Description
1 Classification with CNNs for CIFAR-10
[70 points]
You are going to implement a Convolutional Neural Network (CNN) (class Net) to classify the
CIFAR-10 dataset. The network architecture is not fixed. You will design the network architecture
yourself. You can check out this tutorial to get started.
However, there are some requirements that your network architecture must satisfy:
1. Your network architecture must have at least 4 layers (which is either a convolution or fully
connected layer passed into an activation function like Relu, Tanh, etc). Max pooling or strided
convolution layers used to downsample activation maps do not count. The number of layers
in your network will be very close to the number of activation functions.
2. Must have at least 1 convolution layer.
3. Must have at least 1 max pooling.
4. Must have at least 1 fully connected layer at the end
If your network architecture doesn’t satisfy any of the above requirements, marks will be
deducted. For every requirement that is not satisfied, you lose 10 points. Similar to Assignment
1, observe the metrics on the validation set to check for overfitting.
2 Classification using a pretrained CNN for CIFAR- 10
[30 points]
Implement a CNN model using a pretrained architecture (class PretrainedNet). Examples
include AlexNet, ResNet, etc. This Pytorch documentation might help you out. Save the weights
of the final fine-tuned model as best model.pth (CNN main.py does this for you). You will need
to submit this!
3 Additional Information
3.1 Template Code
You are given CNN main.py and CNN submission.py. Please do not make any changes to CNN main.py
as you will not be submitting this file. CNN submission.py has some template code provided. You
are free to make changes to the functions/classes and add your own, but make sure they return what
CNN main.py is expecting.
You are free to define your own training/validation splits and transforms in the function
load dataset(pretrain). Some default template code has been provided, that you can modify if
you wish to do so. If you need to define your own transformations for the test test, you can pass it
to the main file by editing the variable test transform. You can use the variable pretrain if you
want to define something different for part 2.
1
3.2 Running the Code
If you have paramparse installed, you can use command line arguments.
1. Part 1: python3 CNN main.py
2. Part 2: python3 CNN main.py pretrained=1
or
python3 CNN main.py pretrained=1 load model=1
if you want to load the saved checkpoint instead. The runtime displayed when load model=1
is the time taken for test set inference. (This is how we’ll run your submission for this part
while grading!)
3.3 Grading
Please keep in mind that there are no partial marks in this assignment. Your code will not be
debugged while grading. Any submission that fails to run or does not fall above the Accuracy
lower-bound will get no marks. Exceeding Runtimes will result in penalties. All runtimes are
with respect to Colab GPU.
Runtime penalty: If you exceed the runtime threshold by 10*k%, you will be penalized k%. For
example, if you exceeded the runtime by 20%, the incurred penalty will be 2%.
3.3.1 Part 1
1. Accuracy: should be minimum 65% to get the maximum score. Score will scale linearly from
55-65% on the test set. Any submission with a test accuracy < 55% will get no marks.
2. Runtime: Should be less than 300 seconds.
3.3.2 Part 2
1. Accuracy: should be minimum 90% to get the maximum score. Score will scale linearly from
80-90% on the test set. Any submission with a test accuracy < 80% will get no marks.
2. Runtime (inference on test set only): Should be less than 150 seconds
3.4 Submission Guidelines
Zip CNN submission.py and best model.pth into Assignment2.zip. If you unzip the file, your folder
structure should look like:
Assignment2
CNN submission.py
best model.pth
Do not submit CNN main.py .
3.5 Collaboration Policy
This must be your own work. Do not share or look at the code of other students (whether they are
inside or outside the class). You can talk to others in the class about solution ideas (but detailed
enough that you are verbally sharing, hearing or seeing the code). You must cite online resources
that were referred to and to whom you talked with, in the comments of your programs.
Chatgpt is allowed. However, we reserve the right to evaluate any student’s submission further
through a viva, if we have reason to believe that they do not understand the solution that they have
submitted.