Terorra/wakee-reloaded

Wakee - Emotion Detection Model

Version: baseline_20260216_1243

Model Description

EfficientNet B4 fine-tuned for emotion detection in educational settings.

Predicts 4 emotion intensities (0-3 scale):

• Boredom
• Confusion
• Engagement
• Frustration

Training Data

• Base: DAiSEE dataset
• Fine-tuned: User-validated annotations from Wakee app

Usage

ONNX (Production)

import onnxruntime as ort
import numpy as np
from PIL import Image

# Load model
session = ort.InferenceSession("model.onnx")

# Preprocess image (224x224)
image = Image.open("image.jpg").resize((224, 224))
input_array = np.array(image).transpose(2, 0, 1).astype(np.float32)
input_array = np.expand_dims(input_array, axis=0) / 255.0

# Predict
outputs = session.run(['output'], {'input': input_array})
boredom, confusion, engagement, frustration = outputs[0][0]

PyTorch (Fine-tuning)

import torch
from torchvision import models

# Load checkpoint
model = models.efficientnet_b4()
model.classifier[1] = torch.nn.Linear(model.classifier[1].in_features, 4)
model.load_state_dict(torch.load("model.bin"))

Model Card

• Architecture: EfficientNet B4
• Framework: PyTorch 2.1.2
• Input: RGB images (224x224)
• Output: 4 emotion scores (regression)
• License: MIT

Metrics

See model_versions table in database for evaluation metrics.

Citation

@software{wakee_emotion_detection,
  author = {Terorra},
  title = {Wakee Emotion Detection Model},
  year = {2025},
  version = {baseline_20260216_1243},
}