1 00:00:01,020 --> 00:00:07,500 ‫In the last couple of lectures, we learned how to train and create our model 2 00:00:09,090 --> 00:00:17,370 ‫One thing you have noticed is each time we opened the jupyter notebook, we started creating and 3 00:00:17,370 --> 00:00:20,100 ‫training our models from scratch. 4 00:00:21,500 --> 00:00:23,480 ‫But we don't have to do that. 5 00:00:24,540 --> 00:00:29,410 ‫That is a very simple way to save your model after you have trained it. 6 00:00:30,810 --> 00:00:38,760 ‫Keras will save both the model architecture, including the every layer put parameters and the value 7 00:00:38,760 --> 00:00:45,000 ‫of all the model parameters for each layer that is biases and weights 8 00:00:45,040 --> 00:00:45,420 ‫Value 9 00:00:47,880 --> 00:00:54,690 ‫So instead of training the model each time you open your jupyter notebook or initiate a new session. 10 00:00:55,260 --> 00:01:02,890 ‫You can just load the model that you have previously saved to get all the details of that. 11 00:01:02,910 --> 00:01:03,270 ‫Model 12 00:01:05,440 --> 00:01:07,630 ‫Now, saving model is very easy. 13 00:01:07,990 --> 00:01:10,070 ‫You just have to write, you are more name. 14 00:01:11,370 --> 00:01:18,360 ‫So in the last lecture, we created a model with functional api and object name was model. 15 00:01:18,630 --> 00:01:24,320 ‫So I'm saving that same object so I can write model dot save. 16 00:01:24,390 --> 00:01:25,370 ‫Save is the keyword. 17 00:01:26,100 --> 00:01:31,800 ‫And in the parenthesis, you have to write the name that you want to give to that file. 18 00:01:33,180 --> 00:01:40,710 ‫All this models are saved in hd f5 format. the extension of that is dot h5 19 00:01:42,660 --> 00:01:50,590 ‫So if you just run this, you have saved your model in your present working directory 20 00:01:52,590 --> 00:02:03,140 ‫So if I open my directory, this is this and you can see my model is here. 21 00:02:03,570 --> 00:02:10,860 ‫And the name of model is my_func underscore model is the same as I have mentioned. 22 00:02:12,480 --> 00:02:15,360 ‫You can see your present working directory with this. 23 00:02:15,360 --> 00:02:15,750 ‫Command 24 00:02:16,220 --> 00:02:16,750 ‫Just have to write 25 00:02:16,770 --> 00:02:17,020 ‫. 26 00:02:18,700 --> 00:02:20,150 ‫Percentage symbol and 27 00:02:20,450 --> 00:02:24,030 ‫Pwd it stands for present working directory. 28 00:02:24,440 --> 00:02:27,600 ‫So if you run this, you will get the directory location. 29 00:02:28,500 --> 00:02:32,730 ‫And in that directory you can find the model that you have saved. 30 00:02:34,440 --> 00:02:41,200 ‫Now, in case you want to change your working directory, you can do it with this keyword. 31 00:02:41,820 --> 00:02:46,470 ‫You have to write percentage cd and then the new directory name. 32 00:02:48,180 --> 00:02:52,740 ‫So as we have seen, I have this model present in my folder. 33 00:02:53,550 --> 00:02:57,000 ‫Now let's just delete the model variable. 34 00:03:00,430 --> 00:03:02,930 ‫So I have deleted the model 35 00:03:03,620 --> 00:03:05,870 ‫I'm also clearing the session of keras. 36 00:03:08,860 --> 00:03:15,100 ‫Now, again, I am loading this mortal with Lord underscore model method 37 00:03:16,450 --> 00:03:19,390 ‫And here also you have to just provide the name of the file. 38 00:03:20,530 --> 00:03:24,640 ‫And I'm saving this and another object known as model 39 00:03:25,680 --> 00:03:29,620 ‫Let's just track loading the model that we have recently saved. 40 00:03:32,100 --> 00:03:33,850 ‫Now we can call the summary method. 41 00:03:34,120 --> 00:03:41,840 ‫This should give me the summary that we saw earlier while creating regression neural network with functional 42 00:03:41,840 --> 00:03:42,260 ‫Api. 43 00:03:42,690 --> 00:03:44,990 ‫Just run this. 44 00:03:47,110 --> 00:03:50,770 ‫You can see that we have the same summary. 45 00:03:52,360 --> 00:03:54,170 ‫Remember, we deleted our model. 46 00:03:55,060 --> 00:03:58,450 ‫And then we loaded it using the save version. 