FER2025 – Facial Expression Recognition Dataset

---

Overview

FER2025 is a large-scale, balanced facial emotion dataset designed for deep learning and computer vision research. It contains 1,589,810 images across 7 emotion classes:

Class	Images
Angry	224,624
Disgust	239,366
Fear	223,466
Happy	222,082
Neutral	234,230
Sad	217,884
Surprise	228,158

Image formats: jpg, jpeg, png Balanced: Maximum class difference ≈ 1.3%

FER2025 is suitable for feature extraction, model training, and benchmarkingandTraining deep learning model.

---

Dataset Structure

FER2025/ ➡️ Angry.tar | Disgust.tar | Fear.tar | Happy.tar | Neutral.tar | Sad.tar | Surprise.tar

Each TAR contains images + corresponding .cls label files for efficient streaming.

---

Recommended Usage

• Feature Extraction: ResNet, EfficientNet, ViT embeddings
• Training & Evaluation: Balanced classes remove need for oversampling or class weighting
• Large-Scale Training: Use TAR/WebDataset format for GPU-efficient streaming

---

Example: Loading FER2025 with PyTorch + WebDataset

import webdataset as wds
from torchvision import transforms
import torch

transform = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.ToTensor(),
])

dataset = (
    wds.WebDataset("FER2025/{Angry,Disgust,Fear,Happy,Neutral,Sad,Surprise}.tar")
    .decode("pil")
    .to_tuple("jpg", "cls")
    .map_tuple(transform, int)
)

loader = torch.utils.data.DataLoader(dataset, batch_size=64, num_workers=4, shuffle=True)

for images, labels in loader:
    print(images.shape, labels.shape)
    break

---

License & Ethical Use

License: CC BY-NC 4.0 – Attribution required, non-commercial use

Ethical Use: Images are sourced from publicly available data for research. Users must respect privacy and avoid commercial misuse.

---

Citations

@dataset {FER2025, author = {Adhavan M}, title = {FER2025: Large-Scale Balanced Facial Expression Dataset}, year = {2025}, url = https://huggingface.co/datasets/imadhavan/FER2025 }