WEBVTT 00:02.120 --> 00:03.030 All right, everyone. 00:03.480 --> 00:08.360 So in the last video, we have done all those preprocessing four of our texts in this project. 00:08.940 --> 00:11.120 Next thing we need to build a modern. 00:11.490 --> 00:16.170 So we are starting here like a sequential model and first layer. 00:16.320 --> 00:22.620 We are going to place like embedding layer where our input Deming's and will be of our vocabulary size. 00:22.860 --> 00:28.890 So every single token will be represented by a vector of size, full vocabulary size. 00:29.100 --> 00:38.210 So wherever the values will be, when that particular number will be represented by those token s put 00:38.210 --> 00:39.270 them and will be 50. 00:39.280 --> 00:45.040 So that indicates that that is like in a hyper parameter for this particular neural network, how you 00:45.040 --> 00:51.530 are going to embed your individual token into footy payments in an input line. 00:51.750 --> 00:52.740 So I a one shot. 00:54.140 --> 00:59.000 Before predicting your output, how many values you want to give me as a input? 00:59.040 --> 01:02.470 So a sequence length will be over in my head. 01:02.570 --> 01:05.890 Then I'm just providing to Alistair Lear. 01:06.620 --> 01:07.860 So that will be a first. 01:08.000 --> 01:13.860 Alistair Lever, having got a hundred units and then another hundred units, one more dance, Lear having 01:13.860 --> 01:14.460 100 unit. 01:14.570 --> 01:17.740 But in this case, the activists and I kept like a loop. 01:18.530 --> 01:21.130 And then there will be a one dance layer. 01:21.650 --> 01:28.700 But in this case, the unit will be vocabulary size because we are trying to predict one token. 01:28.880 --> 01:33.440 And as I told you, those token has been represented by a vocabulary size. 01:33.740 --> 01:44.060 So whatever that particular index number will be fired in output layer, we will predict that those 01:44.060 --> 01:47.530 particular token has been predicted and activist unit. 01:47.660 --> 01:55.400 We are predicting like a soft mix because then we are trying to predict the probability of each individual 01:56.210 --> 01:57.670 neuron in a quickly. 01:58.220 --> 02:00.830 So that is our basic model summary. 02:01.130 --> 02:04.610 And it has closed around two point one million parameters. 02:04.610 --> 02:10.910 So that's a too much and might be sometimes I observe that this column environment also didn't support 02:10.910 --> 02:15.380 in terms of AVAM or so, but somehow we managed to run it. 02:15.710 --> 02:20.720 It may happen that you won't be able to run this experiment successfully on a Google column so you can 02:21.110 --> 02:23.110 go with Hyrum on top of Ramsell. 02:23.720 --> 02:28.160 But many times we fail before execution of this code. 02:28.570 --> 02:28.870 All right. 02:28.880 --> 02:36.020 So complacent, arem, optimizer, categorical cross and Kobie, because outputs are multiclass classification 02:36.020 --> 02:41.000 kind of problem and how we are going to find the accuracy. 02:41.030 --> 02:41.750 So that will be a. 02:42.800 --> 02:45.590 Accuracy, Mazarin criteria will take into consideration. 02:46.190 --> 02:55.640 So let just run this input and output on hundred apoc and we'll take a bite size of two fifty six. 02:56.060 --> 02:57.820 So we ran it for 100 people. 02:58.730 --> 03:00.380 And let me go down. 03:02.490 --> 03:09.100 So then even after running under-report with such a big modern vegard accuracy close to around thirty 03:09.100 --> 03:12.110 four point zero percent only religious policy. 03:12.450 --> 03:12.620 Hey. 03:13.900 --> 03:15.580 Now the is created. 03:16.120 --> 03:19.900 Next thing is we need to test this model that how good our model is. 03:20.350 --> 03:22.510 So let's just take any random line. 03:23.020 --> 03:28.210 So from our original dataset, I have taken this particular random line and this particular line. 03:28.240 --> 03:30.430 I have taken it as a shape line. 03:31.740 --> 03:35.430 And I'll just created one function for the prediction function. 03:36.300 --> 03:38.340 What in this particular function will do? 03:38.550 --> 03:43.680 That will take into consideration this sick tax after the SIP tax. 03:43.710 --> 03:46.440 How many total votes or a token? 03:46.620 --> 03:49.500 You want to predict our tokenized? 03:49.560 --> 03:52.390 That is just because while prediction. 03:52.830 --> 03:58.880 Again, you need to convert those Tolkan into kind of numbers and whatever model. 03:59.050 --> 04:03.780 We hope you just now created tax sequence, Len. 04:06.590 --> 04:06.950 All right. 04:07.270 --> 04:14.680 So what we are trying to do, we are just reading over how many times we want to generate the next spread 04:14.680 --> 04:15.010 extent. 04:15.070 --> 04:20.510 So we want to do the Predix and for in and that's codewords now for simple ideas even here, like a 04:20.740 --> 04:21.040 10. 04:22.170 --> 04:23.430 Every day I take. 04:24.680 --> 04:34.040 Those sealed techs padding those sequences and then ready to output classis the moment they predict 04:34.040 --> 04:35.120 the output classes. 04:35.400 --> 04:42.080 I will just create another syntax, which will be a software version of your original syntax. 04:43.010 --> 04:48.860 Plus, your predicted output and from those predicted output, again, we're trying to predict the next 04:48.970 --> 04:49.200 one. 04:49.850 --> 04:57.830 And then if I just supply those things to my January tax sequence function with model tokenized, a 04:57.920 --> 05:00.770 sequence length sequence length will be 50. 05:00.860 --> 05:02.840 And sequence tax and 10. 05:03.200 --> 05:11.190 So based on this particular SEIP tax code of if I help now, I'm not good at how good goes. 05:11.380 --> 05:19.270 Grammar stuff is whatever it is been generated, but you can see it is a reasonably good accuracy. 05:19.760 --> 05:21.050 It looks like an English. 05:21.110 --> 05:26.540 It looks like any grammatical inglis to be steam pooling to choose a brace off. 05:26.930 --> 05:33.610 So based on this particular sig text, it has predicted this time future token. 05:34.190 --> 05:41.810 And that is all about the how you can do the text in recent with this l'estang recurrent neural network. 05:42.650 --> 05:43.040 All right. 05:43.520 --> 05:45.040 That is all about the text. 05:45.050 --> 05:45.530 And listen.