47 00:03:59,980 --> 00:04:02,820 ‫Now summary is just the architecture of that model. 48 00:04:03,910 --> 00:04:12,320 ‫You can also predict using this model because it already have the optimized weights and biases value 49 00:04:13,630 --> 00:04:16,020 ‫So you can use it for prediction also. 50 00:04:16,020 --> 00:04:17,320 ‫So you can see. 51 00:04:17,380 --> 00:04:21,400 ‫This is the same prediction that we were getting earlier. 52 00:04:21,490 --> 00:04:22,780 ‫With functional api. 53 00:04:25,240 --> 00:04:30,940 ‫But you can save a model only after the completion of its training. 54 00:04:33,110 --> 00:04:43,490 ‫But usually we have a very large dataset which can take up to eight to 10 hours just to train the model 55 00:04:44,120 --> 00:04:45,500 ‫for that scenario. 56 00:04:45,890 --> 00:04:47,420 ‫We can use callbacks. 57 00:04:50,330 --> 00:04:58,820 ‫You can think of callbacks as check points, which will allow you to save your model after each epoch. 58 00:05:00,530 --> 00:05:03,380 ‫So you don't have to wait for the complete training. 59 00:05:04,130 --> 00:05:07,850 ‫You can save your model after each epoch as well. 60 00:05:09,800 --> 00:05:14,020 ‫Let's just again delete this model and. 61 00:05:16,010 --> 00:05:18,350 ‫Clear the session 62 00:05:18,580 --> 00:05:22,750 ‫I'm just taking a small example of a sequential model with two dense layer 63 00:05:23,230 --> 00:05:24,550 ‫And 1 output layer 64 00:05:24,790 --> 00:05:26,260 ‫I will using the same data. 65 00:05:27,280 --> 00:05:28,850 ‫Let's just compile it 66 00:05:29,260 --> 00:05:31,860 ‫We are using the same parameters as earlier. 67 00:05:34,040 --> 00:05:40,010 ‫Now, you can provide this callback parameter while training the model. 68 00:05:40,510 --> 00:05:45,460 ‫So earlier we were only providing training, data, epochs and validation dataset. 69 00:05:46,420 --> 00:05:50,590 ‫Now, along with this, we can also provide callbacks. 70 00:05:51,550 --> 00:05:58,750 ‫So you have to write callbacks and then here you can provide the list of callbacks that you have created. 71 00:06:01,150 --> 00:06:02,860 ‫This is how you create a callback. 72 00:06:03,370 --> 00:06:05,350 ‫So this is our variable name that is 73 00:06:05,430 --> 00:06:06,730 ‫Check point callback. 74 00:06:07,890 --> 00:06:12,820 ‫And you have to write keras. callbacks DOT model checkpoint. 75 00:06:14,770 --> 00:06:16,590 ‫Here I can write the file name. 76 00:06:18,310 --> 00:06:19,060 ‫But since. 77 00:06:20,350 --> 00:06:22,970 ‫We are running this model for 10 epochs. 78 00:06:23,150 --> 00:06:27,500 ‫That means we are generating 10 files for each epoch. 79 00:06:30,170 --> 00:06:38,210 ‫So if I provide a constant file names such as my kera model it will just override the file for 80 00:06:38,240 --> 00:06:39,890 ‫each of these 10 epochs. 81 00:06:40,960 --> 00:06:46,600 ‫So, at the end i would be only getting a single file of my 10 epoch 82 00:06:49,210 --> 00:06:50,980 ‫now to resolve this problem. 83 00:06:51,670 --> 00:06:53,860 ‫We can use these variables. 84 00:06:54,550 --> 00:07:05,460 ‫So for one epoch, I want to save my file as model dash zero one dot h5 for the second epoch. 85 00:07:05,890 --> 00:07:08,150 ‫I want to save it as Mordred dash 86 00:07:08,230 --> 00:07:08,920 ‫0 to 87 00:07:09,190 --> 00:07:09,710 ‫Dot h5 88 00:07:10,360 --> 00:07:11,140 ‫And so on. 89 00:07:11,220 --> 00:07:12,310 ‫Till the 10th epoch 90 00:07:13,150 --> 00:07:17,380 ‫So you can use such variable names in your file name also. 91 00:07:18,040 --> 00:07:21,850 ‫So you don't have to manually write each file name. 92 00:07:24,870 --> 00:07:28,590 ‫So here model is the name that I have given. 93 00:07:28,770 --> 00:07:36,160 ‫And this is the variable name which is kept in the curly bracket, so I can write epoch and I want 94 00:07:36,160 --> 00:07:38,070 ‫the two digits of each epoch. 95 00:07:38,310 --> 00:07:40,170 ‫So for one, I want zero one. 96 00:07:40,470 --> 00:07:44,750 ‫So that's why 02 d if I wanted three digits. 