1 00:00:00,960 --> 00:00:08,210 ‫So in this video we will see how functional EPA can help us build complex neural network architectures. 2 00:00:08,450 --> 00:00:14,730 ‫They know we have been building simple architectures that as input goes into foster and layer which 3 00:00:14,730 --> 00:00:21,750 ‫goes in two seconds to deliver the output of which goes into the final output layer. 4 00:00:21,840 --> 00:00:31,530 ‫Now we're slightly modified this architecture will take this input and merged with the output of it 5 00:00:31,530 --> 00:00:40,740 ‫and led to that is will create an additional concatenation layer where the output of the layer too will 6 00:00:40,740 --> 00:00:44,430 ‫be there and the input layer will also be added here. 7 00:00:45,570 --> 00:00:53,220 ‫So in the contract layer we will have 64 values from different layer two because it has 64 units of 8 00:00:53,220 --> 00:01:00,570 ‫neurons and 13 values from the input layer because we have 13 variables. 9 00:01:00,570 --> 00:01:06,600 ‫So in the concatenation layer we have seventy seven values which will be imported into our final output 10 00:01:07,110 --> 00:01:10,710 ‫layer which has a single layer on. 11 00:01:10,740 --> 00:01:17,060 ‫So this kind of architecture is called deep and wide architecture. 12 00:01:18,030 --> 00:01:21,480 ‫And this architecture cannot be built using a sequentially be a 13 00:01:25,650 --> 00:01:29,440 ‫this is where functionally B is used. 14 00:01:29,610 --> 00:01:31,290 ‫Now let's go back to our to do 15 00:01:35,300 --> 00:01:40,940 ‫now let's see how we can create that architecture using functional Libya. 16 00:01:41,230 --> 00:01:43,140 ‫The first step is same. 17 00:01:43,150 --> 00:01:51,490 ‫We have to define the input layer input layer has those 13 variables only so we define the input layer 18 00:01:51,850 --> 00:01:55,950 ‫with a shape of 13 variables. 19 00:01:56,380 --> 00:02:03,520 ‫The first production layer has only two layers after which will come the concatenation layer. 20 00:02:03,550 --> 00:02:08,300 ‫So in the first prediction layer we keep the two hidden layers. 21 00:02:08,590 --> 00:02:13,350 ‫We specified that this layer has the input from inputs Splunk. 22 00:02:13,410 --> 00:02:17,160 ‫This is the rebel that we define for input layer. 23 00:02:17,410 --> 00:02:22,180 ‫Then we have to put the layers after these two hidden layers. 24 00:02:22,180 --> 00:02:27,470 ‫We need a concatenation layer in this layer. 25 00:02:27,550 --> 00:02:29,590 ‫The output of this 26 00:02:32,860 --> 00:02:39,230 ‫first output layer will be concatenated with the initial input layer. 27 00:02:39,980 --> 00:02:41,530 ‫To How do we do that. 28 00:02:41,650 --> 00:02:49,820 ‫We create this new layer mean output in this we concatenate these two parts. 29 00:02:50,020 --> 00:02:53,690 ‫The first part is the prediction func. 30 00:02:53,740 --> 00:03:03,950 ‫This includes these three layers and we also add the input func which is this input layer. 31 00:03:04,300 --> 00:03:13,470 ‫So these do not become the input layer of this main output and this goes into the last layer which is 32 00:03:13,470 --> 00:03:19,600 ‫a dense single unit output layer. 33 00:03:19,780 --> 00:03:23,910 ‫I hope you understand the architecture that we have defined here. 34 00:03:24,040 --> 00:03:32,980 ‫We have this input layer this is we what we stored in input func then we created predictions func which 35 00:03:32,980 --> 00:03:36,450 ‫contained the two layers. 36 00:03:36,880 --> 00:03:47,410 ‫Then we create main output structure in which we take the output of this prediction func and the input 37 00:03:47,410 --> 00:03:56,110 ‫layer and we concatenated to create a can get layer after which we put a output layer of one single 38 00:03:56,110 --> 00:03:56,470 ‫neuron. 39 00:03:58,990 --> 00:04:02,920 ‫So all this is done in this part. 40 00:04:03,670 --> 00:04:09,060 ‫You can notice here that inputs layer has been used twice. 