WEBVTT 00:01.220 --> 00:08.630 High in this session, we will be working with neural network implementation, so let us begin with 00:08.630 --> 00:09.130 the food. 00:09.500 --> 00:15.390 So here we have the import statement. 00:15.480 --> 00:21.800 So first of all, we will import Binder's as we can also import numpties and be. 00:26.680 --> 00:33.430 And after doing the board, we will work on the details of this particular dataset, which I have selected 00:33.730 --> 00:43.210 is a data set of spying, and this will basically help us find out the details of who should be having 00:43.210 --> 00:46.860 our lower back pain and who should not be having a lower back pain. 00:47.140 --> 00:49.790 So it has data related to the same. 00:50.020 --> 00:56.440 So we will import data and read the data by using the CSP function. 00:56.680 --> 01:02.200 And the data underscores fine is the data frame which has been activated. 01:02.470 --> 01:04.930 Now let us have a look at the data set. 01:05.080 --> 01:10.070 So this data set has three hundred and ten rows and 14 columns. 01:10.300 --> 01:18.340 Now we don't really know the levels of the columns, so these it has around 12 columns. 01:18.520 --> 01:26.320 Apart from that, it has class attribute, which shows values like if it is abnormal or normal, and 01:26.320 --> 01:32.580 there is a last column which is unnamed, which does not have much details about that. 01:32.830 --> 01:37.520 And it is saying that the prediction is done by using binary classification. 01:37.780 --> 01:39.550 So now let us go for the. 01:41.400 --> 01:49.440 So from this fine dataset, we will remove the following one and unnamed column 13. 01:49.770 --> 01:55.110 So this column one and column named 13 will be removed. 01:55.770 --> 02:00.230 Now we have data from underscored spine. 02:00.480 --> 02:09.150 So this will have all the details from column two to column number 12, along with the class attribute, 02:09.390 --> 02:12.990 which is declasse Liebl, actually, which we want to predict. 02:13.230 --> 02:18.400 So here we want to predict if the spine is normal or abnormal. 02:19.230 --> 02:24.750 So the first thing that we will have to do is we will have to convert this into a dummy variable where 02:24.870 --> 02:28.470 we will change normal and abnormal to zero and one. 02:29.490 --> 02:34.090 So all the data which we have is numeric in nature. 02:34.320 --> 02:42.630 So all we will have to identify here is if there is any specific correlation is present in this particular 02:42.630 --> 02:43.200 dataset. 02:43.470 --> 02:49.000 So for that, again, we have already stated what other methods you can apply. 02:49.020 --> 02:50.970 You can use find us profiling. 02:51.180 --> 02:55.330 You can use VIFF calculator, you can use correlation matrix. 02:55.330 --> 03:03.720 So any of those could be used for the if you want to use you can use feature importance from random 03:03.720 --> 03:04.380 forest. 03:04.500 --> 03:08.550 So there are well established methods which we have discussed. 03:08.700 --> 03:12.090 So you can use any of those methods which you want. 03:13.750 --> 03:18.110 So now we have this data set and now it has 12 columns. 03:18.430 --> 03:20.970 Now these are the headers which we have. 03:21.160 --> 03:24.810 So we already have the headers for these data. 03:24.910 --> 03:29.910 So we will apply the headers by doing the underscores my DOT columns. 03:30.460 --> 03:35.140 So and then we will assign the header stored until we get the header values. 03:36.280 --> 03:45.850 So it does pelvic incidents, pelvic tilt, then different films related to biology are present. 03:45.870 --> 03:48.370 And finally, the class which you want to predict. 03:49.380 --> 03:55.710 The next thing which we will be doing is we will be checking if there is any data which is not present 03:55.710 --> 03:56.670 or is not. 03:56.910 --> 04:00.950 So we are taking the data from the phone records. 04:00.960 --> 04:01.370 Fine. 04:01.800 --> 04:06.630 We are checking if there are any rules where the spine is not. 04:06.870 --> 04:13.350 And we are checking if there is any of these in axis equal, the one that doesn't any column which has 04:13.620 --> 04:17.210 null value in it and we are getting the go offered so it gives out. 04:17.630 --> 04:20.610 So there are no null values in this particular dataset. 04:21.850 --> 04:29.110 So now we will check the data types so you can easily see that all the data is flawed, so we need not 04:29.110 --> 04:30.320 do anything about it. 04:30.550 --> 04:35.080 The only thing which we need to take care of is the class. 