WEBVTT 00:01.790 --> 00:07.700 In this session, we will discuss about boosting, so let us have a look at this particular. 00:08.050 --> 00:10.960 So here I have imported the packages. 00:10.970 --> 00:17.180 So the first package which I have imported is Mumfie, then Fondas, then McGlothlin, then Seabourne. 00:17.540 --> 00:19.520 And I've imported oil. 00:20.510 --> 00:25.070 Next, I have ignored the warnings so that I don't get extra warnings to. 00:26.040 --> 00:34.250 Here, I've loaded the data set, so my data safe is house prices and train and house prices underscored 00:34.260 --> 00:39.870 this dataset and here I am checking the shape of the data using the chip function. 00:43.980 --> 00:51.420 So training data has one thousand four hundred and sixty rolls, while testing data has one thousand 00:51.420 --> 00:57.780 fifty nine rows, four hundred fifty nine rows, that retail training data has eighty one columns and 00:57.780 --> 00:59.970 testing data has only eight columns. 01:00.270 --> 01:05.500 This will actually help us to find out which is the feature, which we want to predict here. 01:06.270 --> 01:15.090 So I have created a set and removed the training test column from the print column sick. 01:15.510 --> 01:18.830 I could have used all the difference also. 01:19.440 --> 01:22.700 So that would also have given us the value of the missing column. 01:23.990 --> 01:31.490 So the value comes out to be see price, so we have to predict these sales price values, so because 01:31.490 --> 01:33.140 we are predicting sales price. 01:33.320 --> 01:37.270 So it is clear that here we want to solve a regression problem. 01:38.090 --> 01:43.010 So we will be using extra boost repressor for this particular problem. 01:43.550 --> 01:48.500 So whenever you are used to solving a regression problem, you will use it agressor. 01:48.740 --> 01:54.260 And whenever you're solving a classification problem, you will use to classify it. 01:54.830 --> 02:03.530 Each and every library in Ashkelon provides the details of the classified and agressor, the corresponding 02:03.530 --> 02:09.740 hyper parameters and all the details about them and all the things will be in sync with what we have 02:09.740 --> 02:11.010 discussed in the session. 02:11.210 --> 02:14.770 So there should not be any problems in exploring these photo. 02:16.200 --> 02:23.370 So let us have a look at this, so we are checking the prediction column, this particular column, 02:23.370 --> 02:25.270 Visages Seed's price column. 02:25.680 --> 02:28.220 So we have applied describer on top of it. 02:28.470 --> 02:33.510 So here you can see that the minimum value is thirty four thousand nine hundred. 02:33.810 --> 02:37.040 Maximum value is seven fifty five thousand. 02:37.560 --> 02:44.430 Here we have twelve twenty five percentile, which is twelve thousand one twenty nine thousand. 02:46.670 --> 02:53.030 Then we have 60 percentile that is desired, 16 one like sixty three thousand seventy five percentile 02:53.030 --> 02:56.710 at twenty one thousand two fourteen thousand. 02:57.260 --> 03:02.640 So you can see that the maximum value has a great gap and so does the minimum value. 03:02.990 --> 03:08.570 So it shows that there are certain outlier values in the cities, prices which are also present. 03:10.070 --> 03:21.350 Now we will have we will prepare certain plots, so the fullest plot is of the living area, the ground 03:21.350 --> 03:25.460 living area of in comparison with those cities, places. 03:25.970 --> 03:35.260 So here we can see that the living area has seen house prices are having a positive relationship. 03:35.690 --> 03:41.050 And as the living increases decrease, prices also increase. 03:41.510 --> 03:47.690 And after a point of time, the there are these are a few outlier points. 03:49.370 --> 03:56.690 And out of these boys, all the boys look in sync with the relationship, but there are these two data 03:56.690 --> 04:04.220 points which actually do not follow the relationship between where the living area is very high, but 04:04.220 --> 04:06.450 still the price has been set very low. 04:06.800 --> 04:12.620 So this is something which is unusual in nature and might actually impact the performance. 