WEBVTT 00:00.420 --> 00:02.490 -: Hello and welcome to this tutorial. 00:02.490 --> 00:04.320 All right, so today we're gonna make a function 00:04.320 --> 00:05.940 that will save our model. 00:05.940 --> 00:07.890 That is that will save the brain of the core 00:07.890 --> 00:11.430 so that we can reuse it whenever we quit the application. 00:11.430 --> 00:12.690 Thanks to this save function, 00:12.690 --> 00:14.400 we'll be able to save the model, 00:14.400 --> 00:15.960 then quit the application, 00:15.960 --> 00:18.090 and then when we go back to the application, 00:18.090 --> 00:20.430 thanks to another function that is the load function, 00:20.430 --> 00:22.200 which we'll make after this one, 00:22.200 --> 00:24.750 we will be learning the last version of our model 00:24.750 --> 00:25.800 that was trained. 00:25.800 --> 00:27.450 So that will be very practical 00:27.450 --> 00:29.850 and therefore, let's make these two functions, 00:29.850 --> 00:32.160 the save function and the load function. 00:32.160 --> 00:35.490 So let's start with the save function in this tutorial. 00:35.490 --> 00:38.730 So here comes the new dev, then save. 00:38.730 --> 00:41.070 And then it's gonna take one argument 00:41.070 --> 00:43.080 that is going to be self. 00:43.080 --> 00:45.270 And the reason is that the thing we are going 00:45.270 --> 00:48.150 to save is not the whole model here 00:48.150 --> 00:51.390 but our neural network self.model 00:51.390 --> 00:54.690 and our optimizer, self.optimizer 00:54.690 --> 00:56.100 because what we wanna save 00:56.100 --> 00:58.350 is just the last waits 00:58.350 --> 01:00.510 that were updated at the last iteration 01:00.510 --> 01:03.960 because whenever we want to reuse our saved model later, 01:03.960 --> 01:06.090 we want it to predict the action to play 01:06.090 --> 01:09.120 with the waits that were already trained. 01:09.120 --> 01:11.490 So we need to take this last version of the wait 01:11.490 --> 01:14.160 and also, we need to take the last version of the optimizer 01:14.160 --> 01:16.830 because it is connected to these waits. 01:16.830 --> 01:17.940 So let's do this. 01:17.940 --> 01:21.417 We have our self, so we'll be able to take our self.model 01:21.417 --> 01:23.640 and our self.optimizer, 01:23.640 --> 01:25.680 and we will be saving these two objects 01:25.680 --> 01:27.180 in a Python dictionary. 01:27.180 --> 01:28.770 And to save these two objects, 01:28.770 --> 01:32.430 we're gonna use the save function from the Torch module, 01:32.430 --> 01:36.570 so I'm starting here with torch.save. 01:36.570 --> 01:37.830 And then parenthesis, 01:37.830 --> 01:39.960 we are going to input that dictionary. 01:39.960 --> 01:42.720 Brackets, and a dictionary in Python works like that. 01:42.720 --> 01:45.630 You have a key, which is your identifier. 01:45.630 --> 01:46.830 So that's unique. 01:46.830 --> 01:48.750 And for each key, you have the value you want 01:48.750 --> 01:49.860 to give to that key. 01:49.860 --> 01:53.520 So it's like a mapping function from unique identifiers 01:53.520 --> 01:56.880 to a value you want to give these identifiers. 01:56.880 --> 01:59.520 If you take a simple dictionary book, 01:59.520 --> 02:01.830 well, the keys would be the words, 02:01.830 --> 02:04.830 and the values would be the definitions of the words. 02:04.830 --> 02:06.240 Well, here that's the same. 02:06.240 --> 02:07.890 We're gonna make two keys. 02:07.890 --> 02:10.620 One key for the first object we wanna save, 02:10.620 --> 02:12.480 which is self.model. 02:12.480 --> 02:15.060 And one second key for the second thing we want to save, 02:15.060 --> 02:17.850 that is our self.optimizer. 02:17.850 --> 02:20.400 And therefore, let's start with the first key. 02:20.400 --> 02:22.290 So we have to give a name to that key, 02:22.290 --> 02:25.060 and I'm gonna call it state_dict 02:26.070 --> 02:26.903 because then you're gonna see 02:26.903 --> 02:29.520 that I'm gonna use the function state_dict 02:29.520 --> 02:31.710 to save our model in the dictionary. 