04:35.470 --> 04:44.530 So we will import preprocessing and apply preprocessing dot label in order and then transform the class 04:44.830 --> 04:45.730 column here. 04:46.750 --> 04:54.740 So we can see the top then samples, so we are doing the sample and getting the values out of it. 04:55.000 --> 05:00.890 So this is one function which will allow us to take the values in a sample form. 05:01.150 --> 05:09.190 So here you can see the glasses have me created and you can see the glasses which have been generated 05:09.190 --> 05:09.510 here. 05:10.150 --> 05:12.930 So the class values have changed to one and zero. 05:14.800 --> 05:22.030 Now, let us find out the correlation value so we can simply find out the correlation using dot seawater 05:22.300 --> 05:23.330 for correlation. 05:23.590 --> 05:29.250 So these are the values for correlation and we can find out the correlation values. 05:29.590 --> 05:33.530 So let us plot these correlation values. 05:33.560 --> 05:40.350 So here you can see there are no light colored correlation values, which is fine. 05:40.360 --> 05:43.230 So we don't have any highly correlated values. 05:44.350 --> 05:49.300 If we look at the like that and so here we can see there are certain. 05:50.560 --> 05:56.120 Values which are going towards minus four, which is, again, not a critical value. 05:56.260 --> 05:58.470 So we are going to go here. 05:58.480 --> 06:06.010 We don't have any highly correlated problems present so we can go ahead without deleting any particular 06:06.010 --> 06:06.630 columns. 06:07.060 --> 06:10.720 So we are importing the three best split now. 06:11.050 --> 06:15.100 So we have to get the X and Y data frame separated. 06:15.340 --> 06:22.260 So X date of frame will be separated by the ifs fine from the class column and invite. 06:22.270 --> 06:24.760 We will keep only the class column. 06:26.460 --> 06:34.950 Now we will check the data again, so we have all the columns except for the last column, so let us 06:34.950 --> 06:40.830 check that if it is fine so you can see that the glass column is missing from here. 06:40.840 --> 06:42.210 So it is perfectly fine. 06:42.480 --> 06:49.290 And in Divided of the Glass column, should be the only one percent which is rightly shown here. 06:49.680 --> 06:51.350 Now we will split the data. 06:51.360 --> 06:54.000 So we split the data using the split. 06:54.210 --> 06:59.300 So we provide the X and Y data frame and deepest slice, which we want. 06:59.460 --> 07:06.480 We want that size to be on the zero point too, that this 20 percent, because we don't have enough 07:06.480 --> 07:07.430 rules of data. 07:07.620 --> 07:12.540 So we would like to have a majority share of our data for testing. 07:12.900 --> 07:20.080 Now, one thing to note here is, let's see, we have a very large amount of data. 07:20.430 --> 07:22.740 Let's see if I have an. 07:23.690 --> 07:32.170 One hundred thousand rows of data then I could have guessed even a very low amount of data for this, 07:32.400 --> 07:39.440 but maybe 10 percent or maybe five percent because I want to bring my model 50. 07:39.480 --> 07:40.670 So that is the target. 07:40.680 --> 07:46.100 If the data size is very huge, we can reduce the size of testing data. 07:46.320 --> 07:51.120 And again, if the testing if the data size is very small. 07:52.460 --> 07:59.360 Then again, we will like to reduce the size of our testing data and keep more of the data for training 07:59.370 --> 08:07.640 focus, if the size is medium for the data said, you can keep around a ratio of 70, 30 or seventy 08:07.640 --> 08:07.970 five. 08:07.970 --> 08:08.560 Twenty five. 08:08.570 --> 08:10.050 That is completely up to you. 08:10.280 --> 08:18.850 So just make sure that you have enough data points to to train, also enough data points to test it 08:18.860 --> 08:19.340 also. 08:21.480 --> 08:29.900 So we split the data, so after splitting, we get extreme X test by train and rightest. 08:31.020 --> 08:39.450 So now we will import the multileveled Perceptron now this multiplayer Perceptron is nothing but the 08:39.450 --> 08:41.520 neural network which we will be using. 08:42.270 --> 08:49.420 Is it given the name MLT because it is simply multilayered neural network. 08:49.440 --> 08:54.470 So it is just another name which is present for multilayered neural network, which is multilayered 08:54.480 --> 08:55.350 Perceptron. 08:55.740 --> 09:01.570 So Perceptron is of neural network within which all nodes are connected with each other. 09:01.770 --> 09:06.800 That is why we are considering this as a multilayered Perceptron. 09:06.900 --> 09:15.410 If we were supposed to do some changes, which will be a part of a higher topic under the word opaque, 09:15.570 --> 09:22.230 so we can take that up little and like those could be studied later, but that is out of the scope of 09:22.230 --> 09:22.890 this course. 09:23.070 --> 09:30.000 So for now, what you can understand this is that MLP, that this multimeter percent is nothing but 09:30.540 --> 09:33.060 a fully connected neural network. 09:35.250 --> 09:44.040 So in Escalon Metric, we are the driving, the musical and from selection, we are picking up the randomizer 09:44.040 --> 09:46.330 TV because they just want to try. 09:46.350 --> 09:47.950 Only a few combinations. 