04:12.770 --> 04:20.570 So we might decide to get rid of these two data points because these will actually impact the performance 04:20.570 --> 04:25.550 of the model and will try to divide this model towards these two data point. 04:26.970 --> 04:34.640 While here, you can see for a given area, the prices here are very high. 04:36.080 --> 04:43.500 And here again for the these prices are very high in comparison to the usual prices which are there. 04:43.790 --> 04:46.370 So these are, again, certain outlier values. 04:47.270 --> 04:50.770 Next, we will have a look at data preprocessing. 04:51.050 --> 04:55.160 So first of all, we will replace all the numeric missing values. 04:55.380 --> 04:59.990 That is not a number with the zero and non numeric with none. 05:00.890 --> 05:07.200 Then we will create dummy variables for the categorical features and transform the Spearwood diplomatic 05:07.220 --> 05:11.380 features by taking the log of feature plus one. 05:11.390 --> 05:17.120 This will make the features more normal in nature because here you can see that the values are Sadikov. 05:17.150 --> 05:22.380 So we want to remove Desco from it so we can take a log and do that. 05:22.820 --> 05:27.700 So first of all, we are filling on not a number in the values. 05:27.980 --> 05:38.780 So for the training ID you can see we are creating or train ID and from the train dataset we are dropping 05:38.780 --> 05:46.250 the column IDA because it is not the important column and it will not combine the prices as we have 05:46.250 --> 05:53.560 already discussed, that application numbers, IDs, roll numbers have no impact on the predicted variable, 05:53.570 --> 05:55.140 which we are trying to find out. 05:55.640 --> 05:58.780 Next, we are removing the column from the first dataset. 05:59.540 --> 06:07.570 So if you remember, we have done this kind of transformation only a also, but then we combined the 06:07.610 --> 06:09.440 cleaning and testing dataset. 06:09.860 --> 06:11.840 So this is a different way of doing it. 06:11.870 --> 06:16.240 You can either choose that particular way or this way here. 06:16.250 --> 06:22.670 You will have to be more careful about making the conversions while in that scenario you would simply 06:22.670 --> 06:26.420 make the combined the data, do the conversion and simply split the data. 06:26.570 --> 06:30.300 So there is no chance of any discrepancy. 06:30.560 --> 06:34.550 So it is completely up to you which method you are using. 06:36.620 --> 06:40.820 So here I have removed I am from the testing data. 06:42.150 --> 06:50.220 Now from the object column, so I'm creating an object list of object columns by getting the training 06:50.850 --> 06:58.440 data from the columns, I'm taking the data type object so we can also we could have also done select 06:58.440 --> 06:59.010 the type. 06:59.040 --> 07:00.310 So this is another matter. 07:00.510 --> 07:03.730 So there are several methods of doing the same thing. 07:03.960 --> 07:09.840 It is completely up to you will which particular method you are more comfortable with so you can use 07:09.840 --> 07:10.890 whatever you like. 07:10.890 --> 07:16.750 And the main target is how we retrieve the the data and how we process the data. 07:16.950 --> 07:20.920 If the processing is going fine and we are getting the desired. 07:20.940 --> 07:26.020 Then it does not matter if you are using select the type or you are filtering these out. 07:26.280 --> 07:29.340 So these are simple or different methods which we are implementing. 07:29.670 --> 07:32.070 So you can choose whatever you like from these. 07:33.420 --> 07:40.050 So then here I am selecting the Nonobjective, so here I am selecting the unknown object, which will 07:40.050 --> 07:46.500 be either the in type or a float type, so all will be selected in this particular category. 07:46.800 --> 07:54.610 So I am getting the twin columns, which are not objects into the numerical column list from the testing 07:54.630 --> 07:59.760 dataset, also on the training data, but now in the object column. 08:01.010 --> 08:09.950 Whenever we have null values, so wherever the values are not a number, I am basically filtering out 08:09.950 --> 08:13.990 the green object or dataset with the object column. 