GridSearchCV for number of neurons

问题

I am trying to learn by myself how to grid-search number of neurons in a basic multi-layered neural networks. I am using GridSearchCV and KerasClasifier of Python as well as Keras. The code below works for other data sets very well but I could not make it work for Iris dataset for some reasons and I cannot find it why, I am missing out something here. The result I get is:

Best: 0.000000 using {'n_neurons': 3} 0.000000 (0.000000) with: {'n_neurons': 3} 0.000000 (0.000000) with: {'n_neurons': 5}

from pandas import read_csv

import numpy
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import StandardScaler

from keras.wrappers.scikit_learn import KerasClassifier
from keras.models import Sequential
from keras.layers import Dense
from keras.utils import np_utils
from sklearn.model_selection import GridSearchCV

dataframe=read_csv("iris.csv", header=None)
dataset=dataframe.values
X=dataset[:,0:4].astype(float)
Y=dataset[:,4]

seed=7
numpy.random.seed(seed)

#encode class values as integers
encoder = LabelEncoder()
encoder.fit(Y)
encoded_Y = encoder.transform(Y)

#one-hot encoding
dummy_y = np_utils.to_categorical(encoded_Y)

#scale the data
scaler = StandardScaler()
X = scaler.fit_transform(X)

def create_model(n_neurons=1):
    #create model
    model = Sequential()
    model.add(Dense(n_neurons, input_dim=X.shape[1], activation='relu')) # hidden layer
    model.add(Dense(3, activation='softmax')) # output layer
    # Compile model
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model

model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, initial_epoch=0, verbose=0)
# define the grid search parameters
neurons=[3, 5]

#this does 3-fold classification. One can change k. 
param_grid = dict(n_neurons=neurons)
grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1)
grid_result = grid.fit(X, dummy_y)
# summarize results
print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_))
means = grid_result.cv_results_['mean_test_score']
stds = grid_result.cv_results_['std_test_score']
params = grid_result.cv_results_['params']
for mean, stdev, param in zip(means, stds, params):
    print("%f (%f) with: %r" % (mean, stdev, param))

For the purpose of illustration and computational efficiency I search only for two values. I sincerely apologize for asking such a simple question. I am new to Python, switched from R, by the way because I realized that Deep Learning community is using python.

回答1:

Haha, this is probably the funniest thing I ever experienced on Stack Overflow :) Check:

grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=5)

and you should see different behavior. The reason why your model get a perfect score (in terms of cross_entropy having 0 is equivalent to best model possible) is that you haven't shuffled your data and because Iris consist of three balanced classes each of your feed had a single class like a target:

[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 (first fold ends here) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 (second fold ends here)2 2 2 2 2 2 2 2 2 2 2
 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
 2 2]

Such problems are really easy to be solved by every model - so that's why you've got a perfect match.

Try to shuffle your data before - this should result in an expected behavior.

来源：https://stackoverflow.com/questions/47002177/gridsearchcv-for-number-of-neurons

标签

python

neural-network

keras

grid-search