1 00:00:00,450 --> 00:00:05,820 Hi and welcome back to the course in this section, we'll take a look at something called callbacks, 2 00:00:05,820 --> 00:00:11,040 which allows us to implement early stopping, check pointing, learning, read schedules and much more. 3 00:00:11,520 --> 00:00:14,250 They're available in broad Keras and PyTorch. 4 00:00:14,250 --> 00:00:17,100 However, it's much easier to implement and carries it away. 5 00:00:17,520 --> 00:00:23,520 But let's begin so callbacks callbacks are used to perform different actions at different stages in 6 00:00:23,520 --> 00:00:24,000 the training. 7 00:00:24,270 --> 00:00:25,650 But why is this useful? 8 00:00:26,160 --> 00:00:29,910 Well, what if we wanted to see one model after the end of each epoch? 9 00:00:30,360 --> 00:00:33,300 Or what if you create a model for 100 epochs? 10 00:00:33,300 --> 00:00:35,850 But after Tuti so that all of a fit? 11 00:00:36,300 --> 00:00:40,890 Wouldn't you want a way to monitor training so that you can stop trading and stop wasting time at that 12 00:00:40,890 --> 00:00:41,250 point? 13 00:00:41,850 --> 00:00:47,070 Or what if we wanted to log information somewhere during training that can be analyzed elsewhere? 14 00:00:47,520 --> 00:00:49,760 Well, callbacks allow all of those things. 15 00:00:49,770 --> 00:00:50,700 They are the solution. 16 00:00:51,240 --> 00:00:55,200 They can be used to perform a number of things, such as early stopping, which we discussed in our 17 00:00:55,200 --> 00:01:02,250 regularization section model, check pointing, learning, read, scheduler logging and remote monitoring, 18 00:01:02,610 --> 00:01:04,020 as well as a custom function. 19 00:01:04,050 --> 00:01:06,810 So let's take a look at this graph here. 20 00:01:07,350 --> 00:01:11,820 You can see this is an example of overfitting model, and you would want the model to stop here. 21 00:01:12,210 --> 00:01:13,800 This is what early stopping does. 22 00:01:13,800 --> 00:01:15,960 It allows us during the training process. 23 00:01:16,440 --> 00:01:22,660 It allows us to if for validation, los stagnates or stops decreasing and actually can start to increase. 24 00:01:22,680 --> 00:01:29,400 It allows us to do set something called early stopping criteria that allows our model to stop trading 25 00:01:29,400 --> 00:01:32,460 at that point and ends the training process so you don't waste time. 26 00:01:33,060 --> 00:01:35,070 What about model check pointing? 27 00:01:35,550 --> 00:01:38,670 Well, this is how you set the parameters and carrots for model check point. 28 00:01:38,790 --> 00:01:40,140 But what exactly is this? 29 00:01:40,170 --> 00:01:46,320 Well, models are pointing allows us during training to periodically save the weights after each epoch. 30 00:01:46,800 --> 00:01:52,770 That's very good because as you saw in earlier stopping, sometimes the best model isn't the one that 31 00:01:52,770 --> 00:01:54,090 you trained after the most ebooks. 32 00:01:54,090 --> 00:01:59,310 In fact, it could be one of the early models that just got it got to be its right that give us a very 33 00:01:59,310 --> 00:01:59,880 good loss. 34 00:02:00,330 --> 00:02:06,850 So you want to actually store we to monitor the validation loss, and you can see if the best model 35 00:02:06,850 --> 00:02:11,070 of, well, trading so you can train for a thousand epochs if you had all the time in the world and 36 00:02:11,160 --> 00:02:17,720 powerful GPUs, and it would save the best model out of all of those ebooks using the models are pointing, 37 00:02:17,730 --> 00:02:18,660 which is quite cool. 38 00:02:19,110 --> 00:02:25,860 So it's also allows us to resume trading in the event of a crash, and the checkpoint actually contains 39 00:02:25,860 --> 00:02:27,680 the model of weights and the model itself. 40 00:02:27,690 --> 00:02:32,790 So when you see if all those checkpoints that are light allow you to resume training or whatever you 41 00:02:32,790 --> 00:02:37,980 want to do, all those all the checkpoints are basically the model weights at a time, and this is how 42 00:02:37,980 --> 00:02:38,940 it can be configured. 43 00:02:39,750 --> 00:02:42,510 So now let's take a look at lifting weights scheduler. 44 00:02:42,960 --> 00:02:45,820 So this is what it looks like for a living which has Shatila. 45 00:02:45,840 --> 00:02:47,280 But what exactly is it? 46 00:02:47,670 --> 00:02:53,250 Well, we can avoid having a little loss oscillate around a global minimum, which is not what you want. 47 00:02:53,250 --> 00:02:54,780 You wants to find a global minimum. 48 00:02:55,980 --> 00:03:00,600 And we do that by attempting to reduce the learning rate by a specified amount. 49 00:03:01,050 --> 00:03:05,460 So if no improvement is seen, you know, a monitored metric, which is a validation list, typically 50 00:03:06,030 --> 00:03:07,730 we wait a certain number of epochs. 51 00:03:07,740 --> 00:03:12,510 That's the patient setting, which you see here and then does call back and then it reduces learning 52 00:03:12,510 --> 00:03:14,700 rate by a factor that is set here. 53 00:03:14,700 --> 00:03:15,330 Point two. 54 00:03:15,600 --> 00:03:22,920 This is a minimum delta can change, so you can see by using this this function called reduced learning 55 00:03:22,920 --> 00:03:28,950 rate on pilot to just quite descriptive of what it does if you're living, if your loss stops decreasing. 56 00:03:29,460 --> 00:03:33,630 This automatically reduces the learning rate, so it's quite dynamic and quite useful. 57 00:03:34,200 --> 00:03:36,570 You can also do something like logging like this. 58 00:03:36,990 --> 00:03:42,930 You can automatically log your models training stats, using checkpoints, tens of what checkpoints 59 00:03:42,930 --> 00:03:43,200 here. 60 00:03:44,000 --> 00:03:50,000 The stores stausholm model results in the directory and dot logs directory, and you can just load a 61 00:03:50,000 --> 00:03:50,560 tentacle. 62 00:03:50,670 --> 00:03:54,900 Is doing these commands in your Pi Python book and it brings up this. 63 00:03:54,900 --> 00:04:00,750 This brings up this nice web UI where you can actually track and monitor your model as a tree and so 64 00:04:00,750 --> 00:04:02,070 you don't get the graphs afterwards. 65 00:04:02,070 --> 00:04:05,000 You can look at the graphs life, which is quite cool. 66 00:04:05,010 --> 00:04:09,190 I will be doing that in all two models that we will be treating in club. 67 00:04:09,210 --> 00:04:10,620 So stay tuned for that. 68 00:04:11,220 --> 00:04:15,120 So that's it for this section next to what we'll do. 69 00:04:15,420 --> 00:04:18,960 Well, actually look at callbacks in the court and Keras and PyTorch. 70 00:04:19,410 --> 00:04:20,550 So stay tuned for that. 71 00:04:20,790 --> 00:04:21,240 Thank you.