WEBVTT 00:00.080 --> 00:01.040 Welcome back. 00:01.040 --> 00:04.440 Now let's create this function called load model file. 00:04.440 --> 00:07.520 So here private function. 00:07.520 --> 00:15.360 Load model file takes a context object and returns byte buffer. 00:15.480 --> 00:24.760 So the purpose from this function is to load a TensorFlow Lite mode model from assets or from the ML 00:24.800 --> 00:28.600 folder here into a direct byte buffer. 00:28.600 --> 00:34.760 And the input is Android context to access the app context and assets. 00:34.760 --> 00:40.640 And the output would be a buffer, which is a byte buffer containing the model. 00:40.640 --> 00:43.560 Ready for TensorFlow Lite interpreter. 00:43.760 --> 00:49.600 The first step here is to get the model path and name. 00:49.840 --> 00:58.880 So val asset file descriptor equals to context dot assets dot open folder. 00:58.880 --> 01:07.930 And give me the name here which is car mileage prediction.tf light. 01:07.970 --> 01:13.770 Make sure that the name is exactly like you specified here. 01:13.810 --> 01:18.690 Otherwise your your model will not be loaded successfully. 01:18.690 --> 01:21.850 So here car mpg and not mileage. 01:22.130 --> 01:25.050 So you see guys this is a big error. 01:25.170 --> 01:32.090 Car mpg underscore prediction dot df light okay. 01:32.250 --> 01:33.530 And this is good. 01:33.770 --> 01:37.090 We are we get a green highlight. 01:37.090 --> 01:39.370 So it refers to the file. 01:39.410 --> 01:48.050 Now here the context dot asset dot FD gets a file descriptor from the model file. 01:48.250 --> 01:54.690 This is more efficient than just opening a stream because it provides direct file access. 01:54.930 --> 02:04.130 And the model file here which is in the ML car MPG prediction flight should be located in the assets 02:04.130 --> 02:06.770 folder or the ML folder. 02:06.930 --> 02:09.810 Okay, so it's very simple. 02:09.850 --> 02:14.850 Now the second step is to create the input Stream. 02:14.850 --> 02:17.450 So file input stream. 02:17.770 --> 02:19.770 Asset file descriptor. 02:19.810 --> 02:24.090 This is the descriptor disk descriptor dot. 02:24.130 --> 02:26.210 Create input stream. 02:26.370 --> 02:27.170 This. 02:27.370 --> 02:32.330 This line creates the input stream from file descriptor. 02:32.490 --> 02:37.130 This allows you to read the model file as a stream of bytes. 02:37.170 --> 02:40.330 So let me write this note down. 02:40.330 --> 02:46.770 This allows you to read the model file as a stream of bytes. 02:46.930 --> 02:51.170 Then we need to create a file channel. 02:51.170 --> 03:02.690 Or you can simply remove it and use var val bytes equals to input file input stream that we created 03:02.690 --> 03:05.010 before dot read bytes. 03:05.250 --> 03:11.930 Since our function need to return a byte buffer, we need to use bytes. 03:11.930 --> 03:15.610 So file input stream dot read bytes. 03:15.730 --> 03:19.970 This reads all bytes from the stream. 03:20.130 --> 03:24.780 Reads the entire content of the model file into a byte array. 03:25.260 --> 03:31.940 The byte variable now contains the complete binary data of your TensorFlow Lite model. 03:31.980 --> 03:35.100 Now we need to set the model buffer. 03:35.100 --> 03:42.140 So val model buffer equals to byte buffer dot. 03:42.180 --> 03:53.540 Allocate direct and bytes dot size to allocate direct function creates a direct byte buffer outside 03:53.540 --> 03:55.300 the normal Java heat. 03:55.340 --> 04:05.060 This is crucial for TensorFlow Lite because native code C++ TensorFlow Lite Lite engine can access it 04:05.140 --> 04:14.100 directly, avoiding unnecessary memory copies between JVM and native code, also providing better performance 04:14.100 --> 04:16.500 for model inference. 04:16.500 --> 04:18.860 So this is a very important step. 04:19.060 --> 04:22.620 Now we need to continue with model buffer. 04:22.620 --> 04:30.150 So model buffer.org set the byte Order dot native order. 04:30.270 --> 04:38.830 This line of code model buffer dot order sets the byte order to match the device's native byte order. 04:38.870 --> 04:41.630 Little endian and big endian. 04:41.670 --> 04:42.230 The. 04:42.550 --> 04:50.950 This line of code ensures the model data is interpreted correctly by the native TensorFlow Lite code, 04:51.110 --> 05:00.070 and it's very important for cross, cross multiple or cross-platform compatibility or multi-platform 05:00.070 --> 05:01.030 compatibility. 05:01.070 --> 05:07.950 Okay, so we need to use it when we have an interpreter interpreter. 05:08.070 --> 05:11.590 And we need to interpret data correctly from the model. 05:11.630 --> 05:16.710 Then we need to copy the data model data into the buffer. 05:16.710 --> 05:20.990 So copy model data into buffer model buffer dot. 05:21.070 --> 05:25.110 Put the bytes and close the input stream. 05:25.270 --> 05:31.470 This will copy the byte array containing the model into the direct byte buffer. 05:31.710 --> 05:37.910 The buffer now contains the complete TensorFlow Lite model in native memory. 05:37.910 --> 05:42.590 And here returning the model buffer. 05:42.590 --> 05:49.870 So here under the model buffer set return model buffer okay. 05:50.310 --> 05:57.950 Since our application our function need needs the byte buffer as a returning type. 05:57.950 --> 06:05.710 So in this video we created our load model file as we saw it in the previous videos. 06:05.710 --> 06:15.310 And in the previous section, we need to load our prediction model that we created from Colab, and 06:15.310 --> 06:22.790 we need to load it into this file, this this variable and it's of type asset file descriptor. 06:22.790 --> 06:30.070 And here we use the file input stream load loading the asset file descriptor. 06:30.110 --> 06:37.510 Also we transform them into bytes and then buffer and returning as a model buffer.