emotion-vit-model

Facial Emotion Recognition (ViT, fine-tuned on FER2013)

Fine-tuned google/vit-base-patch16-224-in21k for 7-class facial emotion recognition, with class-weighted + label-smoothed loss to handle FER2013's class imbalance (disgust is heavily underrepresented) and data augmentation (flip/rotation/jitter).

Classes

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

Test set results

• Accuracy: 0.7082
• Macro F1: 0.6784

⚠️ Important: input format required

This model was trained on FER2013, which consists of tightly cropped, grayscale, 48x48 face images. Feeding it a full, uncropped, color photo (e.g. straight from your phone or the internet) will give poor, low-confidence results, because that input looks nothing like what the model saw during training.

For good results, always crop to the face and convert to grayscale before running inference. The example below does this automatically using a face detector.

Usage (Google Colab)

facenet-pytorch pins an old numpy<2.0 / pillow combo that no longer has prebuilt wheels on current Colab runtimes, which causes a source-build error. Installing with --no-deps avoids re-resolving those pins and uses the numpy/pillow/torch already present in Colab:

%pip install facenet-pytorch --no-deps
%pip install requests torchvision --no-deps

Restart the session after installing (Runtime → Restart session in the Colab menu), then run:

import torch
import numpy as np
from PIL import Image
from facenet_pytorch import MTCNN
from transformers import AutoImageProcessor, AutoModelForImageClassification

processor = AutoImageProcessor.from_pretrained("bsgcasa/emotion-vit-model")
model = AutoModelForImageClassification.from_pretrained("bsgcasa/emotion-vit-model")
model.eval()

mtcnn = MTCNN(keep_all=False)

def predict_emotion(image_path):
    img = Image.open(image_path).convert("RGB")

    box, _ = mtcnn.detect(img)
    if box is None:
        print("No face detected.")
        return None

    x1, y1, x2, y2 = [int(v) for v in box[0]]
    face_crop = img.crop((x1, y1, x2, y2)).convert("L")  # grayscale, matches training

    face_rgb = Image.fromarray(np.stack([np.array(face_crop)] * 3, axis=-1).astype(np.uint8))

    inputs = processor(images=face_rgb, return_tensors="pt")
    with torch.no_grad():
        logits = model(**inputs).logits

    probs = torch.softmax(logits, dim=-1)[0]
    predicted_id = probs.argmax().item()

    print(f"Predicted: {model.config.id2label[predicted_id]}")
    for i, p in enumerate(probs):
        print(f"  {model.config.id2label[i]:>10}: {p:.3f}")

    return model.config.id2label[predicted_id]

predict_emotion("your_face_image.jpg")  # upload a photo to Colab and update this path

Troubleshooting

• ModuleNotFoundError: No module named 'facenet_pytorch' — the install cell either didn't run or silently failed. Use %pip (not !pip), check the cell's output for errors, and restart the session after installing.
• Build error on numpy/pillow — this is the version-pin issue described above; use --no-deps as shown.
• your_face_image.jpg not found — upload an image to the Colab file browser (folder icon on the left) and update the path, or use from google.colab import files; files.upload() to upload interactively.

Notes

• Trained with class-weighted + label-smoothed cross-entropy loss.
• Data augmentation applied during training (horizontal flip, rotation, color jitter).
• Face detection + grayscale conversion is required at inference time — the saved image processor only handles resizing/normalization, not face cropping. Using the model on full, uncropped, color photos will produce unreliable results.