can see what they are. 01:00:48.470 --> 01:00:50.780 align:middle line:84% I'm already going to tell you, no surprise, 01:00:50.780 --> 01:00:52.160 align:middle line:84% these aren't going to work nearly 01:00:52.160 --> 01:00:55.220 align:middle line:84% as well as the ones we just made. 01:00:55.220 --> 01:00:57.818 align:middle line:84% Note density feature isn't going to help us very much. 01:00:57.818 --> 01:00:59.360 align:middle line:84% I would have thought range might have 01:00:59.360 --> 01:01:02.330 align:middle line:84% told us something about a jig, but maybe it's not. 01:01:02.330 --> 01:01:04.610 align:middle line:84% Some people will tell you, hey, just 01:01:04.610 --> 01:01:07.970 align:middle line:84% throw lots of feature extractors in. 01:01:07.970 --> 01:01:09.330 align:middle line:90% It can't hurt. 01:01:09.330 --> 01:01:10.590 align:middle line:90% It kind of can. 01:01:10.590 --> 01:01:14.080 align:middle line:84% But because we often don't have enough data to make 01:01:14.080 --> 01:01:16.450 align:middle line:90% use of all these features. 01:01:16.450 --> 01:01:17.900 align:middle line:90% But let's give it a shot. 01:01:17.900 --> 01:01:25.150 align:middle line:84% We're going to add these feature extractors to our training set 01:01:25.150 --> 01:01:28.635 align:middle line:90% and to our testing set. 01:01:28.635 --> 01:01:31.960 align:middle line:90% 01:01:31.960 --> 01:01:34.090 align:middle line:84% OK, so our training set and testing 01:01:34.090 --> 01:01:39.020 align:middle line:84% set is now ready to go where any stream we add to these sets, 01:01:39.020 --> 01:01:42.550 align:middle line:84% we're going to have these feature extractors added 01:01:42.550 --> 01:01:43.250 align:middle line:90% to them. 01:01:43.250 --> 01:01:48.970 align:middle line:84% So we'll say is for i, add metadata_entry 01:01:48.970 --> 01:01:54.970 align:middle line:90% in enumerate Ryan, again. 01:01:54.970 --> 01:01:58.960 align:middle line:84% Piece_parsed equals metadata_entry.parse. 01:01:58.960 --> 01:02:02.650 align:middle line:90% 01:02:02.650 --> 01:02:10.960 align:middle line:84% If i is odd, then training or test will be-- 01:02:10.960 --> 01:02:12.780 align:middle line:84% so we only have to write things once, 01:02:12.780 --> 01:02:14.530 align:middle line:84% we'll set our training or test set will be 01:02:14.530 --> 01:02:16.310 align:middle line:90% the training set in this case. 01:02:16.310 --> 01:02:22.570 align:middle line:84% Otherwise our training or test will be our testing set. 01:02:22.570 --> 01:02:24.010 align:middle line:90% Great. 01:02:24.010 --> 01:02:24.970 align:middle line:90% Filename. 01:02:24.970 --> 01:02:27.560 align:middle line:90% Is the piece parsed? 01:02:27.560 --> 01:02:30.610 align:middle line:90% FilePath.name. 01:02:30.610 --> 01:02:31.790 align:middle line:90% Is_jig. 01:02:31.790 --> 01:02:34.870 align:middle line:90% Same thing we're doing before. 01:02:34.870 --> 01:02:36.730 align:middle line:90% Here we're not going to-- 01:02:36.730 --> 01:02:38.300 align:middle line:90% music21 smart enough. 01:02:38.300 --> 01:02:40.940 align:middle line:84% It doesn't need us to cast this to an integer. 01:02:40.940 --> 01:02:43.160 align:middle line:84% It'll figure out what to do itself. 01:02:43.160 --> 01:02:45.580 align:middle line:90% OK. 01:02:45.580 --> 01:02:48.340 align:middle line:84% I'm going to say what row are we, is this a jig, 01:02:48.340 --> 01:02:49.970 align:middle line:90% and what's its filename. 01:02:49.970 --> 01:02:51.200 align:middle line:90% That's easy part. 01:02:51.200 --> 01:02:54.280 align:middle line:84% And then whatever data set we're working with, 01:02:54.280 --> 01:02:59.070 align:middle line:84% training or test, we're going to add the data from-- 01:02:59.070 --> 01:03:01.430 align:middle line:90% here's our stream, our score. 01:03:01.430 --> 01:03:02.450 align:middle line:90% What did we call it? 01:03:02.450 --> 01:03:03.590 align:middle line:90% ClassValue. 01:03:03.590 --> 01:03:05.930 align:middle line:90% What is the right answer? 01:03:05.930 --> 01:03:08.620 align:middle line:84% It's whether or not this is a jig. 01:03:08.620 --> 01:03:13.290 align:middle line:84% And then the identification for this will be the filename. 