Keras Implementation¶

Materials

We will use CIFAR-10 dataset to build a CNN image classifier. CIFAR-10 dataset has 10 different labels

• Airplane
• Automobile
• Bird
• Cat
• Deer
• Dog
• Frog
• Horse
• Ship
• Truck

It has 50,000 training data and 10,000 testing image data. Image size in CIFAR-10 is 32 x 32 x 3. It comes with Keras library

Check GPU¶

In [2]:

import tensorflow as tf
device_name = tf.test.gpu_device_name()
if device_name != '/device:GPU:0':
  raise SystemError('GPU device not found')
print('Found GPU at: {}'.format(device_name))

import tensorflow as tf device_name = tf.test.gpu_device_name() if device_name != '/device:GPU:0': raise SystemError('GPU device not found') print('Found GPU at: {}'.format(device_name))

Found GPU at: /device:GPU:0

Load libraries¶

In [3]:

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import keras
from keras.models import Sequential
from keras.utils import to_categorical
from keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten

import numpy as np import matplotlib.pyplot as plt %matplotlib inline import keras from keras.models import Sequential from keras.utils import to_categorical from keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten

Data Processing¶

In [4]:

from keras.datasets import cifar10
(train_images, train_labels), (test_images, test_labels) = cifar10.load_data()
print('Training data shape : ', train_images.shape, train_labels.shape)

print('Testing data shape : ', test_images.shape, test_labels.shape)

from keras.datasets import cifar10 (train_images, train_labels), (test_images, test_labels) = cifar10.load_data() print('Training data shape : ', train_images.shape, train_labels.shape) print('Testing data shape : ', test_images.shape, test_labels.shape)

Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170498071/170498071 [==============================] - 15s 0us/step
Training data shape :  (50000, 32, 32, 3) (50000, 1)
Testing data shape :  (10000, 32, 32, 3) (10000, 1)

In [5]:

# Find the unique numbers from the train labels
classes = np.unique(train_labels)
nClasses = len(classes)
print('Total number of outputs : ', nClasses)
print('Output classes : ', classes)

# Find the unique numbers from the train labels classes = np.unique(train_labels) nClasses = len(classes) print('Total number of outputs : ', nClasses) print('Output classes : ', classes)

Total number of outputs :  10
Output classes :  [0 1 2 3 4 5 6 7 8 9]

In [6]:

plt.figure(figsize=[4,2])

# Display the first image in training data
plt.subplot(121)
plt.imshow(train_images[0,:,:], cmap='gray')
plt.title("Ground Truth : {}".format(train_labels[0]))

# Display the first image in testing data
plt.subplot(122)
plt.imshow(test_images[0,:,:], cmap='gray')
plt.title("Ground Truth : {}".format(test_labels[0]))

plt.figure(figsize=[4,2]) # Display the first image in training data plt.subplot(121) plt.imshow(train_images[0,:,:], cmap='gray') plt.title("Ground Truth : {}".format(train_labels[0])) # Display the first image in testing data plt.subplot(122) plt.imshow(test_images[0,:,:], cmap='gray') plt.title("Ground Truth : {}".format(test_labels[0]))

Out[6]:

Text(0.5, 1.0, 'Ground Truth : [3]')

Reshape data¶

In [7]:

nRows,nCols,nDims = train_images.shape[1:]
train_data = train_images.reshape(train_images.shape[0], nRows, nCols, nDims)
test_data = test_images.reshape(test_images.shape[0], nRows, nCols, nDims)
input_shape = (nRows, nCols, nDims)

train_data = train_data.astype('float32')
test_data = test_data.astype('float32')

nRows,nCols,nDims = train_images.shape[1:] train_data = train_images.reshape(train_images.shape[0], nRows, nCols, nDims) test_data = test_images.reshape(test_images.shape[0], nRows, nCols, nDims) input_shape = (nRows, nCols, nDims) train_data = train_data.astype('float32') test_data = test_data.astype('float32')