08:14.010 --> 08:21.320 So it is checking all the objects column and wherever we have not a number, it will fill it with none. 08:21.950 --> 08:23.840 Same thing in the best data set. 08:23.840 --> 08:31.610 Whenever we have an object column and we have not a number, we will fill none now in the null value, 08:31.610 --> 08:34.930 we will replace the null values with zero. 08:35.150 --> 08:40.840 So whenever we have null values, we will replace it with zero in the numeric data type. 08:42.100 --> 08:48.980 And they have renamed the date frames to planes zero one zero zero zero zero one. 08:49.780 --> 08:52.710 Now, we have failed the values. 08:52.930 --> 08:55.960 So now we've we've what the ordinal features. 08:55.960 --> 08:58.580 So here are certain ordinal features. 08:58.840 --> 09:02.350 So we have ordinal features as well. 09:02.650 --> 09:09.520 Second, basement quality, basement condition, basement exposure, basement final day. 09:11.680 --> 09:17.700 We've been finishing the basement, finishing day to see garage conditions, so these are different 09:17.950 --> 09:22.540 details of the house and these are have ordinary values. 09:22.780 --> 09:28.330 So for this again, I'm splitting the data and the. 09:29.500 --> 09:38.300 And do X and Y, so from brain zero zero one, I am dropping the prediction column that this is a science 09:38.400 --> 09:47.110 price, I'm creating my X column and similarly from Brain 002, I'm creating by removing, by keeping 09:47.110 --> 09:51.220 only the Sears price column that is F.P. predictable. 09:52.060 --> 09:59.860 Now I'm converting the columns, so I'm generating an object of the one Hawtin Khuda. 10:00.070 --> 10:04.120 So in this we are giving the column name as ordinal feature. 10:04.150 --> 10:11.170 So all the ordinary features will be put into this and these ordinary features will be converted into 10:11.800 --> 10:16.780 the ordinary columns using this one quarter. 10:17.020 --> 10:23.860 So we will simply be the dot vanguard C one not this is the object which we have created and then we 10:23.860 --> 10:25.420 are doing to transform. 10:25.690 --> 10:30.950 So that is we are footing the columns also and then transforming at the same point of time. 10:33.050 --> 10:37.880 So we have transformed 13 zero zero zero two zero zero two. 10:38.780 --> 10:42.530 Similarly, we will be transforming the best zero zero one. 10:43.370 --> 10:44.720 That is the testing data. 10:46.900 --> 10:51.400 And we have created three zero zero three out of it on this 003 11. 10:52.440 --> 11:00.330 Now we are getting the nominal feature, so we will convert the nominal features, so nominal features, 11:01.020 --> 11:02.760 four x four. 11:02.790 --> 11:05.780 We are basically checking for X in object columns. 11:06.000 --> 11:11.790 So we are checking whatever object columns are present and removing all the ordinal features from these 11:11.790 --> 11:12.740 object columns. 11:13.080 --> 11:17.470 So these are the nominal features, which is also known as the categorical features. 11:17.700 --> 11:20.470 So now here have the we have the categorical features. 11:20.490 --> 11:27.570 So one way of converting the categorical features we have seen was using and dot where we have used 11:27.900 --> 11:28.770 as type. 11:28.780 --> 11:38.580 Also, we have also used this condition for converting the dummy variable and we can also use get the 11:38.580 --> 11:39.380 means function. 11:39.600 --> 11:44.940 And here is another method of converting categorical features into. 11:46.530 --> 11:52.200 The one thing that is using the label in the. 11:55.490 --> 12:03.320 So here you can see we have the training and testing data from the original form, so here you can see 12:03.320 --> 12:08.660 different values which are present in the testing and training data, and you can see the numerical 12:08.660 --> 12:09.650 scenes place state. 12:11.510 --> 12:13.490 And here, these place is missing. 12:14.710 --> 12:20.710 Now we have them, the transformation's, so after transformation's, there will be several other changes 12:20.710 --> 12:21.520 in the values. 12:21.730 --> 12:24.160 So I will show you those one by one. 12:26.440 --> 12:34.090 So here we can see the ordinals features, so let us have a look at Extoll call an extension. 