01:03:13.290 --> 01:03:14.260 align:middle line:90% Boom. 01:03:14.260 --> 01:03:15.550 align:middle line:90% It's going to go through. 01:03:15.550 --> 01:03:20.460 align:middle line:84% It has 1,000 of these to run, so I'm not going to wait for all 01:03:20.460 --> 01:03:23.140 align:middle line:90% of it to happen. 01:03:23.140 --> 01:03:27.690 align:middle line:84% I'm going to stop for a second because we don't have 01:03:27.690 --> 01:03:29.610 align:middle line:90% all that time in the world. 01:03:29.610 --> 01:03:31.230 align:middle line:84% That's part of the reason why we also 01:03:31.230 --> 01:03:33.330 align:middle line:90% don't want to do everything. 01:03:33.330 --> 01:03:36.460 align:middle line:84% But that small training set, we're going to process, 01:03:36.460 --> 01:03:39.360 align:middle line:84% and we're going to take that training set 01:03:39.360 --> 01:03:44.280 align:middle line:90% and write it out to train2.tab. 01:03:44.280 --> 01:03:45.240 align:middle line:90% No, we're not. 01:03:45.240 --> 01:03:48.040 align:middle line:84% You should do that because you have time to wait. 01:03:48.040 --> 01:03:50.230 align:middle line:90% But nobody wants to do that. 01:03:50.230 --> 01:03:53.130 align:middle line:84% We'll do the same thing with the testing_set.process 01:03:53.130 --> 01:03:55.830 align:middle line:90% and the testing_set. 01:03:55.830 --> 01:04:00.100 align:middle line:84% We're going to write that to test2.tab. 01:04:00.100 --> 01:04:02.430 align:middle line:84% I'm not doing that right now because they 01:04:02.430 --> 01:04:07.010 align:middle line:84% do in those old cooking shows, I've already done all this 01:04:07.010 --> 01:04:11.730 align:middle line:84% and I'm going to load it from what I've already done. 01:04:11.730 --> 01:04:15.320 align:middle line:90% So you know what I mean. 01:04:15.320 --> 01:04:16.820 align:middle line:84% The old cooking shows where they put 01:04:16.820 --> 01:04:20.460 align:middle line:84% they put the raw turkey in the oven, 01:04:20.460 --> 01:04:24.290 align:middle line:84% and then they just immediately go to the other oven 01:04:24.290 --> 01:04:25.610 align:middle line:90% and the turkey-- 01:04:25.610 --> 01:04:26.160 align:middle line:90% sorry. 01:04:26.160 --> 01:04:29.160 align:middle line:84% I you all can read my mind, right? 01:04:29.160 --> 01:04:32.070 align:middle line:90% 01:04:32.070 --> 01:04:35.370 align:middle line:84% That's what you pay me to do, make you read my mind, 01:04:35.370 --> 01:04:38.060 align:middle line:90% not actually teach. 01:04:38.060 --> 01:04:39.740 align:middle line:90% That's fantastic. 01:04:39.740 --> 01:04:40.990 align:middle line:90% OK. 01:04:40.990 --> 01:04:41.490 align:middle line:90% Yeah. 01:04:41.490 --> 01:04:42.180 align:middle line:90% Anal enough. 01:04:42.180 --> 01:04:48.650 align:middle line:84% I'll make this line up so I don't make any typos. 01:04:48.650 --> 01:04:50.930 align:middle line:90% It's on my desktop. 01:04:50.930 --> 01:04:53.070 align:middle line:90% Test2.tab. 01:04:53.070 --> 01:04:54.260 align:middle line:90% That's why it broke. 01:04:54.260 --> 01:04:55.200 align:middle line:90% OK. 01:04:55.200 --> 01:04:58.610 align:middle line:84% So again, we're going to do everything as before. 01:04:58.610 --> 01:05:04.960 align:middle line:84% Why do everything as before when you can just copy and paste. 01:05:04.960 --> 01:05:09.940 align:middle line:84% Because now we've already written everything out for-- 01:05:09.940 --> 01:05:12.820 align:middle line:84% yeah, this whole process and write thing 01:05:12.820 --> 01:05:16.936 align:middle line:84% takes care of everything we did before where we-- 01:05:16.936 --> 01:05:18.070 align:middle line:90% where is that? 01:05:18.070 --> 01:05:19.810 align:middle line:90% Where we had to define-- 01:05:19.810 --> 01:05:23.600 align:middle line:84% we had to open our datas, we had to define our columns, 01:05:23.600 --> 01:05:26.120 align:middle line:84% we had to write that out, define our descriptions, 01:05:26.120 --> 01:05:29.980 align:middle line:84% write out all our meta things, extract what our feature 01:05:29.980 --> 01:05:35.290 align:middle line:84% extractors do, all that stuff of these little process, 01:05:35.290 --> 01:05:38.390 align:middle line:84% and then write did all that for us. 01:05:38.390 --> 01:05:47.010 align:middle line:84% So now we're actually back onto feature extraction is done. 01:05:47.010 --> 01:05:49.570 align:middle line:90% 01:05:49.570 --> 01:05:53.800 align:middle line:90% Let's do machine learning. 