WEBVTT 00:00.080 --> 00:01.560 Congratulations guys! 00:01.560 --> 00:06.160 We exported our model and load it successfully in TensorFlow. 00:06.280 --> 00:11.520 Now let's export models as TF Lite for Android. 00:11.520 --> 00:17.280 So we're going to use this model to make predictions in Android Studio. 00:17.440 --> 00:27.320 I can't go and paste this model linear regression export or the Savedmodel file in Android Studio because 00:27.680 --> 00:30.000 TensorFlow will not work. 00:30.000 --> 00:34.320 Also, Android Studio will not recognize those files and those models. 00:34.440 --> 00:41.840 We need a specific extension which is the dot light for Android. 00:41.880 --> 00:47.320 Okay, so here let me start converting the current model to TF Lite. 00:47.400 --> 00:54.040 Here we have the loaded model and the model that we build it together. 00:54.040 --> 01:02.600 So let me start exporting the model that we've built together and then move to the loaded model. 01:02.640 --> 01:03.160 Okay. 01:03.350 --> 01:13.550 So here converter equals to tf.light.tf light converter from Keras model and pass the model and not 01:13.550 --> 01:19.430 the current model or the uh, or the tested model or the loaded model. 01:19.470 --> 01:31.110 Okay, so pass the model because this is used to convert our model, our trained model to TF Lite extension, 01:31.110 --> 01:39.630 then we need to convert the model TF Lite model dot equals to converter dot convert. 01:39.910 --> 01:42.190 Also this is the converter. 01:42.350 --> 01:44.750 So we we did the conversion. 01:44.750 --> 01:47.070 We need to save the file. 01:47.070 --> 01:49.710 So I need to get the output file. 01:49.710 --> 01:56.190 So here I need to get an output tf.tf lite file. 01:56.190 --> 02:06.310 For that I need to make the TF lite model path house price model.tf lite extension and with open TF 02:06.310 --> 02:13.350 lite model path we passed here WB as f f dot write tf lite model. 02:13.390 --> 02:16.030 Let me explain what we've done here. 02:16.150 --> 02:22.830 The TF Lite format optimized for mobile and edge devices like microcontrollers. 02:22.950 --> 02:29.790 And here we use the TF Lite converter from Keras model to do this conversion. 02:29.790 --> 02:37.750 So here here we are creating a converter that understands your Keras model structure. 02:37.870 --> 02:46.390 The converter analyzes your model architecture, examines layers, weights and connections, prepares 02:46.550 --> 02:49.470 for optimization and conversion. 02:49.470 --> 02:56.390 By this step, the converter understands the input one feature house size. 02:56.430 --> 03:04.910 The operation dense layer with one neuron output is one value house price and the parameters are weight 03:04.910 --> 03:06.030 and bias. 03:06.030 --> 03:11.220 Then we use TF Lite modal converter dot convert. 03:11.260 --> 03:17.060 The purpose is to perform the actual conversion to two light format. 03:17.100 --> 03:22.580 What happens during conversion optimizes operations for mobile devices. 03:22.620 --> 03:29.620 Quantizers or Quantizers weights, if specified, creates a flat buffer format. 03:29.660 --> 03:34.660 Tflite reduces model size and improves performance. 03:34.780 --> 03:38.620 The last step is to save the tflite file. 03:38.660 --> 03:44.620 WB writes in binary model and in binary mode. 03:44.660 --> 03:49.820 The binary format not human readable, optimized for fast loading. 03:49.820 --> 03:57.580 So WB write in binary mode required for tflite and binary format is not human readable. 03:57.620 --> 04:00.300 Here Tflite model path. 04:00.340 --> 04:04.860 This is the path and the name and WB write in binary. 04:04.900 --> 04:08.260 Go and write it in order to create this file. 04:08.380 --> 04:11.980 So let's run the cell and here we go. 04:12.380 --> 04:13.620 It's executed. 04:13.620 --> 04:16.380 Go open the folders and. 04:16.380 --> 04:18.020 Congratulations, guys. 04:18.020 --> 04:23.980 We have this model called House price model dot flight. 04:24.020 --> 04:25.700 Click download. 04:25.860 --> 04:31.340 And in the next videos we're going to use this model with Android Studio. 04:31.420 --> 04:36.620 So this is the file that we're going to move it to Android Studio Assets folder. 04:36.660 --> 04:41.700 One last thing we need to mention here converting current model to Tflite. 04:41.860 --> 04:46.620 But what about converting the export the loaded model. 04:46.620 --> 04:55.860 So here, as we learned before, we loaded a model from uh, from uh, a file or a folder, how to export 04:55.860 --> 04:57.460 it as Tflite. 04:57.460 --> 04:58.860 It's very simple. 04:58.860 --> 05:07.500 So here as we did in the in the previous conversion, I'll copy this and paste it here. 05:07.700 --> 05:13.730 Converter Tflite dot Tflite converter from Keras model. 05:13.730 --> 05:18.650 It's not from Keras model, it's from saved model. 05:18.690 --> 05:19.210 Okay. 05:19.530 --> 05:22.610 And you pass the name of the model. 05:22.610 --> 05:25.090 For example linear regression. 05:25.370 --> 05:30.250 Export the name of the file a folder that we exported. 05:30.450 --> 05:33.090 And we need to load and convert it. 05:33.090 --> 05:37.610 And then TF Lite model equals to converter dot convert. 05:37.610 --> 05:41.210 Similar to what we've done before with open. 05:41.330 --> 05:43.730 Also go and write it in binary. 05:43.850 --> 05:46.810 Name it as TF Lite. 05:47.050 --> 05:53.210 I'll name it as for example the loaded model.tf Lite. 05:53.250 --> 05:55.330 Okay, it's very simple. 05:55.490 --> 05:58.690 And the print the conversion is done. 05:58.690 --> 05:59.970 Let's run. 05:59.970 --> 06:01.170 And here we go. 06:01.370 --> 06:02.370 Let me check. 06:02.490 --> 06:07.290 And the loaded model.tf Lite is being created. 06:07.530 --> 06:17.290 Okay guys what we've done here we converted our model our model that we learned before our loaded model. 06:17.290 --> 06:19.850 So here are the previous videos. 06:19.850 --> 06:32.050 We worked on a model named as model and we exported to a TF Lite for Android using from Keras model 06:32.170 --> 06:32.970 function. 06:32.970 --> 06:42.650 Also, we learned how to load the exported models, how to load the models from external and and maybe 06:42.650 --> 06:44.890 other developers created. 06:45.090 --> 06:55.290 Um, we, we exported the model and we loaded and we load it here in order to make the conversion. 06:55.290 --> 07:00.410 And we succeeded in converting any loaded model folder to TF Lite. 07:00.410 --> 07:02.650 So this is our current model. 07:02.650 --> 07:08.690 This is our current model that we worked on on the on the previous videos. 07:08.690 --> 07:14.090 And this is the loaded model that we exported from an external.