Functional API - TensorFlow Essentials #2
Chapters
0:00:2 Functional Api
1:30 Input Layer
2:42 Densely Connected Neural Network Layers
00:00:00.000 | Hi, welcome to this video on the Functional API, which
00:00:03.800 | is the second of the two approaches for building models
00:00:06.960 | in TensorFlow Keras.
00:00:09.200 | So we're going to go ahead and start our notebook.
00:00:12.560 | I'm just going to import TensorFlow.
00:00:14.600 | So the Functional API can produce
00:00:20.360 | the same architectures as the sequential model,
00:00:23.200 | plus a few more, thanks to a high level of flexibility.
00:00:27.920 | So the Functional API is ideal for when
00:00:29.920 | we have multiple inputs or outputs,
00:00:32.480 | different streams of information between layers,
00:00:35.560 | or anything else that is more complex than a simple linear
00:00:39.520 | model, where each input and output is
00:00:41.960 | feeding into the next one, and there's nothing fancy going on.
00:00:45.720 | An example of what this might look like
00:00:48.560 | is something like this.
00:00:50.960 | So you can see at the bottom here, we have two inputs.
00:00:54.400 | And then in the middle here, we also
00:00:56.440 | have this stream of data being split into this one layer,
00:01:00.720 | and then also being fed forward into the next layer as well.
00:01:04.720 | This is something that you could not do with the sequential
00:01:08.320 | model.
00:01:09.040 | You would have to use Functional API, which
00:01:10.840 | we're going to go through now.
00:01:13.000 | So rather than defining the model first,
00:01:16.520 | we instead define the input layers, the network layers,
00:01:21.440 | which is the same as what we did with the sequential model.
00:01:24.200 | And then we initialize the model with something
00:01:26.480 | called tf.keras.model.
00:01:28.800 | Let's go ahead and define the input layer first.
00:01:31.760 | So here, we use something different.
00:01:38.600 | We don't use a dense layer.
00:01:40.200 | We're going to use a input layer, which is obviously
00:01:42.320 | specifically for inputs.
00:01:43.640 | Whereas before, we use input_shape
00:01:50.120 | to define the input shape, here we are using shape.
00:01:54.760 | And we're going to build the exact same model
00:01:56.680 | that we did before, but using the functional model instead.
00:02:00.560 | So the only difference here is that we also
00:02:05.840 | need to define the data type going
00:02:08.080 | into our inputs, which in this case will be industry of 32.
00:02:12.680 | If we don't define this, it will default to float with 32.
00:02:17.120 | And then we can execute that.
00:02:18.760 | And then we need to set up our network layers, which
00:02:21.160 | is exactly the same as what we did before.
00:02:23.800 | So if we go over to this image again,
00:02:27.360 | we can see that we have this sort of network architecture.
00:02:30.880 | So we have the 10 inputs, which is what we've already
00:02:33.120 | just defined.
00:02:34.320 | And then we have our 32 units in the middle here and 2 units
00:02:38.280 | on the output layer.
00:02:40.400 | So we need to build that using densely connected neural
00:02:44.720 | network layers.
00:02:46.320 | And we do that just like we did before.
00:02:49.280 | The only difference is that here,
00:02:51.080 | we are passing the previous layer into our new layer.
00:02:57.200 | So our previous layer, we have placed into this variable here,
00:03:00.560 | input sensor.
00:03:03.360 | And we also want to pass it to our new layer here.
00:03:07.840 | And then we are also putting this new layer
00:03:10.600 | or new transformation into a new layer.
00:03:13.240 | So we're just going to pass it into our new layer here.
00:03:17.440 | And we're going to pass this new transformation
00:03:20.040 | into another variable, which is x.
00:03:23.720 | I mean, you can use these names as you want.
00:03:25.920 | But the inputs and the output layer,
00:03:30.240 | they need to have names which are not overwritten
00:03:33.840 | by any of your other layers, because we
00:03:35.640 | need to pass those layers into our model initializer
00:03:39.720 | in a moment, which you will see.
00:03:41.720 | So we will call our output layer y.
00:03:45.200 | And then we're going to use the softmax activation here again.
00:03:53.920 | And we pass x, which is our previous layer.
00:03:57.640 | Now, if we had more than just these layers,
00:04:00.200 | we could use x in the middle and overwrite.
00:04:02.640 | And that wouldn't cause any issues,
00:04:05.040 | just as long as we don't overwrite
00:04:07.600 | the variables which define our output and input.
00:04:12.160 | And the reason that we don't overwrite those
00:04:20.240 | is because we need them to initialize our model, which
00:04:24.360 | we do like this.
00:04:25.120 | So inputs, we have our input tensor.
00:04:35.720 | And outputs, we have y.
00:04:38.960 | So this is where, if we did have multiple inputs or outputs,
00:04:42.400 | we would be able to put them together here.
00:04:45.640 | So rather than having y, we could have y and y2,
00:04:49.840 | for example.
00:04:51.120 | But in our case, we are only using one.
00:04:54.880 | So now we initialize the model.
00:04:56.880 | And if we print our model summary,
00:05:00.800 | we get the same output as we did with the sequential model
00:05:04.560 | with a few small differences.
00:05:06.560 | So at the top here, rather than saying sequential,
00:05:09.480 | it says model, which obviously, this
00:05:11.040 | is because we're using the sequential model
00:05:13.520 | or the functional API.
00:05:14.960 | It will change what we have up here.
00:05:17.400 | And then for the functional API, we also
00:05:19.840 | have a specific definition for the input layer
00:05:23.360 | as well, which is just because we have actually explicitly
00:05:27.080 | defined it here rather than it being an input shape to one
00:05:30.320 | of the other layers.
00:05:31.880 | And that is everything for the functional API.
00:05:35.840 | So I hope that has been useful and I will see you in the next one.
00:05:39.440 | Thanks for watching.