Applied ML
GitHub RepositoryEmotion Recognition Platform
From model experiments to comparable facial-expression predictions in a usable interface.
Context: Applied ML product
Period: 2025-11 to 2026-02
Summary
The project keeps HOG/LBP and SVM baselines beside ResNet and EfficientNet models in one training, validation, comparison, and inference workflow. Evaluation scripts produce reports, while FastAPI and React make the registered models inspectable outside training notebooks.
From a facial image to an emotion prediction
Image input
A sample enters through the browser or evaluation workflow
Preprocessing
Vision transforms prepare consistent model inputs
Model inference
Classical and CNN models produce probabilities
FastAPI
The service exposes model operations
Result
The interface presents the prediction and confidence
What the system handled
Compares classical HOG/LBP + SVM baselines with CNN models.
Uses ResNet and EfficientNet model families.
Supports model training, validation, comparison, prediction, and submission generation.
Turns offline ML experiments into an interactive local demo.
Measured validation results
Repository evaluation artifacts from the reduced validation setup, with accuracy values rounded to two decimal places.
LBP + SVM accuracy
0.32
Reduced validation setup
HOG + SVM accuracy
0.38
Reduced validation setup
ResNet50 accuracy
0.57
Reduced validation setup
EfficientNet-B0 accuracy
0.58
Reduced validation setup
Ensemble accuracy
0.59
Reduced validation setup
Technology
Stack
Areas
Verified implementation
The model work is connected to an evaluation workflow
The repository carries registered models from training through validation, comparison, reporting, and serving in one project structure.
Classical baselines
HOG and LBP feature extractors paired with RBF SVM classifiers.
CNN families
ResNet and EfficientNet training and validation workflows.
Evaluation reports
Comparison scripts write machine-readable and human-readable results.
Usable inference path
FastAPI exposes predictions while React presents comparison and report views.
Conclusions
- The CNN models clearly outperformed the HOG/LBP and SVM baselines, while the ensemble produced a modest improvement over EfficientNet-B0 on the reduced validation setup.
- Serving the models through FastAPI and React extended the comparison into a usable inference workflow.
- The measured accuracy defines the intended interpretation: predictions classify facial expressions. A person's internal emotional state remains outside the model's scope.
Want to inspect the implementation?
The repository contains the source, setup notes, and development history.