lOMoARcPSD|56718048

DEEP Learning LAB Manual

deep learning (Madras Institute of Technology, Anna University)

Scan to open on Studocu

Studocu is not sponsored or endorsed by any college or university

Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

DEPARTMENT OF ARTIFICIAL INTELLIGENCE AND DATA

SCIENCE

MASTER RECORD

COURSE CODE : AD3511

COURSE NAME : DEEP LEARNING LABORATORY
YEAR / SEMESTER : III/ V
FACULTY NAME : NITYA J. V
DESIGNATION : ASSISTANT PROFESSOR
DEPARTMENT : AIDS
REGULATION : R2021
ACADEMIC YEAR : 2024-2025 ODD

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

Syllabus

COURSE OBJECTIVES

• To understand the tools and techniques to implement deep neural networks

• To apply different deep learning architectures for solving problems
• To implement generative models for suitable applications
• To learn to build and validate different models

LIST OF EXPERIMENTS:

1. Solving XOR problem using DNN

2. Character recognition using CNN
3. Face recognition using CNN
4. Language modeling using RNN
5. Sentiment analysis using LSTM
6. Parts of speech tagging using Sequence to Sequence architecture
7. Machine Translation using Encoder-Decoder model
8. Image augmentation using GANs
9. Mini-project on real world applications

2
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Course Outcome

CO Details BTL
Apply deep neural network for simple problems (K3)
CO508.1 K3-Apply
Apply Convolution Neural Network for image processing (K3)
CO508.2 K3-Apply
Apply Recurrent Neural Network and its variants for text analysis
CO508.3 (K4) K4-Analyze

CO508.4 Apply generative models for data augmentation (K4) K4-Analyze

Develop real-world solutions using suitable deep neural networks
CO508.5 K3-Apply
(K3)
Demonstrating an attitude at the level of valuing (attaching values
CO508.6 K3-Apply
and expressing personal opinions by analyzing any dataset).

CO - PO Mapping

Program
Program Outcomes Specific
CO
Outcomes

PO1 PO2 PO3 PO4 PO5 PO6 PO7 PO8 PO9 PO10 PO11 PO12 PSO1 PSO2
CO1 3 - - - - - - - 3 3 - - 3 3
CO2 3 - - - 3 - - 2 3 - 2 - 3 3
CO3 3 2 - 2 - 2 - - - 3 - - - -
CO4 3 - - - - - - - - 3 - - 3 3
CO5 3 - - - - - - - - 3 - 2 2 2
CO6 - - - - - 2 - 2 2 2 - 2 - -
3 2 - 2 3 2 - 2 2.7 2.8 2 2 2.8 2.8
CORRELATION
STRONG S/3 MEDIUM M/2 WEAK W/1

3
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

INDEX

Page
Sl. Name of the Experiment CO No.
No.
6
CO508.1
Solving XOR problem using DNN
1
CO508. 9
2 Character recognition using CNN 2
CO508. 15
3 Face recognition using CNN 2
CO508. 22
4 Language modeling using RNN 3
CO508. 27
5 Sentiment analysis using LSTM 3

CO508. 33
6 Parts of speech tagging using Sequence to Sequence architecture 3

CO508. 43
7 Machine Translation using Encoder-Decoder model 3
CO508. 48
8 4
Image augmentation using GANs

Additional Experiments Beyond the Syllabus

Sl. Name of the Experiment CO Page

No. No.
56
Image detection and classification for Traffic Analysis using
9. CNN CO508.5
61
10. Online Fraud Detection of Share Market Data
CO508.6

4
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

RUBRICS: Grading of Laboratory

exercise
Marks
Marks
5 to 3 to 2 1 to 0 awarded for
Criterion (marks)
the Criterion

1. Correctness of All the steps Some steps are followed Steps are not followed.
the Procedure are sequence. but error occurred. Not showing interest to
for the Knows how to Proceeded the experiment do the experiment.
experiment/ proceed the with the guidance.
Exercise (3) experiment.
8 or 7 or 6 5 or 4 or 3 2 or 1 or 0
2. Skills level in Show excellent Show minimal Show no understanding
performing the understanding of the understanding of the of the experiment. All
experiment/ experiment. All data experiment. All data is data is not recorded and
Exercise (8) is recorded and neatly is not Presented neatly.
recorded and is
presented
not presented neatly
3 2 1
3. Inferences Correct inferences Inferences drawn ar Inferences drawn are
drawn from have been drawn and correctly. But e partially incorrectly or
the presented presented no incorrect.
Professionally professionally
experiment/ t
exercise (3)
3 2 1
4. Presentation of Calculation are Calculation are present Wrong calculations.
results (3) presents neatly with neatly with minor Major mistakes.
accurate results. mistakes

5 to 4 3 to 2 1 to 0
5. Clarity in Almost all the Partially answered Unable to answer.
answering viva questions are
questions. (5) answered.
3 2 1
6. Attitude The experiment is The experiment The experiment is not
reflected in completed on time. is Completed on time.
doing the Manual/ Record note completed on time, Observation / Record
experiment / is submitted on time. but the Manual / note is not submitted
exercise (3) Record note is not on time.
submitted on time.
Total marks out of 25

Lab-Incharge Head of the Department

5
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program: 01: Solving XOR problem using DNN

import numpy as np # For matrix math

import [Link] as plt # For plotting
import sys # For printing
# The training data.
X = [Link]([
[0, 1],
[1, 0],
[1, 1],
[0, 0]
])

# The labels for the training data.

y = [Link]([
[1],
[1],
[0],
[0]
])
num_i_units = 2 # Number of Input units
num_h_units = 2 # Number of Hidden
units num_o_units = 1 # Number of Output
units # The learning rate for Gradient
Descent. learning_rate = 0.01
# The parameter to help with overfitting.
reg_param = 0
# Maximum iterations for Gradient
Descent. max_iter = 5000
# Number of training
ex m = 4
[Link](1)
W1 = [Link](0, 1, (num_h_units, num_i_units)) # 2x2
W2 = [Link](0, 1, (num_o_units, num_h_units)) # 1x2
B1 = [Link]((num_h_units, 1)) # 2x1
B2 = [Link]((num_o_units, 1)) # 1x1
def sigmoid(z, derv=False):
if derv: return z * (1 - z)
return 1 / (1 + [Link](-z))
def forward(x, predict=False):
a1 = [Link]([Link][0], 1) # Getting the training example as a
column vector.
z2 = [Link](a1) + B1 # 2x2 * 2x1 + 2x1 = 2x1
a2 = sigmoid(z2) # 2x1
z3 = [Link](a2) + B2 # 1x2 * 2x1 + 1x1 = 1x1
a3 = sigmoid(z3)
if predict: return
a3 return (a1, a2,
a3)

6
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

dW1 = 0 # Gradient for W1

dW2 = 0 # Gradient for W2
dB1 = 0 # Gradient for B1
dB2 = 0 # Gradient for B2
cost = [Link]((max_iter, 1)) # Column vector to record the cost of the
NN after each Gradient Descent iteration.
def train(_W1, _W2, _B1, _B2): # The arguments are to bypass
UnboundLocalError error
for i in
range(max_iter): c =
0
dW1 = 0
dW2 = 0
dB1 = 0
dB2 = 0

for j in range(m):
[Link]("\rIteration: {} and {}".format(i + 1, j + 1))
# Forward Prop.
a0 = X[j].reshape(X[j].shape[0], 1) # 2x1
z1 = _W1.dot(a0) + _B1 # 2x2 * 2x1 + 2x1 = 2x1
a1 = sigmoid(z1) # 2x1
z2 = _W2.dot(a1) + _B2 # 1x2 * 2x1 + 1x1 = 1x1
a2 = sigmoid(z2) # 1x1
# Back prop.
dz2 = a2 - y[j] # 1x1
dW2 += dz2 * a1.T # 1x1 .* 1x2 = 1x2
dz1 = [Link]((_W2.T * dz2), sigmoid(a1, derv=True)) #
(2x1 * 1x1) .* 2x1 = 2x1
dW1 += [Link](a0.T) # 2x1 * 1x2 = 2x2dB2 += dz2 # 1x1

c = c + (-(y[j] * [Link](a2)) - ((1 - y[j]) * [Link](1 - a2)))

[Link]() # Updating the text.

