| 1 | |
| 00:00:00,300 --> 00:00:06,120 | |
| So in that we move on to a simple step where we actually build uncompiled or model and Karris and this | |
| 2 | |
| 00:00:06,120 --> 00:00:08,950 | |
| is a model we're going to build as you remembered. | |
| 3 | |
| 00:00:09,000 --> 00:00:15,510 | |
| We showed you this before you actually Edley is how we actually do max booing Flaten dance classes and | |
| 4 | |
| 00:00:15,510 --> 00:00:16,080 | |
| stuff. | |
| 5 | |
| 00:00:16,080 --> 00:00:22,080 | |
| This is actually quite useful diagram I created that shows you how to basically layer and add each piece | |
| 6 | |
| 00:00:22,400 --> 00:00:27,090 | |
| here including drop out and including a flattened layer here as well. | |
| 7 | |
| 00:00:27,420 --> 00:00:32,400 | |
| So let's go ahead and Kerrison run this and compile our model. | |
| 8 | |
| 00:00:32,800 --> 00:00:35,110 | |
| OK so welcome back to Python book. | |
| 9 | |
| 00:00:35,140 --> 00:00:40,830 | |
| So I'm pretty sure this code looks familiar because you would have seen it in the presentation slides | |
| 10 | |
| 00:00:40,870 --> 00:00:41,660 | |
| earlier. | |
| 11 | |
| 00:00:42,050 --> 00:00:45,580 | |
| So remember I said we're building a simple convolution neural nets. | |
| 12 | |
| 00:00:45,580 --> 00:00:46,420 | |
| This is how we do it. | |
| 13 | |
| 00:00:46,450 --> 00:00:48,720 | |
| We have two little filters in a convoy. | |
| 14 | |
| 00:00:49,050 --> 00:00:51,730 | |
| Kenneled say as tree by tree activation really. | |
| 15 | |
| 00:00:52,090 --> 00:00:55,450 | |
| We have the input ship which we defined above. | |
| 16 | |
| 00:00:55,450 --> 00:01:01,930 | |
| Here actually it was in this block right here and then moving on we have a convolutional the sim can | |
| 17 | |
| 00:01:01,930 --> 00:01:03,680 | |
| also use activation. | |
| 18 | |
| 00:01:03,760 --> 00:01:06,070 | |
| We have to max beling downsampling liya. | |
| 19 | |
| 00:01:06,280 --> 00:01:09,000 | |
| We now use dropout and dropout as simple to implement. | |
| 20 | |
| 00:01:09,040 --> 00:01:14,190 | |
| You just muddle and dropout and especially if we use Pia's point to 5 here. | |
| 21 | |
| 00:01:14,620 --> 00:01:21,730 | |
| We didn't flatten this layer have a fully connected DENSELOW with 128 nodes which really we add another | |
| 22 | |
| 00:01:21,730 --> 00:01:26,740 | |
| layer of dropout and we have a higher dropout here notes you can play with these values and see what | |
| 23 | |
| 00:01:26,740 --> 00:01:27,940 | |
| gives you the best. | |
| 24 | |
| 00:01:27,940 --> 00:01:33,110 | |
| You tend to stop drop outs smaller on top and get larger but you don't ever really go larger than point | |
| 25 | |
| 00:01:33,110 --> 00:01:33,950 | |
| five. | |
| 26 | |
| 00:01:34,540 --> 00:01:40,840 | |
| And then this is our last Insley which is connected to 10 nodes which is a number of classes and a number | |
| 27 | |
| 00:01:40,840 --> 00:01:43,340 | |
| of classes was defined right here. | |
| 28 | |
| 00:01:43,990 --> 00:01:51,880 | |
| That's basically the linked number of columns I should see in this array and that's it. | |
| 29 | |
| 00:01:51,880 --> 00:01:54,310 | |
| So I talked about compiling a model. | |
| 30 | |
| 00:01:54,310 --> 00:02:00,430 | |
| Now when we when we do model the compile Basically we're taking these layers creating a model and then | |
| 31 | |
| 00:02:00,430 --> 00:02:06,400 | |
| specifying what type of loss we are using what type of optimize are we using and the metrics we need | |
| 32 | |
| 00:02:06,400 --> 00:02:10,780 | |
| to look at and these metrics we look at basically the metrics that will be output. | |
| 33 | |
| 00:02:10,780 --> 00:02:12,460 | |
| When we started to train our model. | |
| 34 | |
| 00:02:12,580 --> 00:02:18,760 | |
| When we start training our model and by doing model print here we can print a model summary and take | |
| 35 | |
| 00:02:18,760 --> 00:02:19,150 | |
| a look. | |
| 36 | |
| 00:02:19,150 --> 00:02:20,460 | |
| So let's check it out. | |
| 37 | |
| 00:02:22,310 --> 00:02:23,830 | |
| This is very cool here. | |
| 38 | |
| 00:02:23,870 --> 00:02:25,540 | |
| Let's go through this quickly. | |
| 39 | |
| 00:02:25,550 --> 00:02:33,880 | |
| So when you print a model we actually see each layer we have is each layer we specified above here. | |
| 40 | |
| 00:02:34,100 --> 00:02:37,580 | |
| And what's cool about this is that it gives you the output shape. | |
| 41 | |
| 00:02:37,580 --> 00:02:43,640 | |
| So we know the input chip coming into this input shape was 28 by 28 dimension. | |
| 42 | |
| 00:02:43,730 --> 00:02:47,230 | |
| Now 81 and we have 22 filters here. | |
| 43 | |
| 00:02:47,520 --> 00:02:48,110 | |
| So. | |
| 44 | |
| 00:02:48,180 --> 00:02:55,610 | |
| But shape if you remember correctly is going to be 26 26 32 would tell you to be number of filters and | |
| 45 | |
| 00:02:55,610 --> 00:02:59,280 | |
| these are a number number of parameters in this lead here. | |
| 46 | |
| 00:02:59,300 --> 00:03:00,680 | |
| So this is pretty cool. | |
| 47 | |
| 00:03:00,680 --> 00:03:03,540 | |
| So now we have second convolutional here. | |
| 48 | |
| 00:03:03,890 --> 00:03:08,150 | |
| Same thing again but no a lot more promises going forward. | |
| 49 | |
| 00:03:08,150 --> 00:03:14,930 | |
| We have Omak spooling here which is this process no parameters drop it again flatten again and flatten | |
| 50 | |
| 00:03:14,930 --> 00:03:19,760 | |
| has an upward shape of this as you can see it's just this expanded into this. | |
| 51 | |
| 00:03:19,990 --> 00:03:25,810 | |
| We have a fully connected densely here and this is where the bulk of the parameters actually exist. | |
| 52 | |
| 00:03:26,240 --> 00:03:30,870 | |
| And we have dropped it again and in a fully Densa connected to attend classes here. | |
| 53 | |
| 00:03:31,460 --> 00:03:36,140 | |
| So this gives us a total number of parameters total number of treatable parameters. | |
| 54 | |
| 00:03:36,140 --> 00:03:38,640 | |
| We have zero nonrenewable parameters. | |
| 55 | |
| 00:03:38,660 --> 00:03:42,420 | |
| We will come to what non-tradable parameters all later on. | |
| 56 | |
| 00:03:42,530 --> 00:03:44,380 | |
| And so we're not ready to train our model. | |
| 57 | |
| 00:03:44,420 --> 00:03:45,410 | |
| So let's get to it. | |