Update hyperparamter tuning section for dollar street dataset

episodes/4-advanced-layer-types.Rmd

@@ -839,7 +839,7 @@ This is called hyperparameter tuning.
 ## Hyperparameter tuning
 ::: instructor
 ## Do a live demo instead of live coding
-You might want to demonstrate this section on hyperparamater tuning instead of doing live coding.
+You might want to demonstrate this section on hyperparameter tuning instead of doing live coding.
 The goal is to show that hyperparameter tuning can be done easily with `keras_tuner`, not to memorize all the exact syntax of how to do it. This will probably save you half an hour of participants typing over code that they already know from before. In addition, on really slow machines running the grid search could possibly take more than 10 minutes.
 :::
 
@@ -865,7 +865,14 @@ def create_nn_with_hp(dropout_rate, n_layers):
 Now, let's find the best combination of hyperparameters using grid search.
 Grid search is the simplest hyperparameter tuning strategy,
 you test all the combinations of predefined values for the hyperparameters that you want to vary.
+
+For this we will make use of the package `keras_tuner`, we can install it by typing in the command line:
+```bash
+pip install keras_tuner
+```
+
 Note that this can take some time to train (around 5 minutes or longer).
+
 ```python
 import keras_tuner
 
@@ -886,11 +893,11 @@ tuner.search(train_images, train_labels, epochs=20,
              validation_data=(val_images, val_labels))
 ```
 ```output
-Trial 6 Complete [00h 00m 46s]
-val_loss: 1.3021799325942993
+Trial 6 Complete [00h 00m 19s]
+val_loss: 2.086069345474243
 
-Best val_loss So Far: 1.2409346103668213
-Total elapsed time: 00h 03m 59s
+Best val_loss So Far: 2.086069345474243
+Total elapsed time: 00h 01m 28s
 ```
 Let's have a look at the results:
 
@@ -903,41 +910,41 @@ Results in ./untitled_project
 Showing 10 best trials
 Objective(name="val_loss", direction="min")
 
-Trial 0004 summary
+Trial 0005 summary
 Hyperparameters:
 n_layers: 2
-dropout_rate: 0.5
-Score: 1.2409346103668213
+dropout_rate: 0.8
+Score: 2.086069345474243
 
-Trial 0003 summary
+Trial 0000 summary
 Hyperparameters:
-n_layers: 2
+n_layers: 1
 dropout_rate: 0.2
-Score: 1.281008005142212
+Score: 2.101102352142334
 
-Trial 0005 summary
+Trial 0001 summary
+Hyperparameters:
+n_layers: 1
+dropout_rate: 0.5
+Score: 2.1184325218200684
+
+Trial 0003 summary
 Hyperparameters:
 n_layers: 2
-dropout_rate: 0.8
-Score: 1.3021799325942993
+dropout_rate: 0.2
+Score: 2.1233835220336914
 
 Trial 0002 summary
 Hyperparameters:
 n_layers: 1
 dropout_rate: 0.8
-Score: 1.3677740097045898
+Score: 2.1370232105255127
 
-Trial 0001 summary
+Trial 0004 summary
 Hyperparameters:
-n_layers: 1
+n_layers: 2
 dropout_rate: 0.5
-Score: 1.3880290985107422
-
-Trial 0000 summary
-Hyperparameters:
-n_layers: 1
-dropout_rate: 0.2
-Score: 1.4468265771865845
+Score: 2.143627882003784
 ```
 
 ::: challenge
@@ -947,11 +954,11 @@ Score: 1.4468265771865845
 1: Looking at the grid search results, select all correct statements:
 
 - A. 6 different models were trained in this grid search run, because there are 6 possible combinations for the defined hyperparameter values
-- B. There are 2 different models trained, 1 for each hyperparameter that we want to change
+- B. 2 different models were trained, 1 for each hyperparameter that we want to change
 - C. 1 model is trained with 6 different hyperparameter combinations
-- D. The model with 1 layer and a dropout rate of 0.2 is the best model with a validation loss of 1.45
-- E. The model with 2 layers and a dropout rate of 0.5 is the best model with a validation loss of 1.24
-- F. We have found the model with the best possible combination of dropout rate and number of layers
+- D. The model with 2 layer and a dropout rate of 0.5 is the best model with a validation loss of 2.144
+- E. The model with 2 layers and a dropout rate of 0.8 is the best model with a validation loss of 2.086
+- F. We found the model with the best possible combination of dropout rate and number of layers
 
 2 (Optional): Perform a grid search finding the best combination of the following hyperparameters: 2 different activation functions: 'relu', and 'tanh', and 2 different values for the kernel size: 3 and 4. Which combination works best?
 
@@ -995,52 +1002,52 @@ def build_model(hp):
     compile_model(model)
     return model
 
-tuner = keras_tuner.GridSearch(build_model, objective='val_loss')
+tuner = keras_tuner.GridSearch(build_model, objective='val_loss', project_name='new_project')
 tuner.search(train_images, train_labels, epochs=20,
              validation_data=(val_images, val_labels))
 ```
 ```output
-Trial 4 Complete [00h 00m 56s]
-val_loss: 1.44037926197052
+Trial 4 Complete [00h 00m 25s]
+val_loss: 2.0591845512390137
 
-Best val_loss So Far: 1.2700632810592651
-Total elapsed time: 00h 04m 01s
+Best val_loss So Far: 2.0277602672576904
+Total elapsed time: 00h 01m 30s
 ```
 Let's look at the results:
 ```python
 tuner.results_summary()
 ```
 ```output
 Results summary
-Results in ./untitled_project
+Results in ./new_project
 Showing 10 best trials
 Objective(name="val_loss", direction="min")
 
-Trial 0000 summary
-Hyperparameters:
-kernel_size: 3
-activation: relu
-Score: 1.2700632810592651
-
 Trial 0001 summary
 Hyperparameters:
 kernel_size: 3
 activation: tanh
-Score: 1.2945374250411987
+Score: 2.0277602672576904
 
-Trial 0002 summary
+Trial 0003 summary
 Hyperparameters:
 kernel_size: 4
+activation: tanh
+Score: 2.0591845512390137
+
+Trial 0000 summary
+Hyperparameters:
+kernel_size: 3
 activation: relu
-Score: 1.431167483329773
+Score: 2.123767614364624
 
-Trial 0003 summary
+Trial 0002 summary
 Hyperparameters:
 kernel_size: 4
-activation: tanh
-Score: 1.44037926197052
+activation: relu
+Score: 2.150160551071167
 ```
-A kernel size of 3 and `relu` as activation function is the best tested combination.
+A kernel size of 3 and `tanh` as activation function is the best tested combination.
 
 ::::
 :::
@@ -1057,7 +1064,7 @@ Let's save our model
 model.save('cnn_model')
 ```
 
-## Conclusion and Next Steps
+## Conclusion and next steps
 How successful were we with creating a model here?
 With ten image classes, and assuming that we would not ask the model to classify an image that contains none of the given classes of object, a model working on complete guesswork would be correct 10% of the time.
 Against this baseline accuracy of 10%, and considering the diversity and relatively low resolution of the example images, perhaps our last model's validation accuracy of ~30% is not too bad.