_W1 = _W1 - learning_rate * (dW1 / m) + ( (reg_param / m) *

_W1)
_W2 = _W2 - learning_rate * (dW2 / m) + ( (reg_param / m) *
_W2)

_B1 = _B1 - learning_rate * (dB1 / m)

_B2 = _B2 - learning_rate * (dB2 / m)
cost[i] = (c / m) + (
(reg_param / (2 * m))
*(
[Link]([Link](_W1, 2)) +
[Link]([Link](_W2, 2))
)
)
dB1 += dz1 # 2x1

7
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

return (_W1, _W2, _B1, _B2)

W1, W2, B1, B2 = train(W1, W2, B1, B2)
# Assigning the axes to the different elements.
[Link](range(max_iter), cost)
# Labelling the x axis as the iterations axis.
[Link]("Iterations")
# Labelling the y axis as the cost axis.
[Link]("Cost")
# Showing the
plot. [Link]()

Output:

8
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 02: Character recognition using CNN

pip install opencv-python

pip install keras
pip install tensorflow
import cv2
import numpy as
np import pandas
as pd
import [Link] as plt
from sklearn.model_selection import
train_test_split from [Link] import shuffle
from [Link] import Sequential
from [Link] import Dense, Flatten, Conv2D, MaxPool2D, Dropout
from [Link] import SGD, Adam
from [Link] import ReduceLROnPlateau, EarlyStopping
from [Link] import to_categorical
data = pd.read_csv(r"A_Z Handwritten [Link]").astype('float32')
X = [Link]('0',axis = 1)
y = data['0']
train_x, test_x, train_y, test_y = train_test_split(X, y, test_size =
0.2) train_x = [Link](train_x.values, (train_x.shape[0], 28,28))
test_x = [Link](test_x.values, (test_x.shape[0], 28,28))
word_dict =
{0:'A',1:'B',2:'C',3:'D',4:'E',5:'F',6:'G',7:'H',8:'I',9:'J',10:'K',11:'L',12:'M',13:'N',14:'O',
15:'P',16:'Q',17:'R',18:'S',19:'T',20:'U',21:'V',22:'W',23:'X', 24:'Y',25:'Z'}
y_int = np.int0(y)
count = [Link](26,
dtype='int') for i in y_int:
count[i] +=1

alphabets = []
for i in word_dict.values():

[Link](i)

fig, ax = [Link](1,1, figsize=(10,10))

[Link](alphabets, count)
# naming the x axis
[Link]("Number of elements")
# naming the y axis
[Link]("Alphabets")
# giving a title
9
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

[Link]("Plotting the number of alphabets")

# Turn on the minor TICKS, which are required for the minor GRID
plt.minorticks_on()
# Customize the major grid
[Link](which='major', linestyle='-', linewidth='0.5',
color='red') # Customize the minor grid
[Link](which='minor', linestyle=':', linewidth='0.5', color='black')

[Link]()

uff = shuffle(train_x[:100])
fig, ax = [Link](3,3, figsize = (10,10))
axes = [Link]()
for i in range(9):
_, shu = [Link](shuff[i], 30, 200, cv2.THRESH_BINARY)
axes[i].imshow([Link](shuff[i], (28,28)), cmap=plt.get_cmap('gray'))
[Link]()
# Reshape data for model creation
train_X = train_x.reshape(train_x.shape[0],train_x.shape[1],train_x.shape[2],1)
print("The new shape of train data: ", train_X.shape)

test_X = test_x.reshape(test_x.shape[0], test_x.shape[1], test_x.shape[2],1)

print("The new shape of train data: ", test_X.shape)

train_yOHE = to_categorical(train_y, num_classes = 26, dtype='int')

print("The new shape of train labels: ", train_yOHE.shape)

test_yOHE = to_categorical(test_y, num_classes = 26, dtype='int')

print("The new shape of test labels: ", test_yOHE.shape)
model = Sequential()

[Link](Conv2D(filters=32, kernel_size=(3, 3),

activation='relu', input_shape=(28,28,1)))
[Link](MaxPool2D(pool_size=(2, 2), strides=2))

[Link](Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding

= 'same'))
[Link](MaxPool2D(pool_size=(2, 2), strides=2))

[Link](Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding =

'valid'))
[Link](MaxPool2D(pool_size=(2, 2), strides=2))

10
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

[Link](Flatten())

[Link](Dense(64,activation ="relu"))
[Link](Dense(128,activation ="relu"))

[Link](Dense(26,activation ="softmax"))
[Link](optimizer = Adam(learning_rate=0.001),
loss='categorical_crossentropy', metrics=['accuracy'])

history = [Link](train_X, train_yOHE, epochs=1, validation_data =

(test_X,test_yOHE))
[Link]()
[Link](r'model_hand.h5')

print("The validation accuracy is :", [Link]['val_accuracy'])

print("The training accuracy is :", [Link]['accuracy'])
print("The validation loss is :", [Link]['val_loss'])
print("The training loss is :", [Link]['loss'])
# Prediction on test data
fig, axes = [Link](3,3, figsize=(8,9))
axes = [Link]()

for i,ax in enumerate(axes):

img = [Link](test_X[i], (28,28))
[Link](img, cmap=plt.get_cmap('gray'))

pred = word_dict[[Link](test_yOHE[i])]
ax.set_title("Prediction: "+pred)
# Predection on External Image

img = [Link](r'test_image.jpg')
img_copy = [Link]()

img = [Link](img,
cv2.COLOR_BGR2RGB) img = [Link](img,
(400,440))

img_copy = [Link](img_copy, (7,7), 0)

img_gray = [Link](img_copy, cv2.COLOR_BGR2GRAY)
_, img_thresh = [Link](img_gray, 100, 255, cv2.THRESH_BINARY_INV)
img_final = [Link](img_thresh, (28,28))
img_final =[Link](img_final, (1,28,28,1))

img_pred = word_dict[[Link]([Link](img_final))]

11
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

[Link](img, "Image Data", (100,25), cv2.FONT_HERSHEY_DUPLEX ,

fontScale= 1, thickness=2, color = (255,0,0))
[Link](img, "Character Prediction: " + img_pred, (10,410),
cv2.FONT_HERSHEY_SIMPLEX, fontScale= 1, thickness=2, color = (0,0,255))
[Link]('Character Recognition', img)

while (1):
k = [Link](1) & 0xFF
if k == 27:
break
[Link]()

Output:

12
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

The new shape of train data: (297960, 28, 28, 1)

The new shape of train data: (74490, 28, 28, 1) The
new shape of train labels: (297960, 26) The new shape of
test labels: (74490, 26)

9312/9312 [==============================] - 85s 9ms/step - loss

: 0.1440

0.9761 - accuracy: 0.9595 - val_loss: 0.0853 - val_accuracy:

model: "sequential"

Layer (type) Output Shape Param #

=================================================================
conv2d (Conv2D) (None, 26, 26, 32) 320

max_pooling2d (MaxPooling2D (None, 13, 13, 32) 0

)

conv2d_1 (Conv2D) (None, 13, 13, 64) 18496

max_pooling2d_1 (MaxPooling 2D) (None, 6, 6, 64) 0

conv2d_2 (Conv2D) (None, 4, 4, 128) 73856

13
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

max_pooling2d_2 (MaxPooling 2D) (None, 2, 2, 128) 0

flatten (Flatten) (None, 512) 0

dense (Dense) (None, 64) 32832

dense_1 (Dense) (None, 128) 8320

dense_2 (Dense) (None, 26) 3354

=================================================================
Total params: 137,178
Trainable params: 137,178
Non-trainable params: 0

The validation accuracy is : [0.9760907292366028] The training

accuracy is : [0.9595012664794922] The validation loss is :
[0.08530429750680923] The training loss is :
[0.1440141350030899]

1/1 [==============================] - 0s 79ms/step

14
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 03 : Face recognition using CNN

import numpy as
np import pandas
as pd
from [Link] import fetch_lfw_people

faces = fetch_lfw_people(min_faces_per_person=100, resize=1.0, slice_=(slice(60,

188), slice(60, 188)), color=True)
class_count = len(faces.target_names)

print(faces.target_names)
print([Link])
%matplotlib inline
import [Link] as
plt import seaborn as sns
[Link]()

fig, ax = [Link](3, 6, figsize=(18, 10))

for i, axi in enumerate([Link]):

[Link]([Link][i] / 255) # Scale pixel values so Matplotlib doesn't clip
everything above 1.0
[Link](xticks=[], yticks=[], xlabel=faces.target_names[[Link][i]])
from collections import Counter
counts = Counter([Link])
names = {}

for key in [Link]():

names[faces.target_names[key]] = counts[key]

df = [Link].from_dict(names,
orient='index') [Link](kind='bar')
mask = [Link]([Link], dtype=[Link])

for target in [Link]([Link]):

mask[[Link]([Link] == target)[0][:100]] = 1

x_faces = [Link][mask]
y_faces = [Link][mask]
x_faces = [Link](x_faces, (x_faces.shape[0], [Link][1],
[Link][2], [Link][3]))
x_faces.shape
from [Link] import to_categorical
from sklearn.model_selection import train_test_split