01:05:53.800 --> 01:05:56.260 align:middle line:84% Anybody who tells you that the machine learning is 01:05:56.260 --> 01:05:59.830 align:middle line:84% the only hard part and that feature extraction is 01:05:59.830 --> 01:06:04.150 align:middle line:84% easy stuff is lying to you or doesn't know what they're 01:06:04.150 --> 01:06:09.160 align:middle line:84% talking about or whatever because feature extraction 01:06:09.160 --> 01:06:12.555 align:middle line:84% and doing it right is really, really hard. 01:06:12.555 --> 01:06:16.300 align:middle line:84% So we're going to basically copy and paste everything 01:06:16.300 --> 01:06:21.130 align:middle line:90% that we did before. 01:06:21.130 --> 01:06:22.030 align:middle line:90% Totals. 01:06:22.030 --> 01:06:24.850 align:middle line:90% Something like that. 01:06:24.850 --> 01:06:25.540 align:middle line:90% Great. 01:06:25.540 --> 01:06:30.070 align:middle line:84% We're going to run this again and then 01:06:30.070 --> 01:06:37.540 align:middle line:84% we're going to print everything afterwards. 01:06:37.540 --> 01:06:38.380 align:middle line:90% Great. 01:06:38.380 --> 01:06:42.400 align:middle line:84% And that's going to go through and it's 01:06:42.400 --> 01:06:44.740 align:middle line:90% going to print a couple things. 01:06:44.740 --> 01:06:45.450 align:middle line:90% How are we doing? 01:06:45.450 --> 01:06:49.250 align:middle line:90% 01:06:49.250 --> 01:06:53.510 align:middle line:84% We've given it a lot more things to train on, so a lot more data. 01:06:53.510 --> 01:06:55.180 align:middle line:84% These are much bigger files because 01:06:55.180 --> 01:06:58.580 align:middle line:84% of all the different feature extractors, so it takes a while. 01:06:58.580 --> 01:07:04.980 align:middle line:84% You can see in all this time, it's just done one row. 01:07:04.980 --> 01:07:05.800 align:middle line:90% Two rows. 01:07:05.800 --> 01:07:07.023 align:middle line:90% Are we going to wait for-- 01:07:07.023 --> 01:07:09.750 align:middle line:90% 01:07:09.750 --> 01:07:13.320 align:middle line:84% I think it's, like, an hour and a half. 01:07:13.320 --> 01:07:15.130 align:middle line:84% I ran this at 2:00 in the morning. 01:07:15.130 --> 01:07:17.640 align:middle line:84% Are we going to wait an hour and a half? 01:07:17.640 --> 01:07:18.520 align:middle line:90% We can all do that. 01:07:18.520 --> 01:07:19.053 align:middle line:90% No. 01:07:19.053 --> 01:07:21.220 align:middle line:84% I'm going to hit Stop for a second on the recording, 01:07:21.220 --> 01:07:24.990 align:middle line:84% and then I'm going to load up our results. 01:07:24.990 --> 01:07:25.690 align:middle line:90% OK. 01:07:25.690 --> 01:07:28.500 align:middle line:90% Here are the results I got. 01:07:28.500 --> 01:07:31.810 align:middle line:84% Again, our majority learner got 77% right. 01:07:31.810 --> 01:07:35.970 align:middle line:84% And our k nearest neighbors super intelligent artificial 01:07:35.970 --> 01:07:40.235 align:middle line:90% intelligence only got 4% better. 01:07:40.235 --> 01:07:41.610 align:middle line:84% I mean, I don't even know if that 01:07:41.610 --> 01:07:44.950 align:middle line:84% meets the margin of error in a political poll or something. 01:07:44.950 --> 01:07:49.990 align:middle line:84% So all that extra time didn't actually improve anything. 01:07:49.990 --> 01:07:54.330 align:middle line:84% What's important is always to choose your-- 01:07:54.330 --> 01:07:57.940 align:middle line:84% choose your feature extractors well. 01:07:57.940 --> 01:08:03.560 align:middle line:84% The fact that I didn't include that one simple extractor 01:08:03.560 --> 01:08:07.620 align:middle line:84% about the time signature made all the difference in the world, 01:08:07.620 --> 01:08:09.350 align:middle line:90% I think. 