When 73% was state-of-the-art for reading faces

What it does Trains and evaluates VGG19 and ResNet18 on two standard facial-expression datasets: FER2013 (48×48 grayscale images, seven emotions) and CK+ (981 frames extracted from video sequences). The repo includes pretrained weights, confusion-matrix plotting, and a script to visualize predictions on a single test image. The README reports a private-test accuracy of 73.112% on FER2013 and 94.646% on CK+ using VGG19.

The interesting bit This is less a framework and more a preserved benchmark snapshot. Its value lies in packaging 2017-era baselines with downloadable weights, making it a quick sanity-check reference for emotion-classification experiments without reading a dozen papers.

Key highlights

• Supports both FER2013 and CK+ with dedicated training and evaluation scripts.
• Provides pretrained VGG19 and ResNet18 models via external download links.
• Includes utilities for confusion-matrix generation and single-image visualization.
• Reports accuracies of 73.112% on FER2013 (private test) and 94.646% on CK+ (10-fold cross-validation).

Caveats

• Targets Python 2.7 and PyTorch >=0.2.0, so expect significant compatibility friction with modern stacks.
• The “state-of-the-art” claim is historical and the README does not date it or compare against newer methods.

Verdict Worth a look if you need a minimal, reproducible baseline for classic FER datasets or want to compare against 2017-era scores. Skip it if you need modern architectures, current data pipelines, or production-ready code.