41 00:04:09,460 --> 00:04:17,770 ‫This has been made possible because in functional API we are having each layer as individual part of 42 00:04:17,770 --> 00:04:25,050 ‫the structure so we can use those parts many times in the structure. 43 00:04:25,240 --> 00:04:29,550 ‫So this input func became a part of the predictions one variable also. 44 00:04:30,310 --> 00:04:37,100 ‫And it became part of the main output structure also so next on this 45 00:04:45,010 --> 00:04:52,690 ‫and now we will define the architecture of our model using kid US model function in which we will call 46 00:04:52,690 --> 00:05:02,170 ‫the inputs are the inputs funk and output is to be taken from the main output structure because this 47 00:05:02,170 --> 00:05:10,060 ‫structure has everything it has predictions func also an input Fung also as a concatenated layer in 48 00:05:10,060 --> 00:05:13,030 ‫IT structure. 49 00:05:13,210 --> 00:05:25,270 ‫When we done this we have model form ready with us is a dysfunctional API model and it has that complex 50 00:05:25,270 --> 00:05:26,990 ‫architecture stored in it. 51 00:05:27,310 --> 00:05:35,320 ‫Now we can figure this model again we will use outermost properties optimizer lost is messy and the 52 00:05:35,320 --> 00:05:38,800 ‫matrix to be recorded is mean absolute error 53 00:05:41,860 --> 00:05:49,720 ‫you can see the structure of this model func also using this somebody come on man and on this you can 54 00:05:49,720 --> 00:06:00,190 ‫see here that the first layer is the input layer with 13 variables then comes a dense 64 neutron layer 55 00:06:00,700 --> 00:06:09,270 ‫which is decoding Layer 1 it is connected to the input layer then comes another dense layer which tells 56 00:06:09,270 --> 00:06:18,430 ‫sixty 64 neurons and it is connected to firstly the layer after this we have a concatenation layer in 57 00:06:18,430 --> 00:06:25,210 ‫which we can get energy the output of this dense layer with the 13 variables from input. 58 00:06:25,840 --> 00:06:31,420 ‫So this concatenation layer is connected to layer 2 and the input layer 59 00:06:35,760 --> 00:06:42,570 ‫then we have the last layer which is the output layer which has only one neuron and it is connected 60 00:06:42,570 --> 00:06:46,130 ‫to the concatenation. 61 00:06:46,130 --> 00:06:48,830 ‫Now we can bring this model using different function 62 00:07:00,510 --> 00:07:07,610 ‫and we can compare the best performance of our normal model with a normal deep architecture. 63 00:07:07,620 --> 00:07:13,410 ‫What is this complex functional model with wide and deep architecture. 64 00:07:13,420 --> 00:07:15,330 ‫So let's run these command 65 00:07:18,700 --> 00:07:25,570 ‫to see that the functional model that is the complex model has a test loss of twenty nine whereas a 66 00:07:25,580 --> 00:07:31,010 ‫normal model has had a test loss of thirty two. 67 00:07:31,090 --> 00:07:37,240 ‫Similarly they mean absolute error for our functional model that is this complex model that we have 68 00:07:37,240 --> 00:07:39,390 ‫created is four point three. 69 00:07:39,910 --> 00:07:45,730 ‫Whereas the mean absolute better for the simple model that we created earlier was four point forty seven. 70 00:07:46,810 --> 00:07:53,990 ‫So there is slight improvement on both of these parameters. 71 00:07:54,130 --> 00:08:01,750 ‫Now you have seen how to use functional API to create complex neural network architecture because functional 72 00:08:01,750 --> 00:08:08,640 ‫API enables us to use several layers or several structures multiple times. 73 00:08:09,280 --> 00:08:13,490 ‫Also in that example this time we solve the immigration problem. 74 00:08:13,810 --> 00:08:19,910 ‫The few differences that we have integration problem is instead of using accuracy. 75 00:08:19,990 --> 00:08:21,670 ‫We were using MSE 76 00:08:24,790 --> 00:08:31,440 ‫for the lost function and the metric we were observing integration problem was mean absolute. 77 00:08:31,620 --> 00:08:43,000 ‫Edit in the architecture of integration problem the output layer has no activation function. 78 00:08:43,570 --> 00:08:48,400 ‫So with this we conclude our lecture on functional API and regression problems.