Normalize the data¶

In [8]:

train_data /= 255
test_data /= 255

train_labels_one_hot = to_categorical(train_labels)
test_labels_one_hot = to_categorical(test_labels)

train_data /= 255 test_data /= 255 train_labels_one_hot = to_categorical(train_labels) test_labels_one_hot = to_categorical(test_labels)

In [9]:

## Display the change for category label using one-hot encoding.
print('Original label 0 : ', train_labels[0])
print('After conversion to categorical ( one-hot ) : ', train_labels_one_hot[0])

## Display the change for category label using one-hot encoding. print('Original label 0 : ', train_labels[0]) print('After conversion to categorical ( one-hot ) : ', train_labels_one_hot[0])

Original label 0 :  [6]
After conversion to categorical ( one-hot ) :  [0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]

Create Model¶

• We will add up Convo layers followed by pooling layers.
• Then we will connect Dense(FC) layer to predict the classes.
• Input data fed to first Convo layer, output of that Convo layer acts as input for next Convo layer and so on.
• Finally data is fed to FC layer which try to predict the correct labels.

In [10]:

def createModel():
    model = Sequential()
    # The first two layers with 32 filters of window size 3x3
    model.add(Conv2D(32, (3, 3), padding='same', activation='relu', input_shape=input_shape))
    model.add(Conv2D(32, (3, 3), activation='relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dropout(0.25))

    model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))
    model.add(Conv2D(64, (3, 3), activation='relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dropout(0.25))

    model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))
    model.add(Conv2D(64, (3, 3), activation='relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dropout(0.25))

    model.add(Flatten())
    model.add(Dense(512, activation='relu'))
    model.add(Dropout(0.5))
    model.add(Dense(nClasses, activation='softmax'))
    
    return model

def createModel(): model = Sequential() # The first two layers with 32 filters of window size 3x3 model.add(Conv2D(32, (3, 3), padding='same', activation='relu', input_shape=input_shape)) model.add(Conv2D(32, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Conv2D(64, (3, 3), padding='same', activation='relu')) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Conv2D(64, (3, 3), padding='same', activation='relu')) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(512, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(nClasses, activation='softmax')) return model

Initialize all parameters and specify loss function and optimizer¶

In [11]:

model1 = createModel()
batch_size = 256
epochs = 50
model1.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

model1 = createModel() batch_size = 256 epochs = 50 model1.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

model.summary() is used to see all parameters and shapes in each layers in our models.
You can observe that total parameters are 276, 138 and total trainable parameters are 276, 138.
Non-trainable parameter is 0.

In [12]:

model1.summary()

model1.summary()

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 32, 32, 32)        896       
                                                                 
 conv2d_1 (Conv2D)           (None, 30, 30, 32)        9248      
                                                                 
 max_pooling2d (MaxPooling2D  (None, 15, 15, 32)       0         
 )                                                               
                                                                 
 dropout (Dropout)           (None, 15, 15, 32)        0         
                                                                 
 conv2d_2 (Conv2D)           (None, 15, 15, 64)        18496     
                                                                 
 conv2d_3 (Conv2D)           (None, 13, 13, 64)        36928     
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 6, 6, 64)         0         
 2D)                                                             
                                                                 
 dropout_1 (Dropout)         (None, 6, 6, 64)          0         
                                                                 
 conv2d_4 (Conv2D)           (None, 6, 6, 64)          36928     
                                                                 
 conv2d_5 (Conv2D)           (None, 4, 4, 64)          36928     
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 2, 2, 64)         0         
 2D)                                                             
                                                                 
 dropout_2 (Dropout)         (None, 2, 2, 64)          0         
                                                                 
 flatten (Flatten)           (None, 256)               0         
                                                                 
 dense (Dense)               (None, 512)               131584    
                                                                 
 dropout_3 (Dropout)         (None, 512)               0         
                                                                 
 dense_1 (Dense)             (None, 10)                5130      
                                                                 
=================================================================
Total params: 276,138
Trainable params: 276,138
Non-trainable params: 0
_________________________________________________________________

