CS 6476 Assignment

CS 6476

Assignment Instructions:
In this assignment, you will implement Fully-Connected Neural Networks and
Convolutional Neural Networks for image classification models. The goals of this
assignment are as follows:
● Understand the layered architecture of Neural Networks and its components
like activation functions, loss functions, and optimizers.
● Understand the architecture of Convolutional Neural Networks and get
practice with training these models on data.
There are four parts to this assignment:
P0: Simple Classifiers
The notebook simple_models.ipynb will walk you through basic experiments
(already coded) of linear classifiers and KNN classifiers. You are supposed to
understand and analyze the results of those experiments. You will be asked to
answer a few questions on the same. Note: You can modify the hyper-parameters to
see if your results improve and make that a part of your analysis.
P1: Fully Connected Neural Network
The notebook fully_connected_networks.ipynb will walk you through
implementing Fully-Connected Neural Networks. We have used a modular approach
to code each essential element of the fully connected neural networks. You are also
supposed to read the supplementary reading provided throughout the notebook. It
will help you in the Q/A part of the assignment. You are supposed to complete all the
TODOs in the fully_connected_networks.py file and submit the file and model
weights to Gradecope.
P2: Convolutional Neural Network
The notebook convolutional_networks.ipynb will walk you through implementing
Convolutional Neural Networks. Same as part 1, first you’ll implement components of
the convolution network which will be followed by implementing a three-layered
network for classification. You are supposed to complete all the TODOs in the
convolutional_networks.py file and submit the file and model weights to
Gradecope.
P3: Fine-tuning Pretrained Network
The notebook transfer_learning.ipynb will demonstrate transfer learning using
ResNet-18 on the CIFAR-10 dataset. You are supposed to upload the notebook to
Colab, complete the TODOs in the notebook and upload only the trained model
weight to Gradescope.
Steps:
1. Download the Assignment2 folder from Canvas.
2. Unzip the file. All files should appear in a folder named Assignment2.
3. Upload this ‘Assignment2’ folder on your Google Drive.
4. Double-click on the simple_models.ipynb, it should open this notebook in
Google Colab. (If it doesn’t then click on ‘connect more apps’ and install
Collaboratory). Go through this notebook.
5. Next, you open the fully_connected_networks.ipynb the same way. Follow
the notebook line by line. The first step in this is to mount Google Drive. Once
you do that you can access and edit your ‘.py’ files on Colab.
6. Open the fully_connected_networks.py file in the colab workspace
side-by-side and proceed with the assignment
7. Once you are done with Part 1, follow steps 4 to 6 for the Part 3 as well.
8. Lastly, open the transfer_learning.ipynb, there is no ‘.py’ file for this section.
Just complete the TODO in the notebook
Submission
You are supposed to submit in total 5 files for this assignment.
1. fully_connected_networks.py
2. convolutional_networks.py
3. fcn_model.pth
4. final_threeconvnet.pth
5. resnet.pth
6. Q/A report (PDF file containing your answers to the given below questions)
IMPORTANT NOTE: Do not change the names of either of the above files. For ‘.py’
files, put your code only inside the given TODO space.
Grading: (of Total 16)
Individual component grading schema within each part of the assignment would be
present with the assignment. The final scores from each part will be scaled to the
final scores as below.
P0: Ungraded (results might be helpful in Q/A)
P1: 5/16
P2: 6/16
P3: 2/16
Q/A: 3/16
Questions
You need to answer these questions in brief, using evidence from the output of your
assignment, and submit it as a Q/A report. Please add your GTID and GaTech mail
on top of the report and name the file “lastname_assignment2.pdf” (example:
“burdell_assignment2.pdf”). PDF is the only accepted upload format for the report.
The report should be strictly less than equal to 4 pages including any references you
may want to provide. Feel free to use any template you like as long as the report is
readable.
1. (1 point) Compare and Contrast the results of the Linear Classifier, KNN
CLassifier, Fully Connected Classifier (your optimal solution), CNN Classifier
(your optimal solution), and Transfer Learning (using ResNet-18). Why did
you get the results you got (Write one line each about the classifier, as to why
they performed well and didn’t perform well)
2. (1 Point) With respect to Part 1, briefly explain how different regularization
parameters and different optimization techniques affect the results of your
training and validation set. (Mention 2 points each on your findings with
regularization and optimization techniques. You can use any relevant part of
the assignment for context)
3. (1 Point) With respect to the final model that you submitted in Part 2, how
many algebraic operations (additions and multiplications) were performed in
the forward pass for the first layer (Conv_ReLU_Pool)? Explicitly mention
input image dimensions, number of filters, and size of the filter along with your
calculations. For simplicity assume only one image is passed.
4. (1 Point) For Part 2, justify your choice of hyperparameters based on the
observations you made while playing with the model.
5. (1 Point) In Part 3, what changes do you observe when you continue training
for longer? Comment on the training and validation losses and accuracies.
What is this observed phenomenon called? What would your approach be to
help with this?
6. (1 Point) You have learned about regularization in CNNs. However, not all
networks use the same regularization parameters. Comment on the strength
of regularization needed to train a model effectively, as the size of the model
increases.