CH06_SEC05_1_DeepCNN.ipynb

Kernel: Python 3

import numpy as np
import matplotlib.pyplot as plt
from matplotlib import rcParams
from scipy import io
import os
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Dense, Conv2D, Flatten, MaxPool2D
from keras import optimizers


rcParams.update({'font.size': 18})
plt.rcParams['figure.figsize'] = [12, 12]

letters_train_mat = io.loadmat(os.path.join('..','DATA','lettersTrainSet.mat'))
letters_test_mat = io.loadmat(os.path.join('..','DATA','lettersTestSet.mat'))
XTrain = letters_train_mat['XTrain']
TTrain = letters_train_mat['TTrain_cell']
XTest = letters_test_mat['XTest']
TTest = letters_test_mat['TTest_cell']

perm = np.random.permutation(1500)[:20]


# By default, Keras expects data in form (batch, height, width, channels) 
XTrain = np.transpose(XTrain,axes=[3,0,1,2])
XTest = np.transpose(XTest,axes=[3,0,1,2])



fig,axs = plt.subplots(4,5)
axs = axs.reshape(-1)

for j in range(len(axs)):
    axs[j].imshow(np.squeeze(XTrain[perm[j],:,:,:]),cmap='gray')
    axs[j].axis('off')

classes = np.unique(TTrain)
y_train_label = np.zeros_like(TTrain)
y_test_label = np.zeros_like(TTest)
for nc in range(len(classes)):
    y_train_label[TTrain == classes[nc]] = nc
    y_test_label[TTest == classes[nc]] = nc

y_train_label = y_train_label.reshape(-1)
y_test_label = y_test_label.reshape(-1)
    
#one-hot encode categorical classes
y_train = to_categorical(y_train_label)
y_test = to_categorical(y_test_label)

#create model
model = Sequential()

#add model layers
model.add(Conv2D(filters=16, kernel_size=5, activation='relu', input_shape=(28,28,1)))
model.add(MaxPool2D(pool_size=2, strides=2))
model.add(Flatten())
model.add(Dense(len(classes), activation='softmax'))

sgd_optimizer = optimizers.SGD(momentum=0.9)
model.compile(optimizer=sgd_optimizer, loss='categorical_crossentropy')
model.fit(XTrain, y_train, epochs=30)

Epoch 1/30
1500/1500 [==============================] - 0s 183us/step - loss: 0.4938
Epoch 2/30
1500/1500 [==============================] - 0s 66us/step - loss: 0.0512
Epoch 3/30
1500/1500 [==============================] - 0s 66us/step - loss: 0.0173
Epoch 4/30
1500/1500 [==============================] - 0s 69us/step - loss: 0.0108
Epoch 5/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0080
Epoch 6/30
1500/1500 [==============================] - 0s 67us/step - loss: 0.0060
Epoch 7/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0050
Epoch 8/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0040
Epoch 9/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0035
Epoch 10/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0030
Epoch 11/30
1500/1500 [==============================] - 0s 66us/step - loss: 0.0026
Epoch 12/30
1500/1500 [==============================] - 0s 72us/step - loss: 0.0025
Epoch 13/30
1500/1500 [==============================] - 0s 66us/step - loss: 0.0021
Epoch 14/30
1500/1500 [==============================] - 0s 69us/step - loss: 0.0019
Epoch 15/30
1500/1500 [==============================] - 0s 67us/step - loss: 0.0018
Epoch 16/30
1500/1500 [==============================] - 0s 62us/step - loss: 0.0017
Epoch 17/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0015
Epoch 18/30
1500/1500 [==============================] - 0s 67us/step - loss: 0.0015
Epoch 19/30
1500/1500 [==============================] - 0s 67us/step - loss: 0.0014
Epoch 20/30
1500/1500 [==============================] - 0s 72us/step - loss: 0.0013
Epoch 21/30
1500/1500 [==============================] - 0s 71us/step - loss: 0.0012
Epoch 22/30
1500/1500 [==============================] - 0s 78us/step - loss: 0.0011
Epoch 23/30
1500/1500 [==============================] - 0s 68us/step - loss: 0.0011
Epoch 24/30
1500/1500 [==============================] - 0s 70us/step - loss: 0.0010
Epoch 25/30
1500/1500 [==============================] - 0s 68us/step - loss: 9.5643e-04
Epoch 26/30
1500/1500 [==============================] - 0s 66us/step - loss: 9.0637e-04
Epoch 27/30
1500/1500 [==============================] - 0s 70us/step - loss: 8.7863e-04
Epoch 28/30
1500/1500 [==============================] - 0s 75us/step - loss: 8.3013e-04
Epoch 29/30
1500/1500 [==============================] - 0s 75us/step - loss: 7.9356e-04
Epoch 30/30
1500/1500 [==============================] - 0s 75us/step - loss: 7.6706e-04

<keras.callbacks.History at 0x12140b09940>

YPredict = np.argmax(model.predict(XTest),axis=1)
# argmax reverses the one-hot encoding scheme

accuracy = np.sum(YPredict == y_test_label)/len(y_test_label)
print('Accuracy = {}'.format(accuracy))

Accuracy = 0.8853333333333333