15
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

face_images = x_faces / 255 # Normalize pixel values

face_labels = to_categorical(y_faces)

x_train, x_test, y_train, y_test = train_test_split(face_images, face_labels,

train_size=0.8, stratify=face_labels, random_state=0)
from [Link] import Dense
from [Link] import
Sequential from [Link] import
Conv2D
from [Link] import
MaxPooling2D from [Link]
import Flatten

model = Sequential()
[Link](Conv2D(32, (3, 3),
activation='relu',
input_shape=(face_images.shape[1:])))
[Link](MaxPooling2D(2, 2))
[Link](Conv2D(64, (3, 3), activation='relu'))
[Link](MaxPooling2D(2, 2))
[Link](Conv2D(64, (3, 3), activation='relu'))
[Link](MaxPooling2D(2, 2))
[Link](Flatten())
[Link](Dense(128, activation='relu'))
[Link](Dense(class_count, activation='softmax'))
[Link](optimizer='adam', loss='categorical_crossentropy',
metrics=['accuracy'])
[Link]()
hist = [Link](x_train, y_train, validation_data=(x_test, y_test), epochs=20,
batch_size=25)
acc = [Link]['accuracy']
val_acc = [Link]['val_accuracy']
epochs = range(1, len(acc) + 1)
[Link](epochs, acc, '-', label='Training Accuracy')
[Link](epochs, val_acc, ':', label='Validation Accuracy')
[Link]('Training and Validation Accuracy')
[Link]('Epoch')
[Link]('Accuracy')
[Link](loc='lower right')
[Link]()
from [Link] import confusion_matrix

y_predicted = [Link](x_test)
mat = confusion_matrix(y_test.argmax(axis=1), y_predicted.argmax(axis=1))

[Link](mat.T, square=True, annot=True, fmt='d', cbar=False,

cmap='Blues', xticklabels=faces.target_names,
yticklabels=faces.target_names)

16
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

[Link]('Predicted label')
[Link]('Actual label')
import [Link] as image

x = image.load_img('[Link]', target_size=(face_images.shape[1:]))
[Link]([])
[Link]([])
[Link](x)

x = image.img_to_array(x) /
255 x = np.expand_dims(x,
axis=0) y = [Link](x)[0]

for i in range(len(y)):
print(faces.target_names[i] + ': ' +
str(y[i]))

Output:
['Colin Powell' 'Donald Rumsfeld' 'George W Bush' 'Gerhard Schroeder'
'Tony Blair']
(1140, 128, 128, 3)

17
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

(500, 128, 128, 3)

Model: "sequential"

Layer (type) Output Shape Param #

=================================================================
conv2d (Conv2D) (None, 126, 126, 32) 896

max_pooling2d (MaxPooling2D (None, 63, 63, 32) 0

)

conv2d_1 (Conv2D) (None, 61, 61, 64) 18496

max_pooling2d_1 (MaxPooling 2D) (None, 30, 30, 64) 0

conv2d_2 (Conv2D) (None, 28, 28, 64) 36928

max_pooling2d_2 (MaxPooling 2D) (None, 14, 14, 64) 0

flatten (Flatten) (None, 12544) 0

dense (Dense) (None, 128) 1605760

dense_1 (Dense) (None, 5) 645

=================================================================
Total params: 1,662,725
Trainable params: 1,662,725
Non-trainable params: 0

Epoch 1/20

18
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

16/16 [==============================] - 2s 123ms/step - loss: 1.6558 - accuracy: 0.1925 -

val_loss: 1.6038 - val_accuracy: 0.2000
Epoch 2/20
16/16 [==============================] - 2s 110ms/step - loss: 1.5860 - accuracy: 0.3175 -
val_loss: 1.5416 - val_accuracy: 0.3200
Epoch 3/20
16/16 [==============================] - 2s 112ms/step - loss: 1.4851 - accuracy: 0.3675 -
val_loss: 1.3706 - val_accuracy: 0.4500
Epoch 4/20
16/16 [==============================] - 2s 110ms/step - loss: 1.1602 - accuracy: 0.5775 -
val_loss: 1.0931 - val_accuracy: 0.5900
Epoch 5/20
16/16 [==============================] - 2s 112ms/step - loss: 0.8385 - accuracy: 0.7000 -
val_loss: 0.8494 - val_accuracy: 0.6700
Epoch 6/20
16/16 [==============================] - 2s 111ms/step - loss: 0.5011 - accuracy: 0.8275 -
val_loss: 0.8085 - val_accuracy: 0.6900
Epoch 7/20
16/16 [==============================] - 2s 111ms/step - loss: 0.3819 - accuracy: 0.8550 -
val_loss: 0.7241 - val_accuracy: 0.7200
Epoch 8/20
16/16 [==============================] - 2s 110ms/step - loss: 0.3558 - accuracy: 0.8950 -
val_loss: 0.5499 - val_accuracy: 0.7800
Epoch 9/20
16/16 [==============================] - 2s 114ms/step - loss: 0.1407 - accuracy: 0.9575 -
val_loss: 0.7090 - val_accuracy: 0.8000
Epoch 10/20
16/16 [==============================] - 2s 115ms/step - loss: 0.0869 - accuracy: 0.9875 -
val_loss: 0.6296 - val_accuracy: 0.8400
Epoch 11/20
16/16 [==============================] - 2s 111ms/step - loss: 0.0413 - accuracy: 0.9950 -
val_loss: 0.5816 - val_accuracy: 0.8300
Epoch 12/20
16/16 [==============================] - 2s 110ms/step - loss: 0.0325 - accuracy: 0.9950 -
val_loss: 0.5888 - val_accuracy: 0.8300
Epoch 13/20
16/16 [==============================] - 2s 110ms/step - loss: 0.0359 - accuracy: 0.9900 -
val_loss: 0.6945 - val_accuracy: 0.8100
Epoch 14/20
16/16 [==============================] - 2s 110ms/step - loss: 0.0085 - accuracy: 1.0000 -
val_loss: 0.5278 - val_accuracy: 0.8600
Epoch 15/20
16/16 [==============================] - 2s 111ms/step - loss: 0.0048 - accuracy: 1.0000 -
val_loss: 0.5697 - val_accuracy: 0.8500
Epoch 16/20
16/16 [==============================] - 2s 111ms/step - loss: 0.0032 - accuracy: 1.0000 -
val_loss: 0.6065 - val_accuracy: 0.8500
Epoch 17/20
16/16 [==============================] - 2s 110ms/step - loss: 0.0022 - accuracy: 1.0000 -
val_loss: 0.6007 - val_accuracy: 0.8500
Epoch 18/20
16/16 [==============================] - 2s 112ms/step - loss: 0.0017 - accuracy: 1.0000 -
val_loss: 0.6242 - val_accuracy: 0.8500
Epoch 19/20

19
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

16/16 [==============================] - 2s 118ms/step - loss: 0.0013 - accuracy: 1.0000 -

val_loss: 0.6333 - val_accuracy: 0.8500
Epoch 20/20
16/16 [==============================] - 2s 111ms/step - loss: 0.0011 - accuracy: 1.0000 -
val_loss: 0.6541 - val_accuracy: 0.8500

4/4 [==============================] - 0s 26ms/step

Text(89.18, 0.5, 'Actual label')

<[Link] at 0x1ec80d4d910>

20
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

1/1 [==============================] - 0s 48ms/step

Colin Powell: 0.20101844
Donald Rumsfeld: 0.20214622 George W
Bush: 0.2216323 Gerhard Schroeder:
0.21147959
Tony Blair: 0.16372345

21
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 04: Language modeling using RNN

from future import unicode_literals, print_function,
division from io import open
import glob
import os
import unicodedata
import string
all_letters = string.ascii_letters + " .,;'-"
n_letters = len(all_letters) + 1 # Plus EOS marker
def findFiles(path): return [Link](path)

# Turn a Unicode string to plain ASCII, thanks to

[Link]
def unicodeToAscii(s):
return ''.join(
c for c in [Link]('NFD',
s) if [Link](c) != 'Mn'
and c in all_letters
)

# Read a file and split into lines

def readLines(filename):
with open(filename, encoding='utf-8') as some_file:
return [unicodeToAscii([Link]()) for line in some_file]

