Home assignment - Deep Learning (ITLB359, MIB)

Task

 

description

Challenge: Get the best results on the malaria dataset!

 

"Malaria contains images from various cells. Some cells are infected with malaria, other cells are not. You can download the dataset. Each image is a different shape in 24-bit colour."

 

Your task is as follows:

 

Download

 

the

 

dataset

 

from

 

google

drive like:

!wget "https://drive.google.com/uc?id=1dl6KXbpRrUNrH5xC26MpYq3rHjOewUTF&export=download&authuser=0" -O malaria.zip   !unzip malaria.zip   %cd malaria

 

Convert

from

image

dataset

 

to

 

NumPy

 

array

 

for

 

training

:

list_of_class1 = os.listdir("Parasitized") list_of_class0 = os.listdir("Uninfected")   labels1 = np.ones(len(list_of_class1)) labels0 = np.zeros(len(list_of_class0)) labels = np.concatenate((labels1, labels0), axis=0)   dataset = [] dim = (100, 100) for img in list_of_class1: image = cv2.imread("Parasitized/" + str(img)) resized = cv2.resize(image, dim, interpolation = cv2.INTER_AREA) dataset.append(resized)   for img in list_of_class0: image = cv2.imread("Uninfected/" + str(img)) resized = cv2.resize(image, dim, interpolation = cv2.INTER_AREA) dataset.append(resized)

Split

to

 

training

,

validation

and test

sets

.

Challenge

:

Get

 

the

 

best

 

results

 

on

 

the

 

malaria

 

dataset

!

The

challenge

of

this

 

task

is

to

 

get

 

the

 

best

 

results

 

on

 

the

 

malaria

 

dataset

 

by

 

tuning

 

hyperparameters

of a NN

model

and

observing

 

convergence

 

behavior

.

Best -

for

 

simplicity

-

means

 

the

 

highest

 

accuracy

 

on

 

the

 

validation

 

set

.

Reflection

Critically

 

evaluate

 

your

 

work

,

including

 

other

 

approaches

.

Answer

in a markdown

cell

.

Added constraint: The model with the "best performance" has to be saved, so it should not be just a printout happening once during training!

You may NOT manipulate the validation set (that is, you cannot pick and choose which samples belong to the validation set to make your model “perform” better on validation data)!

Please observe the following:

You

must

use

a

single

 

standalone

 

Jupyter

Notebook

to

 

solve

 

the

 

task

and

submit

 

the

.

ipynb

file.

Note for those working on Google Colab: a link to your notebook will not suffice: you have to download and submit the file itself.

Follow

 

the

 

principle

of

literate

 

programming

, and

make

 

use

of

the

markdown

cells

of

the

notebook.

Deadline

Please refer to the Moodle page of the module.

Assessment

The assignment will be assessed based on the following criteria (see the grid on Moodle):

·  Specification fulfilment and Conceptual grounding (60%)

·  Literate programming and markdown cells (20%)

·  Clean code (20%)

The resit arrangement for the assignment is the same as above; you may resubmit the same paper, with corrections, that you submitted by the original deadline. The resubmission deadline will be specified on Moodle after the grades for the original submission are published.

Upload your file (.ipynb) to Moodle.

Academic

 

conduct

 

notice

Where the Academic Conduct Officer has reason to suspect that a piece of work submitted by a student was wholly or in part written by someone other than the student who submitted it, and this has not been disclosed by the student, they may call for the student to defend the work in viva or a written comprehension test. The burden of proof in such a viva or test will be upon the student to demonstrate to the examination panel’s satisfaction his/her full comprehension of the work s/he has submitted. Failure to appear without satisfactory explanation will result in immediate failure of that assessment, with consequences of academic misconduct and application of sanctions.