pat229988/AffectNet-face-to-emotion-tflite

Emo-AffectNet Face Emotion Recognition — TensorFlow Lite

This repository contains a TensorFlow Lite float32 conversion of the static image face emotion recognition model from ElenaRyumina/face_emotion_recognition.

The model predicts one of seven facial emotion classes from an RGB face image.

Files

File	Description
emo_affectnet_static_float32.tflite	Converted TensorFlow Lite model.
labels.txt	Output labels in model order.
conversion_report.json	Conversion metadata and PyTorch-vs-TFLite parity check.
infer_tflite.py	Minimal Python inference example.
requirements.txt	Python dependencies for the inference example.

Model details

• Source model: ElenaRyumina/face_emotion_recognition
• Original project: ElenaRyumina/EMO-AffectNetModel
• Task: Face emotion / facial expression classification
• Format: TensorFlow Lite, float32
• Input: RGB image tensor, shape [1, 224, 224, 3], float32, pixel range 0..255
• Output: Logits tensor, shape [1, 7]; apply softmax outside the model

Labels

The output logits use this label order:

1. Neutral
2. Happiness
3. Sadness
4. Surprise
5. Fear
6. Disgust
7. Anger

Quick start

Install the example dependencies:

pip install -r requirements.txt

Run inference on a face image:

python infer_tflite.py --image path/to/face.jpg

Example output:

{
  "top_label": "Happiness",
  "top_score": 0.9123,
  "predictions": [
    {"label": "Happiness", "score": 0.9123},
    {"label": "Neutral", "score": 0.0531},
    {"label": "Surprise", "score": 0.0214}
  ]
}

Preprocessing

The TFLite model expects:

1. A face image in RGB order.
2. Resize to 224 × 224 pixels.
3. Convert to float32.
4. Keep pixel values in the 0..255 range.
5. Add a batch dimension to produce shape [1, 224, 224, 3].

Do not normalize to 0..1 unless you reconvert the model with a different input convention.

Conversion validation

The conversion report records a parity check between the converted model and the source PyTorch model:

• Samples checked: 3
• Maximum absolute difference: 1.4901161193847656e-06
• Mean absolute difference: 5.336921825952837e-07
• Top-1 match rate: 1.0

See conversion_report.json for the full metadata.

Publishing to Hugging Face

This folder is ready to upload as a Hugging Face model repository.

Recommended upload flow:

pip install -U huggingface_hub
hf auth login
hf repos create <namespace>/<repo-name>
hf upload <namespace>/<repo-name> . . --repo-type model --exclude ".serena/*" --exclude "__pycache__/*"

If you publish with Git instead of huggingface-cli upload, install Git LFS first so the .tflite file is stored correctly:

git lfs install
git init
git add .
git commit -m "Add TFLite face emotion recognition model"
git remote add origin https://huggingface.co/<namespace>/<repo-name>
git push -u origin main

Limitations and intended use

• This model is intended for facial expression recognition research and application prototyping.
• Predictions can be sensitive to face detection quality, crop alignment, lighting, occlusion, pose, and image quality.
• Emotion labels are simplified categories and should not be treated as definitive measurements of a person's internal state.
• Evaluate the model carefully before using it in production or high-impact settings.

Citation

If you use the original Emo-AffectNet model in research, cite the source paper:

@article{RYUMINA2022,
  title   = {In Search of a Robust Facial Expressions Recognition Model: A Large-Scale Visual Cross-Corpus Study},
  author  = {Elena Ryumina and Denis Dresvyanskiy and Alexey Karpov},
  journal = {Neurocomputing},
  year    = {2022},
  doi     = {10.1016/j.neucom.2022.10.013},
  url     = {https://www.sciencedirect.com/science/article/pii/S0925231222012656}
}

License

The source model and original project are licensed under the MIT License. This converted TFLite repository keeps the same license. See LICENSE and NOTICE.md.