# Build the category_lines dictionary, a list of lines per category

category_lines = {}
all_categories = []
for filename in findFiles('data/names/*.txt'):
category = [Link]([Link](filename))[0]

22
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

all_categories.append(category)
lines = readLines(filename)
category_lines[category] = lines

n_categories = len(all_categories)

if n_categories == 0:
raise RuntimeError('Data not found. Make sure that you downloaded data
' 'from [Link] and extract it to '
'the current directory.')

print('# categories:', n_categories,

all_categories) print(unicodeToAscii("O'Néàl"))

pip install torch

import torch
import [Link] as nn

class RNN([Link]):
def init (self, input_size, hidden_size, output_size):
super(RNN, self). init ()
self.hidden_size = hidden_size

self.i2h = [Link](n_categories + input_size + hidden_size, hidden_size)

self.i2o = [Link](n_categories + input_size + hidden_size, output_size)
self.o2o = [Link](hidden_size + output_size, output_size)
[Link] = [Link](0.1)
[Link] = [Link](dim=1)

def forward(self, category, input, hidden):

input_combined = [Link]((category, input, hidden), 1)
hidden = self.i2h(input_combined)
output = self.i2o(input_combined)
output_combined = [Link]((hidden, output), 1)
output = self.o2o(output_combined)
output = [Link](output)
output = [Link](output)
return output, hidden

def initHidden(self):
return [Link](1, self.hidden_size)
import random

# Random item from a list

def randomChoice(l):
return l[[Link](0, len(l) - 1)]

# Get a random category and random line from that category

def randomTrainingPair():
category = randomChoice(all_categories)
line = randomChoice(category_lines[category])
return category, line
# One-hot vector for category

23
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

def categoryTensor(category):
li = all_categories.index(category)
tensor = [Link](1,
n_categories) tensor[0][li] = 1
return tensor

# One-hot matrix of first to last letters (not including EOS) for input
def inputTensor(line):
tensor = [Link](len(line), 1, n_letters)
for li in range(len(line)):
letter = line[li] tensor[li][0]
[all_letters.find(letter)] = 1
return tensor

# ``LongTensor`` of second letter to end (EOS) for target

def targetTensor(line):
letter_indexes = [all_letters.find(line[li]) for li in range(1, len(line))]
letter_indexes.append(n_letters - 1) # EOS
return [Link](letter_indexes)
# Make category, input, and target tensors from a random category, line pair
def randomTrainingExample():
category, line = randomTrainingPair()
category_tensor = categoryTensor(category)
input_line_tensor = inputTensor(line)
target_line_tensor = targetTensor(line)
return category_tensor, input_line_tensor,
target_line_tensor criterion = [Link]()
learning_rate = 0.0005

def train(category_tensor, input_line_tensor,

target_line_tensor): target_line_tensor.unsqueeze_(-1)
hidden = [Link]()
rnn.zero_grad()
loss = 0
for i in range(input_line_tensor.size(0)):
output, hidden = rnn(category_tensor, input_line_tensor[i], hidden)
l = criterion(output, target_line_tensor[i])
loss += l

[Link]()
for p in [Link]():
[Link].add_([Link], alpha=-learning_rate)

return output, [Link]() / input_line_tensor.size(0)

import time
import math

def timeSince(since):
now = [Link]()
s = now - since
m = [Link](s / 60)
s -= m * 60
return '%dm %ds' % (m, s)

24
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

rnn = RNN(n_letters, 128, n_letters)

n_iters = 100000
print_every = 5000
plot_every = 500
all_losses = []
total_loss = 0 # Reset every ``plot_every`` ``iters``

start = [Link]()
for iter in range(1, n_iters + 1):
output, loss =
train(*randomTrainingExample()) total_loss +=
loss
if iter % print_every == 0:
print('%s (%d %d%%) %.4f' % (timeSince(start), iter, iter / n_iters * 100, loss))

if iter % plot_every == 0:
all_losses.append(total_loss / plot_every)
total_loss = 0
import [Link] as plt

[Link]()
[Link](all_losses)
max_length = 20

# Sample from a category and starting letter

def sample(category, start_letter='A'):
with torch.no_grad(): # no need to track history in sampling
category_tensor = categoryTensor(category)
input = inputTensor(start_letter)
hidden = [Link]()

output_name = start_letter

for i in range(max_length):
output, hidden = rnn(category_tensor, input[0],
hidden) topv, topi = [Link](1)
topi = topi[0][0]
if topi == n_letters - 1:
break
else:
letter = all_letters[topi]
output_name += letter
input = inputTensor(letter)

return output_name

# Get multiple samples from one category and multiple starting letters
def samples(category, start_letters='ABC'):
for start_letter in start_letters:
print(sample(category, start_letter))
samples('Russian', 'RUS')
samples('German', 'GER')
samples('Spanish', 'SPA')
samples('Chinese', 'CHI')

25
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Output:
# categories: 18 ['Arabic', 'Chinese', 'Czech', 'Dutch', 'English', 'French', 'German',
'Greek', 'Irish', 'Italian ', 'Japanese', 'Korean', 'Polish', 'Portuguese', 'Russian ',
'Scottish', 'Spanish', 'Vietnamese']
O'Neal
0m 5s (5000 5%) 2.6595
0m 11s (10000 10%) 2.9644
0m 16s (15000 15%) 3.3754
0m 22s (20000 20%) 2.0799
0m 27s (25000 25%) 2.6884
0m 33s (30000 30%) 2.2509
0m 38s (35000 35%) 2.3497
0m 43s (40000 40%) 2.5290
0m 49s (45000 45%) 2.9439
0m 54s (50000 50%) 2.7406
0m 59s (55000 55%) 3.0044
1m 4s (60000 60%) 2.5765
1m 10s (65000 65%) 2.3694
1m 15s (70000 70%) 2.2810
1m 20s (75000 75%) 2.2660
1m 26s (80000 80%) 2.1720
1m 31s (85000 85%) 2.4900
1m 36s (90000 90%) 2.0302
1m 42s (95000 95%) 1.8320
1m 47s (100000 100%) 2.4904

[<[Link].Line2D at 0x1e56757bcd0>]

Rovonov
Uarakov
Shavanov
Gerre Eeren
Roure Salla
Para Allana
Cha
Han
Iun

26
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 05 : Sentiment analysis using LSTM

pip install Keras-Preprocessing

import re
import pandas as
pd import numpy
as np
from [Link] import LabelEncoder
from sklearn.model_selection import
train_test_split from [Link]
import Tokenizer
from keras_preprocessing.sequence import
pad_sequences import keras
from [Link] import
classification_report from [Link]
import accuracy_score import math
import nltk
data = pd.read_csv('IMDB
[Link]') data
def remove_tags(string):
removelist = ""
result = [Link]('','',string) #remove HTML
tags result = [Link]('[Link] #remove
URLs
result = [Link](r'[^w'+removelist+']', ' ',result) #remove non-alphanumeric
characters
result = [Link]()
return result
data['review']=data['review'].apply(lambda cw : remove_tags(cw))
[Link]('stopwords')
from [Link] import stopwords
stop_words = set([Link]('english'))
data['review'] = data['review'].apply(lambda x: ' '.join([word for word in [Link]() if

word not in (stop_words)])) import nltk [Link]()

#we want to download 'wordnet' and 'omw-1.4' from nltk w_tokenizer =
[Link]() lemmatizer =
[Link]() def lemmatize_text(text): st = ""
for w in w_tokenizer.tokenize(text):
st = st + [Link](w) + " " return st
data['review'] = [Link](lemmatize_text)
data

reviews = data['review'].values
labels = data['sentiment'].values
encoder = LabelEncoder()
encoded_labels = encoder.fit_transform(labels)

27
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

train_sentences, test_sentences, train_labels, test_labels =

train_test_split(reviews, encoded_labels, stratify = encoded_labels)

# Hyperparameters of the model

vocab_size = 3000 # choose based on statistics
oov_tok = ''
embedding_dim = 100
max_length = 200 # choose based on statistics, for example 150 to 200
padding_type='post'
trunc_type='post'
# tokenize sentences
tokenizer = Tokenizer(num_words = vocab_size, oov_token=oov_tok)
tokenizer.fit_on_texts(train_sentences)
word_index = tokenizer.word_index
# convert train dataset to sequence and pad sequences
train_sequences = tokenizer.texts_to_sequences(train_sentences)
train_padded = pad_sequences(train_sequences, padding='post',
maxlen=max_length)
# convert Test dataset to sequence and pad sequences
test_sequences = tokenizer.texts_to_sequences(test_sentences)
test_padded = pad_sequences(test_sequences, padding='post', maxlen=max_length)

