Computer Vision

This project teaches a computer to recognise seven facial expressions (angry, disgust, fear, happy, neutral, sad, surprise) from small grayscale photos, and compares three ways of building that model.

I built and trained all three approaches in PyTorch, evaluated them on 7,178 held-out test images, and compared the results.

Knowing when a big pretrained model helps, and when it doesn't, saves real compute and time. This experiment showed that the off-the-shelf approach alone wasn't enough.

FER2013 is hard: 48×48 grayscale faces, noisy labels, and heavy class imbalance (436 'disgust' vs 7,215 'happy' training images).

PyTorch on a Colab T4 GPU. A custom convolutional network is trained from scratch at native resolution, then compared with EfficientNet-B0 pretrained on ImageNet (images resized to 224×224, grayscale to 3 channels), first as a frozen feature extractor and then with its last blocks unfrozen.

1. 01FER201328,709 train / 7,178 test
2. 02Custom CNNfrom scratch
3. 03EfficientNet-B0frozen backbone
4. 04Partial fine-tunelast blocks unfrozen
5. 05Compare on test set

1. 01Baseline before transfer learningTraining a small CNN first set an honest baseline to judge whether a pretrained backbone actually helps.
2. 02Partial unfreezingImageNet features transfer poorly to tiny grayscale faces, so the last blocks plus the classifier were unfrozen with a 90/10 train/validation split.

• -A frozen ImageNet backbone barely beat the scratch CNN (44.57% vs 44.25%), which pointed to domain mismatch rather than model capacity.

Custom CNN

44.25%

test accuracy, 7 classes, 5 epochs

EfficientNet-B0 (frozen)

44.57%

test accuracy, 20 epochs

Python · PyTorch · torchvision · EfficientNet-B0 · Google Colab · NumPy · Matplotlib

• View in the terminal →