WEBVTT 00:01.230 --> 00:01.710 Hey, everyone. 00:01.920 --> 00:08.790 So now we are going to play this in recurrent neural network concept on our same text classification 00:08.800 --> 00:14.210 dataset, which we have applied on this last section of spam Nixon. 00:14.760 --> 00:22.500 So same standard CSP file, we are going to use it and all those pre processing step will always remain 00:22.500 --> 00:22.890 the same. 00:22.950 --> 00:26.630 So I'm just going to execute the code till this particular part. 00:27.060 --> 00:33.120 And if you're not referred to my earlier videos of this, Pam Dixon with CNN, you can always go back 00:33.120 --> 00:39.900 and refer for the data importing part, data cleaning part and data processing, plus converting all 00:39.900 --> 00:47.130 your taxes into all kind of sequences because that is a basic pre processing step before building your 00:47.220 --> 00:47.550 model. 00:47.820 --> 00:52.100 So first, let me import spam dot CSB file. 00:53.900 --> 00:54.290 All right. 00:54.320 --> 00:56.330 So you can see sex is pretty important. 00:57.140 --> 01:02.030 And let me run all those necessary inputs plus pre processing step. 01:04.990 --> 01:09.250 Let me drop unnecessary all column renaming column. 01:10.360 --> 01:14.230 Converting your output video will have to zero in spam to one. 01:15.220 --> 01:21.430 Splitting your dataset named, we are going to convert this sentence to the sequences. 01:22.800 --> 01:28.440 This is for our vocabulary size and then we are going to pack this sequence. 01:29.600 --> 01:33.900 So we already explained this part earlier to CNN with you. 01:34.520 --> 01:39.980 And as we all take on exactly same dataset, I just didn't go through every single step. 01:40.520 --> 01:46.630 Now that we have a dataset available, Vitas, we are going to start building the model. 01:47.150 --> 01:52.580 And for that, we are going to use this Alistaire a long sought term memory. 01:52.820 --> 01:53.180 Martin. 01:54.750 --> 02:01.830 So if you go, Eliot, compared to the earlier case, we have imported the layers like an extra Eliot 02:02.670 --> 02:03.240 Alistair. 02:05.980 --> 02:11.060 In case of CNN, it was like a convolution one million, as we are dealing with the text that I feel 02:11.110 --> 02:14.040 dealing with the video retailer, a mistake. 02:14.110 --> 02:16.420 I hope to use this convolution truly. 02:17.270 --> 02:17.630 All right. 02:18.490 --> 02:20.980 So I would define the two variables. 02:21.280 --> 02:25.210 And that is called as in a kind of hyper parameter for this particular model. 02:25.840 --> 02:32.540 So first one is for the while, creating the word vector or embedding first live. 02:32.980 --> 02:34.100 What is the diamonds? 02:34.260 --> 02:38.350 And so we are converting every single token into printed. 02:38.370 --> 02:38.890 I'm insane. 02:39.010 --> 02:43.360 Racked up our hidden state rector will be, let's say, 50. 02:43.870 --> 02:45.730 And we'll see how we are going to use this. 02:45.890 --> 02:46.090 And. 02:47.420 --> 02:48.900 This is our first layer. 02:49.140 --> 02:50.740 So we define our booklet. 02:51.410 --> 02:52.530 And from input layer. 02:52.850 --> 02:53.900 The first ever layer. 02:54.440 --> 02:55.420 We are going to add it. 02:55.610 --> 02:56.150 That is nothing. 02:56.180 --> 02:57.440 But I am very glad. 02:57.760 --> 03:00.530 So now the story was almost similar. 03:00.890 --> 03:06.340 But now, instead of adding this convolution 1000000, we are going to add the Palestinian layer. 03:06.740 --> 03:09.690 Now, this particular less than layer has how many hidden state? 03:10.100 --> 03:11.330 So that will be a fitting. 03:11.840 --> 03:14.790 And I just kept the return sequence will be true. 03:15.230 --> 03:20.900 That indicates that we are not dealing here with the encoder decoder kind of representation. 03:21.620 --> 03:26.660 Instead of that, the moment something came after every first iteration. 03:27.080 --> 03:28.210 After every second reason. 03:28.400 --> 03:31.970 After the third iteration will preserve's those output. 03:32.630 --> 03:35.900 And one more layers like a global max pulling. 03:36.290 --> 03:43.200 So that is just trying to take the maximum out of those particular Vigneault headbang. 03:43.340 --> 03:44.130 Same legal layer. 03:44.210 --> 03:52.070 We have a Sigmoid Downstair VI segment because we are dealing here with the binary classification problem. 03:52.490 --> 03:58.220 And at the end, we have defined this model which accept the input and all those layers. 03:58.250 --> 04:02.560 So let me define this model object. 04:05.450 --> 04:08.970 So model, what define next is we need to compile this model. 04:09.050 --> 04:13.440 So just like earlier for optimization, we are going to use this Skydome optimizer. 04:13.770 --> 04:16.480 Lost will be binary and it's called cross entropy. 04:16.490 --> 04:18.770 Loss and metrics will be accuracy. 04:18.830 --> 04:25.000 So it will indicate then how could quite accurately our models while doing the testing on testing dataset? 04:25.630 --> 04:28.460 Let us fit our model with Tiny Poch. 04:32.290 --> 04:34.780 So training got started on Époque. 04:34.840 --> 04:40.050 Now, in this case of hot, and obviously it will take a good amount of time. 04:42.130 --> 04:49.990 Accuracy losses keep decreasing and aiding in the valuation accuracy is quite similar. 04:50.920 --> 04:52.240 It didn't increase too much. 04:53.050 --> 04:57.550 But when we tried to display this thing as a plot, we'll get much more idea. 04:57.970 --> 05:00.350 So let's see how lost got decrease. 05:01.210 --> 05:08.350 And you can see lost was instantly decreased from the first three percent to second increase in almost 05:08.800 --> 05:10.260 close to zero point thirty two. 05:10.300 --> 05:16.270 But then it didn't increase, whereas in case of relevation lost, it all was I mean, remains same 05:16.270 --> 05:18.790 throughout the our training process. 05:19.180 --> 05:20.650 And if you see the accuracy. 05:22.240 --> 05:23.930 Our accuracy is a. 05:25.160 --> 05:28.100 Not very much high compared to our CNN model. 05:28.640 --> 05:31.370 So close to around 86, six percent accuracy. 05:31.460 --> 05:35.520 We got, you know, first iteration only, but then it didn't increase. 05:35.960 --> 05:38.780 So what you can do, you can experiment with a different. 05:40.730 --> 05:45.030 Hyper parameters like a hand beating dimension and the other state. 05:45.140 --> 05:48.900 So just keep changing and try to see whether it will increase or not. 05:48.920 --> 05:53.630 Otherwise, this spam did extend, which CNN was very much trying for us. 05:54.050 --> 05:56.750 So that is about the demo of spam detection. 05:56.880 --> 05:59.320 We are seeing the next video.