Problem statement :
In this Section we are implementing Convolution Neural Network(CNN) Classifier for Classifying dog and cat images. The Total number of images available for training is 25,000 and final testing is done on seperate 10,000 images. Note:This problem statement and dataset is taken from this Kaggle competition. Dependencies
Google colab
Tensorflow 1.10
Matplotlib
Seaborn
Scikit-Learn
Pandas
Numpy
cv
Install dependencies Test Train Split
The image training set contains 20,000 images for each category. I split those into 80% train and 20% means test Split each class image into 18,000 for train and 2000 for the test.
Architecture
import tensorflow as tf from tensorflow import keras from keras import Sequential from keras.layers import Dense,Conv2D,MaxPooling2D,Flatten,BatchNormalization,Dropout
train_ds = keras.utils.image_dataset_from_directory( directory = '/content/train', labels='inferred', label_mode = 'int', batch_size=32, image_size=(256,256) )
validation_ds = keras.utils.image_dataset_from_directory( directory = '/content/test', labels='inferred', label_mode = 'int', batch_size=32, image_size=(256,256) )
def process(image,label): image = tf.cast(image/255. ,tf.float32) return image,label
train_ds = train_ds.map(process) validation_ds = validation_ds.map(process)
Rectifier Linear Unit
Adam optimizer
Sigmoid on Final output
Binary CrossEntropy loss
model = Sequential()
model.add(Conv2D(32,kernel_size=(3,3),padding='valid',activation='relu',input_shape=(256,256,3))) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2,2),strides=2,padding='valid'))
model.add(Conv2D(64,kernel_size=(3,3),padding='valid',activation='relu')) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2,2),strides=2,padding='valid'))
model.add(Conv2D(128,kernel_size=(3,3),padding='valid',activation='relu')) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2,2),strides=2,padding='valid'))
model.add(Flatten())
model.add(Dense(128,activation='relu')) model.add(Dropout(0.1)) model.add(Dense(64,activation='relu')) model.add(Dropout(0.1)) model.add(Dense(1,activation='sigmoid'))
Using some Data Augmentation techniques for more data and Better results.
Shearing of images
Random zoom
Horizontal flips
from tensorflow.keras.preprocessing.image import ImageDataGenerator batch_size=32 train_datagene=ImageDataGenerator(rescale=1./255, shear_range=0.2, zoom_range=0.2, horizontal_flip=True) test_datagene=ImageDataGenerator(rescale=1./255) test_datagene test_datagen = ImageDataGenerator(rescale=1./255)
#Training Set
train_generate = train_datagene.flow_from_directory( directory = '/content/train', batch_size=32, target_size=(256,256), class_mode='binary' )
valid_generate = train_datagene.flow_from_directory( directory = '/content/test', batch_size=32, target_size=(256,256), class_mode='binary' )
%%capture classifier.fit_generator(train_set, steps_per_epoch=800, epochs = 200, validation_data = test_set, validation_steps = 20, #callbacks=[tensorboard] );
#Some Helpful Instructions:
#finetune you network parameter in last by using low learning rate like 0.00001 #classifier.save('resources/dogcat_model_bak.h5') #from tensorflow.keras.models import load_model #model = load_model('partial_trained1') #100 iteration with learning rate 0.001 and after that 0.0001
from tensorflow.keras.models import load_model classifier = load_model('resources/dogcat_model_bak.h5')
Prediction of Single Image
#Prediction of image %matplotlib inline import tensorflow from tensorflow.keras.preprocessing import image import matplotlib.pyplot as plt import numpy as np img1 = image.load_img('test/Cat/10.jpg', target_size=(64, 64)) img = image.img_to_array(img1) img = img/255
img = np.expand_dims(img, axis=0) prediction = classifier.predict(img, batch_size=None,steps=1) #gives all class prob. if(prediction[:,:]>0.5): value ='Dog :%1.2f'%(prediction[0,0]) plt.text(20, 62,value,color='red',fontsize=18,bbox=dict(facecolor='white',alpha=0.8)) else: value ='Cat :%1.2f'%(1.0-prediction[0,0]) plt.text(20, 62,value,color='red',fontsize=18,bbox=dict(facecolor='white',alpha=0.8))
plt.imshow(img1) plt.show()
png
import pandas as pd test_set.reset ytesthat = classifier.predict_generator(test_set) df = pd.DataFrame({ 'filename':test_set.filenames, 'predict':ytesthat[:,0], 'y':test_set.classes })
pd.set_option('display.float_format', lambda x: '%.5f' % x) df['y_pred'] = df['predict']>0.5 df.y_pred = df.y_pred.astype(int) df.head(10)
filename predict y y_pred
0 Cat/0.jpg 0.00000 0 0 1 Cat/1.jpg 0.00000 0 0 2 Cat/10.jpg 0.00000 0 0 3 Cat/100.jpg 0.99970 0 1 4 Cat/10001.jpg 0.00000 0 0 5 Cat/10009.jpg 0.02340 0 0 6 Cat/1001.jpg 0.00001 0 0 7 Cat/10012.jpg 0.00000 0 0 8 Cat/10017.jpg 0.00000 0 0 9 Cat/10018.jpg 0.00000 0 0
