ViT Emotion Recognition (FER2013)

Vision Transformer fine-tuned for 7-class facial emotion recognition on the FER2013 Enhanced dataset.

Best for in-the-wild / webcam faces.

Labels

angry, disgust, fear, happy, sad, surprise, neutral

Usage

from transformers import pipeline

clf = pipeline("image-classification", model="ChristopherLi/vit-fer2013-emotion")
print(clf("face.jpg"))

Or with this project's webcam app:

python webcam_app.py --model ChristopherLi/vit-fer2013-emotion

Training

  • Base: google/vit-base-patch16-224-in21k
  • 3 epochs, fp16, lr 5e-5, effective batch size 32
  • Input: 48ร—48 grayscale faces resized to 224ร—224 and replicated to 3 channels

Results (test set)

Metric Value
Accuracy 69.4%
Macro F1 0.656

FER2013 is a noisy benchmark (human agreement ~65%). See the project README for details.

Downloads last month
173
Safetensors
Model size
85.8M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ChristopherLi/vit-fer2013-emotion

Finetuned
(2,589)
this model

Dataset used to train ChristopherLi/vit-fer2013-emotion

Evaluation results