# Homework 3

In this homework, we will use all the trick at our disposal to tune a classification network on supertux. We then make this classification network fully convolutional and solve a semantic labeling task (labeling every pixel in the image).

This assignment should be solved individually. No collaboration, sharing of solutions, or exchange of models is allowed. Please, do not directly copy existing code from anywhere other than your previous solutions, or the previous master solution. We will check assignments for duplicates. See below for more details.

In the first part of this homework, you’ll tune your classifier, or the classifier of the master solution 2 to achieve as high an accuracy as possible. You may directly copy any code from your solution or the master solution to homework 2. The goal of this first part is to tune the classifier to an accuracy of $90\%$. This might require several tricks:

• Input normalization
• Residual blocks
• Dropout
• Data augmentations (Both geometric and color augmentations are important. Be aggressive here. Different levels of supertux have radically different lighting.)
• Weight regularization
• Early stopping

We managed to tune our classifier to a $94\%$ accuracy with a training time of 10 GPU minutes. If you manage to reach this accuracy you’ll get 10 extra credits.

python3 -m grader homework


## Dense prediction (semantic segmentation)

In the second part of the homework, you’ll make your CNN fully convolutional. Instead of predicting a single output per image, you’ll now predict an output per pixel.

Our current dataset does not support this output, we thus switch to a new dense prediction dataset. Here all images have a $128 \times 96$ resolution, and the labels are of the same size. We have 5 labels here: background, kart, track, bomb/projectile, pickup/nitro. We merged the last two classes, as they already have very few labeled instances. The class distribution for those labels is quite bad. background and track make up $96\%$ of the labeled area! We will address this issue later for evaluation.

To see some examples from the dense dataset run

python3 -m homework.utils


### FCN design

Design your FCN by writing the model in models.py. Make sure to use only convolutional operators, pad all of them correctly and match strided operators with up-convolutions. Use skip and residual connections.

Make sure your FCN handles an arbitrary input resolution and produces an output of the same shape as the input. Use output_padding=1 if needed. Crop the output if it is too large.

python3 -m grader homework


### FCN Training

To train your FCN you’ll need to modify your CNN training code a bit. First, you need to use the DenseSuperTuxDataset. This dataset accepts a data augmentation parameters transform. Most standard data augmentation in torchvision do not directly apply to dense labeling tasks. We thus provide you with a smaller subset of useful augmentations that properly work with a pair of image and label in dense_transforms.py.

You will need to use the same bag of tricks as for classification to make the FCN train well.

Since the training set has a large class imbalance, it is easy to cheat in a pixel-wise accuracy metric. Predicting only track and background gives a $96\%$ accuracy. We additionally measure the Intersection-over-Union evaluation metric. This is a standard semantic segmentation metric that penalizes largely imbalanced predictions. This metric is harder to cheat, as it computes $$\frac{\text{true positives}}{\text{true positives} + \text{false positives} + \text{false negatives}}.$$ You might need to change the class weights of your torch.nn.CrossEntropyLoss, although our master solution did not require this. You can compute the IoU and accuracy using the ConfusionMatrix class.

python3 -m grader homework


### Relevant Operations

python3 -m grader homework -v


will run a subset of test cases we use during the actual testing. The point distributions will be the same, but we will use additional test cases. More importantly, we evaluate your model on the test set. The performance on the test grader may vary. Try not to overfit to the validation set too much.

## Submission

Once you finished the assignment, create a submission bundle using

python3 bundle.py homework [YOUR UT ID]


python3 -m grader [YOUR UT ID].zip


We will use an automated grader through canvas to grade all your submissions. There is a soft limit of 5 submisisons per assignment. Please contact the course staff before going over this limit, otherwise your submission might be counted as invalid.

The online grading system will use a slightly modified version of python and the grader:

• Please do not use the exit or sys.exit command, it will likely lead to a crash in the grader
• Please do not try to access, read, or write files outside the ones specified in the assignment. This again will lead to a crash. File writing is disabled.
• Network access is disabled. Please do not try to communicate with the outside world.
• Forking is not allowed!
• print or sys.stdout.write statements from your code are ignored and not returned.

Please do not try to break or hack the grader. Doing so will have negative consequences for your standing in this class and the program.

You might need a GPU to train your models. You can get a free one on google colab. We provide you with a ipython notebook that can get you started on colab for each homework.

If you’ve never used colab before, go through colab notebook (tutorial)
When you’re comfortable with the workflow, feel free to use colab notebook (shortened)

Follow the instructions below to use it.

• Select the upload tab then select the .ipynb file.
• Follow the instructions on the homework notebook to upload code and data.

## Honor code

This assignment should be solved individually.

What interaction with classmates is allowed?

• Talking about high-level concepts and class material
• Talking about the general structure of the solution (e.g. You should use convolutions and ReLU layers)
• Looking at online solutions, and pytorch samples without directly copying or transcribing those solutions (rule of thumb, do not have your coding window and the other solution open at the same time). Always cite your sources in the code (put the full URL)!
• Using any of your submissions to prior homework
• Using the master solution to prior homework
• Using ipython notebooks from class

What interaction is not allowed?

• Exchange of code
• Exchange of architecture details
• Exchange of hyperparameters
• Directly (or slightly) modified code from online sources
• Any collaboration
• Putting your solution on a public repo (e.g. github). You will fail the assignment if someone copies your code.

Ways students failed in past years (do not do this):

• Student A has a GPU, student B does not. Student B sends his solution to Student A to train 3 days before the assignment is due. Student A promises not to copy it but fails to complete the homework in time. In a last-minute attempt, Student A submits a slightly modified version of Student B’s solution. Result: Both students fail the assignment.

• Student A struggles in class. Student B helps Student A and shows him/her his/her solution. Student A promises to not copy the solution but does it anyway. Result: Both students fail the assignment.

• Student A sits behind Student B in class. Student B works on his homework, instead of paying attention. Student A sees Student B’s solution and copies it. Result: Both students fail the assignment.

• Student A and B do not read the honor code and submit identical solutions for all homework. Result: Both students fail the class.

## Installation and setup

### Installing the dependencies

Install all dependencies using

python3 -m pip install -r requirements.txt


Note: On some systems, you might be required to use pip3 instead of pip for python 3.

If you’re using conda use

conda env create environment.yml


The test grader will not have any dependencies installed, other than native python3 libraries and libraries mentioned in requirements.txt. This includes packages like pandas. If you use additional dependencies ask on piazza first, or risk the test grader failing.

### Manual installation of pytorch

Go to https://pytorch.org/get-started/locally/ then select the stable Pytorch build, your OS, package (pip if you installed python 3 directly, conda if you installed Anaconda), python version, cuda version. Run the provided command. Note that cuda is not required, you can select cuda = None if you don’t have a GPU or don’t want to do GPU training locally. We will provide instruction for doing remote GPU training on Google Colab for free.

### Manual installation of the Python Imaging Library (PIL)

The easiest way to install the PIL is through pip or conda.

python3 -m pip install -U Pillow


There are a few important considerations when using PIL. First, make sure that your OS uses libjpeg-turbo and not the slower libjpeg (all modern Ubuntu versions do by default). Second, if you’re frustrated with slow image transformations in PIL use Pillow-SIMD instead:

CC="cc -mavx2" python3 -m pip install -U --force-reinstall Pillow-SIMD


The CC="cc -mavx2" is only needed if your CPU supports AVX2 instructions. pip will most likely complain a bit about missing dependencies. Install them, either through conda, or your favorite package manager (apt, brew, …).