1 00:00:00,170 --> 00:00:01,070 At this point. 2 00:00:01,100 --> 00:00:09,530 One question you may ask yourself is we are having this, um, outputs which are being generated from 3 00:00:09,530 --> 00:00:13,520 the making use of this Z vector right here. 4 00:00:14,450 --> 00:00:20,300 But how does the Z vector this Z vector here? 5 00:00:21,260 --> 00:00:27,860 Of the different images vary with the particular digits in those images. 6 00:00:27,860 --> 00:00:36,320 So how does a vector in say let's take this for example in six vary with that of zero uh, with two, 7 00:00:36,530 --> 00:00:39,080 uh, and all other digits. 8 00:00:40,480 --> 00:00:50,620 To experiment with this would copy out this code we had already and then get back here and now try to 9 00:00:50,620 --> 00:00:58,000 pass in the different digits into our encoder and then plot out the positions of. 10 00:00:59,840 --> 00:01:10,220 Z vectors and then see whether there is some correlation with the Z vector position and the input we 11 00:01:10,220 --> 00:01:11,010 have here. 12 00:01:11,030 --> 00:01:13,490 Write your other specific input digit. 13 00:01:14,510 --> 00:01:17,840 That said, we're going to now make use of this. 14 00:01:17,840 --> 00:01:19,460 So we have that. 15 00:01:19,460 --> 00:01:23,000 And then here we're going to have our Y train. 16 00:01:23,000 --> 00:01:30,050 So this time around we'll be needing the Y train because we need to know exactly what digit we're dealing 17 00:01:30,050 --> 00:01:33,950 with, unlike here, where we just needed to have the input images. 18 00:01:34,160 --> 00:01:35,380 Okay, so that's it. 19 00:01:35,390 --> 00:01:43,400 We have X train, y train, we're going to reload this and then we do the usual pre-processing. 20 00:01:43,550 --> 00:01:51,110 From here, we run the cell and then we move straight away to pass in this images in our encoder. 21 00:01:52,320 --> 00:01:54,480 Now similar to what we had here. 22 00:01:54,510 --> 00:02:01,080 Now we're going to be making use of our layers and then picking out the first index. 23 00:02:01,080 --> 00:02:04,650 So here we have our output. 24 00:02:04,680 --> 00:02:12,270 We could call this or basically we have Z, then we have the mean and the variance and then this is 25 00:02:12,270 --> 00:02:13,070 equal. 26 00:02:13,080 --> 00:02:14,790 Let's copy this here. 27 00:02:15,420 --> 00:02:17,400 Um, there we go. 28 00:02:17,430 --> 00:02:18,860 We have that. 29 00:02:18,870 --> 00:02:20,630 Then this is now one. 30 00:02:20,640 --> 00:02:21,270 See that? 31 00:02:21,270 --> 00:02:26,310 And then from here we pass in this x train. 32 00:02:26,730 --> 00:02:27,660 So that's it. 33 00:02:27,690 --> 00:02:34,320 We pass in the X train and then we could go ahead to start with the plotting, uh, plot figure. 34 00:02:34,350 --> 00:02:36,720 Specify the figure size. 35 00:02:37,020 --> 00:02:41,220 Figure size equals 12 by 12. 36 00:02:41,550 --> 00:02:42,600 There we go. 37 00:02:42,600 --> 00:02:45,300 And then we do a scatter plot. 38 00:02:45,330 --> 00:02:51,450 Now, the scatter plot will show us the different positions of the values of Z. 39 00:02:51,480 --> 00:02:53,020 That's the. 40 00:02:55,810 --> 00:02:59,530 Z points in a two dimensional space. 41 00:02:59,680 --> 00:03:00,740 Let's take this off. 42 00:03:00,760 --> 00:03:02,860 This different Z points. 43 00:03:02,860 --> 00:03:07,090 And then also we're going to color this such that. 44 00:03:08,200 --> 00:03:10,690 Digits belonging to the same level. 45 00:03:10,690 --> 00:03:12,430 Have the same color. 