02:31.710 --> 02:33.300 So that's our first key. 02:33.300 --> 02:36.510 Then to give the value we want to attribute 02:36.510 --> 02:39.150 for that first key, well, as you can see, 02:39.150 --> 02:41.400 I added a little colon here 02:41.400 --> 02:43.440 and here I'm gonna add the object, 02:43.440 --> 02:44.970 the object I wanna save. 02:44.970 --> 02:48.690 So the first object I wanna save is self.model. 02:48.690 --> 02:50.850 So we can just copy this. 02:50.850 --> 02:55.620 Self.model and paste it as the value of our first key. 02:55.620 --> 02:56.850 Self.model. 02:56.850 --> 03:00.450 Then we add .state_dict. 03:00.450 --> 03:02.160 Here we go, the first one. 03:02.160 --> 03:04.140 And then we add some parenthesis. 03:04.140 --> 03:06.270 And that will save the parameters 03:06.270 --> 03:09.930 of your model in this first key state_dict. 03:09.930 --> 03:12.330 And now let's save our optimizer. 03:12.330 --> 03:15.480 So we're gonna add a second key in the dictionary. 03:15.480 --> 03:17.610 And to do this, we add here a comma, 03:17.610 --> 03:21.240 then press Enter and there we go with our second key. 03:21.240 --> 03:23.160 So the second key we're gonna call it, 03:23.160 --> 03:26.160 well, we can call it optimizer. 03:26.160 --> 03:28.500 Then colon, and then we just need 03:28.500 --> 03:31.230 to add the name of the object we wanna save 03:31.230 --> 03:33.360 and that is self. 03:33.360 --> 03:34.770 That is our optimizer. 03:34.770 --> 03:37.317 So we add here self.optimizer. 03:39.180 --> 03:40.680 And then again to save the parameters 03:40.680 --> 03:44.083 of this optimizer, we have here again .state_dict. 03:45.180 --> 03:46.013 And there we go. 03:46.013 --> 03:48.330 We have our model saved 03:48.330 --> 03:49.650 with all the wait saved, 03:49.650 --> 03:52.080 and our optimizer saved. 03:52.080 --> 03:53.160 Perfect. 03:53.160 --> 03:56.490 And then we will save all this into a file 03:56.490 --> 03:59.040 and to do this, I'm going to add a second argument 03:59.040 --> 04:00.660 to the save function, 04:00.660 --> 04:04.200 which is going to be the name of this file 04:04.200 --> 04:05.787 where we want to have our model 04:05.787 --> 04:07.650 and our optimizer saved. 04:07.650 --> 04:09.720 So remember, I showed you a quick demo 04:09.720 --> 04:13.380 in the first section of this first module self-driving car. 04:13.380 --> 04:16.590 That was the demo where we just had some random actions. 04:16.590 --> 04:18.750 So that was not yet a self-driving car 04:18.750 --> 04:20.850 but then remember, I clicked on the save button 04:20.850 --> 04:22.410 to save the model, 04:22.410 --> 04:26.160 and this created the last_brain.pth file, 04:26.160 --> 04:28.530 which is the file that contains the saved version 04:28.530 --> 04:29.520 of your model. 04:29.520 --> 04:34.520 So I'm going to add here last_brain.pth 04:36.180 --> 04:38.970 so that your model and your optimizer 04:38.970 --> 04:43.470 will be saved into this created file last_brain.pth. 04:43.470 --> 04:44.820 So you don't have it yet 04:44.820 --> 04:47.250 but as soon as you save your model in the application, 04:47.250 --> 04:51.303 this val will be created thanks to this code we just added. 04:52.230 --> 04:53.820 And so now perfect, 04:53.820 --> 04:56.550 we have a save function that will save your model, 04:56.550 --> 04:57.720 save the brain of your core 04:57.720 --> 05:00.510 by saving the waits and the optimizer 05:00.510 --> 05:03.930 of the neural network that is in fact the brain of the car. 05:03.930 --> 05:08.220 So perfect, we now have only one function to create left. 05:08.220 --> 05:09.870 That's the load function. 05:09.870 --> 05:11.850 And that's because the save function never goes 05:11.850 --> 05:13.560 without a load function. 05:13.560 --> 05:15.480 There's no purpose saving your model 05:15.480 --> 05:17.820 if you cannot load what you save afterwards. 05:17.820 --> 05:19.290 So that's the last step in our journey 05:19.290 --> 05:20.940 before the exciting demo, 05:20.940 --> 05:24.090 and we will make this load function in the last tutorial 05:24.090 --> 05:25.170 of this section. 05:25.170 --> 05:26.880 So I'll see you in this next tutorial 05:26.880 --> 05:28.713 and until then, enjoy AI.