12:39.840 --> 12:40.770 Or. 12:48.500 --> 12:55.640 Here you can see the you see value has been changed, emetic and similarly other values have been changed 12:55.640 --> 12:56.600 to new type. 12:57.680 --> 13:04.310 For the let us have a look at the phenomenon of baby boomers that this categorical variables, so what 13:04.310 --> 13:11.390 we are doing here is we are importing the label encoder and using the label and we are looping on the 13:11.390 --> 13:16.910 different nominal features president in this particular dataset. 13:17.240 --> 13:23.150 And we are applying the label in order to transform or transform one of these. 13:24.730 --> 13:31.870 And after that, we are getting the dummy variables for these nominal features by using we don't get 13:31.870 --> 13:35.460 dummies and we are dropping the first one from these. 13:36.040 --> 13:42.250 So let us have a look at the data center, which has been generated so you can see we have all the columns. 13:43.760 --> 13:47.390 Which have been converted into numerical columns. 13:49.590 --> 13:55.010 And here you can see all the columns have been converted into one hot including. 13:58.050 --> 14:04.050 So this is how all the values have been changed now for those, we will apply a fee to engineering. 14:04.080 --> 14:12.690 So here we are generating new columns because we will have to create new columns to find out new relationships. 14:13.140 --> 14:20.580 So here we are creating a column as a total board, which will have a combination of full board hoverboard 14:20.680 --> 14:22.740 and basement full board and basement hub. 14:22.770 --> 14:32.910 But then we are going to work on creating a complete or total Baulch column, which has a sum of all 14:32.910 --> 14:36.540 the open porch, different type of porch and deck. 14:38.190 --> 14:48.300 Then we are checking if there has been a recent remodel by subtracting the years old from the remodel 14:48.300 --> 14:48.680 at. 14:50.010 --> 14:58.280 And similarly, we are checking if the house is new or not or if the house has a pool or not, then 14:58.290 --> 15:02.780 if the house has a fireplace, has second floor, his garage has basements. 15:02.780 --> 15:08.820 So these are different conditions which we have created using the lambda functions and comforters and 15:08.820 --> 15:09.450 operators. 15:11.190 --> 15:19.670 Next, we will be fitting the sort of the model, first of all, let us divide the in to prevent a split. 15:20.070 --> 15:24.600 So we are dividing the data into X and Y volumes. 15:24.930 --> 15:34.270 We are getting the X values in two three zero zero five by dropping the C splice column and then by 15:34.320 --> 15:36.100 keeping only the Sainsbury's column. 15:36.300 --> 15:42.230 After that, we are splitting the data with the testing size as zero point to five. 15:42.240 --> 15:47.100 So we are keeping twenty five percent of data in two of the. 15:49.120 --> 15:50.140 Thing said and. 15:51.150 --> 15:53.310 Seventy five percent did in decorating. 15:54.480 --> 16:03.130 Next, we are importing the packages, so here we are importing the extubate agressor and Igby orif 16:03.130 --> 16:12.480 regressive and the grid search c.v and we're importing the mean, absolute better and mean square error 16:12.480 --> 16:13.930 for the matrix. 16:14.520 --> 16:17.430 So these are the default parameter values. 16:17.580 --> 16:24.480 So the default parameters are maxed out as three learning rate, as one number of estimate, those as 16:24.520 --> 16:32.070 hundred Mobilicity as one would be basically defines if we want certain messages to be printed or not, 16:32.280 --> 16:35.380 which will show you the progress of the learning. 16:35.550 --> 16:42.720 So in case you want to have a view of the progress of the learning by training the model, you can increase 16:42.720 --> 16:44.190 the value of Bilbo's D. 16:44.670 --> 16:48.680 I usually define it as to be the silent level. 16:48.690 --> 16:56.010 None of this actually means that if you want to have the message is silent, then objective is the. 16:57.470 --> 17:04.100 I get that you want to have that is we want to apply a regression and the regression method is squared 17:04.100 --> 17:09.350 error, we want to find out this squared away from this particular model. 17:09.650 --> 17:16.520 Now, similarly, if you are applying classification, then for classification, there will be a different 17:16.850 --> 17:18.980 type of barometer. 