46 00:03:13,720 --> 00:03:14,620 There we go. 47 00:03:14,620 --> 00:03:20,500 We're going to plot this as our X and then our Y. 48 00:03:20,740 --> 00:03:23,110 So we have that. 49 00:03:23,110 --> 00:03:28,990 And then now we specify this Y train as our labels. 50 00:03:28,990 --> 00:03:30,100 So that's it. 51 00:03:30,100 --> 00:03:37,090 And then we can now go ahead to plot out this color bar and show. 52 00:03:37,660 --> 00:03:38,770 Let's run that. 53 00:03:39,460 --> 00:03:41,470 And here's what we get. 54 00:03:41,710 --> 00:03:42,970 As you could see. 55 00:03:44,210 --> 00:03:49,250 The labels here go from zero right up to nine. 56 00:03:49,520 --> 00:03:53,990 And then you could notice this different clusters we have in here. 57 00:03:54,170 --> 00:03:56,030 See these different clusters? 58 00:03:56,270 --> 00:03:57,080 See that? 59 00:03:57,830 --> 00:03:59,030 This shows that. 60 00:04:00,510 --> 00:04:10,140 The encoder part of our variational auto encoder has been trained such that now it's able to generate 61 00:04:10,140 --> 00:04:12,990 values of Z where. 62 00:04:13,880 --> 00:04:15,260 This point. 63 00:04:17,660 --> 00:04:27,830 Is output at such that two inputs belonging to the same digit will be closer to each other as compared 64 00:04:27,830 --> 00:04:31,850 to two inputs belonging to two different digits. 65 00:04:31,880 --> 00:04:40,940 So what we've been doing so far is we've been training the variational autoencoder model with our own 66 00:04:40,940 --> 00:04:43,130 custom training block. 67 00:04:43,160 --> 00:04:51,310 You see right here, we make use of this gradient tape and we carry out the gradient descent manually 68 00:04:51,320 --> 00:04:58,100 as compared to when you just have, say, model fit and then you pass in your train data set. 69 00:04:58,100 --> 00:05:03,250 And basically, uh, this takes care of doing the training. 70 00:05:03,260 --> 00:05:11,510 Now the advantage of working with this fit method is that it's quite simple to use, but uh, compared 71 00:05:11,510 --> 00:05:17,160 to this custom training loops, it doesn't give you that much freedom. 72 00:05:17,160 --> 00:05:25,380 And that's why we're making use of this, um, gradient tape here to train our model. 73 00:05:25,650 --> 00:05:33,570 Now, with TensorFlow, it's possible to actually get the best of both worlds. 74 00:05:33,600 --> 00:05:43,110 Now, this means that we can be able to train or get this custom training right here and then still 75 00:05:43,110 --> 00:05:45,660 make use of the fit method. 76 00:05:45,780 --> 00:05:55,470 The way this is done is by overriding the trainstep method of our model, which we shall define. 77 00:05:55,770 --> 00:06:01,950 So that said, let's get right here and define this VA model. 78 00:06:01,980 --> 00:06:09,180 Then we go ahead and define this model as we would do with a usual Keras model. 79 00:06:09,720 --> 00:06:18,060 But just as we had done right here, that's in this model we had built here, we need to take into consideration 80 00:06:18,230 --> 00:06:25,810 in our take us argument here the encoder model and the decoder model. 81 00:06:25,820 --> 00:06:26,750 So that's it. 82 00:06:26,780 --> 00:06:35,960 Now, once we have this two models we have here our encoder, which is our encoder model, and we do 83 00:06:35,960 --> 00:06:37,610 the same for the decoder. 