This project was developed as part of the Deep Learning course at UPF. Our goal was to explore the trade-off between model size and performance for the task of Facial Emotion Recognition on the FER2013 dataset. In particular, we focused on designing lightweight architectures (under 500k parameters) suitable for low-resource environments, while achieving fair performance across all emotion classes — especially in the presence of significant class imbalance.

The experiments performed here provide valuable insights into the behavior of compact CNN architectures such as ResNet and MobileNetV2 on a challenging imbalanced dataset. While several limitations remain, including the difficulty of distinguishing subtle emotions and the tendency of some models to overfit, the project highlights promising directions for building efficient FER systems.

This project consists of several Jupyter notebooks that guide you through the full experimentation pipeline: from exploratory data analysis (EDA) to model training and result visualization.

We recommend running the project in a virtual environment. Install the dependencies using:

pip install -r requirements.txt

After having installed the dependencies, you should be able to run any notebook in this repository. You can now open and run the notebooks in your preferred Jupyter environment.

Please note that the data folder provided in this repository is empty due to size restrictions. To run the notebooks, you should place your dataset inside the data folder with the following structure:

data/
├── train/
│   ├── angry/
│   ├── disgust/
│   ├── fear/
│   ├── happy/
│   ├── neutral/
│   ├── sad/
│   ├── surprise/
├── test/
│   ├── angry/
│   ├── disgust/
│   ├── fear/
│   ├── happy/
│   ├── neutral/
│   ├── sad/
│   ├── surprise/

Each subfolder should contain the corresponding images for each emotion class.

• EDA.ipynb
  Performs Exploratory Data Analysis on the dataset, including visualizations of class distributions and sample images.

• KerasBaseline.ipynb
  Trains and evaluates the VGG16 baseline model.

• ResNet.ipynb
  Defines, trains, and evaluates the custom ResNet variant.

• MobileNetV2.ipynb
  Defines, trains, and evaluates the custom MobileNetV2 variant.

• plots-resnet.ipynb and plots-mobv2.ipynb
  Generate plots for visualizing performance metrics and embedding spaces of the trained ResNet and MobileNetV2 models.

All experimental results (metrics, plots, intermediate model outputs) are saved in the results/ folder.