17:19.050 --> 17:26.360 So for you that you will have another objective function, which you can take from the classification 17:28.280 --> 17:31.880 documentation which we have in the in the library. 17:32.240 --> 17:39.020 So booster is the geometry, which means we want to have a gradient boosting tree and job is the number 17:39.020 --> 17:40.540 of traits which you want to have. 17:40.550 --> 17:46.160 So it was always suggested to have in jobs as more than one. 17:46.160 --> 17:53.950 If you are working on a multiple computer or if you are working on, let's say, a GBU or Onwubiko lab. 17:54.260 --> 17:58.040 So on those platforms you can have more number of jobs running. 17:58.220 --> 18:05.840 But usually when we are working on our Lupo's systems and using Jupiter more for training, then increasing 18:05.840 --> 18:08.380 in jobs might not be beneficial enough. 18:08.930 --> 18:10.540 So it is completely up to you. 18:10.820 --> 18:15.320 You can try increasing in job if actually it works out on your system. 18:15.320 --> 18:20.600 Otherwise, in other foster systems with processors, it would work faster. 18:21.350 --> 18:22.820 The more value, the better. 18:22.820 --> 18:24.290 It will run faster. 18:24.360 --> 18:29.090 Then then Intrade is the number of trades you are trying to run together. 18:29.480 --> 18:31.260 Guerma is the gamma value. 18:31.280 --> 18:33.320 Minimum wage is one maximum. 18:33.320 --> 18:36.890 Leptis Zeitels subsample one column sample by three one. 18:37.250 --> 18:40.820 So these are different values which we are keeping as the. 18:41.980 --> 18:49.630 Default parameters, so there are certain important parameters that is booster, the type of booster 18:49.630 --> 18:55.930 which we want to use, that is the small model which we want to use for this particular regression. 18:56.260 --> 18:57.820 So there are two options. 18:57.820 --> 18:59.050 That is geometry. 18:59.050 --> 19:07.090 That is, if you want to have a three based model or a linear model of Kloner, then the end is the 19:07.090 --> 19:10.370 default number of the maximum number of trades available. 19:10.540 --> 19:16.630 So if we don't define Intrade, then it will by default to the maximum number available and then choose 19:16.630 --> 19:21.010 that objective defines the loss function to be more minimized. 19:21.430 --> 19:28.300 Then we have subsample going on by three, which is to provide the randomness and estimate the is the 19:28.300 --> 19:35.230 number of threes which we will be generating or the number of sequential Vitullo knows we will be generating 19:35.230 --> 19:36.430 for this particular. 19:38.310 --> 19:45.030 Then we have low grade, maximum depth of the three, then the minimum wage, that is a minimum some 19:45.030 --> 19:47.460 of the above all observations required in the. 19:48.000 --> 19:50.580 So these are different type of barometers. 19:50.760 --> 19:56.880 So for during these high bhavana meters, we will use a grid, such CV. 19:57.210 --> 20:00.410 So Ingrid TV again, we have the meter. 20:00.450 --> 20:06.600 That is the model which we will be using Vadum grid, which will be these parameters listed out, which 20:06.600 --> 20:12.240 we will be selecting from scoring, is the scoring method which we will be using in job is the number 20:12.240 --> 20:18.720 of jobs which we will be running in parallel then CVS's the cross-validation which we will be using. 20:18.730 --> 20:21.110 We have been using this for a long time now. 20:22.240 --> 20:28.900 So these are different barometers, which we have so far from using these bottom feeders, you can find 20:28.900 --> 20:31.200 out the number of models which will be generated. 20:31.450 --> 20:34.260 So if you want, you can use Garito TV also. 20:34.510 --> 20:36.730 And if you want, you can use random fortressed. 20:36.730 --> 20:39.280 Also, it is completely up to you. 20:41.260 --> 20:49.630 So we will simply do grid search, dog food and then find out the best barometers from the best barometers, 20:49.630 --> 20:54.620 we will find the best model again based on the best barometers which we have been. 20:55.060 --> 20:58.020 And finally, we will check the value. 