# model initialization
model =
[Link]([
[Link](vocab_size, embedding_dim,
input_length=max_length),
[Link]([Link](64)),
[Link](24, activation='relu'),
[Link](1, activation='sigmoid')
])
# compile model
[Link](loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# model summary
[Link]()

num_epochs = 5
history = [Link](train_padded, train_labels,
epochs=num_epochs, verbose=1,
validation_split=0.1)
prediction = [Link](test_padded)
# Get labels based on probability 1 if p>= 0.5 else 0
pred_labels = []
for i in prediction:
if i >= 0.5:
pred_labels.append(1)
else:

28
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

pred_labels.append(0)
print("Accuracy of prediction on test set :
", accuracy_score(test_labels,pred_labels))

# reviews on which we need to predict

sentence = ["The movie was very touching and heart
whelming", "I have never seen a terrible movie like
this",
"the movie plot is terrible but it had good
acting"] # convert to a sequence
sequences = tokenizer.texts_to_sequences(sentence)
# pad the sequence
padded = pad_sequences(sequences, padding='post', maxlen=max_length)
# Get labels based on probability 1 if p>= 0.5 else 0
prediction = [Link](padded)
pred_labels = []
for i in
prediction: if i
>= 0.5:
pred_labels. else:
pred_labels.append(0)
for i in
range(len(sentence)):
print(sentence[i])
if pred_labels[i] ==
1: s = 'Positive'
else:
s = 'Negative'
print("Predicted sentiment : ",s)

Output:

review sentiment
0 One of the other reviewers has mentioned that ... positive
1 A wonderful little production. <br /><br />The... positive
2 I thought this was a wonderful way to spend ti... positive
3 Basically there's a family where a little boy ... negative
4 Petter Mattei's "Love in the Time of Money" is... positive
... ... ...
49995 I thought this movie did a down right good job... positive
49996 Bad plot, bad dialogue, bad acting, idiotic di... negative
49997 I am a Catholic taught in parochial elementary... negative
49998 I'm going to have to disagree with the previou... negative
49999 No one expects the Star Trek movies to be high... negative

[nltk_data] Downloading package stopwords to

[nltk_data] C:\Users\[Link]\AppData\Roaming\nltk_data... [nltk_data]
Package stopwords is already up-to-date!
append(1)

29
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

showing info [Link] [Link]

Out[6]:
True

review sentiment

0 w w w w w w w w w w w w w w w w w w w w w w w ... positive

1 wwwwwwwwwwwwwww positive

2 wwwwwwwwwwwwwwwwwww positive

3 wwwwwwwwwww negative

4 wwwwwwwwwwwwwwwwww positive

review sentiment

... ... ...

49995 wwwwwwwwwwwwwwwwwww positive

49996 wwwwwwww negative

49997 wwwwwwwwwwww negative

49998 wwwwwwwwwwwwwww negative

49999 wwwwwwwwwwww negative

50000 rows × 2 columns

Average length of each review : 18.714

Percentage of reviews with positive sentiment is 50.0% Percentage of reviews with
negative sentiment is 50.0%

Model: "sequential"

30
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Layer (type) Output Shape Param #

===========================================================
======
embedding (Embedding) (None, 200, 100) 300000

bidirectional (Bidirectiona (None, 128) 84480

l)

dense (Dense) (None, 24) 3096

dense_1 (Dense) (None, 1) 25

===========================================================
======
Total params: 387,601
Trainable params: 387,601
Non-trainable params: 0

Epoch 1/5
1055/1055 [==============================] - 60s 55ms/step - loss: 0.69
32 - accuracy: 0.5021 - val_loss: 0.6925 - val_accuracy: 0.5205 Epoch 2/5
1055/1055 [==============================] - 58s 55ms/step - loss: 0.69
26 -
Epoch 3/5
1055/1055 [==============================] - 59s 56ms/step - loss: 0.69
26 - accuracy: 0.5129 - val_loss: 0.6924 - val_accuracy: 0.5171 Epoch
4/5
1055/1055 [==============================] - 59s 56ms/step - loss: 0.69
23 - accuracy: 0.5166 - val_loss: 0.6927 - val_accuracy: 0.4965 Epoch
5/5
1055/1055 [==============================] - 58s 55ms/step - loss: 0.69
25 - accuracy: 0.5141 - val_loss: 0.6924 - val_accuracy: 0.5173

391/391 [==============================] - 6s 14ms/step

31
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Accuracy of prediction on test set : 0.5148

1/1 [==============================] - 0s 23ms/step

The movie was very touching and heart whelming Predicted sentiment
: Positive
I have never seen a terrible movie like this Predicted
sentiment : Positive
the movie plot is terrible but it had good acting Predicted sentiment
: Positive

accuracy: 0.5094 - val_loss: 0.6925 - val_accuracy: 0

32
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 06 : Parts of speech tagging using Sequence to Sequence architecture

import numpy as np
import pandas as
pd import json
import functools as fc
from [Link] import accuracy_score

Task 1: Vocabulary Creation

#train = pd.read_csv('data1\train', sep='\t', names=['index', 'word', 'POS'])
train = pd.read_csv('data1/train', sep='\t', names=['index', 'word', 'POS'])
[Link]()
word = train['word'].[Link]()
index = train['index'].[Link]()
pos = train['POS'].[Link]()
vocab = {}

for i in
range(len(word)): if
word[i] in vocab:
vocab[word[i]] += 1
else:
vocab[word[i]] = 1
# replace rare words with <unk> (threshold = 3)
vocab2 = {}
num_unk = 0

for w in vocab:
if vocab[w] >= 3:
vocab2[w] = vocab[w]
else:
num_unk += vocab[w]

# sort the vocabulary by occurrences of words

vocab_sorted = sorted([Link](), key=lambda item: item[1], reverse=True)
# write the sorted vocabulary to vocab file
#with open('recap/[Link]', 'w') as vocab_file:
with open('output/vocab_frequent', 'w') as vocab_file:
# the format of the vocab is word index occurrence
# we add <unk> to the top of the vocabulary manually
vocab_file.write('<unk>' + '\t' + str(0) + '\t' + str(num_unk) + '\n')
for i in range(len(vocab_sorted)):
vocab_file.write(vocab_sorted[i][0] + '\t' + str(i+1) + '\t' + str(vocab_sorted[i][1]) +'\n')

print(f'The total size of my vocabulary is {len(vocab_sorted)}\n')

print(f'The total occurrences of <unk> is {num_unk}\n')

Task 2: Model Learning

# build a vocabulary list with only frequent words (i.e. occur no less than 3 times)
vocab_ls = list([Link]())

# write the frequent words into a json file

#with open('recap/vocab_frequent.txt', 'w') as
output: with open('output/vocal_frequent', 'w') as
output:
for word in vocab_ls:
[Link](word + '\n')

33
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

for i in range(len(word)):
if word[i] not in vocab_ls:
word[i] == '<unk>'

# count (s, s') and (s, x) pairs

ss = {}
sx = {}
for i in range(len(word)-1):
# make sure the index of the current word is less than the next
# ss = {pos[i+1]|pos[i]: count}
# we are not using the format {(pos[i], pos[i+1]): count} because
# json doesn't support tuple
if index[i] < index[i+1]:
if str(pos[i+1]) + '|' + str(pos[i]) in ss:
ss[str(pos[i+1]) + '|' + str(pos[i])] +=1
else:
ss[str(pos[i+1]) + '|' + str(pos[i])] = 1

if str(word[i]) + '|' + str(pos[i]) in sx:

sx[str(word[i]) + '|' + str(pos[i])] +=1
else:
sx[str(word[i]) + '|' + str(pos[i])] = 1

# for ss, we need to count the times that a pos tag occurs at the beginning
# of a sequence (i.e. (s|<s>))
for i in
range(len(word)): if
index[i] == 1:
if str(pos[i]) + '|' + '<s>' in ss:
ss[str(pos[i]) + '|' + '<s>'] += 1
else:
ss[str(pos[i]) + '|' + '<s>'] = 1

# build an emission and a transition dictionaries

emission = {}
transition = {}

# count occurrences of pos tags

count_pos = {}

for p in pos:
if p in count_pos:
count_pos[p] += 1
else:
count_pos[p] = 1

# don't forget to count the occurrences of <start>

count_pos['<s>'] = 0
for i in
index: if i
== 1:
count_pos['<s>'] += 1

# emission dictionary {(s, x): count(s, x) / count(s)}

for sx_pair in sx:
emission[sx_pair] = sx[sx_pair] / count_pos[sx_pair.split('|')[1]]

