WEBVTT 00:00.080 --> 00:01.040 Welcome back. 00:01.080 --> 00:03.320 I want from you to focus with me. 00:03.320 --> 00:05.440 This is a very important step. 00:05.600 --> 00:10.880 We need to understand the input and output of our model. 00:11.040 --> 00:15.520 For that, let's go to a website called natron. 00:15.520 --> 00:19.080 So natron dot app, upload your model. 00:19.080 --> 00:24.160 And here we have a hierarchy for our model. 00:24.160 --> 00:31.040 You are not um like forced to understand all of those. 00:31.320 --> 00:43.120 We need to start with the first one and understand that it takes an image with 300 width and 300 height 00:43.280 --> 00:44.760 times three. 00:45.000 --> 00:48.160 This is the shape, this is the input shape. 00:48.440 --> 00:56.200 So if we go to here click on this normalized input image tensor you get inputs and outputs. 00:56.520 --> 01:00.280 The inputs normalized input image tensor. 01:00.320 --> 01:04.930 We need to convert it to a unit eight unified integer. 01:05.050 --> 01:12.010 You int eight with this array one 300 303. 01:12.010 --> 01:15.130 So keep this in your mind. 01:15.290 --> 01:21.690 The outputs would be for the TF light detection post process. 01:21.770 --> 01:23.650 Its type is Float32. 01:23.970 --> 01:32.850 Another output would be detection process, post process one Postprocess two, and post process three. 01:33.010 --> 01:44.410 All of those are of type float32 so the input is unit eight and unit or you int eight. 01:44.570 --> 01:50.130 With this shape this array one 300 303. 01:50.170 --> 01:52.730 Now let's understand the input. 01:52.850 --> 02:01.090 So this is the shape and data type of that data your model expects to receive. 02:01.210 --> 02:07.740 Copy it And let me move back to Android Studio and write the notes here. 02:07.780 --> 02:08.100 Okay. 02:08.140 --> 02:12.900 I want from you to write them down because those are very important notes. 02:13.020 --> 02:17.140 The input you want and it's not unified. 02:17.140 --> 02:25.180 It's unsigned integer, unsigned integer, unsigned integer eight bit. 02:25.340 --> 02:27.860 Those values are always positive. 02:27.860 --> 02:34.260 So think about it as an number between 0 and 255 okay. 02:34.300 --> 02:38.540 So this is a positive eight bit integer. 02:38.580 --> 02:46.980 Eight bit means it uses eight bit one byte of memory per number, allowing to two to power eight equals 02:46.980 --> 02:49.140 to 256. 02:49.140 --> 02:50.740 Possible values is. 02:51.060 --> 02:55.260 For this purpose we start from 0 to 255. 02:55.580 --> 03:01.820 This data type is extremely common for image data, because the standard image pixel values range from 03:01.820 --> 03:10.350 0 to 250, where zero typically represents the absence of color, which is black, and 250, which is 03:10.350 --> 03:14.510 full intensity of a color white or a primary color. 03:14.590 --> 03:25.710 In short, your model expects image pixel values as integer between 0 and 55 255. 03:25.750 --> 03:30.590 Now let's take this array, copy it and paste it here. 03:30.710 --> 03:32.230 The first dimension. 03:32.430 --> 03:34.310 This is the batch size. 03:34.550 --> 03:43.470 Deep learning models are often designed to process multiple inputs simultaneously for efficiency. 03:43.670 --> 03:49.550 This dimension tells us how many images are in a single batch. 03:49.670 --> 03:57.990 A one here means the model is set up to process one image at a time. 03:58.150 --> 04:00.510 303 hundred. 04:00.830 --> 04:10.410 The first 300 is the height, and the second is the width, which is more common in deep learning frameworks 04:10.410 --> 04:11.610 like TensorFlow. 04:11.810 --> 04:17.810 Those represents the height and width of the image in pixels. 04:17.930 --> 04:21.370 And the last dimension is three. 04:21.570 --> 04:25.890 This represents the color channels of the image. 04:26.050 --> 04:28.930 It's a standard RGB image. 04:29.090 --> 04:36.930 Red values channel one, green values, channel two B values and blue values channel three. 04:37.170 --> 04:44.090 So for a single pixel, the model sees three numbers r, g, b. 04:44.610 --> 04:50.730 So this number represents the r, g b color channels of the image. 04:50.850 --> 04:51.850 A tensor. 04:51.890 --> 05:03.570 The unsigned integer describes a batch containing a single 300 by 300 pixels color RGB image. 05:03.610 --> 05:13.300 You can think of it as a stack of three mattresses, each 300 by 300. 05:13.500 --> 05:22.620 One for each color channel red, blue and green, all wrapped up in a single container. 05:22.780 --> 05:23.660 The batch. 05:23.820 --> 05:26.140 So this is for image one. 05:26.460 --> 05:31.460 This is a batch of one image one 300 pixel height. 05:31.620 --> 05:41.260 The first row the RGB, RGB, RGB, and uh till till the end row number two the same. 05:41.580 --> 05:50.300 This is a matrix and a matrix, which is an array of arrays, and this is repeated 300 times for width 05:50.340 --> 05:53.060 till reaching row number 300. 05:53.100 --> 05:53.620 Okay. 05:53.860 --> 06:02.980 So when you want to use a model with this input specification you must pre-process your image to match 06:02.980 --> 06:04.180 it exactly. 06:04.300 --> 06:08.030 Resize your image to 300 by 300 pixels. 06:08.030 --> 06:09.550 This is what we're going to do. 06:09.590 --> 06:20.470 Ensures it is an RGB format, not BRG or BGR, which OpenCV sometimes uses, or grayscale. 06:20.750 --> 06:28.470 Do not normalize the pixel values to 0 to 1 or -1 to 1. 06:28.670 --> 06:34.270 Leave them as integers in 0 to 255. 06:34.470 --> 06:40.510 Range because the data type is an unsigned integer eight. 06:40.670 --> 06:51.190 Some models have a normalization step built in, but given this specific model, it expects the row 06:51.590 --> 06:54.630 unsigned integer eight values. 06:54.710 --> 06:57.590 This is a very important lesson. 06:57.750 --> 07:04.150 You should pay attention to it when dealing with different models and different combinations, different 07:04.150 --> 07:07.040 input and different output shapes. 07:07.040 --> 07:08.280 The output is. 07:08.320 --> 07:09.920 Float 300. 07:10.080 --> 07:22.000 Float 32, which is the float 32 provides much higher numerical precision you than, uh, unsigned integer 07:22.040 --> 07:22.520 eight. 07:22.760 --> 07:33.760 So the result would be a 3 or 4 of those tflite detection postprocess, which are of type Float32. 07:33.800 --> 07:40.760 Training and inference require fine grained values for accurate calculations. 07:41.000 --> 07:42.600 Deep learning operations. 07:42.600 --> 07:44.160 Metric metrics. 07:44.640 --> 07:51.560 Multiplication activations work much better with floating point numbers. 07:51.600 --> 08:00.120 Many models internally normalize pixel values to range from 0 to 1 or -1 to 1, but in this example, 08:00.120 --> 08:02.640 we're not normalizing anything. 08:02.760 --> 08:03.240 Okay.