Wiki Report for ICP 11 - NagaSurendraBethapudi/Python-ICP GitHub Wiki

Video Link : https://drive.google.com/file/d/1R8FtRQP7L13345J-XSAmIdweB8PDs2gH/view?usp=sharing

---

Question 1 :

By using the source code in https://umkc.box.com/s/0jkz2eljon8v374xgy1f6b4ooni0bx3t , Follow the instruction below and then report how the performance changed.

• Applied all of them

1. Convolutional input layer, 32 feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(32, (3, 3), input_shape=(32, 32, 3), padding='same', activation='relu', kernel_constraint=maxnorm(3)))

2. Dropout layer at 20%.

model.add(Dropout(0.2))

3. Convolutional layer, 32 feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(32, (3, 3), activation='relu', padding='same', kernel_constraint=maxnorm(3)))

4. Max Pool layer with size 2×2.

model.add(MaxPooling2D(pool_size=(2, 2)))

5. Convolutional layer, 64 feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(64, (3, 3), activation='relu', padding='same', kernel_constraint=maxnorm(3)))

6. Dropout layer at 20%.

model.add(Dropout(0.2))

7. Convolutional layer, 64 feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(64, (3, 3), activation='relu', padding='same', kernel_constraint=maxnorm(3)))

8. Max Pool layer with size 2×2.

model.add(MaxPooling2D(pool_size=(2, 2)))

9. Convolutional layer, 128feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(128, (3, 3), activation='relu', padding='same', kernel_constraint=maxnorm(3)))

10. Dropout layer at 20%.

model.add(Dropout(0.2))

11. Convolutional layer,128 feature maps with a size of 3×3 and a rectifier activation function.

model.add(Conv2D(128, (3, 3), activation='relu', padding='same', kernel_constraint=maxnorm(3)))

12. Max Pool layer with size 2×2.

model.add(MaxPooling2D(pool_size=(2, 2)))

13. Flatten layer.

model.add(Flatten())

14. Dropout layer at 20%.

model.add(Dropout(0.2))

15. Fully connected layer with 1024units and a rectifier activation function.

model.add(Dense(1024, activation='relu', kernel_constraint=maxnorm(3)))

16. Dropout layerat 20%.

model.add(Dropout(0.2))

17. Fully connected layer with 512units and a rectifier activation function.

model.add(Dense(512, activation='relu', kernel_constraint=maxnorm(3)))

18. Dropoutlayer at 20%.

model.add(Dropout(0.2))

19. Fully connected output layer with 10 units and a Softmax activation function

model.add(Dense(10, activation='softmax'))

Accuracy and Loss of the model after adding layers

1. Loss of the model is decreased by adding more layers.
2. Accuracy of the model is increased by adding more layers.

---

Question 2 :

Predict the first 4 images of the test data using the above model. Then, compare with the actual label for those 4 images to check whether or not the model has predicted correctly.

Explanation :

By using the above model and by using below logic predicted the labels

# Predict function will predict the number of images as per the request by user
def predict(x):
    for i in range(x):
        image = X_test[i] #assigning images to image
        plt.imshow(image)
        plt.show() #displaying the image
        print('Actual label    : ', y_test[i])
        print('Predicted label : ' , model.predict_classes(image.reshape(1, 32, 32, 3)))
        i = i + 1 #iterating for exit when it reaches max

predict(4) #predicting first four images

Model detected all labels correctly

---

Question 3 :

Visualize Loss and Accuracy using the history object.

Explanation :

plotted the loss and accuracy plots

Observations:

After adding more layers accuracy was increased and loss was decreased.

---

Bonus Question :

Explanation :

Saved the model and reloaded the model by using below logics

model.save('convolution_model.h5')
model = load_model('convolution_model.h5')

Changed code a little bit for finding the labels, this time in output how much percent model is predicted the image is displayed .

predictions = model.predict(X_test) #passing all the test data to the model for predictions

#Predict_image function will take below inputs:
# i - no of images to be passed
# predictions_array - no of images to be predicted
# actual label - y_test
# img - X_test

def predict_image(i, predictions_array, actual_label, img):
    predictions_array, actual_label, img = predictions_array, actual_label[i], img[i]
    predicted_label = np.argmax(predictions_array) #predicting the labels
    plt.xlabel(labels[int(actual_label)]) #naming the xlabel with name of the label
    print('Actual label    : ', y_test[i], ' --- ', labels[int(actual_label)] )
    print('Predicted label : ' , model.predict_classes(img.reshape(1, 32, 32, 3)), ' --- ' , f"{labels[int(predicted_label)]}")
    print('Model detected the image as' , labels[int(actual_label)], 'with' , f"{100*np.max(predictions_array):2.0f}%",'accuracy')
    plt.imshow(img)
    plt.show()
    
#predict function will call the predict_image function as per the requested times
    
def predict(x):
    for i in range(x):
        predict_image(i,predictions[i], y_test, X_test)
        i = i + 1

predict(4)

Output :

---

Learnings :

Learned about Convolution Neural Networks and also about traditional vs convolutional .

Challenges :

Every thing looks good.