CS 334 Section 2 — Machine Learning
Fall 2024

Homework 5: The Convoluted Case of Ken’s Canine Companions
Issued: Fri. 11/08, Due: Fri. 11/22 at 11:59pm

Dog-gone! The neighborhood kennel owner, Ken, is having a “ruff” time: during his morning rounds to check on his canine companions, he found to his horror that they had all dug their way under the fence and escaped! Knowing how much these dogs like to play tricks on him, he thought of a trick of his own to help find his furry friends: placing cameras around town and using image recognition techniques to detect passing dogs. He’s come to the CS-334-2 team in search of help implementing this system.

Ken’s kennel houses dogs from 10 distinct breeds: Samoyeds, Miniature Poodles, Golden Retrievers, Great Danes, Dalmatians, Collies, Siberian Huskies, Yorkshire Terriers, Chihuahuas, and Saint Bernards. He has provided us with images of all of his dogs — a few of which are shown below.

Figure 1: Sample images from the the dataset.

THIS CODE IS MY OWN WORK, IT WAS WRITTEN WITHOUT CONSULTING CODE

The main part of this HW requires you to write <30 lines of code. We’ve designed the HW to run in a reasonable time (<5min) on standard CPU machines; however, if you have access to a compatible GPU and wish to use it, please specify if you have used any CUDA/MPS GPU device. Challenge submissions that have used GPU hardware, additional data, or pretrained models will be graded separately.2

Getting Started

Dataset

1 Data Preprocessing [20 pts]

While you don’t need to write any code for this question, you are encouraged to study the implementation of the preprocessing functions as well as how they are used. Then, answer the following questions:

Save and include at least 3 sets of images in your write-up (each set is 4 pairs). What are the visible effects of preprocessing on the image data?

2 Convolutional Neural Networks [60 pts]

We have provided the training framework necessary to complete this problem in train_cnn.py. All of the following problems that involve implementation will require you to use the PyTorch API. We’ve provided some tips in Appendix A, and you are encouraged to explore the documentation and online tutorials.3

Now that you’ve implemented the architecture, let’s train the neural net. Review the content of the trainer script train_cnn.py. In the main function, we have a training loop that loops through the dataset for a prespecified number of times as defined in the config.json file. After each complete pass of the training set (aka “epoch”), we evaluate the model on the validation set, and save model parameters as a checkpoint.

Run train_cnn.py to train the model. If you run the script more than once, you will be prompted to enter the epoch number at which the model parameters from the saved checkpoint will be restored and training will continue from that epoch. If you enter 0, then the model will be trained from scratch and the old checkpoints will be deleted. This script will also create two graphs that are updated every epoch to monitor the model’s performance in terms of loss and accuracy on the train and validation set.

(e) Include the final learning curve plots [3 pts] in your write-up and answer the following questions.

In a convolutional layer, every filter has different weights and generates a distinct activation map given the same input. For example, the first convolutional layer has 16 filters, thus we have 16 different activation maps. These activation maps are single-channel images, and we can plot them using a continuum of colors, where darker color indicates small values and lighter color indicates large values. Filters in one layer are sorted by the mean of each filter’s weights to improve consistency of results.

Answer the following questions:

3 Challenge [20 pts]

Once your model has been trained, ensure that its checkpoint exists under checkpoints/challenge/ and run the script challenge_predict.py. This will load your model from your saved checkpoint and generate predictions.csv that contains your model’s predictions on the test set.

Your grade for this question will be assessed by the following two components:

Note that this means that the output test predictions follow a specified order. For that reason, please do not shuffle the test set data points during prediction.

This class implements the 2D convolutional layer we have learned in lecture. Create the layer as follows: torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding). Use default settings for all other arguments. For example dilation=1 means no dilation is used, and bias=True adds a learnable bias term to the neuron.

With Pytorch’s default setting padding=0, if we apply a 3 × 3 consolutional filter to a 4 × 4 input image, the output would be 2 × 2. As we keep applying convolutional layers in this way, the spatial dimensionswill keep decreasing. However, in the early layers of a CNN, we want to preserve as much informationabout the original input volume so that we can extract those low level features. To do this, we can pad the borders of the input image with zeros in a way such that the output dimension is the SAME as the input (assuming unit strides). If the filter size is odd k = 2m + 1, then this amount of zero padding on each side is p = ⌊k/2⌋ = m. For example, in Figure 3, since the filter size is k = 3, the appropriate padding size is p = ⌊3/2⌋ = 1.

Let ⃗z be the network output logits and yˆc the probability that the network assigns to the input example as belonging to class c. Let ⃗y ∈ R D be a one-hot vector with a one in the position of the true label and zero otherwise (i.e., if the true label is t, then yc = 1 if t = c and yc = 0 if t = c). D is the number of classes.

For a multiclass classification problem, we typically apply a soft max activation at the output layer to generate a probability distribution vector. Each entry of this vector can be computed as:

These two steps can be done together in PyTorch using the CrossEntropyLoss class, which combines the softmax activation with negative log likelihood loss and improves computational efficiency and numerical stability. We don’t need to add a separate soft max activation at the output layer.