misclassified = df[df['y']!=df['y_pred']] print('Total misclassified image from 5000 Validation images : %d'%misclassified['y'].count())
Total misclassified image from 5000 Validation images : 122
#Prediction of test set from sklearn.metrics import confusion_matrix import matplotlib.pyplot as plt import seaborn as sns
conf_matrix = confusion_matrix(df.y,df.y_pred) sns.heatmap(conf_matrix,cmap="YlGnBu",annot=True,fmt='g'); plt.xlabel('predicted value') plt.ylabel('true value');
png
#Some of Cat image misclassified as Dog. import matplotlib.image as mpimg
CatasDog = df['filename'][(df.y==0)&(df.y_pred==1)] fig=plt.figure(figsize=(15, 6)) columns = 7 rows = 3 for i in range(columns*rows): #img = mpimg.imread() img = image.load_img('test/'+CatasDog.iloc[i], target_size=(64, 64)) fig.add_subplot(rows, columns, i+1) plt.imshow(img)
plt.show()
png
#Some of Dog image misclassified as Cat. import matplotlib.image as mpimg
DogasCat = df['filename'][(df.y==1)&(df.y_pred==0)] fig=plt.figure(figsize=(15, 6)) columns = 7 rows = 3 for i in range(columns*rows): #img = mpimg.imread() img = image.load_img('test/'+DogasCat.iloc[i], target_size=(64, 64)) fig.add_subplot(rows, columns, i+1) plt.imshow(img) plt.show()
png
classifier.summary()
conv2d_6 (Conv2D) (None, 62, 62, 32) 896
max_pooling2d_6 (MaxPooling2 (None, 31, 31, 32) 0
conv2d_7 (Conv2D) (None, 29, 29, 32) 9248
max_pooling2d_7 (MaxPooling2 (None, 14, 14, 32) 0
flatten_3 (Flatten) (None, 6272) 0
dense_6 (Dense) (None, 128) 802944
Total params: 813,217 Trainable params: 813,217 Non-trainable params: 0
Visualization of Layers Ouptut
#Input Image for Layer visualization img1 = image.load_img('test/Cat/14.jpg') plt.imshow(img1); #preprocess image img1 = image.load_img('test/Cat/14.jpg', target_size=(64, 64)) img = image.img_to_array(img1) img = img/255 img = np.expand_dims(img, axis=0)
png
model_layers = [ layer.name for layer in classifier.layers] print('layer name : ',model_layers)
layer name : ['conv2d_6', 'max_pooling2d_6', 'conv2d_7', 'max_pooling2d_7', 'flatten_3', 'dense_6', 'dense_7']
from tensorflow.keras.models import Model conv2d_6_output = Model(inputs=classifier.input, outputs=classifier.get_layer('conv2d_6').output) conv2d_7_output = Model(inputs=classifier.input,outputs=classifier.get_layer('conv2d_7').output)
conv2d_6_features = conv2d_6_output.predict(img) conv2d_7_features = conv2d_7_output.predict(img) print('First conv layer feature output shape : ',conv2d_6_features.shape) print('First conv layer feature output shape : ',conv2d_7_features.shape)
First conv layer feature output shape : (1, 62, 62, 32) First conv layer feature output shape : (1, 29, 29, 32)
Single Convolution Filter Output
plt.imshow(conv2d_6_features[0, :, :, 4], cmap='gray')
<matplotlib.image.AxesImage at 0x7f3b1c90f978>
png First Covolution Layer Output
import matplotlib.image as mpimg
fig=plt.figure(figsize=(14,7)) columns = 8 rows = 4 for i in range(columns*rows): #img = mpimg.imread() fig.add_subplot(rows, columns, i+1) plt.axis('off') plt.title('filter'+str(i)) plt.imshow(conv2d_6_features[0, :, :, i], cmap='gray') plt.show()
png Second Covolution Layer Output
fig=plt.figure(figsize=(14,7)) columns = 8 rows = 4 for i in range(columns*rows): #img = mpimg.imread() fig.add_subplot(rows, columns, i+1) plt.axis('off') plt.title('filter'+str(i)) plt.imshow(conv2d_7_features[0, :, :, i], cmap='gray') plt.show()
png Model Performance on Unseen Data
fig=plt.figure(figsize=(15, 6)) columns = 7 rows = 3 for i in range(columns*rows): fig.add_subplot(rows, columns, i+1) img1 = image.load_img('test1/'+test_set1.filenames[np.random.choice(range(12500))], target_size=(64, 64)) img = image.img_to_array(img1) img = img/255 img = np.expand_dims(img, axis=0) prediction = classifier.predict(img, batch_size=None,steps=1) #gives all class prob. if(prediction[:,:]>0.5): value ='Dog :%1.2f'%(prediction[0,0]) plt.text(20, 58,value,color='red',fontsize=10,bbox=dict(facecolor='white',alpha=0.8)) else: value ='Cat :%1.2f'%(1.0-prediction[0,0]) plt.text(20, 58,value,color='red',fontsize=10,bbox=dict(facecolor='white',alpha=0.8)) plt.imshow(img1)
png
%%capture
x1 = classifier.evaluate_generator(train_set) x2 = classifier.evaluate_generator(test_set)
print('Training Accuracy : %1.2f%% Training loss : %1.6f'%(x1[1]*100,x1[0])) print('Validation Accuracy: %1.2f%% Validation loss: %1.6f'%(x2[1]*100,x2[0]))
Training Accuracy : 99.96% Training loss : 0.002454 Validation Accuracy: 97.56% Validation loss: 0.102678
The Architecture and parameter used in this network are capable of producing accuracy of 97.56% on Validation Data which is pretty good. It is possible to Achieve more accuracy on this dataset using deeper network and fine tuning of network parameters for training.