34
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

# transition dictionary {(s, s'): count(s, s') / count(s)}

for ss_pair in ss:
transition[ss_pair] = ss[ss_pair] / count_pos[ss_pair.split('|')[1]]

print(f'There are {len(transition)} transition parameters in my HMM\n')

print(f'There are {len(emission)} emission parameters in my HMM\n')

# write the emission and transition dictionaries into a json file

emission_transition = [emission, transition]
#with open('recap/[Link]', 'w') as output:
with open('output/[Link]', 'w') as output:
[Link](emission_transition, output)
# build a list of distinct pos
pos_distinct = list(count_pos.keys())

# write the pos_distinct into a txt file

#with open('recap/[Link]', 'w') as pos_output:
with open('output/[Link]', 'w') as pos_output:
for _, pos in enumerate(pos_distinct):
pos_output.write(pos + '\n')
Task 3: Greedy Decoding with HMM
# load txt file vocab
vocab_frequent = []
#with open('recap/vocab_frequent.txt', 'r') as vocab_txt:
with open('output/vocab_frequent.txt', 'r') as vocab_txt:
for word in vocab_txt:
word = [Link]('\n')
vocab_frequent.append(word)
vocab_frequent

# load txt file pos

pos_distinct = []

#with open('recap/[Link]', 'r') as pos_txt:

with open('output/[Link]', 'r') as pos_txt:
for pos in pos_txt:
pos = [Link]('\n')
pos_distinct.append(pos)

# load json file hmm

#with open('recap/[Link]', 'r') as hmm:
with open('output/[Link]', 'r') as hmm:
json_data = [Link](hmm)

emission, transition = json_data[0], json_data[1]

#dev = pd.read_csv('data1/dev', sep='\t', names=['index', 'word', 'POS'])

dev = pd.read_csv('data1/dev', sep='\t', names=['index', 'word', 'POS'])
[Link]()
index_dev = [Link][:,
'index'].[Link]() word_dev = [Link][:,
'word'].[Link]() pos_dev = [Link][:,
'POS'].[Link]()

35
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

# split dev lists (index, word and pos) to individual samples (list --> list of sublists)
word_dev2 = []
pos_dev2 = []
word_sample = []
pos_sample = []
for i in range(len(dev)-1):
if index_dev[i] < index_dev[i+1]:
word_sample.append(word_dev[i])
pos_sample.append(pos_dev[i])
else:
word_sample.append(word_dev[i])
word_dev2.append(word_sample)
word_sample = []

pos_sample.append(pos_dev[i])
pos_dev2.append(pos_sample)
pos_sample = []

def greedy(sentence):
# initialize a dictionary to keep track of the pos for each position
pos = []

# predict the pos of the first word in the sentence

# we need to make sure the first word is in the vocabulary. If not,

replace # with <unk>
if sentence[0] not in vocab_frequent:
sentence[0] = '<unk>'
# predict pos based on the product of the emission and transition
max_prob = 0
p0 = 'UNK'

for p in
pos_distinct: try:
temp = emission[sentence[0] + '|' + p] * transition[p + '|' + '<s>']
if temp > max_prob:
max_prob = temp
p0 = p
except:
pass

[Link](p0)

# predict the pos of the remaining words

for i in range(1, len(sentence)):

# again, we need to check the existence of the word in the vocabulary.
if sentence[i] not in vocab_frequent:
sentence[i] = '<unk>'

max_prob = 0
pi = 'UNK'

36
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

for p in
pos_distinct: try:
temp = emission[sentence[i] + '|' + p] * transition[p + '|' + pos[-
1]] if temp > max_prob:
max_prob = temp
pi = p
except:
pass

[Link](pi)

return pos
pos_greedy = [greedy(s) for s in word_dev2]
# concatenate the list of sublists into one single list
pos_greedy = [Link](lambda a, b: a + b, pos_greedy)
pos_dev = [Link](lambda a, b: a + b, pos_dev2)

acc = accuracy_score(pos_dev, pos_greedy)

print('The prediction accuracy on the dev data is {:.2f}%'.format(acc * 100))

Task 4: Viterbi Decoding with HMM

# load txt file vocab
vocab_frequent = []
#with open('recap/vocab_frequent.txt', 'r') as vocab_txt:
with open('output/vocab_frequent.txt', 'r') as vocab_txt:
for word in vocab_txt:
word = [Link]('\n')
vocab_frequent.append(word)
# load txt file pos
pos_distinct = []

#with open('recap/[Link]', 'r') as pos_txt:

with open('output/[Link]', 'r') as pos_txt:
for pos in pos_txt:
pos = [Link]('\n')
pos_distinct.append(pos)
# load json file hmm
#with open('recap/[Link]', 'r') as hmm:
with open('output/[Link]', 'r') as hmm:
json_data = [Link](hmm)

emission, transition = json_data[0], json_data[1]

#dev = pd.read_csv('data1/dev', sep='\t', names=['index', 'word', 'POS'])

dev = pd.read_csv('data1/dev', sep='\t', names=['index', 'word', 'POS'])
[Link]()

index_dev = [Link][:,
'index'].[Link]() word_dev = [Link][:,
'word'].[Link]() pos_dev = [Link][:,
'POS'].[Link]()

# split dev lists (index, word and pos) to individual samples (list --> list of sublists)
word_dev2 = []

37
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

pos_dev2 = []
word_sample = []
pos_sample = []
for i in range(len(dev)-1):
if index_dev[i] < index_dev[i+1]:
word_sample.append(word_dev[i])
pos_sample.append(pos_dev[i])
else:
word_sample.append(word_dev[i])
word_dev2.append(word_sample)
word_sample = []

pos_sample.append(pos_dev[i])
pos_dev2.append(pos_sample)
pos_sample = []

# define a function to predict the pos for an input sentence

def viterbi(sentence):
# initialize a dictionary that keeps track of the highest cumulative probability of each
possible
# pos at each position of the input sentence
seq = {i:{} for i in range(len(sentence))}
# also initialize a dictionary that keeps track of the pos of the previous pos that leads to the
# highest cumulative probability of each possible pos at each position of the input sentence
# for instance, for a pos of NNP at position i, we want to know which pos of position i-1
leads to
# the highest cumulative probability of NNP at position i.
pre_pos = {i:{} for i in range(len(sentence))}

# for the first position, the highest cumulative probability of each possible pos would be
# emission[sentence[0]|pos] * transition[pos|<s>]

# check if the first word is in the vocabualry. If not, replace with '<unk>'
if sentence[0] not in vocab_frequent:
sentence[0] = '<unk>'

for p in pos_distinct:
if p + '|' + '<s>' in transition:
try:
seq[0][p] = transition[p + '|' + '<s>'] * \
emission[sentence[0] + '|' + p]
except:
seq[0][p] = 0
# set <s> as the previous pos of each possible pos at the first position
for p in seq[0].keys():
pre_pos[0][p] = '<s>'

# for position i > 0, the highest cumulative probability of each possible pos would be
# emission[sentence[i]|pos[i]] * transition[pos[i]|pos[i-1]] * seq[i-1][pos]
for i in range(1, len(sentence)):
# still, check if the word is in the vocabulary
if sentence[i] not in vocab_frequent:
sentence[i] = '<unk>'

38
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

for p in seq[i-1].keys():
for p_prime in pos_distinct:
if p_prime + '|' + p in transition:
if p_prime in seq[i]:
try:
temp = seq[i-1][p] * \
transition[p_prime + '|' + p] * \
emission[sentence[i] + '|' + p_prime]
if temp > seq[i][p_prime]:
seq[i][p_prime] = temp
pre_pos[i][p_prime] = p
except:
pass
else:
try:
seq[i][p_prime] = seq[i-1][p] * \
transition[p_prime + '|' + p] * \
emission[sentence[i] + '|' + p_prime]
pre_pos[i][p_prime] = p
except:
seq[i][p_prime] = 0
# after we get the maximum probability for every possible pos at every position of a
sentence,
# we can trace backward to find out our prediction on the pos for the sentence.
seq_predict = []

# The pos of the last word in the sentence is the one with the highest probability
# after predicting the pos of the last word in the sentence, we can iterate through pre_pos
to predict
# the pos of the remaining words in the input sentence in the reverse order

# the highest probability

prob_max = max(seq[len(sentence)-1].values())
# the index of the highest probability
index_max = list(seq[len(sentence)-1].values()).index(prob_max)
# the pos of the highest probability
pos_max = list(seq[len(sentence)-1].keys())[index_max]
seq_predict.append(pos_max)

