Mood Journal

Offline-first Progressive Web App (PWA) for multi-label emotion classification and journaling. Emotion classification is powered by a fine-tuned SetFit model using MiniLM embeddings, with inference running entirely client-side via ONNX Runtime Web and Transformers.js

FastAPI serves the web application and model assets; all inference is executed locally in the browser.

Key Features

• Offline-first PWA
  • All inference and journaling run fully client-side via a service worker and Transformers.js
  • Emotion analysis is performed locally in the browser so no user data or journal entries ever leave the device
• Service worker powered by Workbox
  • Pre-caches the ONNX model and tokenizer assets required for offline inference
  • After first load, all features (including ML) work fully offline (offline-first caching strategy)
• Multi-label emotion classification
  • Fine-tuned SetFit (MiniLM) classifier predicts multiple emotions from a single journal entry
  • SetFit was selected because it provides strong performance on limited labeled datasets

Tech Stack

Backend/ML: Python, FastAPI, PyTorch, Transformers, ONNX Runtime

Frontend: React, TypeScript, Vite, Tailwind CSS, Workbox

Demo

• Live demo: https://mood-tracker.projects.jaisawhney.me/

Project Structure

├── apps/
│   ├── api/            # FastAPI backend
│   └── web/            # React/Vite frontend
├── docker/             # Container configuration
├── images/             # README assets
├── ml/
│   ├── config.py       # Typed config loader
│   ├── config.yaml     # Project configuration
│   ├── data.py         # Dataset loading
│   ├── datasets/       # Train/validation/test datasets
│   ├── export.py       # ONNX export pipeline
│   ├── inference.py    # Document-level inference
│   ├── evaluate.py     # Model benchmarking
│   ├── trainer.py      # SetFit training
│   └── validate.py     # Model evaluation
├── notebooks/          # Experimentation
├── tests/              # Test suite
├── pyproject.toml
└── README.md

Model Performance

The dataset is highly imbalanced across emotion labels. Despite class rebalancing and hyperparameter optimization, minority classes remain the main source of prediction error.

The underlying classifier predicts a broader set of emotion labels than are exposed in the web app. To improve usability, the web app presents a curated subset of emotions that are most relevant to mood journaling while continuing to predict the full set internally.

Datasets

• Lemotif (Li & Parikh, arXiv 2019): Journaling dataset derived from personal journal entries, used for fine-tuning and evaluation

Data Preprocessing:

• Undersampling of majority classes to reduce class imbalance

Notebooks

• 01_dataset_exploration.ipynb: Exploratory Data Analysis
• 02_data_preprocessing.ipynb: Data Preprocessing & Class Rebalancing
• 03_hyperparameter_optimization.ipynb: Hyperparameter Optimization (Optuna)
• 04_final_model_evaluation.ipynb: Final Model Evaluation

Installation & Usage

Prerequisites

• Python 3.10+
• Node.js 18+ (for frontend dev)

Install Python dependencies

pip install -e ".[api,ml,dev]"

Train the model

python -m ml.trainer

FastAPI (development)

uvicorn apps.api.main:app --reload

API docs: http://localhost:8000/docs

Frontend (development)

cd apps/web
npm install
npm run dev

Docker

Start locally with docker-compose (models mounted from MODELS_DIR):

export MODELS_DIR=/path/to/models
docker compose up --build -d

Model files from MODELS_DIR are served inside the container at /api/models.

Testing

pytest

References

• Dataset: lemotif (Li & Parikh, arXiv 2019)