84 00:06:38,000 --> 00:06:42,050 Now, once we're done with this decoder, let's type this out. 85 00:06:42,080 --> 00:06:48,710 We have your decoder, Uh, as we're saying, once we're done with this decoder, the next thing we 86 00:06:48,710 --> 00:06:51,740 want to do is define some loss. 87 00:06:51,740 --> 00:06:54,080 So here we have our loss. 88 00:06:54,230 --> 00:06:56,830 Um, it's actually a loss tracker. 89 00:06:56,840 --> 00:07:00,710 Now, we'll get to understand better why we're calling this a loss tracker. 90 00:07:00,710 --> 00:07:05,780 So here we have the average or the mean of the different losses which you shall get. 91 00:07:06,890 --> 00:07:13,460 And then from here, we'll go ahead to this main section where we shall override the Trainstep method. 92 00:07:13,460 --> 00:07:20,570 So, uh, just like, let's add this here, just like we used to working with methods like model, like 93 00:07:20,570 --> 00:07:24,050 the compile method, like the fit method. 94 00:07:24,050 --> 00:07:32,180 Here we have this Trainstep method, which is called each time we call on this fit method right here. 95 00:07:33,120 --> 00:07:40,200 But since we are now going to make use of the default, um, fit or the default trainstep method, we'll 96 00:07:40,200 --> 00:07:48,000 have to override this and write out our own training block, which is basically this block we have here. 97 00:07:48,000 --> 00:07:55,350 So what we could do here is copy this out so we could simply copy this out and then paste it out here. 98 00:07:55,380 --> 00:07:57,890 Now let's just copy all this. 99 00:07:57,900 --> 00:07:58,890 There we go. 100 00:07:58,890 --> 00:08:01,770 And, uh, that's it. 101 00:08:01,800 --> 00:08:02,880 We paste this out. 102 00:08:02,880 --> 00:08:04,860 So here we have our training block. 103 00:08:04,860 --> 00:08:07,650 We'll call this, um, trainstep. 104 00:08:07,650 --> 00:08:14,250 So we override this Trainstep method, which is a keras model method. 105 00:08:14,430 --> 00:08:16,890 So here we have our trainstep. 106 00:08:16,890 --> 00:08:19,170 We're taking in our batch. 107 00:08:19,470 --> 00:08:20,760 Uh, there we go. 108 00:08:20,760 --> 00:08:21,930 And then. 109 00:08:23,130 --> 00:08:26,670 Is going to be similar to what we've had already in the training block. 110 00:08:26,670 --> 00:08:33,180 So you see here we have our Z mean log VAR, which is gotten from the encoder model. 111 00:08:33,180 --> 00:08:37,710 And from here we pass Z and then we obtain Y pred. 112 00:08:37,740 --> 00:08:38,820 We have our Y. 113 00:08:38,850 --> 00:08:40,500 True, we obtain our loss. 114 00:08:40,500 --> 00:08:44,970 We've already defined our custom loss method and that's it. 115 00:08:45,920 --> 00:08:48,920 Then from here, we're going to update our loss metric. 116 00:08:48,920 --> 00:08:54,730 So here we have, um, loss tracker, which we defined already here. 117 00:08:54,740 --> 00:08:56,690 We're going to update its state. 118 00:08:56,690 --> 00:09:02,630 So we update our loss state with the value of this loss. 119 00:09:02,960 --> 00:09:03,830 So that's it. 120 00:09:03,980 --> 00:09:07,970 Now we are going to return the result. 121 00:09:07,970 --> 00:09:17,450 So we have this dictionary, we have loss and then we have our loss tracker dot result. 122 00:09:17,450 --> 00:09:18,170 See that? 123 00:09:18,170 --> 00:09:19,040 We have that. 124 00:09:19,040 --> 00:09:20,780 So there we go. 125 00:09:20,780 --> 00:09:24,350 We have updated our loss, we output the loss. 126 00:09:25,220 --> 00:09:31,640 But one important method we need to call again is the metrics method. 127 00:09:31,640 --> 00:09:33,770 So here we have metrics. 