Mood-Detector

An end-to-end emotion classification pipeline: from raw FER2013 data to containerized REST API, training and evaluation reproduction using DVC and expermentation tracking using MLFlow.

DEMO

docker build -t mood-detector:v1 .
docker run -p 8000:8000 mood-detector:v1

then open http://localhost:8000 in your browser

Introduction

Most Deep Learning projects live in notebooks. This project demonstrates running experimentations, writing modular code, training custom architectural models from scratch, making it reproducible, using MLOps tools and techniques to make this project close to production ready, versioned and served. Closing the gap between "this works on my machine" and "this works in production".

Architecture Overview

Pipeline Diagram:

flowchart TD
    A[FER2013 Dataset\nKaggle API] --> B[data_ingestion.py\nDownload + Extract]
    B --> C[Preprocessing\nDrop classes 1,2,5,6\nRemap 0→0, 3→1, 4→2]
    C --> D[Train / Val / Test Split\n80% / 10% / 10%]
    D --> E[(data/processed/\ntrain.pkl\nval.pkl\ntest.pkl)]
    E --> F[model_training.py\nResidual CNN\nAdam + CrossEntropyLoss\nMLflow Tracking]
    F --> G[(model/\nbest_emotion_model.pth\nnormalization_stats.json\nclass_weights_stats.json\nrun_info.json)]
    G --> H[model_evaluation.py\nTest Set Evaluation\nClassification Report]
    H --> I[(reports/\neval_metrics.json)]
    H --> J[(mlflow.db\nTest Metrics\nArtifacts)]
    F --> J

    style A fill:#4A90D9,color:#fff
    style E fill:#2ECC71,color:#fff
    style G fill:#2ECC71,color:#fff
    style I fill:#2ECC71,color:#fff
    style J fill:#E67E22,color:#fff

Loading

---

Serving Diagram:

flowchart LR
    A[User Uploads\nJPEG / PNG] --> B[FastAPI\n/predict]
    B --> C[Read Image Bytes\nDecode via OpenCV]
    C --> D[Haar Cascade\nFace Detection]
    D -->|No face found| E[Return\n404 message]
    D -->|Face ROI| F[Preprocess\nGrayscale → Resize 48x48\nToTensor → Normalize]
    F --> G[emotionModel\nResidual CNN\n3-class output]
    G --> H[Softmax\nPer-class probabilities]
    H --> I[JSON Response\nemotion\nconfidence\nprobabilities]

    B2[FastAPI\n/health] --> J[Return\nstatus: healthy\nmodel_loaded: true]

    style A fill:#4A90D9,color:#fff
    style E fill:#E74C3C,color:#fff
    style I fill:#2ECC71,color:#fff
    style J fill:#2ECC71,color:#fff
    style G fill:#9B59B6,color:#fff

Loading

MLflow Experiment Results

Metrics across multiple runs

Metrics	RUN 1	RUN 2	RUN 3	RUN 4	RUN 5
Best Validation F1 Score	0.759	0.676	0.745	0.735	0.751
Train Accuracy	89.94	90.42	95.05	93.02	89.45
Train Loss	0.275	0.266	0.135	0.191	0.271
Validation Accuracy	71.95	66.26	67.24	73.43	74.56
Validation Loss	0.637	0.938	1.043	0.695	0.608
Test Accuracy	78.49	69.17	74.13	75.78	78.21
Test Loss	0.582	0.713	0.612	0.616	0.587
Test F1 Score	0.756	0.675	0.728	0.74	0.755

Parameters across multiple runs

RUNS	Batch Size	Epochs	Learning Rate	Optimizer	Weight Decay	Class Weights
RUN 1	32	70	0.001	Adam	0.001	True
RUN 2	32	70	0.0001	Adam	0.001	True
RUN 3	32	70	0.001	Adam	0.0001	True
RUN 4	64	70	0.001	Adam	0.001	True
RUN 5	32	70	0.001	Adam	0.001	False

• Run 1 (baseline with class weights, lr=0.001, batch_size=32) achieved the best validation F1 score of 0.759.
• Run 2 (lr=0.0001, batch_size=32) shows clear overfitting - high training accuracy with significantly worse validation and test accuracy scores, confirming learning rate is too conservative for this architecture.
• Run 3 (lower weight decay) overfits too with training accuracy being 95.05% but validation and test accuracy only being 67.24% and 74.13% respectively.
• Run 4 (increased batch_size to 64) performed significantly well compared to Runs 2 and 3, converged earlier than others, but still show slight overfitting compared to baseline (Run1) and Run 5.
• Run 5 (same parameters as Baseline, but without class weights(for loss function)), had similar results to Run 1(baseline) with slightly less overfitting and hence being used for the final application.

