WEBVTT 00:01.940 --> 00:02.460 Here, everyone. 00:02.620 --> 00:09.950 So in the last video between our own way to work model and meeting Martin and we tested and somewhere 00:09.990 --> 00:16.330 we found when a good result or so, but now we are going to use this pre train model and this pre train 00:16.330 --> 00:18.370 model is trained by Google. 00:18.460 --> 00:20.140 So obviously you can think about that. 00:20.290 --> 00:26.680 The amount of data they use for the training purpose, that will be a nominal sum on top data. 00:27.010 --> 00:32.560 So in this model, we are just going to import it and we will see how we can. 00:34.000 --> 00:39.440 Logos model, and then we will based on our testing dataset. 00:40.050 --> 00:43.950 And let's see what results we'll be getting on exactly. 00:44.100 --> 00:45.030 Similar stuff. 00:48.150 --> 00:55.550 Let me clear all the outputs and let us start working on so import all those lively Jansing tenths of 00:55.550 --> 00:55.820 law. 00:56.940 --> 00:59.640 One more is a Kagan library. 01:00.390 --> 01:04.080 We need to make the Kaggle directly because same thing. 01:04.080 --> 01:06.270 We are going to download it from the Kegl only. 01:06.720 --> 01:10.500 Let me go to this link and you will get idea. 01:11.530 --> 01:14.740 Next, things we need to applaud this careful dog. 01:15.520 --> 01:18.690 Now, as we are dealing with a different notebook. 01:19.030 --> 01:20.680 Every time we have to do this thing. 01:21.430 --> 01:23.310 So Kagle Dargis on got uploaded. 01:23.830 --> 01:24.910 Let me just copy. 01:24.950 --> 01:27.470 Do not kegl for that. 01:29.250 --> 01:33.850 And let me be severely and then let just download this data set. 01:34.520 --> 01:35.780 Now it's a little bigger. 01:35.850 --> 01:41.450 Martin almost drowned PTB, so it will take a little bit amount of time. 01:42.230 --> 01:43.940 So I'm just fast forwarding my video. 01:43.970 --> 01:47.300 Tilda's download will complete. 01:51.710 --> 01:57.920 All right, so you can see model is successfully downloaded and it's almost around PGD, so it took 01:57.920 --> 01:58.490 a little time. 01:59.060 --> 02:00.860 Next phase, we need to inject this model. 02:00.950 --> 02:05.240 Now, unzipping definitely is going to take a good amount of time. 02:06.340 --> 02:07.590 So I ended up in the file. 02:09.990 --> 02:12.900 We can even go to this contain. 02:15.160 --> 02:20.640 Let me go one step up and instead of containers, start extracting all those things. 02:23.750 --> 02:28.870 This particular bin file, we hope, downloading Jim File, VEO downloaded. 02:28.910 --> 02:31.270 We are expecting into this bin fine. 02:32.420 --> 02:33.330 Let me collapse. 02:36.630 --> 02:42.540 All right, so now you can see model is completely unzip and it is available. 02:43.450 --> 02:46.710 Think, you know, many people file for Mix Bender Dodgier. 02:46.890 --> 02:48.070 And again, it will be me. 02:48.660 --> 02:50.070 So it's a quite big model. 02:51.350 --> 02:55.750 Next is we need to load this model, and for that we are going to use this vector. 02:56.660 --> 02:59.330 We are limited to one Lack's reactor only. 02:59.330 --> 03:00.160 We are going to load it. 03:00.170 --> 03:01.910 Otherwise it will be too much heavy. 03:02.420 --> 03:05.840 So let us lower fully model inside of it. 03:06.090 --> 03:07.130 Modern variable. 03:07.760 --> 03:11.360 And we are not doing any kind of model building. 03:11.390 --> 03:19.020 We are just loading the existing retrain model and that is being pre trained by the global organization. 03:19.610 --> 03:23.540 And that is Google News based on the news data available. 03:24.000 --> 03:25.340 And they train this model. 03:25.860 --> 03:28.160 But like, let's go with the. 03:29.470 --> 03:35.650 Man, so let me execute it and it will give us the vector representational man. 03:36.650 --> 03:40.460 Now you can see the size of this vector is much, much bigger. 03:40.970 --> 03:45.530 If you try to find, let's say, land function. 03:48.960 --> 03:52.290 Earlier case, we had a just 32 diamond set. 03:52.380 --> 03:58.430 But this Google retrain model, that is a 300 dimension vector. 03:58.460 --> 04:01.790 So every single token in this case will be presented by. 04:02.880 --> 04:04.570 Three hundred different numbers. 04:04.960 --> 04:07.360 Now let's try to find the same thing. 04:07.570 --> 04:14.140 Like a most similar with this pre train modern and we'll try to compare what results we will get. 04:14.620 --> 04:21.280 So in this case, also, we got the woman, but the boys first one compared to earlier, we had a couple 04:21.550 --> 04:22.720 girls and then boy. 04:23.380 --> 04:25.990 So a little better riding man is related to women. 04:26.170 --> 04:27.730 And then Boyd and teenager again. 04:28.340 --> 04:31.100 That's OK, but good results. 04:31.120 --> 04:31.570 We are right. 04:32.200 --> 04:36.880 But the real test will come when we test on this king minus man. 04:37.120 --> 04:37.910 Plus remain. 04:38.530 --> 04:39.780 So that is electric magic. 04:40.420 --> 04:42.460 And we got the king. 04:42.460 --> 04:45.610 So first when we got the king, but then we have a queen. 04:45.940 --> 04:48.730 So that's quite embracing and quite good results we got. 04:49.120 --> 04:56.660 So as the queen is to man, woman as queen, let's drive with the Germany capital, Berlin. 04:56.710 --> 05:01.930 So Paris is the capital of France and we got, yes, France. 05:01.940 --> 05:03.820 So in case those of you go to France. 05:03.830 --> 05:06.730 So what this model also, it works fine. 05:07.090 --> 05:10.570 But this particular case, we got a very good, quite impressive results. 05:11.050 --> 05:13.190 Let me go with the last one. 05:13.210 --> 05:15.400 Massey football and cricket. 05:15.460 --> 05:16.930 Now we have a good amount of data. 05:17.500 --> 05:20.020 So obviously, first one we automatically because. 05:21.110 --> 05:27.720 This particular actor is quite close to Massey, but the second one you ought to consider like a sidewalk. 05:27.830 --> 05:30.250 So Sela Tendulkar, Dravid. 05:30.730 --> 05:38.540 They needed the like a king or I would say the gall of this cricket tournament or a cricket sport sameway 05:38.600 --> 05:41.920 like a Massee in case of football kind of goes haywire. 05:42.020 --> 05:43.070 In case of cricket. 05:43.340 --> 05:49.010 So those of you don't know about cricket, I'm just getting idea that these two guys are the God of 05:49.010 --> 05:49.820 the cricket. 05:51.220 --> 05:57.640 Spots like that, then the conclusion is that this preprint model looks very much based, even though. 05:57.700 --> 06:02.830 And one more thing I want to highlight that you just reported, one lacks reactor only. 06:02.950 --> 06:07.720 But there are many, many more vectors is embedded inside this pretty model. 06:08.290 --> 06:13.270 So that is all about how to use this retrain model for the embedding purpose. 06:13.630 --> 06:14.860 See you in the next video.