01:08:09.350 --> 01:08:10.070 align:middle line:90% OK. 01:08:10.070 --> 01:08:14.870 align:middle line:84% I just want to show that you can print 01:08:14.870 --> 01:08:18.260 align:middle line:84% all of the different features onto various charts and zoom in 01:08:18.260 --> 01:08:21.029 align:middle line:84% and stuff like that if people want to do that. 01:08:21.029 --> 01:08:24.140 align:middle line:84% But that'll be something for another time. 01:08:24.140 --> 01:08:27.859 align:middle line:84% This is what is available to you now. 01:08:27.859 --> 01:08:30.450 align:middle line:84% Maybe not this kind of low level one, 01:08:30.450 --> 01:08:35.810 align:middle line:84% but using these feature extractors, simple ones 01:08:35.810 --> 01:08:39.229 align:middle line:90% like sharps and meter and-- 01:08:39.229 --> 01:08:39.830 align:middle line:90% what was that? 01:08:39.830 --> 01:08:43.649 align:middle line:84% Percentage of compound time, range features, 01:08:43.649 --> 01:08:47.359 align:middle line:84% things like that, in order to classify pieces 01:08:47.359 --> 01:08:49.800 align:middle line:90% is incredibly powerful. 01:08:49.800 --> 01:08:53.090 align:middle line:84% And when put together with the right machine 01:08:53.090 --> 01:08:56.370 align:middle line:84% learning and artificial intelligence algorithm, 01:08:56.370 --> 01:08:59.109 align:middle line:84% there's a lot of possibilities for the future 01:08:59.109 --> 01:09:03.970 align:middle line:90% of computational music analysis. 01:09:03.970 --> 01:09:07.279 align:middle line:84% This lecture has run pretty darn long, 01:09:07.279 --> 01:09:12.933 align:middle line:84% so I'm going to say for a second little supplement on what 01:09:12.933 --> 01:09:14.350 align:middle line:84% are some of the things that I have 01:09:14.350 --> 01:09:18.700 align:middle line:84% learned through using computational artificial 01:09:18.700 --> 01:09:19.760 align:middle line:90% intelligence? 01:09:19.760 --> 01:09:23.020 align:middle line:84% What are some of the projects other people have done, 01:09:23.020 --> 01:09:29.060 align:middle line:84% and why the world state of the art black box deep learning 01:09:29.060 --> 01:09:31.870 align:middle line:84% or other sophisticated algorithms 01:09:31.870 --> 01:09:34.960 align:middle line:84% might not always be the best ones that we want 01:09:34.960 --> 01:09:37.960 align:middle line:84% to use when we're doing artificial intelligence 01:09:37.960 --> 01:09:42.800 align:middle line:84% and machine learning techniques on musical scores. 01:09:42.800 --> 01:09:44.720 align:middle line:84% But we'll leave that for another time. 01:09:44.720 --> 01:09:46.760 align:middle line:90% Thank you so much for listening. 01:09:46.760 --> 01:09:49.819 align:middle line:84% Go ahead and play around with some data sets. 01:09:49.819 --> 01:09:52.420 align:middle line:84% You've learned how to use corpus search to be 01:09:52.420 --> 01:09:55.240 align:middle line:90% able to find various pieces. 01:09:55.240 --> 01:10:00.580 align:middle line:84% there's a music21, I think user's guide chapter 01:10:00.580 --> 01:10:04.130 align:middle line:84% and tutorial on how to do more interesting things with that. 01:10:04.130 --> 01:10:08.020 align:middle line:84% There's the paper written with Chris Ariza and Lisa Friedland 01:10:08.020 --> 01:10:11.540 align:middle line:84% on using feature extractors with music21. 01:10:11.540 --> 01:10:16.870 align:middle line:84% And there's also some other music21 feature extractor papers 01:10:16.870 --> 01:10:20.770 align:middle line:84% that were written with MIT undergraduate UROPs on how 01:10:20.770 --> 01:10:24.340 align:middle line:84% we can use feature extraction to go beyond what we can do 01:10:24.340 --> 01:10:28.540 align:middle line:84% with midi, to really look at the nuance of musical scores 01:10:28.540 --> 01:10:30.080 align:middle line:90% in feature extraction. 01:10:30.080 --> 01:10:31.880 align:middle line:84% I'll be sending all those around. 01:10:31.880 --> 01:10:36.370 align:middle line:84% The meantime, please go back through the most interesting 01:10:36.370 --> 01:10:40.540 align:middle line:84% parts of this lecture to you, hit pause, and play around 01:10:40.540 --> 01:10:42.850 align:middle line:90% with your own data sets. 01:10:42.850 --> 01:10:48.270 align:middle line:84% Good luck in applying this to future projects. 01:10:48.270 --> 01:10:54.000 align:middle line:90%