97 00:07:44,790 --> 00:07:50,880 ‫So suppose if I was running for 300 epochs I would want three digits. 98 00:07:50,910 --> 00:07:55,750 ‫Zero zero one or a hundred or a hundred and fifty to ninety nine. 99 00:07:56,280 --> 00:07:59,480 ‫In that case I would have written zero 3D. 100 00:08:02,380 --> 00:08:06,740 ‫So let's just cleared this checkpoint, call back again. 101 00:08:07,000 --> 00:08:12,210 ‫Keras.callbacks Dot model checkpoint and then the variable name and inside variable name. 102 00:08:12,220 --> 00:08:14,410 ‫Also, we can give variables. 103 00:08:15,560 --> 00:08:17,730 ‫You mentioned different epoch values. 104 00:08:18,260 --> 00:08:18,520 ‫Just 105 00:08:18,570 --> 00:08:19,200 ‫Run this. 106 00:08:20,260 --> 00:08:22,180 ‫And now just train the model. 107 00:08:22,500 --> 00:08:28,240 ‫We are using the same method model. fit training epoch, validation data. 108 00:08:28,410 --> 00:08:30,090 ‫And then the callback. 109 00:08:35,580 --> 00:08:40,890 ‫So with each of this epoch, I'm also saving a model file for that epoch. 110 00:08:43,320 --> 00:08:50,310 ‫So if we just go back to our directory, we should have 10 model files. 111 00:08:50,640 --> 00:08:54,570 ‫Here you can see we have a model dash zero one more dash 02 112 00:08:54,570 --> 00:08:56,430 ‫Zero to model. 113 00:08:56,720 --> 00:08:57,660 ‫Dash 10 114 00:08:58,990 --> 00:09:00,150 ‫This for First epoch. 115 00:09:00,260 --> 00:09:03,680 ‫Second epoch third epoch and so on till the 10th epoch. 116 00:09:04,660 --> 00:09:13,330 ‫If I would ever use the single file name, such as model, then at the end I would be getting only 117 00:09:13,330 --> 00:09:15,710 ‫single file because after each epoch 118 00:09:15,760 --> 00:09:17,420 ‫That file was overwrited 119 00:09:17,450 --> 00:09:19,030 ‫By the next epoch value 120 00:09:20,530 --> 00:09:25,760 ‫So remember to use the variable names of Epoch while creating a model file for Epoch. 121 00:09:27,400 --> 00:09:34,210 ‫Now, I can load this files similar to the previous time. 122 00:09:34,640 --> 00:09:35,620 ‫We just have to. 123 00:09:35,620 --> 00:09:36,020 ‫Write 124 00:09:36,440 --> 00:09:37,680 ‫Keras.model.loadmodel 125 00:09:37,790 --> 00:09:38,250 ‫. 126 00:09:38,590 --> 00:09:41,290 ‫And the file, which we want to load. 127 00:09:43,120 --> 00:09:50,740 ‫So just to demonstrate that, let's delete the model and clear the session. 128 00:09:50,890 --> 00:09:57,620 ‫So suppose if I want five of my 10th epoch, so model.10 129 00:10:02,290 --> 00:10:04,780 ‫So this is the model I want to export. 130 00:10:09,330 --> 00:10:11,850 ‫You can see i have imported that model. 131 00:10:12,990 --> 00:10:17,700 ‫So in future, also, if I'm opening this, you buy a notebook of that three or four days. 132 00:10:18,000 --> 00:10:19,920 ‫I can just write this model. 133 00:10:20,400 --> 00:10:22,440 ‫Keras.model.loadmodel 134 00:10:22,590 --> 00:10:25,230 ‫And then the model nale to load that model 135 00:10:25,620 --> 00:10:32,580 ‫There is absolutely no need of creating, compiling and tarining the model once again. 136 00:10:33,690 --> 00:10:35,470 ‫We can also use evaluate method. 137 00:10:36,390 --> 00:10:39,660 ‫You can see this model object is working. 138 00:10:41,250 --> 00:10:45,330 ‫Now there is no need of saving the model. 139 00:10:45,630 --> 00:10:55,200 ‫At each epoch, if you are using the regulations set during the training, you can set the safe best 140 00:10:55,230 --> 00:10:58,980 ‫only equal to true value, creating a checkpoint called the back. 141 00:11:01,770 --> 00:11:07,450 ‫It will only save your more than when its performance on the validation set is best. 142 00:11:10,210 --> 00:11:13,870 ‫So suppose out of this 10 epoch values. 143 00:11:15,070 --> 00:11:20,440 ‫We were getting the best validation is score at suppose, epoch 8 144 00:11:21,400 --> 00:11:26,520 ‫Then in the final 5, we will be getting the model corresponding to Epoch eight. 145 00:11:27,100 --> 00:11:32,260 ‫There is no need to create a separate model for all this epoch values. 