Facial Expression Classification

Deep learning model created for the Facial Expression Classification competition on Kaggle. Reference research_poster.pdf for implementation details.

Results

Metric	Validation (with TTA)
Accuracy	0.836
Macro precision	0.821
Macro recall	0.819
Macro F1	0.819

Kaggle test accuracy: 0.837

Note: runs are not bit-for-bit reproducible across hardware and library versions, so expect small differences (a few tenths of a point) from the numbers above.

Settings used

Setting	Value
Random seed	42 (SEED), split random_state=42
Data split	80/20 train/validation, stratified by class
Input	48×48 grayscale scaled to 0–1, resized to 96×96
Model	ResNet-50 (1-channel conv1), pretrained weights remapped from final_model.pth, new head: Dropout(0.5) + Linear(2048, 3)
Trained layers	layer3, layer4, and the new head (earlier layers frozen, kept in eval mode)
Optimizer	AdamW, weight decay 0.02; learning rates layer3 3e-5, layer4 7e-5, head 1e-4
Schedule	Cosine annealing over 20 epochs
Loss	Cross-entropy with label smoothing 0.1
Batch size	64, plain shuffling
Training augmentation	Random horizontal flip, rotation up to 10°, shift up to 10%, scale 0.9–1.1
Model selection	Best validation accuracy across epochs
Inference	Flip test-time augmentation (average of image and mirrored image)

Requirements

• Python 3.10+ (developed on <your Python version>)
• Packages (install with pip install -r requirements.txt):

torch>=2.6
torchvision>=0.21
numpy
pandas
scikit-learn
matplotlib
jupyter

Files needed

The files train_data.csv, train_target.csv, and test_data.csv are too large to commit to GitHub. Download them from the Data section of the Facial Expression Classification competition on Kaggle.

.
├── code.ipynb                  # the notebook
├── base_model.pth              # pretrained ResNet-50 checkpoint (source: <link or where you got it>)
└── data/
    ├── train_data.csv          # training images: one row per image, 2304 pixel values (48x48), no header
    ├── train_target.csv        # training labels: one column, no header
    └── test_data.csv           # test images, same format as train_data.csv, no labels provided

Labels: 0 = angry, 1 = happy, 2 = neutral.

Note the absence of test_target.csv. The Kaggle competition facilitators excluded this file from the Data section in order to prevent cheating.

How to run

1. Install the requirements.
2. Install the necessary data files.
3. Open code.ipynb in Jupyter Notebook.
4. Restart the kernel and run all cells in order (the model and optimizer are created in specific cells, so rerunning individual cells continues from already-trained weights).
5. After training, the notebook writes submission_full.csv (columns Id, Category) and saves the best weights to model.pth.