Project description

---

Goal

This project is undertaken by Group 50 as part of the 02476 Machine Learning Operations course at DTU. The primary objective is to apply the principles and techniques learned during the course to solve a machine learning problem. We aim to leverage one of the provided frameworks to build a scalable and well-structured solution. As an outcome, we could expect to be gained a profficiency in MLOps tools such as, Docker, Hydra, WanDB, Cloud services and many more.

Framework

For this project, we have chosen the TIMM (PyTorch Image Models) framework, specifically designed for computer vision tasks. TIMM offers a wide selection of pre-trained models that can be fine-tuned to suit our needs. This framework is particularly advantageous due to its robust integration with PyTorch, streamlined APIs, and access to cutting-edge architectures. Using TIMM, we will construct and experiment with state-of-the-art deep learning models. As part of this process, we will set up the TIMM framework within our development environment, ensuring that our codebase is clean, modular, and well-managed through version control.

Data

The dataset selected for this project is the Cats vs Dogs dataset from Kaggle. This dataset consists of 25,000 labeled images of cats and dogs, making it ideal for a binary image classification task. The dataset is well-suited for our goal of exploring and benchmarking convolutional neural networks (CNNs). The images are diverse in terms of size, lighting, and orientation, which presents unique challenges for model training. Before feeding the data into our models, we will perform preprocessing steps such as resizing, normalization, and augmentation to enhance model performance.

Models

Our task is to perform binary image classification on the Cats vs Dogs dataset. Initially, we will develop a baseline CNN model to establish a point of reference. Subsequently, we will explore and compare the performance of advanced models available in the TIMM framework, including ConvNeXt, MobileNet V3, ResNet, and VGG. This comparison will help us understand the strengths and trade-offs of each model in terms of accuracy, computational efficiency, and scalability.

---

Instructions

Setup

1. Clone the repository

2. Create a virtual environment

conda create -n mlops python=3.11 -y
conda activate mlops

3. Install invoke

pip install invoke

4. Install the required packages

invoke requirements

or if in development mode

invoke dev-requirements

5. Install the pre-commit hooks

pre-commit install

Data

The dataset is available on Kaggle at the following link: Cats-vs-Dogs.

To download it, first you need a Kaggle token. To this go to your settings page, and in API section click on Create New Token. This will download a kaggle.json file. Depending on your OS, you will need to place this file in the following directories:

• Windows: C:\Users\<Windows-username>\.kaggle\kaggle.json
• Linux, OSX, and other UNIX-based OS: ~/.kaggle/kaggle.json

Then, on the terminal, run the following command:

invoke preprocess-data

This will download the dataset and preprocess it. The number of images to use per class, image size and test size can be modified (TODO: add configuration).

Training

To train the model, run the following command:

invoke train

Also, parameters can be set:

invoke train --lr 0.01 --batch-size 32 --epochs 10

Results will be saved in the reports directory.

Evaluation

To evaluate the model, run the following command:

invoke evaluate --model-path <path-to-model>

Unit testing

To run the unit tests, run the following command:

invoke test

Good coding practices

Styling

We use 'ruff' to enforce code styling. To check the code styling, run the following command:

ruff check .

Some styling issues can be fixed automatically by running:

ruff check . --fix

Other issues will need to be fixed manually.

To format the code, run the following command:

ruff format .