4/29/25, 1:06 PM cnnquiz.
md
CNN Quiz Summary
The CNN quiz will ask you to
1. obtain a fresh copy of the pytorch tutorial,
2. apply a series of steps, and
3. report your accuracy statistics after each step.
Description of Steps
Each quiz question will ask you to perform a combination of the following steps, and then
report your accuracy. You do not need to perform all the following steps in their order
here.
Step A1: Obtaining a Fresh Copy of the Assignment
If you do not want to lose your current progress in the assignment, look for the
following question "How can I get a fresh copy of the assignment while preserving
the old one?" question in the FAQ about Jupyter Notebooks article.
Please note that you should not use or build upon your submitted notebook for this
quiz, and you need to start from scratch for this quiz.
Step A2: Copying the PyTorch Tutorial
You can copy and paste the content from
https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html into your jupyter
notebook one cell at a time just like you did when you started working on the
assignment.
You only need to do this for the CIFAR portion of the assignment.
Upon finishing the copy & paste operation, locate the following two lines (under the
cell titled "1.1 Loading the Data")
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 1/6
4/29/25, 1:06 PM cnnquiz.md
testset = torchvision.datasets.CIFAR10(root='./data', train=False,
download=True, transform=transform)
and modify their root argument as the following:
root_data = '/home/jovyan/work/release/CNN-lib/data_cifar'
trainset = torchvision.datasets.CIFAR10(root=root_data, train=True,
download=True, transform=transform)
testset = torchvision.datasets.CIFAR10(root=root_data, train=False,
download=True, transform=transform)
Important Note: Make sure you copy every single piece of evaluation or visualization
code, even if it had nothing to do with the training. This is important so that you do
not have a hard time reproducing the results.
Important Note: Here is a print of a notebook with the tutorial material copied into it.
Make sure that (1) you have all the cells this template has, and (2) the cell contents are
identical before proceeding to the next part.
Step A3: Controlling the Randomized Effects
Copy and paste the following 3 lines of code right at the top of the cell titled "1.1 Loading
the Data":
np.random.seed(12345)
random.seed(12345)
torch.manual_seed(12345)
After pasting these 3 lines, your cell should look like the following
np.random.seed(12345)
random.seed(12345)
torch.manual_seed(12345)
message = 'You can implement the pre-processing transformations, data sets, data loa
message = message + '**Important Note**: Read the "Grading Reference Pre-processing"
message = message + ' test pre-processing transformations in the "Autograding and Fi
message = message + ' training models for long periods of time.'
print(message)
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 2/6
4/29/25, 1:06 PM cnnquiz.md
...
Why are we doing this?
For maximum reproducibility and better comparison, it is often a good idea to fix your
random number generator seeds to a fixed number.
This way you can achive "deterministic stochasticity", meaning that while you did not
specify the random numbers in your code, your code will generate the same exact
output every time you run it from scratch.
That being said, trying to exploit the statistics by finding a random seed that increases
your accuracy is an unforgivable sin.
Step A4: Restart the Python Kernel
On the top left of your screen, there is a "Kernel " menu. Click on it.
Click on either the "Restart" or the "Restart and Clear Output" options, and wait a few
seconds for the restart request to go through.
Sometimes the restart request does not execute properly or require more time to
finish.
Do not perform this step too quickly.
If you see a small circle filled with black next to the "Python 3 (Threads: 2)"
message in the top right corner of your screen, this means that the kernel is busy
executing your request, and you need to wait for the circle to be unfilled.
Step A5: Running the CIFAR Portion of the Notebook
Start running the notebook cells one-by-one from the top of the notebook.
Make sure that no cell raises a pytho error, and the trained network gets stored at
cifar_net.pth .
You may want to back-up your old cifar_net.pth file that you used for your
assignment submission, but we will not ask you for it.
Run all the cells before the MNIST part starts, including the cell titled "1.8 Autograding
and Final Tests"
Step A6: Reporting the Final Training Loss and Overall Testing
Accuracy
After running the CIFAR part of the notebook without an error, the cell titled "1.8
Autograding and Final Tests" will print a final testing accuracy value at the end of its
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 3/6
4/29/25, 1:06 PM cnnquiz.md
output.
Look for the phrase "Overall Testing Accuracy" as the following:
----------
Overall Testing Accuracy: XX.YY %%
The number XX.YY will be your final testing accuracy that should be reported in your
quiz.
For the final training loss, look for the output of the training cell (i.e., the cell under the
title "1.5 Training the Model"). For example, suppose the following is the entire
training log:
[1, 2000] loss: 2.222
[1, 4000] loss: 1.869
[1, 6000] loss: 1.679
[1, 8000] loss: 1.588
[1, 10000] loss: 1.511
[1, 12000] loss: 1.445
[2, 2000] loss: 1.400
[2, 4000] loss: 1.363
[2, 6000] loss: 1.335
[2, 8000] loss: 1.297
[2, 10000] loss: 1.292
[2, 12000] loss: 1.345
Finished Training
The value 1.345 is the last printed loss, and you are supposed to enter this value as
your final training loss.
Step A7: Updating the Pre-processing transformations
Locate the following lines from the pytorch tutorial in the cell titled "1.1 Loading the Data"
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
and change it to the following
transformation_list = [transforms.RandomAffine(degrees=30, translate=(0.01, 0.01),
scale=(0.9, 1.1), shear=None,
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 4/6
4/29/25, 1:06 PM cnnquiz.md
resample=0, fillcolor=0),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
transform = transforms.Compose(transformation_list)
This way you will be using a matching pre-processing for both the training and testing
phases.
Step A8: Increasing the Number of Optimization Epochs
Locate the following for loop under the title "1.5 Training the Model"
for epoch in range(2): # loop over the dataset multiple times
and change the number of epochs to 10:
for epoch in range(10): # loop over the dataset multiple times
It is worth noting that only fixing the pre-processing transformation (Step A6) and
training the model for 80 epochs (i.e., for about an hour and half) would have gotten a
full score upon submitting the assignment.
Step A9: Increasing the Second Convolutional Layer's Channels
Locate the following Neural Network constructor lines
class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
...
and increase the number 16 to 64 in both self.conv2 and self.fc1 definition lines:
class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 5/6
4/29/25, 1:06 PM cnnquiz.md
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 64, 5)
self.fc1 = nn.Linear(64 * 5 * 5, 120)
...
https://courses.grainger.illinois.edu/cs441/fa2023/secure/9_cnnquiz/cnnquiz.html 6/6