# iterate through pre_pos

for i in range(len(sentence)-1, 0, -1):
# for some rare ss or sx pairs, there is no corresponding key in the
# transition or emission dictionary. In this case, we need to set manually
# the pos to 'UNK' at those positions
try:
pos_max = pre_pos[i][pos_max]
seq_predict.append(pos_max)
except:
seq_predict.append('UNK')

# The final seq_predict should be the reverse of the original

seq_predict = [seq_predict[i] for i in range(len(seq_predict)-1, -1, -1)]
return seq_predict

39
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

# use viterbi to predict pos for dev

pos_viterbi = [viterbi(s) for s in word_dev2]

# merge the list of sublists to a single list

pos_viterbi = [Link](lambda a, b: a + b, pos_viterbi)
pos_dev = [Link](lambda a, b: a + b, pos_dev2)

acc = accuracy_score(pos_dev, pos_viterbi)

print('The prediction accuracy on the dev data is {:.2f}%'.format(acc * 100))

Output:

index word POS

0
1 Pierre NNP

1
2 Vinken NNP

2
3 , ,

3
4 61 CD

4 5 years NNS

index word POS

0 1 The DT

1 2 Arizona NNP

2 3 Corporations NNP

3 4 Commission NNP

4 5 authorized VBD

40
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

The total size of my vocabulary is 43193 The total

occurrences of <unk> is 32537

There are 7 transition parameters in my HMM There are 6

emission parameters in my HMM
['Pierre',
',',

41
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

'61', 'years', 'old', 'will', 'join', 'the', 'board', 'as', 'a', 'nonexecutive',
'director', 'Nov.', '29', '.', 'Mr.', 'is', 'chairman', 'of', 'N.V.', 'Dutch',
'publishing', 'group', 'Rudolph', 'Agnew', '55', 'and', 'former', 'Consolidated',
'Gold', 'Fields', 'PLC', 'was', 'named', 'this', 'British', 'industrial', 'conglomerate', 'A', 'form',
'asbesto s', 'once', 'used', 'to', 'make', 'Kent', 'cigarette',
'filters', 'has', 'caused', 'high', 'percentage', 'cancer', 'deaths', 'among',
'workers', 'exposed', 'it', 'more',]

The prediction accuracy on the dev data is 0.00%

index word POS

0 1 The DT
1 2 Arizona NNP
2 3 Corporations NNP
3 4 Commission NNP
4 5 authorized VB
D

The prediction accuracy on the dev data is 0.01%

42
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 07 : Machine Translation using Encoder-Decoder model

import numpy as np # linear algebra

import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)

# Input data files are available in the read-only "../input/" directory

# For example, running this (by clicking run or pressing Shift+Enter) will list all files underinput
directory
import os
for dirname, _, filenames in [Link]('/kaggle/input'):
for filename in filenames:
print([Link](dirname, filename))

# You can write up to 5GB to the current directory (/kaggle/working/) that gets
preserved as output
when you create a version using "Save & Run All"
# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of

current session
from [Link] import Model
from [Link] import Input,LSTM,Dense

batch_size=64
epochs=100
latent_dim=256 # here latent dim represent hidden state or cell state
num_samples=10000

data_path='[Link]'
# Vectorize the data. input_texts = [] target_texts = [] input_characters = set() target_characters = set()
with open(data_path, 'r', encoding='utf-8') as f: lines = [Link]().split('\n')
for line in lines[: min(num_samples, len(lines) - 1)]: input_text, target_text, _ = [Link]('\t')
# We use "tab" as the "start sequence" character
# for the targets, and "\n" as "end sequence" character. target_text = '\t' + target_text + '\n'
input_texts.append(input_text) target_texts.append(target_text)
for char in input_text:
if char not in input_characters: input_characters.add(char)
for char in target_text:
if char not in target_characters: target_characters.add(char)
input_characters=sorted(list(input_characters)) target_characters=sorted(list(target_characters))

num_encoder_tokens=len(input_characters) num_decoder_tokens=len(target_characters)
max_encoder_seq_length=max([len(txt) for txt in input_texts]) max_decoder_seq_length=max([len(txt) for
txt in target_texts]) print('Number of samples:', len(input_texts))
print('Number of unique input tokens:', num_encoder_tokens) print('Number of unique output tokens:',
num_decoder_tokens) print('Max sequence length for inputs:', max_encoder_seq_length) print('Max
sequence length for outputs:', max_decoder_seq_length)

43
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

input_token_index=dict(
[(char,i) for i, char in enumerate(input_characters)])
target_token_index=dict(
[(char,i) for i, char in enumerate(target_characters)])

encoder_input_data = [Link](
(len(input_texts), max_encoder_seq_length, num_encoder_tokens),
dtype='float32')
decoder_input_data = [Link](
(len(input_texts), max_decoder_seq_length, num_decoder_tokens),
dtype='float32')
decoder_target_data = [Link](
(len(input_texts), max_decoder_seq_length, num_decoder_tokens),
dtype='float32')

for i, (input_text, target_text) in enumerate(zip(input_texts, target_texts)):

for t, char in enumerate(input_text):
encoder_input_data[i, t, input_token_index[char]] = 1.encoder_input_data[i, t + 1: for t, char in
enumerate(target_text):
# decoder_target_data is ahead of decoder_input_data by one timestep
decoder_input_data[i, t, target_token_index[char]] = 1.
if t > 0:
# decoder_target_data will be ahead by one
timestep # and will not include the start character.
decoder_target_data[i, t - 1, target_token_index[char]] = 1.
decoder_input_data[i, t + 1:, target_token_index[' ']] = 1.
decoder_target_data[i, t:, target_token_index[' ']] = 1.
encoder_inputs = Input(shape=(None,
num_encoder_tokens)) encoder = LSTM(latent_dim,
return_state=True) encoder_outputs, state_h, state_c =
encoder(encoder_inputs) # We discard `encoder_outputs`
and only keep the states. encoder_states = [state_h, state_c]
, input_token_index[' ']] = 1.
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,
num_decoder_tokens)) # We set up our decoder to return full
output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(decoder_inputs,
initial_state=encoder_states)
decoder_dense = Dense(num_decoder_tokens,
activation='softmax') decoder_outputs =
decoder_dense(decoder_outputs)

44
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)

# Run training
[Link](optimizer='rmsprop', loss='categorical_crossentropy',
metrics=['accuracy'])
[Link]([encoder_input_data, decoder_input_data], decoder_target_data,
batch_size=batch_size,
epochs=epochs,
validation_split=0.2)

[Link]('eng2french.h5')

encoder_model = Model(encoder_inputs, encoder_states)

decoder_state_input_h = Input(shape=(latent_dim,))
decoder_state_input_c = Input(shape=(latent_dim,))
decoder_states_inputs = [decoder_state_input_h,
decoder_state_input_c] decoder_outputs, state_h, state_c =
decoder_lstm(
decoder_inputs, initial_state=decoder_states_inputs)
decoder_states = [state_h, state_c]
decoder_outputs = decoder_dense(decoder_outputs)
decoder_model = Model(
[decoder_inputs] + decoder_states_inputs,
[decoder_outputs] + decoder_states)

# Reverse-lookup token index to decode sequences back

to # something readable.
reverse_input_char_index = dict(
(i, char) for char, i in input_token_index.items())
reverse_target_char_index = dict(
(i, char) for char, i in target_token_index.items())
def decode_sequence(input_seq):
# Encode the input as state vectors.
states_value = encoder_model.predict(input_seq)

# Generate empty target sequence of length 1.

target_seq = [Link]((1, 1, num_decoder_tokens))
# Populate the first character of target sequence with the start character.
target_seq[0, 0, target_token_index['\t']] = 1.

# Sampling loop for a batch of sequences

# (to simplify, here we assume a batch of size 1).
stop_condition = False
decoded_sentence = ''
while not
stop_condition:
output_tokens, h, c =
decoder_model.predict( [target_seq] +
states_value)

# Sample a token
sampled_token_index = [Link](output_tokens[0, -1, :])
sampled_char = reverse_target_char_index[sampled_token_index]
decoded_sentence += sampled_char

45
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

# Exit condition: either hit max

length # or find stop character.
if (sampled_char == '\n' or
len(decoded_sentence) > max_decoder_seq_length):
stop_condition = True

# Update the target sequence (of length 1).

target_seq = [Link]((1, 1, num_decoder_tokens))
target_seq[0, 0, sampled_token_index] = 1.