20:58.040 --> 21:05.020 So the mean absolute error comes out of fifteen thousand three hundred and thirty a thirty seven, which 21:05.020 --> 21:10.740 is the price difference, which is present out of the oil prices. 21:11.050 --> 21:14.840 So we are plotting the result of between different values. 21:15.100 --> 21:16.830 So here is the plot. 21:17.020 --> 21:23.110 So the blue ones are the real values and red ones are the predicted values. 21:23.350 --> 21:33.430 So you can see that it has predicted the values pretty nicely, but only has defaulted in those two 21:33.430 --> 21:34.820 points which were coming here. 21:34.840 --> 21:39.970 So one of that has already been actually predicted nicely, but that is one value which has not been 21:39.970 --> 21:40.960 predicted correctly. 21:41.500 --> 21:44.260 Then we have predicted result down submission. 21:44.290 --> 21:52.930 So if you want to submit to the predictions or send out these submissions to any other person in the 21:52.930 --> 22:00.700 team, or if you want to export the predictions so you can simply use or predict to predict the values 22:00.700 --> 22:06.820 and then simply send it to the door to KVI method to convert it into a certified. 22:08.270 --> 22:16.220 Now, there is one more thing which I want to capture here, that is by training your model while running 22:16.220 --> 22:23.740 this great TV, when we are training with these small number of barometers, it is completely fine. 22:24.260 --> 22:31.140 But when we are training with a huge number of barometer's, then it might become difficult for us. 22:31.730 --> 22:40.880 So what we can do with that is no particular level of hyper barometer that this hypovolemic that is 22:40.880 --> 22:44.480 more important or this hyper badme that is more important. 22:44.870 --> 22:49.760 And there is nothing like which hyperbola meter has to be reinforced or lead. 22:50.570 --> 22:54.950 So keeping this in mind, what you can do is let us see. 22:54.950 --> 22:57.290 I will not train these. 22:59.390 --> 23:02.900 Hyper barometer's, so I just command these. 23:04.020 --> 23:04.560 And. 23:07.660 --> 23:14.830 I can simply provide the objective function and I can provide the investigators, let us keep the investigators 23:14.830 --> 23:16.210 focused and keep the learning. 23:16.990 --> 23:25.600 So what I can do is I can provide the investigators and the objective function for us and give several 23:25.600 --> 23:32.800 other values for the investigators so I can give seven hundred here and thousand here and fifteen hundred 23:32.800 --> 23:33.100 here. 23:36.300 --> 23:39.870 And then I can run this particular dorning. 23:41.370 --> 23:48.510 Once I've done this, I will get to know the values for an estimated and objective function, using 23:48.510 --> 23:56.530 my best model, using the best barometer, I will be able to obtain the best values for an estimated 23:56.580 --> 23:57.840 and objective function. 23:58.380 --> 24:02.100 So then what I can do is I can simply comment this. 24:03.670 --> 24:11.770 Or instead of commenting this, I can simply go ahead and remove all of these values and keep only that 24:11.770 --> 24:14.920 one value, which has given me the best result. 24:15.310 --> 24:18.220 So let's say I get rid of this entire thing. 24:21.170 --> 24:28.790 And I keep only led to save my best estimate that came out to be the one with five hundred meters so 24:28.790 --> 24:30.200 I could put five hundred. 24:31.160 --> 24:40.390 Now, once I have fixed to these values, then I can again bring my A model by changing different learning 24:40.400 --> 24:48.020 rate so I can test a different learning rate and then find out the best value of the learning and then 24:48.020 --> 24:49.140 updated here. 24:49.160 --> 24:52.000 So let's say it comes out to be say zero point five. 24:52.310 --> 24:53.530 So I fix the learning. 24:54.140 --> 25:01.850 Now I will once I have fixed the learning rate, then again I will train this great sword and I will 25:01.850 --> 25:04.020 train in four different max depth values. 25:04.940 --> 25:08.320 Now, using this, I will be able to find out. 25:08.330 --> 25:11.090 One makes the max depth, which is working best for me. 25:11.480 --> 25:15.200 Then again, I can figure this out and put only one value here. 25:15.560 --> 25:21.270 And once I fix the value of max depth, I'm learning Grayden and an estimate, an objective function. 