How to use scikit-learn optimize in a class (especially the use_named_args decorator)?

Question

I am using the scikit-learn optimize package to tune the hyperparameters of my model. For performance and readability reasons (I am training several models with the same process), I want to structure the whole hyperparameter-tuning in a class:

...
import tensorflow as tf
from tensorflow.keras import backend as K
from tensorflow.keras.preprocessing import sequence
from tensorflow.keras.models import Sequential, load_model
from tensorflow.keras.layers import InputLayer, Input, Dense, Embedding, BatchNormalization, Dropout
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import TensorBoard, EarlyStopping
from sklearn.preprocessing import MinMaxScaler, OneHotEncoder
from sklearn.model_selection import train_test_split

import skopt
from skopt import gp_minimize
from skopt.space import Real, Categorical, Integer
from skopt.plots import plot_convergence
from skopt.plots import plot_objective, plot_evaluations
from skopt.utils import use_named_args

class hptuning:
   def __init__(self, input_df):
         self.inp_df = input_df
         self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(...)
         self.param_space = self.dim_hptuning()
         self.best_loss = 10000

   def dim_hptuning(self):
         dim_layers = Integer(low=0, high=7, name='layers')
         dim_nodes = Integer(low=2, high=90, name='num_nodes')
         dimensions = [dim_layers, dim_nodes]
         return dimensions

   def create_model(self, layers, nodes):
         model = Sequential()
         for layer in range(layers):
             model.add(Dense(nodes))
         model.add(Dense(1,activation='sigmoid'))
         optimizer = Adam
         model.compile(loss='mean_absolute_error',
                  optimizer=optimizer,
                  metrics=['mae', 'mse'])
         return model
         
   @use_named_args(dimensions=self.param_space)
   def fitness(self,nodes, layers):
         model = self.create_model(layers=layers, nodes=nodes)
         history = model.fit(x=self.X_train.values,y=self.y_train.values,epochs=200,batch_size=200,verbose=0)
         loss = history.history['val_loss'][-1]
         if loss < self.best_loss:
             model.save('model.h5')
             self.best_loss = loss
         del model
         K.clear_session()
         return loss

   def find_best_model(self):
         search_result = gp.minimize(func=self.fitness, dimensions=self.param_space,acq_func='EI',n_calls=10)
         return search_result
hptun = hptuning(input_df=df)
search_result = hptun.find_best_model()
print(search_result.fun)

Now I get the problem that the decorator @use_named_args is not working within a class as he should be (example code of scikit-optimize). I get the error message

Traceback (most recent call last):
File "main.py", line 138, in <module>
class hptuning:
File "main.py", line 220, in hptuning
@use_named_args(dimensions=self.param_space)
NameError: name 'self' is not defined

which is obviously about the misuse of the decorator in this scenario.

Probably due to my missing understanding of the functionality of such decorators, I am not able to get this running. Could someone help me on this one?

Thank you all in advance for the support!

maarten · Accepted Answer · 2020-11-06T09:39:19.470

1

The problem of self not being defined is unrelated to scikit.learn. You cannot use self to define a decorator, because it is only defined inside the method you are decorating. But even if you sidestep this issue (e.g. by providing param_space as a global variable) I expect the next problem will be that self will be passed to the use_named_args decorator, but it expects only arguments to be optimized.

The most obvious solution would be to not use the decorator on the fitness method but to define a decorated function that calls the fitness method, inside the find_best_model method, like this:

   def find_best_model(self):
         @use_named_args(dimensions=self.param_space)
         def fitness_wrapper(*args, **kwargs):
             return self.fitness(*args, **kwargs)
         search_result = gp.minimize(func=fitness_wrapper, dimensions=self.param_space,acq_func='EI',n_calls=10)
         return search_result

edited Nov 06 '20 at 09:39

answered Nov 05 '20 at 13:45

maarten

412
5
7

Thank you very much! I think I am a step closer, but this solution still does not work, I get the following error: ` File "main.py", line 166, in find_best_model search_result = gp_minimize(func=self.fitness, dimensions=self.param_space, ... TypeError: fitness() missing 2 required positional arguments: 'layers', 'nodes',` Right now it seems it only gets the first argument (learning rate). Do I have to adjust the other fitness function or even move it inside the find_best_model method? – Kai Schelthoff Nov 05 '20 at 15:48
1

The error message seems to indicate that you are still passing the fitness method, not the decorated wrapper function around the method. Use `gp.minimize(func=fitness, ...)` instead of `gp.minimize(func=self.fitness, ...)` – maarten Nov 05 '20 at 21:34
1

I changed the name of the wrapper function to `fitness_wrapper`, to avoid confusion. Indeed `fitness_wrapper` is a nested function, it is defined inside `find_best_model`, and you should pass `fitness_wrapper` as the `func` argument to `gp.minimize`. – maarten Nov 06 '20 at 09:41

How to use scikit-learn optimize in a class (especially the use_named_args decorator)?

1 Answers1