# Update states
states_value = [h, c]

return decoded_sentence

for seq_index in range(100):

# Take one sequence (part of the training set)
# for trying out decoding.
input_seq = encoder_input_data[seq_index: seq_index + 1]
decoded_sentence = decode_sequence(input_seq)
print('-')
print('Input sentence:', input_texts[seq_index])print('Decoded sentence:', decoded_

Output:
Number of samples: 10000
Number of unique input tokens: 71 Number of
unique output tokens: 93 Max sequence length
for inputs: 15 Max sequence length for outputs:
59
Epoch 1/100
125/125 [==============================] - 15s 105ms/step - loss:
1.2150 - accuracy: 0.7315 - val_loss: 1.0873 - val_accuracy: 0.7068
Epoch 2/100
125/125 [==============================] - 13s 106ms/step - loss:
0.9334 - accuracy: 0.7490 - val_loss: 0.9959 - val_accuracy: 0.7128
Epoch 3/100
125/125 [==============================] - 13s 105ms/step - loss:
0.8396 - accuracy: 0.7679 - val_loss: 0.9039 - val_accuracy: 0.7500
…
Epoch 98/100
125/125 [==============================] - 13s 107ms/step - loss:
0.1532 - accuracy: 0.9531 - val_loss: 0.5529 - val_accuracy: 0.8705
Epoch 99/100
125/125 [==============================] - 13s 108ms/step - loss:
0.1517 - accuracy: 0.9533 - val_loss: 0.5561 - val_accuracy: 0.8697
Epoch 100/100
125/125 [==============================] - 13s 108ms/step - loss:
0.1497 - accuracy: 0.9543 - val_loss: 0.5522 - val_accuracy: 0.8706
sentence)

46
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Input sentence: Smile.

Decoded sentence: Pours pres votr.

1/1 [==============================] - 0s 14ms/step

1/1 [==============================] - 0s 22ms/step
1/1 [==============================] - 0s 18ms/step
1/1 [==============================] - 0s 23ms/step
1/1 [==============================] - 0s 15ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 20ms/step
1/1 [==============================] - 0s 16ms/step
1/1 [==============================] - 0s 21ms/step
1/1 [==============================] - 0s 18ms/step
-
Input sentence: Sorry?
Decoded sentence: Pardon
?

1/1 [==============================] - 0s 20ms/step

1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 20ms/step
1/1 [==============================] - 0s 19ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 17ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 19ms/step
1/1 [==============================] - 0s 19ms/step
1/1 [==============================] - 0s 23ms/step
1/1 [==============================] - 0s 17ms/step
1/1 [==============================] - 0s 13ms/step
-

47
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Program 8: Image augmentation using GANs

import os
import numpy as np
import [Link] as
image
import [Link] as plt
%matplotlib inline

def load_images_from_path(path, label):

images = []
labels = []

for file in [Link](path):

img = image.load_img([Link](path, file), target_size=(224, 224, 3))
[Link](image.img_to_array(img))
[Link]((label))

return images, labels

def show_images(images):
fig, axes = [Link](1, 8, figsize=(20, 20), subplot_kw={'xticks': [], 'yticks': []})

for i, ax in enumerate([Link]):
[Link](images[i] / 255)

x_train = []
y_train = []
x_test = []
y_test = []

images, labels = load_images_from_path('arctic-wildlife/train/arctic_fox', 0) show_images(images)

x_train += images y_train += labels

images, labels = load_images_from_path('arctic-wildlife/train/walrus', 2) show_images(images)

x_train += images y_train += labels

images, labels = load_images_from_path('arctic-wildlife/test/polar_bear', 1) show_images(images)

x_test += images y_test += labels

48
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

from [Link] import to_categorical

from [Link].resnet50 import preprocess_input
x_train = preprocess_input([Link](x_train))
x_test = preprocess_input([Link](x_test))
y_train_encoded = to_categorical(y_train)
y_test_encoded = to_categorical(y_test)
from [Link] import ResNet50V2
base_model = ResNet50V2(weights='imagenet', include_top=False)
for layer in base_model.layers:
[Link] = False
from [Link] import Sequential
from [Link] import Flatten, Dense, Dropout
from [Link] import Rescaling, RandomFlip, RandomRotation,
RandomTranslation, RandomZoom
model = Sequential()
[Link](Rescaling(1./255))
[Link](RandomFlip(mode='horizontal'))
[Link](RandomTranslation(0.2, 0.2))
[Link](RandomRotation(0.2))
[Link](RandomZoom(0.2))

[Link](base_model)
[Link](Flatten())
[Link](Dense(1024, activation='relu'))
[Link](Dropout(0.2))
[Link](Dense(3, activation='softmax'))
[Link](optimizer='adam', loss='categorical_crossentropy',
metrics=['accuracy'])
hist = [Link](x_train, y_train_encoded, validation_data=(x_test,
y_test_encoded), batch_size=10, epochs=25)

49
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

acc = [Link]['accuracy']
val_acc = [Link]['val_accuracy']
epochs = range(1, len(acc) + 1)
[Link](epochs, acc, '-', label='Training Accuracy')
[Link](epochs, val_acc, ':', label='Validation Accuracy')

[Link]('Training and Validation

Accuracy') [Link]('Epoch')
[Link]('Accuracy')
[Link](loc='lower right')
[Link]()

from [Link] import

confusion_matrix import seaborn as sns
[Link]()
y_predicted = [Link](x_test)
mat = confusion_matrix(y_test_encoded.argmax(axis=1),
y_predicted.argmax(axis=1))
class_labels = ['arctic fox', 'polar bear', 'walrus']
[Link](mat, square=True, annot=True, fmt='d', cbar=False,
cmap='Blues', xticklabels=class_labels,
yticklabels=class_labels)
[Link]('Predicted label')

[Link]('Actual label')
x = image.load_img('arctic-wildlife/samples/arctic_fox/arctic_fox_140.jpeg',
target_size=(224, 224))
[Link]([])
[Link]([])
[Link](x)

50
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

x = image.img_to_array(x)
x = np.expand_dims(x,
axis=0) x =
preprocess_input(x)
predictions = [Link](x)
for i, label in
enumerate(class_labels):
print(f'{label}: {predictions[0][i]}')
x = image.load_img('arctic-wildlife/samples/walrus/walrus_143.png',
target_size=(224, 224))
[Link]([])
[Link]([])
[Link](x)
x = image.img_to_array(x)
x = np.expand_dims(x,
axis=0) x =
preprocess_input(x)
predictions = [Link](x)

for i, label in enumerate(class_labels):

print(f'{label}: {predictions[0][i]}')

51
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Output:
Train :

Test :

Train a network based on ResNet-50V2 with data augmentation

Epoch 1/25
30/30 [==============================] - 27s 848ms/step - loss: 31.6208
- accuracy: 0.7400 - val_loss: 8.3153 - val_accuracy: 0.8667
Epoch 2/25
30/30 [==============================] - 25s 848ms/step - loss: 6.1519 -
accuracy: 0.8900 - val_loss: 2.5947 - val_accuracy: 0.9583
Epoch 3/25
30/30 [==============================] - 26s 870ms/step - loss: 3.0037 -
accuracy: 0.9467 - val_loss: 2.7972 - val_accuracy: 0.9750
.
.
.
.

52
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Epoch 24/25
30/30 [==============================] - 27s 896ms/step - loss: 0.5841 -
accuracy: 0.9633 - val_loss: 0.5701 - val_accuracy: 0.9667
Epoch 25/25
30/30 [==============================] - 25s 844ms/step - loss: 0.7861 -
accuracy: 0.9500 - val_loss: 0.5762 - val_accuracy: 0.9667

4/4 [==============================] - 3s 635ms/step

Text(89.18, 0.5, 'Actual label')

<[Link] at 0x2c5496dfa00>

53
Downloaded by trivenij 24cse (trivenij.24cse@[Link])
 lOMoARcPSD|56718048

Preprocess the image and submit it to the network for classification.

1/1 [==============================] - 0s 57ms/step

arctic fox: 1.0
polar bear:
1.1824497264458205e-28 walrus:
0.0
Now try it with a walrus image that the network hasn't seen before. Start by
loading the image.
<[Link] at 0x2c546cb7790>

Preprocess the image and make a prediction.

1/1 [==============================] - 0s 51ms/step
arctic fox: 0.0
polar bear: 0.0
walrus: 1.0

Program 9 : Mini-project on real world applications

54

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

CONTENT BEYOND

SYLLABUS

55

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

PROJECT 1

56

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

57

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

58

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

59

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

60

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

PROJECT 2

61

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

62

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

63

Downloaded by trivenij 24cse (trivenij.24cse@[Link])

 lOMoARcPSD|56718048

64

Downloaded by trivenij 24cse (trivenij.24cse@[Link])