146 00:11:34,590 --> 00:11:39,150 ‫Because generally, we want the model corresponding to the best score only 147 00:11:42,940 --> 00:11:47,030 ‫So let us delete the model and creating a structure. 148 00:11:47,740 --> 00:11:56,440 ‫And here, while creating the callback, I am also setting the parameter value of save best only 149 00:11:56,650 --> 00:11:57,560 ‫equal to true 150 00:11:58,310 --> 00:12:03,640 ‫At the end I would be getting only a single model with the name 151 00:12:03,760 --> 00:12:04,480 ‫Best model. 152 00:12:05,550 --> 00:12:09,900 ‫The model, which is corresponding to the best value of validation score. 153 00:12:12,570 --> 00:12:15,760 ‫So let's just create this checkpoint. 154 00:12:16,140 --> 00:12:24,360 ‫And then we are training the mofel, again, x train epochs validation set and then the callbacks. 155 00:12:36,030 --> 00:12:41,160 ‫Now, to get the best model, we can just import the best model that we have saved. 156 00:12:41,820 --> 00:12:43,770 ‫So if you see in the directory also 157 00:12:43,800 --> 00:12:47,720 ‫So we have saved this best model.h5 . 158 00:12:51,510 --> 00:12:53,950 ‫We are just loading and using that model. 159 00:12:54,270 --> 00:12:56,550 ‫To evaluate the performance on test set 160 00:13:00,360 --> 00:13:05,950 ‫Now one of the best method to train your model is to use early stopping. 161 00:13:07,470 --> 00:13:14,280 ‫What I'm referring by early stopping, is that you run your model for a very large number of epoch 162 00:13:14,280 --> 00:13:14,790 ‫values. 163 00:13:15,240 --> 00:13:17,910 ‫Suppose here I am using two hundred epochs. 164 00:13:19,190 --> 00:13:23,900 ‫And then I will keep an eye on my validation, set score . 165 00:13:25,060 --> 00:13:33,820 ‫So suppose if after 60 epochs, my validation said accuracy is not improving, then I will stop the 166 00:13:33,820 --> 00:13:40,760 ‫training at that point and I will be using the model with the best validation score, so far. 167 00:13:41,620 --> 00:13:43,330 ‫So this is an ideal situation. 168 00:13:43,570 --> 00:13:46,840 ‫We don't have to worry about how many epochs we want to run. 169 00:13:47,500 --> 00:13:49,750 ‫We just set a very high value of epoch. 170 00:13:50,320 --> 00:13:56,990 ‫We monitor the valuation set score and then select the best model that we have got so far. 171 00:13:59,360 --> 00:14:06,980 ‫So to achieve that, we will be needing and another callback, which we generally call early stopping 172 00:14:06,980 --> 00:14:07,440 ‫callback. 173 00:14:08,390 --> 00:14:16,190 ‫And we will use this callback along with our normal callback of save best only model. 174 00:14:18,170 --> 00:14:22,310 ‫So in early stopping callback, we have to provide the patients value. 175 00:14:23,240 --> 00:14:31,730 ‫Patients value is the number of epochs after which our training will be interrupted if there is no progress 176 00:14:31,760 --> 00:14:34,160 ‫on the validation set score. 177 00:14:36,410 --> 00:14:44,210 ‫So suppose after 53 epochs, we are not getting any improvement on our validation set scores. 178 00:14:44,870 --> 00:14:49,510 ‫Then keras will stop the training at 63 epochs 179 00:14:51,020 --> 00:14:54,420 ‫So the difference between this 63 and 53 is 10 180 00:14:54,930 --> 00:14:55,950 ‫This is the patience. 181 00:14:56,450 --> 00:15:02,380 ‫Thus the number of epochs keras us will wait for any improvement in validation score. 182 00:15:04,520 --> 00:15:08,150 ‫So earlier you saw while creating our functional model. 183 00:15:08,180 --> 00:15:14,990 ‫We have we have to look at our valuation score graph to check out whether our model is converging or 184 00:15:14,990 --> 00:15:15,380 ‫not. 185 00:15:16,100 --> 00:15:18,070 ‫Now, we don't have to worry about that. 186 00:15:18,500 --> 00:15:27,110 ‫We can early stopping along with other callbacks, which will automatically save the best model 187 00:15:27,140 --> 00:15:28,110 ‫that is available. 