Model Details

Architecture

The model is a custom Residual Convolutional Neural Network (CNN) designed specifically for edge deployment on grayscale facial crops (48 x 48 pixels). It leverages residual connections to prevent vanishing gradients during backpropagation and heavily features dropout to counteract the noisy nature of the FER2013 dataset.

Tech Stack

Tool	Why
DVC	Data Versioning, pipeline reproducibility - dvc repro reruns only changed stages
MLFlow	Experiment tracking and comparison across runs, model artifacts storage
FastAPI	Typed, async-capable inference serving with auto-generated docs
Docker	Eliminates environment dependency - docker run -p 8000:8000 mood-detector:v1
Haar Cascade	Lightweight face detection without a second model dependency

Project Structure

Mood-Detector/
├── .dvc
├── assets/
│   ├── demo.gif
│   └── model_summary.png
├── experiments/
│   └── experimentation.ipynb
├── reports/
│   ├── .gitignore
├── src/
│   ├── haarcascades/
│   │   └── haarcascade_frontalface_default.xml
│   ├── static/
│   │   └── index.html
│   ├── app.py
│   ├── custom_dataset.py
│   ├── custom_model.py
│   ├── data_ingestion.py
│   ├── engine.py
│   ├── model_evaluation.py
│   ├── model_training.py
│   ├── predict.py
│   └── utils.py
├── .dvcignore
├── .gitignore
├── .python-version
├── dockerfile
├── dvc.lock
├── dvc.yaml
├── pyproject.toml
├── README.md
├── requirements-docker.txt
└── uv.lock

How to Run

• Setup Kaggle API

  • Go to the 'Account' tab on your Kaggle Profile.
  • Click 'Create New Token'. This will download a file named kaggle.json containing your API credentials.
  • Move this file to the appropriate location:
    • Linux/OSX: ~/.kaggle/kaggle.json
    • Windows: C:\Users\Windows-username\.kaggle\kaggle.json

• Clone the GitHub Repository

  git clone https://github.com/hari-vash/Mood-Detector
  cd Mood-Detector
  

• Configure virtual environment

  # create the virtual envrionment
  uv venv .
  
  # activate the environment
  # macOS/linux
  source .venv/bin/activate
  # Windows
  .venv\Scripts\activate
  
  # install the dependencies
  uv sync
  

• Install PyTorch with CUDA support (skip --index-url for CPU-only)

  uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
  

  (Note:- The link provided for PyTorch installation is for Cuda Version: 12.6)

• Reproduce Training from Scratch

  dvc repro
  
  # View experiment results after training
  mlflow ui --backend-store-uri sqlite:///mlflow.db
  

• Running the API locally

  cd src
  uvicorn app:app --port 8000
  

• Run via Docker(recommended) - from the root project directory

  # build image
  docker build -t mood-detector:v1 .
  
  # run as a container
  docker run -p 8000:8000 mood-detector:v1
  

  then open http://localhost:8000

What I'd Add for Production

• Model Registry with promotion gates - Currently the models go straight from training to serving. In production, a new model would have to beat the current production model's F1 on a held-out eval dataset.
• Data Drift Detection - FER2013 being a lab-collected dataset, the real world data would be different, hence input distribution will diverge over time.
• CI/CD Pipeline - GitHub actions to run tests, validate model performance, build and push docker image.
• Retraining Trigger - Current retraining is manual. In production it would be triggered by drift detection or on a schedule.
• GPU inference optimization - Current serving uses CPU. For latency-sensitive production use "ONNX + TensorRT" quantization to reduce inference time significantly.

Known Limitations

• FER2013 is a lab-collected dataset with large class imbalance and noise.
• Haar cascade face detection is inconsistent and sometimes fails with non-frontal faces, poor lighting etc.
• 3-class scope is deliberate but limits applicability.
• Model's performance and metrics reflects dataset noise, not architectural failure.

Acknowledgement/Data

Dataset used is FER2013, downloaded from kaggle. You can find the dataset here.