Add K-Fold Validation to neural network

Last article , I have descussed how to use sequential neural network to train the digit recognition , but re we didn't add cross validation in , so the result may varies . Here I will describe some params to better tune your model . by the way, I will add more hidden layers this time . This is also a way to improve your socre

import keras
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation

Define the model : 

model  = Sequential() # adopt sequential neural model

#input layer
model.add(Dense(64, input_shape=(64,))) # Dense with 64 features
model.add(Activation("relu"))

#hidden layer 1 
model.add(Dense(160))
model.add(Activation("relu"))

#hidden layer 2 
model.add(Dense(160)) 
model.add(Activation("relu"))

#hidden layer 3
model.add(Dense(160))
model.add(Activation("relu"))

#hidden layer 4
model.add(Dense(160))
model.add(Activation("relu"))

# output layer
model.add(Dense(10)) # 10 categories
model.add(Activation("softmax")) # only can output 1 & 0

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=["accuracy"])

Read data & train :

from sklearn.datasets import load_digits
from keras.utils import np_utils

digit = load_digits() # here we use another embeded dataset called digits 

data_x = digit.data
data_y = np_utils.to_categorical(digit.target, 10) # convert to one-hot-encoder

train_x = data_x
train_y = data_y

# k-fold crossvalidation is added with a split point at 20% 80%
history = model.fit(train_x, train_y, epochs=200,  verbose=1, validation_split= 0.2) 

print (history.history['acc'])
print (history.history['loss'])

Here we all one parameter called validation_split, it will split the dataset into 80% training set and 20% test set to do the corss validation , It is not using the normal k-fold , it is using one called StratifiedKFold so that the accurancy will be more accurate.

Visualise to Tune :

%matplotlib inline

import matplotlib
import matplotlib.pyplot as plt


plt.xlabel('x')
plt.ylabel('y')
plt.title('acc graph')

plt.plot(range(len(history.history['val_acc'])), history.history['val_acc'])

plt.show()

This graph shown is the corss- validation graph which will be more precise .

In the end , you may think whether there exist a way to persist the model which i can use next time . What I can tell you is Yes.

Persist the model :

import h5py
from keras.models import load_model

model.save('model.h5') # persist the model to external file

model2 = load_model('model.h5') # load the persistent model from external file

Compare the  result  from loaded model with the previous :

y_pred2 = model2.predict(train_x)

y_pred1 = model.predict(train_x)

y_pred1 == y_pred2 # compare the difference, no different

Revert data from one-hot-vector back :

import numpy as np

# revert one-hot-encoder back to the original formatnp.argmax(y_pred1[1500], axis=0)  # revert one-hot-encoder back to normal digit

1

If anything is unclear for you, please leave a msg.