Fit data in¶

In [13]:

history = model1.fit(train_data, 
                     train_labels_one_hot, 
                     batch_size=batch_size, 
                     epochs=epochs, 
                     verbose=1, 
                     validation_data=(test_data, test_labels_one_hot))

model1.evaluate(test_data, test_labels_one_hot)

history = model1.fit(train_data, train_labels_one_hot, batch_size=batch_size, epochs=epochs, verbose=1, validation_data=(test_data, test_labels_one_hot)) model1.evaluate(test_data, test_labels_one_hot)

Epoch 1/50
196/196 [==============================] - 23s 21ms/step - loss: 2.0635 - accuracy: 0.2349 - val_loss: 1.7800 - val_accuracy: 0.3685
Epoch 2/50
196/196 [==============================] - 3s 13ms/step - loss: 1.7274 - accuracy: 0.3723 - val_loss: 1.7093 - val_accuracy: 0.3618
Epoch 3/50
196/196 [==============================] - 3s 14ms/step - loss: 1.5212 - accuracy: 0.4532 - val_loss: 1.3849 - val_accuracy: 0.5216
Epoch 4/50
196/196 [==============================] - 3s 13ms/step - loss: 1.3829 - accuracy: 0.5051 - val_loss: 1.3491 - val_accuracy: 0.5238
Epoch 5/50
196/196 [==============================] - 3s 13ms/step - loss: 1.2761 - accuracy: 0.5458 - val_loss: 1.1934 - val_accuracy: 0.5774
Epoch 6/50
196/196 [==============================] - 3s 13ms/step - loss: 1.1873 - accuracy: 0.5760 - val_loss: 1.3081 - val_accuracy: 0.5609
Epoch 7/50
196/196 [==============================] - 3s 13ms/step - loss: 1.1227 - accuracy: 0.6057 - val_loss: 1.1414 - val_accuracy: 0.6008
Epoch 8/50
196/196 [==============================] - 3s 13ms/step - loss: 1.0520 - accuracy: 0.6304 - val_loss: 1.0671 - val_accuracy: 0.6210
Epoch 9/50
196/196 [==============================] - 3s 13ms/step - loss: 0.9989 - accuracy: 0.6498 - val_loss: 0.9351 - val_accuracy: 0.6670
Epoch 10/50
196/196 [==============================] - 3s 13ms/step - loss: 0.9496 - accuracy: 0.6690 - val_loss: 0.8911 - val_accuracy: 0.6865
Epoch 11/50
196/196 [==============================] - 3s 13ms/step - loss: 0.9022 - accuracy: 0.6851 - val_loss: 0.8515 - val_accuracy: 0.7009
Epoch 12/50
196/196 [==============================] - 3s 13ms/step - loss: 0.8585 - accuracy: 0.6996 - val_loss: 0.9883 - val_accuracy: 0.6675
Epoch 13/50
196/196 [==============================] - 3s 14ms/step - loss: 0.8270 - accuracy: 0.7120 - val_loss: 0.7599 - val_accuracy: 0.7358
Epoch 14/50
196/196 [==============================] - 3s 14ms/step - loss: 0.7909 - accuracy: 0.7251 - val_loss: 0.8385 - val_accuracy: 0.7140
Epoch 15/50
196/196 [==============================] - 3s 13ms/step - loss: 0.7656 - accuracy: 0.7326 - val_loss: 0.7481 - val_accuracy: 0.7385
Epoch 16/50
196/196 [==============================] - 3s 14ms/step - loss: 0.7421 - accuracy: 0.7410 - val_loss: 0.8492 - val_accuracy: 0.7176
Epoch 17/50
196/196 [==============================] - 3s 13ms/step - loss: 0.7221 - accuracy: 0.7511 - val_loss: 0.6704 - val_accuracy: 0.7696
Epoch 18/50
196/196 [==============================] - 3s 13ms/step - loss: 0.7005 - accuracy: 0.7556 - val_loss: 0.7235 - val_accuracy: 0.7519
Epoch 19/50
196/196 [==============================] - 3s 13ms/step - loss: 0.6782 - accuracy: 0.7647 - val_loss: 0.7859 - val_accuracy: 0.7450
Epoch 20/50
196/196 [==============================] - 3s 14ms/step - loss: 0.6688 - accuracy: 0.7704 - val_loss: 0.7000 - val_accuracy: 0.7591
Epoch 21/50
196/196 [==============================] - 3s 14ms/step - loss: 0.6495 - accuracy: 0.7747 - val_loss: 0.8003 - val_accuracy: 0.7359
Epoch 22/50
196/196 [==============================] - 3s 14ms/step - loss: 0.6343 - accuracy: 0.7792 - val_loss: 0.6582 - val_accuracy: 0.7777
Epoch 23/50
196/196 [==============================] - 3s 13ms/step - loss: 0.6170 - accuracy: 0.7868 - val_loss: 0.6441 - val_accuracy: 0.7852
Epoch 24/50
