MiniProject 3: Multi-label Classification of Image Data

 
Problem definition

In this mini-project, we will develop models to classify image data. We will use a ”modified” MNIST dataset: https://drive.google.com/file/d/1LcKqf1d7bctw5lx0YZf31kCUF0zEYOsi/view?usp=sharing where each image contains 1-5 digits. The task is to implement a model for this multi-label classification task.

If the number of digits in an image is less than 5 then the remaining labels are associated with a special class called ”no-digit”. As can be seen in the figure-1, no-digit is class ”10”. This means we have a total of 11 classes (0-10 digits and no-digit), and each image is associated with 5 such classes. The dataset contains both training and test data, where the label for the test set is not provided.

You are free to use any Python libraries you like to extract features and pre-process the data, evaluate your model, and to tune the hyper-parameters, etc. Your model should be a neural network and it should be trained by automatic differentiation.

This project can be divided into two part:

    1. Implement a deep neural network model for the task and report the followings:

        ◦ Describe the proposed model and training objective for the task in detail.

        ◦ Report the best performing model architecture and its hyper-parameter in a table in the Appendix.

        ◦ Justify the of choice of hyper-parameters (model architecture, optimizer and its parameters) that lead to this final model.

        ◦ Compare and report the validation and train performance of your best performing model as a function of training epochs.

    2. Participate in COMP 551 Kaggle competition:

        ◦ Once you have formed a group in MyCourses, go to the Multi-label Classification of Image Data page below and join the competition using the same group name: https://www.kaggle.com/t/162b731e49e44d7e914aabc04f449a8f

        ◦ 50% of the grade is allocated based on the performance on the kaggle competition. Once you have a trained model, you need to predict the test label and submit at the Kaggle competition.

        ◦ The evaluation is split private and public score (based on a further split of the test data). The private will be revealed after the deadline and it will be used for grading.

        ◦ Here’s how the submission file should look like, ”sample.csv”: https://drive.google.com/file/d/1xcmkOcshHTtROxYf9ca8Q7WFJ6R6OHIa/view? usp=sharing.

It should have two columns (”Id”,”Label”), where the Label will be the sequence of predicted labels. For example, have a look at the sample table, where predicted [0; 1; 2; 8; 10] has been converted as a sequence of [012810].

Id
Label
0
14923
1
27238
2
65192
3
012810

Table 1: An example of submission file


Evaluation

Evaluation has two components, where the first component is similar to previous mini-projects:

    • 10% completeness

    • 15% correctness

    • 10% writing and code quality

    • 15% originality and creativity

The second component is your based on your group’s performance on the private part of the test set in the Kaggle competition (50%):

    • 5% for successful submission of your results in the competition

    • 20% for obtaining the performance better than the baseline

    • 25% based on ranking

Deliverables

In addition to participating in the competition, please submit:

    1. code.zip: Your data processing, classification and evaluation code (as a combination of .py and .ipynb files).

    2. writeup.pdf: Your (max 2-page) project write-up as a pdf (appendices may be used for additional graphs and tables if really needed). In contrast to previous mini-projects there is no need to analyze the dataset unless you find it important to discuss to justify your proposed method.