128 00:09:34,280 --> 00:09:44,720 You could define several metrics, but here we just have this loss and then we return our loss tracker. 129 00:09:44,720 --> 00:09:49,140 So we have loss tracker returned. 130 00:09:49,170 --> 00:09:52,050 Now notice how this is actually a list. 131 00:09:52,050 --> 00:09:58,170 So you see, this is a list because we could have, um, say different losses which we could return. 132 00:09:58,170 --> 00:10:04,560 And then we have the property decorator which will place right here. 133 00:10:04,560 --> 00:10:12,930 So once we have this set, you see, we do not need anymore to make use of this or to write this as, 134 00:10:13,080 --> 00:10:14,130 as it was reading. 135 00:10:14,130 --> 00:10:20,910 And so this means that we, we are not going to be writing out, for example, training start for epoch. 136 00:10:20,940 --> 00:10:24,330 This, uh, the training loss is this and all of that. 137 00:10:24,330 --> 00:10:28,740 Now all we need to do is just call on our model.fit method. 138 00:10:28,740 --> 00:10:33,480 So right here, let's do, uh, let's run this first. 139 00:10:33,480 --> 00:10:35,700 Let's run this right here. 140 00:10:36,300 --> 00:10:37,320 There we go. 141 00:10:37,320 --> 00:10:39,240 And then we go ahead. 142 00:10:39,240 --> 00:10:46,100 We're getting an error Now from here we define our var. 143 00:10:46,100 --> 00:10:51,500 So it's quite similar to to what we had already seen here. 144 00:10:51,680 --> 00:10:52,370 See that? 145 00:10:52,370 --> 00:10:57,530 Now we have this Var, which is now this which we've created here. 146 00:10:57,530 --> 00:11:07,940 So let's have that equal var E, and then it takes in the encoder model, encoder model and the decoder 147 00:11:07,940 --> 00:11:08,780 model. 148 00:11:08,780 --> 00:11:09,500 That's it. 149 00:11:09,500 --> 00:11:13,010 So once we have this now we could do just as we are used to doing. 150 00:11:13,010 --> 00:11:15,710 Let's, let's even say model, let's just say model. 151 00:11:15,710 --> 00:11:23,360 And then here we have model.compile as usual, and then here we're going to pass in our optimizer. 152 00:11:23,360 --> 00:11:26,000 So here we have Optimizer. 153 00:11:26,240 --> 00:11:29,300 Uh, guess we define this optimizer already. 154 00:11:29,300 --> 00:11:30,590 So let's check here. 155 00:11:30,590 --> 00:11:32,150 Yeah, we have defined this optimizer. 156 00:11:32,150 --> 00:11:41,540 So here we just have optimizer, um, equal Optimizer, which we've defined already, and then we do 157 00:11:41,690 --> 00:11:43,580 our model.fit. 158 00:11:43,610 --> 00:11:44,390 See that? 159 00:11:44,390 --> 00:11:45,350 So that's all. 160 00:11:45,350 --> 00:11:49,590 We just need to, um, now make or call this fit method. 161 00:11:49,590 --> 00:11:56,040 And then we have our train data set which we pass in the number of epochs. 162 00:11:56,400 --> 00:12:04,770 We'll define this to be 20 and then the batch size, um, equal say 128. 163 00:12:04,770 --> 00:12:05,490 So that's it. 164 00:12:05,490 --> 00:12:12,840 So here, um, unlike before, where we needed, for example, to specify the optimizer in this way 165 00:12:12,840 --> 00:12:21,360 and then actually, uh, get in-depth to understand that this, uh, gradient descent process needs, 166 00:12:21,360 --> 00:12:25,740 for example, the partial derivatives and the trainable weights. 167 00:12:25,770 --> 00:12:32,970 Now you'll see that unlike before, where if you were, for example, needing to make use of some callbacks, 168 00:12:32,970 --> 00:12:39,120 you would have to write that or include that in your with some custom code. 169 00:12:39,120 --> 00:12:43,650 But now all you need to do is just have callbacks. 