196/196 [==============================] - 3s 13ms/step - loss: 0.6101 - accuracy: 0.7902 - val_loss: 0.7041 - val_accuracy: 0.7629
Epoch 25/50
196/196 [==============================] - 3s 14ms/step - loss: 0.5984 - accuracy: 0.7928 - val_loss: 0.7197 - val_accuracy: 0.7627
Epoch 26/50
196/196 [==============================] - 3s 13ms/step - loss: 0.5909 - accuracy: 0.7960 - val_loss: 0.6819 - val_accuracy: 0.7719
Epoch 27/50
196/196 [==============================] - 3s 13ms/step - loss: 0.5738 - accuracy: 0.8020 - val_loss: 0.6671 - val_accuracy: 0.7772
Epoch 28/50
196/196 [==============================] - 3s 13ms/step - loss: 0.5652 - accuracy: 0.8044 - val_loss: 0.6551 - val_accuracy: 0.7806
Epoch 29/50
196/196 [==============================] - 4s 19ms/step - loss: 0.5619 - accuracy: 0.8059 - val_loss: 0.6860 - val_accuracy: 0.7753
Epoch 30/50
196/196 [==============================] - 4s 20ms/step - loss: 0.5555 - accuracy: 0.8077 - val_loss: 0.6611 - val_accuracy: 0.7873
Epoch 31/50
196/196 [==============================] - 4s 20ms/step - loss: 0.5438 - accuracy: 0.8097 - val_loss: 0.6509 - val_accuracy: 0.7866
Epoch 32/50
196/196 [==============================] - 4s 20ms/step - loss: 0.5352 - accuracy: 0.8161 - val_loss: 0.6431 - val_accuracy: 0.7892
Epoch 33/50
196/196 [==============================] - 4s 21ms/step - loss: 0.5333 - accuracy: 0.8176 - val_loss: 0.6354 - val_accuracy: 0.7891
Epoch 34/50
196/196 [==============================] - 4s 23ms/step - loss: 0.5279 - accuracy: 0.8176 - val_loss: 0.6336 - val_accuracy: 0.7940
Epoch 35/50
196/196 [==============================] - 5s 24ms/step - loss: 0.5218 - accuracy: 0.8202 - val_loss: 0.6622 - val_accuracy: 0.7851
Epoch 36/50
196/196 [==============================] - 5s 23ms/step - loss: 0.5197 - accuracy: 0.8204 - val_loss: 0.7469 - val_accuracy: 0.7623
Epoch 37/50
196/196 [==============================] - 5s 24ms/step - loss: 0.5127 - accuracy: 0.8248 - val_loss: 0.7208 - val_accuracy: 0.7579
Epoch 38/50
196/196 [==============================] - 5s 23ms/step - loss: 0.5022 - accuracy: 0.8273 - val_loss: 0.7328 - val_accuracy: 0.7765
Epoch 39/50
196/196 [==============================] - 5s 23ms/step - loss: 0.5008 - accuracy: 0.8272 - val_loss: 0.6307 - val_accuracy: 0.7917
Epoch 40/50
196/196 [==============================] - 5s 24ms/step - loss: 0.5010 - accuracy: 0.8279 - val_loss: 0.6767 - val_accuracy: 0.7827
Epoch 41/50
196/196 [==============================] - 4s 22ms/step - loss: 0.4983 - accuracy: 0.8272 - val_loss: 0.5953 - val_accuracy: 0.8034
Epoch 42/50
196/196 [==============================] - 4s 19ms/step - loss: 0.4932 - accuracy: 0.8306 - val_loss: 0.6459 - val_accuracy: 0.7940
Epoch 43/50
196/196 [==============================] - 4s 22ms/step - loss: 0.4859 - accuracy: 0.8322 - val_loss: 0.6623 - val_accuracy: 0.8014
Epoch 44/50
196/196 [==============================] - 4s 21ms/step - loss: 0.4866 - accuracy: 0.8341 - val_loss: 0.6432 - val_accuracy: 0.7969
Epoch 45/50
196/196 [==============================] - 5s 24ms/step - loss: 0.4788 - accuracy: 0.8355 - val_loss: 0.8475 - val_accuracy: 0.7470
Epoch 46/50
196/196 [==============================] - 4s 22ms/step - loss: 0.4803 - accuracy: 0.8332 - val_loss: 0.7082 - val_accuracy: 0.7832
Epoch 47/50
196/196 [==============================] - 5s 23ms/step - loss: 0.4734 - accuracy: 0.8353 - val_loss: 0.6438 - val_accuracy: 0.7987
Epoch 48/50
196/196 [==============================] - 4s 22ms/step - loss: 0.4705 - accuracy: 0.8363 - val_loss: 0.6317 - val_accuracy: 0.8033
Epoch 49/50
196/196 [==============================] - 5s 24ms/step - loss: 0.4670 - accuracy: 0.8402 - val_loss: 0.7534 - val_accuracy: 0.7684
Epoch 50/50
196/196 [==============================] - 4s 22ms/step - loss: 0.4637 - accuracy: 0.8415 - val_loss: 0.6584 - val_accuracy: 0.7933
313/313 [==============================] - 1s 3ms/step - loss: 0.6584 - accuracy: 0.7933

