1 00:00:11,690 --> 00:00:15,590 In this lecture we are going to answer one of the most common questions I get. 2 00:00:15,830 --> 00:00:19,910 That is once we've trained the model how do we save it for later use. 3 00:00:19,910 --> 00:00:25,850 For example after creating a model in Google collab I might want to copy that model into my production 4 00:00:25,850 --> 00:00:32,020 server so that other services in my company can make use of that model that I just trained. 5 00:00:32,120 --> 00:00:37,540 This lecture will show you how to save a model to a file and then load it back at a later time. 6 00:00:37,550 --> 00:00:41,630 In addition we'll see how to download the saved model from Google Colette 7 00:00:47,190 --> 00:00:47,760 first. 8 00:00:47,760 --> 00:00:51,970 Note that this is a continuation of a notebook we were previously looking at. 9 00:00:52,050 --> 00:00:57,690 As usual you can determine what notebook we are currently looking at using the title of the notebook. 10 00:00:58,050 --> 00:01:01,560 So if you left this notebook previously you may want to reopen it. 11 00:01:01,560 --> 00:01:05,660 At this point so you can follow along. 12 00:01:05,830 --> 00:01:08,700 Let's start with the central objects that we need for this process. 13 00:01:08,710 --> 00:01:11,790 The state dict as you can tell by its name. 14 00:01:11,830 --> 00:01:13,240 It's a dictionary. 15 00:01:13,570 --> 00:01:18,940 If you look at what we get when we call model that state dict we can see that it's not just any dictionary 16 00:01:19,090 --> 00:01:21,260 but rather an ordered dictionary. 17 00:01:21,400 --> 00:01:25,780 That kind of makes sense since if you're iterating through the dictionary to populate the weights of 18 00:01:25,780 --> 00:01:32,380 a neural network you would want to do that in the order that your model expects as you can see it's 19 00:01:32,380 --> 00:01:37,300 nothing but the parameters of the model since we just have one layer in our model. 20 00:01:37,300 --> 00:01:43,650 The first element is the weight matrix and the second element is the bias vector. 21 00:01:43,810 --> 00:01:45,430 The next step is to save the model 22 00:01:50,420 --> 00:01:56,120 we can accomplish that in just one line of code torch dot save for the first argument we pass in the 23 00:01:56,120 --> 00:02:02,590 model's state dict and for the second argument we pass in a file path I've just passed in my model dot 24 00:02:02,710 --> 00:02:09,470 Petey so the file my model that Petey will be saved to the current directory you can see that after 25 00:02:09,470 --> 00:02:17,570 saving the model when I do an L S the file I just saved appears as expected the next step is to load 26 00:02:17,600 --> 00:02:20,600 and evaluate the model to ensure that we saved it correctly 27 00:02:25,110 --> 00:02:30,480 in order to do this we first have to recreate the model the same way it was created originally. 28 00:02:30,480 --> 00:02:34,500 That might seem strange since we still have to replicate the code we had earlier. 29 00:02:34,800 --> 00:02:36,700 This will probably make more sense to you. 30 00:02:36,840 --> 00:02:42,390 Once you've seen more complex models which will involve defining our own classes in those cases you 31 00:02:42,390 --> 00:02:47,730 would just instantiate an object of that class which is one line of code which would be more compact 32 00:02:47,730 --> 00:02:48,820 than this. 33 00:02:48,900 --> 00:02:54,400 In any case passing code around is very lightweight because code files are just text anyway. 34 00:02:54,410 --> 00:02:59,050 We'll call this model too so as not to interfere with our original model. 35 00:02:59,070 --> 00:03:06,250 Next we call model to load state debt and parse in the state dict that we just saved we get that by 36 00:03:06,250 --> 00:03:11,960 calling torture out load and passing in the file path where our model file is located. 37 00:03:12,010 --> 00:03:16,360 You can see that if your load is successful it will say all keys match successfully 38 00:03:21,410 --> 00:03:22,070 at this point. 39 00:03:22,070 --> 00:03:26,960 We would like to know is our reloaded model the same as the model we saved. 40 00:03:26,960 --> 00:03:31,040 Luckily this is an easy check and we already know how to write this code. 41 00:03:31,040 --> 00:03:36,500 We just take our code from earlier to compute the accuracy of the model on the train set and test sets. 42 00:03:36,560 --> 00:03:40,750 But this time we'll use a model 2 instead of the original model. 43 00:03:41,000 --> 00:03:45,070 As you can see we get precisely the same accuracy as we did earlier. 44 00:03:53,070 --> 00:03:54,810 At this point you might be wondering. 45 00:03:54,810 --> 00:03:56,380 So I just saved my model. 46 00:03:56,400 --> 00:04:02,430 It's in my local directory but it's not my local directory it's my Google drives local directory. 47 00:04:02,460 --> 00:04:04,540 What if my teammate asked me for this model. 48 00:04:04,560 --> 00:04:05,400 How can I share it. 49 00:04:06,060 --> 00:04:08,990 And that is the next question we are going to discuss. 50 00:04:09,060 --> 00:04:13,040 In fact this is quite easy and it's another reason I really like collab. 51 00:04:13,380 --> 00:04:14,950 So over in left here. 52 00:04:15,000 --> 00:04:18,180 You may have noticed this sidebar with some icons. 53 00:04:18,420 --> 00:04:24,120 If you click the folder icon you'll see all the files in your current directory including my model that 54 00:04:24,130 --> 00:04:32,890 Petey the model file we just saved so let's right click that and now you can see that there's this option 55 00:04:32,890 --> 00:04:38,020 to download and of course that's how you can download this file to your local harddrive 56 00:04:45,680 --> 00:04:49,650 another option is to use Python code to download the file. 57 00:04:49,700 --> 00:04:54,020 Now I've heard that this only works for Chrome but of course that may change in the future. 58 00:04:54,350 --> 00:05:00,300 We can use the same module we used earlier for uploading files so we import files from the Google Docs 59 00:05:00,300 --> 00:05:01,460 code that module. 60 00:05:01,670 --> 00:05:06,940 Then we say files that download and pass in the file path when we run the cell. 61 00:05:06,950 --> 00:05:09,260 The file is downloaded by our browser as usual 62 00:05:14,720 --> 00:05:15,030 all right. 63 00:05:15,040 --> 00:05:16,440 So that works as expected.