Machine Learning pipeline

Hyperparameter Tuning --> Change hyperparameters can improve performance.

 

## Is this architectual optimal?

Why 512 neurons for dense? why dropout is 0.2? 

* Do the model need more or less hidden units to perfrom well?

* How does model size affect the convergence speed?

* Is there any trade off between converenge speed... etc

 

* Keras tuner built in search functionality.

16 --> 256 etc what is the best result?

 

Pip install -q -U keras -tuner

import kt

instead of 512, min_value = 16, max_value = 512 and step =16. Tuned by Keras tuner.

model.add(keras.layers.Dense(units=hp_units, activation='relu'))
Instead of 512. The optimized value is 64 so we can change it to 64. However, it might give you less accuracy.

 

 

define search strategy. 

kt.Hyperband strategy. search and baysian optimation and sklearn startegies.

Objective = 'val_accuracy'

 

stop_early = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=5)