188 00:15:28,970 --> 00:15:36,950 ‫And it will also train our model for optimal amount of time needed to train that model to its best 189 00:15:36,950 --> 00:15:37,880 ‫accuracy score. 190 00:15:39,740 --> 00:15:42,680 ‫So let's just run this. 191 00:15:42,860 --> 00:15:49,510 ‫We are first creating this structure now, along with early stopping callback. 192 00:15:49,550 --> 00:15:52,940 ‫We have to use our the normal callbacks also to save the model. 193 00:15:54,380 --> 00:15:59,960 ‫And here we are also using save best only parameter as well. 194 00:16:01,760 --> 00:16:09,020 ‫So at the end of all this training, we would be getting on the single model file on which we would 195 00:16:09,020 --> 00:16:11,850 ‫be getting the best validation set score. 196 00:16:13,550 --> 00:16:16,960 ‫Now our second callback is early stopping callback. 197 00:16:17,480 --> 00:16:23,740 ‫Here you have to provide patience parameters for the number of epochs. 198 00:16:25,780 --> 00:16:26,360 ‫Run this. 199 00:16:27,710 --> 00:16:28,610 ‫Now we have to train 200 00:16:28,640 --> 00:16:33,020 ‫Our model, we will be passing training dataset, epochs, values. 201 00:16:34,040 --> 00:16:36,710 ‫Remember to increase epoch values. 202 00:16:38,630 --> 00:16:42,150 ‫Then the validation data and then callbacks. 203 00:16:43,070 --> 00:16:46,100 ‫Again, we are passing callbacks in the form of list. 204 00:16:46,280 --> 00:16:52,250 ‫We are first passing out checkpoint callback and then we are passing our early stopping callback. 205 00:16:52,880 --> 00:16:53,900 ‫Just run this. 206 00:17:21,250 --> 00:17:25,860 ‫So the training has stopped after 87 epoch. 207 00:17:27,690 --> 00:17:33,150 ‫This means that there is no need to train it, till 200 epoch values. 208 00:17:34,950 --> 00:17:40,890 ‫Our model has already converged and there is no way to further decrease the validation loss. 209 00:17:41,970 --> 00:17:51,270 ‫You can see the best value we are getting is at 77th epoch, and it waited for 10 more epochs to get 210 00:17:51,700 --> 00:17:53,840 ‫the better value than zero point three two. 211 00:17:54,990 --> 00:17:57,250 ‫And you can see up to 10 epoch values. 212 00:17:58,590 --> 00:18:02,550 ‫The value is still hovering over zero point three three. 213 00:18:02,820 --> 00:18:06,290 ‫That's why it is stopped at 87 epoch value. 214 00:18:08,150 --> 00:18:10,360 ‫Let's just look at our directory. 215 00:18:11,330 --> 00:18:17,580 ‫We should have a model with file name early stop model here 216 00:18:18,500 --> 00:18:26,990 ‫You can see keras has already saved the best model here with the name of early stop model. 217 00:18:28,240 --> 00:18:29,860 ‫So now the training has stop. 218 00:18:30,080 --> 00:18:35,900 ‫We can use this early stop model to predict and evaluate our future values. 219 00:18:37,550 --> 00:18:45,710 ‫You can see the loss on tests set we are getting is zero point two three one as compared to our previous 220 00:18:45,710 --> 00:18:49,820 ‫loss values of zero point four three seven, eight and so on. 221 00:18:51,830 --> 00:18:58,820 ‫So using early stop callback is a very good way to optimize performance of your model. 222 00:18:59,150 --> 00:19:02,570 ‫You don't have to worry about the number of epochs. 223 00:19:05,330 --> 00:19:09,200 ‫So let's have a quick recap of what we learned in this lecture. 224 00:19:09,680 --> 00:19:19,890 ‫First, we learn how to save and load our model tan we used callbacks to save our model at each 225 00:19:19,890 --> 00:19:20,600 ‫epoch value. 226 00:19:21,890 --> 00:19:28,970 ‫Then we learned about a parameter that is saved best only, which will save the model for which we are 227 00:19:28,970 --> 00:19:32,870 ‫getting the best accuracy score or lowest validation loss. 228 00:19:34,010 --> 00:19:43,550 ‫And after that, we saw how to use early stop callback to terminate the training if no progress on 229 00:19:43,640 --> 00:19:46,070 ‫validation loss is occurring. 230 00:19:46,940 --> 00:19:48,440 ‫So that's all for this lecture. 231 00:19:48,860 --> 00:19:49,310 ‫Thank you.