SEARO1/FER_model

Facial Expression Recognition for Mental Health Detection

Model Description

This model is a Swin Transformer fine-tuned for facial expression recognition (FER) with applications in mental health detection. It can classify facial expressions into 7 categories and provide depression risk analysis based on emotional patterns.

Model Architecture

• Base Model: Swin Transformer (swin_base_patch4_window7_224)
• Custom Classifier:
  • Linear layer (backbone features → 512)
  • ReLU activation
  • Dropout (p=0.6)
  • Linear layer (512 → 7 classes)

Emotion Classes

The model predicts 7 facial expressions:

1. Angry 😠
2. Disgust 🤢
3. Fear 😨
4. Happy 😊
5. Neutral 😐
6. Sad 😢
7. Surprise 😲

Training Details

Dataset

• Name: FER2013 (Facial Expression Recognition 2013)
• Size: ~35,000 grayscale images (48x48 pixels)
• Split: Train/Validation/Test

Training Configuration

• Optimizer: AdamW
• Learning Rate: 1e-4 with cosine annealing
• Batch Size: 32
• Epochs: 5
• Image Size: 224x224
• Data Augmentation: Random horizontal flip, rotation, color jitter
• Loss Function: Cross-Entropy Loss

Usage

Installation

pip install torch torchvision timm huggingface_hub

Load Model

import torch
import timm
from huggingface_hub import hf_hub_download

class CustomSwinTransformer(torch.nn.Module):
    def __init__(self, pretrained=True, num_classes=7):
        super(CustomSwinTransformer, self).__init__()
        self.backbone = timm.create_model('swin_base_patch4_window7_224', 
                                         pretrained=pretrained, num_classes=0)
        self.classifier = torch.nn.Sequential(
            torch.nn.Linear(self.backbone.num_features, 512),
            torch.nn.ReLU(),
            torch.nn.Dropout(p=0.6),
            torch.nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.backbone(x)
        return self.classifier(x)

# Download and load model
model_path = hf_hub_download(repo_id="SEARO1/FER_model", filename="best_model.pth")
model = CustomSwinTransformer(pretrained=False, num_classes=7)
model.load_state_dict(torch.load(model_path, map_location='cpu'), strict=False)
model.eval()

Inference Example

from torchvision import transforms
from PIL import Image

# Prepare image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

image = Image.open("face.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0)

# Predict
with torch.no_grad():
    output = model(input_tensor)
    probabilities = torch.nn.functional.softmax(output, dim=1)
    predicted_class = torch.argmax(probabilities, dim=1)

emotions = ['Angry', 'Disgust', 'Fear', 'Happy', 'Neutral', 'Sad', 'Surprise']
print(f"Predicted Emotion: {emotions[predicted_class.item()]}")
print(f"Confidence: {probabilities[0][predicted_class].item()*100:.2f}%")

Mental Health Application

This model can be used for depression risk analysis by analyzing emotional patterns:

Depression Risk Calculation

def analyze_depression_risk(emotion_probs):
    sad_score = emotion_probs[5]  # Sad
    fear_score = emotion_probs[2]  # Fear
    angry_score = emotion_probs[0]  # Angry
    happy_score = emotion_probs[3]  # Happy
    
    negative_emotions = (sad_score * 0.4 + fear_score * 0.3 + angry_score * 0.3)
    positive_emotions = happy_score
    
    depression_risk = (negative_emotions * 100) - (positive_emotions * 20)
    depression_risk = max(0, min(100, depression_risk))
    
    if depression_risk < 30:
        return "Low Risk"
    elif depression_risk < 60:
        return "Moderate Risk"
    else:
        return "High Risk"

⚠️ Important: This is an educational tool and should NOT replace professional medical advice or diagnosis.

Performance

The model achieves competitive performance on the FER2013 dataset. See the training logs for detailed metrics.

Limitations

• Trained on FER2013 dataset which may not represent all demographics equally
• Performance may vary with different lighting conditions, angles, and image quality
• Should not be used as the sole basis for mental health diagnosis
• Requires frontal face images for best results

Citation

If you use this model, please cite:

@misc{fer-mental-health-2024,
  author = {Your Name},
  title = {Facial Expression Recognition for Mental Health Detection},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/SEARO1/FER_model}}
}

License

MIT License - See LICENSE file for details

Contact

For questions or issues, please open an issue on the model repository.

---

Developed for educational and research purposes in mental health technology.