← anamaya.fyi On-device
FER-2013 · convolutional network · ONNX Runtime Web

Seven expressions

No server. The image is read into a canvas and turned into numbers in this tab.

01

What the model sees

A convolutional network trained on FER-2013, exported to ONNX and executed here. Nothing is uploaded, and there is no backend to upload it to.

What the model sees · 48×48 grayscale

Pick a face below, upload one, or use your camera.

Loading the model…

02

Faces it has never seen

Held-out FER-2013 test images with their true labels. Some of these it gets wrong — they are included on purpose.

03

How good is it, honestly

Human agreement on FER-2013 is roughly 65% and published state of the art is near 73%. A model small enough to download lands below both.

Per-class recall matters more than the headline: the training set has 7,215 happy faces and 436 disgusted ones, so a model can ignore disgust completely and barely be punished for it by accuracy alone.

04

Confusion matrix

Rows are the true expression, columns what the model said. Darker is more.