Out[13]:

[0.6583714485168457, 0.7932999730110168]

Evaluate results¶

In [14]:

plt.figure(figsize=[8,6])
plt.plot(history.history['loss'],'r',linewidth=3.0)
plt.plot(history.history['val_loss'],'b',linewidth=3.0)
plt.legend(['Training loss', 'Validation Loss'],fontsize=18)
plt.xlabel('Epochs ',fontsize=16)
plt.ylabel('Loss',fontsize=16)
plt.title('Loss Curves',fontsize=16)

plt.figure(figsize=[8,6]) plt.plot(history.history['loss'],'r',linewidth=3.0) plt.plot(history.history['val_loss'],'b',linewidth=3.0) plt.legend(['Training loss', 'Validation Loss'],fontsize=18) plt.xlabel('Epochs ',fontsize=16) plt.ylabel('Loss',fontsize=16) plt.title('Loss Curves',fontsize=16)

Out[14]:

Text(0.5, 1.0, 'Loss Curves')

In [18]:

plt.figure(figsize=[8,6])
plt.plot(history.history['accuracy'],'r',linewidth=3.0)
plt.plot(history.history['val_accuracy'],'b',linewidth=3.0)
plt.legend(['Training Accuracy', 'Validation Accuracy'],fontsize=18)
plt.xlabel('Epochs ',fontsize=16)
plt.ylabel('Accuracy',fontsize=16)
plt.title('Accuracy Curves',fontsize=16)

plt.figure(figsize=[8,6]) plt.plot(history.history['accuracy'],'r',linewidth=3.0) plt.plot(history.history['val_accuracy'],'b',linewidth=3.0) plt.legend(['Training Accuracy', 'Validation Accuracy'],fontsize=18) plt.xlabel('Epochs ',fontsize=16) plt.ylabel('Accuracy',fontsize=16) plt.title('Accuracy Curves',fontsize=16)

Out[18]:

Text(0.5, 1.0, 'Accuracy Curves')