25:21.500 --> 25:25.430 Now I can go to the next photometer and then print. 25:26.060 --> 25:27.680 Now why do I do this? 25:27.710 --> 25:29.990 Why have I told you to do this? 25:30.260 --> 25:35.660 Is because it actually helps us in reducing the training thing. 25:36.290 --> 25:43.140 Because let us consider we are training only these many we only these many barometers. 25:43.880 --> 25:46.130 So what will happen is. 25:48.400 --> 25:57.580 It will actually cause a problem for me because I will have too many parameters to do and every time 25:57.580 --> 26:00.700 I the barometer here it. 26:01.590 --> 26:05.590 Actually increases the number of models exponentially. 26:05.970 --> 26:11.240 So let us consider the number of Morbius, let me do a little calculation for you. 26:11.460 --> 26:18.240 So let's see, I had zero point one year and zero point zero one zero point. 26:20.850 --> 26:23.180 And is 0.01 here? 26:26.700 --> 26:33.000 So these were the different values, so what are the combinations I have I have these values. 26:34.520 --> 26:36.770 So for these, I have. 26:41.170 --> 26:41.620 To. 26:43.050 --> 26:44.190 Cross for. 26:45.750 --> 26:46.770 Across three. 26:48.120 --> 26:49.140 Cross to. 26:50.780 --> 26:52.070 Cross St.. 26:53.330 --> 26:55.490 Clause six. 26:57.470 --> 26:58.070 Values. 26:58.100 --> 27:00.710 So this comes out to be five seventy six. 27:01.860 --> 27:06.310 And if you will notice, I am creating a hundred models here. 27:06.630 --> 27:08.810 I am creating 200 models here. 27:09.430 --> 27:18.900 So fifteen hundred models being created for each one of this in 215 hundred, if it is generating only 27:18.900 --> 27:21.660 fifteen hundred models are not generating all of these. 27:23.030 --> 27:29.580 Only this much then if I'll just get rid of this so you can see for only this much I am creating. 27:31.230 --> 27:38.760 One forty four thousand Modern's, so this is such a large number of models, which I am reading, but 27:38.880 --> 27:43.140 if you will be using the method I just told you, what you will be doing is. 27:44.120 --> 27:46.190 You are just training for. 27:47.500 --> 27:53.290 In the very first go, it will take some time because now you are training for a more number of values, 27:53.290 --> 27:57.970 so you forced you will be competing between one hundred five hundred and thousand, literally. 27:57.970 --> 28:01.360 You will just compare between one hundred five hundred ten thousand. 28:01.750 --> 28:09.400 Then you imbued values for to wonder if the value comes out to be, let us say five hundred, then you 28:09.400 --> 28:14.140 can change one hundred and thousand two hundred and seven hundred and see what values comes up. 28:14.440 --> 28:20.040 If it comes up as two hundred then you can find out another value between two hundred and five and that 28:20.050 --> 28:20.960 is how you can decide. 28:21.580 --> 28:28.870 Now once I have found out the estimate, then what will happen is I will be tuning for lesson number 28:28.870 --> 28:29.380 of values. 28:29.380 --> 28:38.440 So I will be training only, let's say 10 or 15 models act one thing, which will be a very huge reduction 28:38.440 --> 28:40.260 in the model training thing. 28:40.840 --> 28:47.980 So instead of planning a lot of barometer's together, you can use these sequential approach which will 28:47.980 --> 28:54.350 allow you to be will only fuel barometer's at one point of time and you will be winning them. 28:54.820 --> 29:01.330 Now, one more benefit which you will be getting by using this mechanism is that you can actually. 29:03.560 --> 29:10.310 Get the result out of it, sometimes what happens is that are training with so many barometer's, so 29:10.310 --> 29:19.670 what happens is that somehow the the running, it just gets stuck or it hands the computer or something 29:19.670 --> 29:23.210 like that happened because not everyone is using BU's. 29:23.210 --> 29:23.550 Right. 29:23.810 --> 29:29.450 So in that case, it will take a lot of training time and sometimes it may happen that you will get 29:29.450 --> 29:35.780 an error and all the things which you have done, it will be completely lost, but then you will be 29:35.780 --> 29:38.270 training with only one photometer. 