Assignment I-4

EE452 Computer Vision (Fall 2024)
Assignment 1
Training Neural Networks for Image Classification
Release date: Monday, 07 October 2024
Due date: Sunday, 03 November 2024
Maximum Marks: 100 (Weight in Final grade: 8%)
Objectives:
• To understand the usage of numpy for vectorized coding and perform preprocessing of data set
• To perform feature extraction and apply Support Vector Machine (SVM) classifier on extracted features
• To create dense and sparse neural network classifiers from scratch and apply on CIFAR-10 dataset
• To perform regularization to improve the trained networks
• To create neural network classifier using transfer learning (trained on ImageNet) and apply on CIFAR-10 dataset
• To perform hyperparameter tuning
Instructions:
• This assignment is individual task where only collaboration in terms of discussing and idea sharing is allowed.
• A python notebook (.ipynb) is to be submitted on LMS describing the different phases of the work including what
worked, what did not work, what you learned, what extra things you tried etc. Link to the notebook is not
acceptable and would give you straight 0.
Tasks:
1. Load the Intel Image Classification Dataset and briefly describe what the dataset is about and the constituents
of it. You can view the documentation for Intel Image Classification dataset. Also mention some of the
preprocessing steps which can beapplied while dealing with image data in general. [10]
2. Write a function "filt" whose aim would be to perform the task of filtering. The function should take the
following arguments: input, filter, padding, normalization. The 'input' would be the numpy representation of an
image over which filtering is to be performed - this will either have a single channel or three channels, 'filter' would
be a 2D numpy array representing the filter (for example: np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]]), 'padding' would
be a boolean value which would determine as to whether the output image should have the same size as the input
image, and 'normalization' would again be a boolean value which would determine as to whether the filtering
operation needs to be normalized or not. The output of this function should be the filtered image. Also note that
the only external library which can be used in this function is “numpy". [10]
3. Extract HOG features from the Intel Image Classification dataset, and train a Linear SVM using those
features. You areexpected to reach around 60% accuracy on test data using this method. [20]
4. This part requires you to use only pytorch. In each of the following subparts (a, b, and c) you are required to
do the following: [20x3]
i. Create a network architecture.

ii. Load the Fashion MNIST dataset.
iii. Preprocess the dataset accordingly.
iv. Train your architecture on the dataset for a reasonable number of epochs (maximum is 100) using the testing
data as your validation data.
v. Report both your training accuracy as well as validation accuracy.
vi. Plot both the training loss as well as the validation loss on the same plot.
vii. Plot both the training accuracy as well as the validation accuracy on the same plot.
For each subpart you are encouraged to experiment both with the hyperparameters as well as different
architecture designs for your network.
a. Create a feed forward neural network consisting of only fully connected layers; the network should have at least
two hidden layers. You are allowed to experiment with regularization techniques and are expected to reach above
90% accuracy on test data.
b. Create a convolutional neural network consisting of at least two convolutional layers. You are allowed to
experiment with regularization techniques and are expected to reach above 92% accuracy on test data.
c. Use the pre-trained VGG-16 network (trained on ImageNet) as a feature extractor, and connect it to a feed
forward network consisting of fully connected layers. You should try to fine tune the network and are expected
to reach above 85% accuracy on test data.
Submission Guidelines:
The entire assignment is to be done as a python notebook. Once you are done, you should upload the notebook,
not the link to it, to LMS. Create sections within the notebook in order to answer each question. For the second
question, you should also display some results of using your function on custom images (which you may obtain
from the web). For the fourth question, you should only submit one model architecture for each part, and in case
you have experimented with other architectures, you should mention those in a dedicated paragraph.
It is preferred that you do this assignment on google colab: https://colab.research.google.com/. Colab allows you
to utilize its GPU, which can be accessed as follows: Runtime => Change runtime type => Change Hardware
accelerator to GPU. Once you have completed your assignment on colab, you can download it as a notebook as
follows: File => Download.ipynb

Assignment I-4

Uploaded by

Copyright:

Available Formats

Assignment I-4

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Assignment I-4

Uploaded by

Copyright:

Available Formats

EE452 Computer Vision (Fall 2024)

i. Create a network architecture.

You might also like