09:47.970 --> 09:48.150 You. 09:50.480 --> 09:58.790 So I'm giving the barometer's so the parameters are great if I want to have different variants in my 09:58.790 --> 09:59.490 learning rate. 09:59.840 --> 10:07.730 So these are if I want the learning to be constant, if I want my learning, they be changing. 10:07.970 --> 10:14.000 And if I want by learning to be actually adaptive to how my learning is going. 10:14.000 --> 10:15.510 So I choose adaptive for that. 10:15.920 --> 10:18.340 These are the hidden layer sizes. 10:18.560 --> 10:22.070 What this means is this is a couple which I'm giving it. 10:23.170 --> 10:31.180 So this means that the forest here, the left will have five nodes, the next Hitler will have their 10:31.180 --> 10:34.330 nodes and the next to the left will have five. 10:36.190 --> 10:40.490 So how will this look like so let me show this to you. 10:41.080 --> 10:43.270 So it will look something like this. 10:43.360 --> 10:44.800 Apologies for the drawing. 10:45.350 --> 10:55.390 It will have 12 in less than 12 input points, one output point, then five nodes in the layer one, 10:55.870 --> 11:00.390 then nodes in the layer two and five nodes in the layer three. 11:00.550 --> 11:05.110 And all of these points, all of these points will be connected to these points. 11:05.470 --> 11:11.110 The all of layer one point will be connected with the layer to point. 11:11.410 --> 11:16.300 All the layer to points will be connected with the layer three points and so on. 11:18.090 --> 11:20.800 Now, let us go further with the training. 11:20.820 --> 11:25.650 So these are different features, so we have four values. 11:25.650 --> 11:28.320 So these are different values which we will be having. 11:28.890 --> 11:36.330 This is the activation functions which we are having logistic bendu and then we adjust to creating the 11:36.330 --> 11:43.560 object of my dealer, Perceptron, and creating an object of randomizer TV in which we have provided 11:43.560 --> 11:50.430 the model name, the number of models we want to select out of these then the cross-validation which 11:50.430 --> 11:51.390 we want to have. 11:52.890 --> 11:58.950 I'm the scoring method, which you want to use, then we have done random search dog search, so it 11:58.950 --> 12:03.320 is fitting five candidates out of all the combinations that are present here. 12:03.670 --> 12:10.400 You can run a grid search also so that you get an extensive result, a better result out of this. 12:10.920 --> 12:14.390 So here you can see these are the results which I have got. 12:14.400 --> 12:15.980 Let me get the best estimate. 12:16.440 --> 12:24.360 So my best estimate has the activation value logistic and for value zero point one that says auto B.W. 12:24.360 --> 12:25.450 one zero point nine. 12:25.740 --> 12:27.110 And these are different values. 12:27.360 --> 12:30.860 So for this, I will simply read the report. 12:30.870 --> 12:37.740 So the best model has accuracy level of zero point eight seven six. 12:39.130 --> 12:50.050 And this what is Elby If Jesus is adaptive, Hitler is 21, Martin, and Alpha value is zero point one, 12:50.050 --> 12:52.390 activation for action is logistic. 12:54.620 --> 13:02.780 So you can see that this is the best performing model and the worst one is zero point six nine percentage 13:02.780 --> 13:10.220 atkins', so now I will simply get the details with the model again and I will make the predictions 13:10.220 --> 13:13.700 and show you the predictions in the dorms of crosstab. 13:14.820 --> 13:20.550 So this is the cross matrix for this, so you can see. 13:21.770 --> 13:30.260 These are the confusion matrix deal, so you can see that Ford predicted values and actual value one. 13:31.820 --> 13:36.090 This is forward and these are different values which have been theater. 13:36.440 --> 13:40.000 Now let us have a look at the classification report. 13:40.030 --> 13:48.650 So here the precision for a precision is zero point nine legal is zero point seventy three point seven 13:48.650 --> 13:49.030 five. 13:49.490 --> 13:53.280 And if one score is zero point eight four point sixty. 13:53.600 --> 14:00.170 Now, as you can see, the precision value is zero point nine zero and for one it is zero point five 14:00.170 --> 14:00.470 five. 14:00.480 --> 14:05.380 So it is biased towards one zero plus zero. 14:05.390 --> 14:11.620 So you can change the model and train it again so that you can get a better result. 14:11.720 --> 14:17.240 So for that, probably you can use Garozzo TV so that obtain better results from it. 14:17.630 --> 14:25.610 So this is just an implementation which is not really fine tuned and in case of neural networks. 14:26.640 --> 14:34.090 You will have to try different combinations of different layers and see how it works out for you. 14:34.380 --> 14:37.280 So that is the process which we have for neural networks. 14:37.290 --> 14:42.630 You will have to try different combinations and find out the one which works the best. 14:44.450 --> 14:44.960 Thank you.