Final project for DTU 02476 Machine Learning Operations course
By: Kevin Moore, Mario Medoni, Angeliki-Artemis Doumeni and Konstantina Freri
Our project leverages natural language processing (NLP) to identify mental health disorders from Reddit posts across five major mental health subreddits. By building and fine-tuning a machine learning model, we aim to classify Reddit posts into categories that reflect what the author might suffer from or seek help with. This approach showcases the potential of NLP in analyzing user-generated content and offers a step toward developing tools to enhance mental health awareness and support systems.
We will use the Hugging Face Transformers library, a powerful tool for NLP tasks. It offers pre-trained transformer models such as BERT and RoBERTa, known for their strong performance in understanding textual data. These models will be fine-tuned using both the title and selftext (body) of Reddit posts to classify them into five specific mental health disorder categories. This framework enables us to utilize state-of-the-art NLP techniques efficiently while focusing on our specific classification problem.
Additionally, we will incorporate PyTorch Lightning to streamline code and reduce boilerplate, enabling faster experimentation and improved project organization.
The project will utilize the “Mental Disorders Identification (Reddit)” dataset from Kaggle. This dataset contains approximately 700k rows, including:
- Post title (string)
- Post body (selftext) (string)
- Creation time (timestamp)
- Over 18 tag (boolean)
- Subreddit name (mental disorder label, string)
Note: We will exclude the mentalillness subreddit, as it represents a generalized disorder, reducing the classification task to five subreddits.
Link to the Kaggle Dataset: Mental Disorders Identification (Reddit)
We will fine-tune transformer models such as BERT and RoBERTa. As a baseline, we will use random guessing, assuming a balanced dataset where each category has a 20% probability of being predicted. This baseline will provide a clear comparison to demonstrate the performance of our models.
The directory structure of the project looks like this:
├── .github/ # Github actions and dependabot
│ ├── dependabot.yaml
│ └── workflows/
│ └── tests.yaml
├── configs/ # Configuration files
├── data/ # Data directory
│ ├── processed
│ └── raw
├── dockerfiles/ # Dockerfiles
│ ├── api.Dockerfile
│ └── train.Dockerfile
├── docs/ # Documentation
│ ├── mkdocs.yml
│ └── source/
│ └── index.md
├── models/ # Trained models
├── notebooks/ # Jupyter notebooks
├── reports/ # Reports
│ └── figures/
├── src/ # Source code
│ ├── project_name/
│ │ ├── __init__.py
│ │ ├── api.py
│ │ ├── data.py
│ │ ├── evaluate.py
│ │ ├── models.py
│ │ ├── train.py
│ │ └── visualize.py
└── tests/ # Tests
│ ├── __init__.py
│ ├── test_api.py
│ ├── test_data.py
│ └── test_model.py
├── .gitignore
├── .pre-commit-config.yaml
├── LICENSE
├── pyproject.toml # Python project file
├── README.md # Project README
├── requirements.txt # Project requirements
├── requirements_dev.txt # Development requirements
└── tasks.py # Project tasksCreated using mlops_template, a cookiecutter template for getting started with Machine Learning Operations (MLOps)
-
Build docker
docker build -t api-model -f dockerfiles/api.dockerfile . -
Run container
docker run --name mycontainer -p 8000:8000 api-model -
Open
http://localhost:8000/docs -
Profit
-
Build docker (remember to specify x86_64 if youre on Apple silicon)
docker buildx build --platform linux/amd64 -t api-model -f dockerfiles/api.dockerfile . -
(optional) set project to the correct in gcloud
gcloud config set project dtumlops-448110 -
Tag the image
docker tag api-model gcr.io/dtumlops-448110/api-model:latest -
Push docker to gcloud
docker push gcr.io/dtumlops-448110/api-model:latest -
Run/Deploy gcloud service
gcloud run deploy api-model-service \
--image gcr.io/dtumlops-448110/api-model:latest \
--platform managed \
--region europe-west1 \
--allow-unauthenticated \
--port 8000