29:38.270 --> 29:40.410 I think it will be running faster. 29:40.610 --> 29:47.660 So after like a seven or 15, 20 minutes or 30 minutes, you will have a desire for at least INDISTINCTIVE. 29:48.080 --> 29:50.520 Then after that you will run for the learning date. 29:50.990 --> 29:59.000 So after fixing the number of an estimated Pacifics fix the estimate or two hundred now I will I will 29:59.000 --> 29:59.900 train for learning. 29:59.940 --> 30:00.200 Great. 30:00.200 --> 30:06.050 And I will decide, OK, these are the learning rates and after training I get a particular value of 30:06.050 --> 30:06.640 learning grade. 30:06.650 --> 30:08.030 So I will fix that. 30:08.030 --> 30:09.140 Learning great value. 30:09.530 --> 30:16.220 And now after 15, 20 or 30 minutes, I will be getting the results for my next learning barometer. 30:16.610 --> 30:22.610 So this is how I will eventually keep on getting the values of my hyperbola meters instead of losing 30:22.610 --> 30:29.660 out on a lot of time and again losing out of my world because of some particular reason, maybe the 30:29.660 --> 30:36.250 laptop was unplugged, maybe something was not going well, maybe the laptop restarted, or somehow 30:36.250 --> 30:42.800 all of this got interrupted or something just happened, or maybe the controller stopped this one doing 30:43.130 --> 30:47.030 so because of all these reasons, the world can actually get lost. 30:47.300 --> 30:49.010 So we don't want that to happen. 30:49.010 --> 30:52.280 So we can use the sequential approach of what you can simply do is. 30:53.490 --> 30:58.340 First of all, government everything keep only fuel barometer. 30:58.410 --> 31:01.000 So these are the two Hyppönen we designed Duni. 31:01.410 --> 31:07.410 I have fuel for these high barometer's using my guess the best badam. 31:07.770 --> 31:09.990 I will get the best value for these. 31:09.990 --> 31:12.640 Let's say the best value comes out to be 500. 31:13.380 --> 31:16.000 So I will change this to five hundred. 31:16.230 --> 31:19.180 Now I know an estimated value has to be five hundred. 31:19.410 --> 31:27.570 Now I will bring this and add certain values to learning by and we do see a zero point zero two or zero 31:27.570 --> 31:30.860 point zero five and zero point zero seven. 31:31.920 --> 31:37.140 So I have these values which I want to buy right now out of all of these values. 31:37.140 --> 31:39.940 One, I all of these values I really get to get. 31:39.960 --> 31:42.690 This is one value which is giving me the best results. 31:43.140 --> 31:48.950 So I will keep only that one particular value and now I will bring for my next hypovolemic. 31:49.440 --> 31:53.220 So now I will check which maximum depth is giving me the best result. 31:53.700 --> 32:01.200 So again, then I will select the best Maxted and then go ahead with tuning the next model so you can 32:01.200 --> 32:05.460 iteratively keep doing this until and unless you get the best hypovolemic. 32:05.910 --> 32:12.270 And this will also show you if this learning, this barometer tuning is actually improving the model 32:12.270 --> 32:19.080 or not, because if that changes, which you are getting in the model improvement are not enough, then 32:19.080 --> 32:22.760 probably you don't really need to do the immediately anymore. 32:23.700 --> 32:29.450 So it will actually tell you that exactly if if the model is improving by this tuning or not. 32:29.820 --> 32:37.650 So you can actually judge those things without running that extreme board for 10 or 20 hours and then 32:37.650 --> 32:40.170 getting the conclusion that it did not really. 32:41.040 --> 32:45.290 So you can use this particular method and it is a very useful method. 32:45.300 --> 32:51.540 I have been using it for a very long time and it really helps a lot in reducing the training thing. 32:51.930 --> 32:58.200 So please try this particular method as it does a very useful and very efficient method of training. 33:00.310 --> 33:03.920 So that's it for extra boost. 33:04.360 --> 33:12.910 So next, we will be learning of the next model that is, I think, Canada's name, but we will be taking 33:12.910 --> 33:14.700 up the nearest neighbor next. 33:15.190 --> 33:15.660 Thank you.