170 00:12:43,650 --> 00:12:47,400 So we, we, we both will. 171 00:12:47,400 --> 00:12:57,960 Now how to write our own custom training loops while still taking advantage of everything that comes 172 00:12:57,960 --> 00:12:59,550 with the fit method. 173 00:12:59,550 --> 00:13:04,050 So right here you could have some callbacks and that will be it. 174 00:13:04,050 --> 00:13:10,050 Anyway, let's take this off and then let's get back here where there's this slight error. 175 00:13:10,080 --> 00:13:13,650 You see here we have this v dot trainable weights and that. 176 00:13:13,650 --> 00:13:20,910 Now this was in the case where we had defined our this way, we had this model and then get it into 177 00:13:20,910 --> 00:13:23,670 our custom training loop right here. 178 00:13:23,700 --> 00:13:25,050 Let's get down here. 179 00:13:25,050 --> 00:13:32,670 We had to actually, uh, we, the way we got the trainable weights was from our model, which we had 180 00:13:32,670 --> 00:13:33,390 defined. 181 00:13:33,420 --> 00:13:42,390 Now, given that we have this model right here, the way we access this trainable weights is by using 182 00:13:42,390 --> 00:13:43,170 self. 183 00:13:43,170 --> 00:13:45,290 So we have self dot trainable weights. 184 00:13:45,320 --> 00:13:48,980 And then here we have soft, dot trainable weights. 185 00:13:49,940 --> 00:13:59,030 And this is simply possible because we are inheriting from a Keras model which already has this different 186 00:13:59,840 --> 00:14:01,940 attributes which have been defined. 187 00:14:01,970 --> 00:14:04,440 Now, that said, let's run this. 188 00:14:04,460 --> 00:14:05,750 That should be fine. 189 00:14:05,960 --> 00:14:07,250 Training are complete. 190 00:14:07,250 --> 00:14:08,330 You can see that. 191 00:14:08,690 --> 00:14:09,230 Um. 192 00:14:10,510 --> 00:14:15,700 We getting our usual outputs when we making use of the fit method? 193 00:14:16,640 --> 00:14:19,100 And then we could go ahead and start with the testing. 194 00:14:19,100 --> 00:14:24,200 So we run the cells and then in here we have some modifications to make. 195 00:14:24,200 --> 00:14:32,090 So right here or previously we had this layers and the predict and we took in the inputs and we got 196 00:14:32,090 --> 00:14:32,810 some output. 197 00:14:32,840 --> 00:14:39,200 Now here we have our model model dot predict, see that? 198 00:14:39,200 --> 00:14:42,410 But this model is, is a whole model. 199 00:14:42,410 --> 00:14:48,500 So since we want to make use of our decoder, we're going to say model dot, decoder model. 200 00:14:48,500 --> 00:14:49,490 There we go. 201 00:14:49,520 --> 00:14:56,930 Model dot decoder actually dot decoder because here you've defined this decoder attribute. 202 00:14:56,930 --> 00:15:00,650 So we have model dot decoder. 203 00:15:00,680 --> 00:15:01,670 There we go. 204 00:15:01,700 --> 00:15:04,190 Dot predict and that should be fine. 205 00:15:04,340 --> 00:15:07,430 Okay, so let's run this and then see what we get. 206 00:15:09,020 --> 00:15:09,770 There we go. 207 00:15:09,770 --> 00:15:18,290 You see, we have exact same kind of output as when we were having our own custom training loop right 208 00:15:18,290 --> 00:15:18,830 here. 209 00:15:19,700 --> 00:15:21,190 And that's it for this section. 210 00:15:21,200 --> 00:15:22,220 Hope you enjoyed it. 211 00:15:22,250 --> 00:15:23,090 See you in the next section.