Version: v1.0
Framework: TensorFlow 2.x
Optimized format: TensorFlow Lite
Input: 478 Mediapipe Face Mesh landmarks per frame (up to 300 frames)
Output: 6-class emotion prediction (Angry, Disgust, Fear, Happy, Neutral, Sad)

🧠 Model Overview

The Emotion Sequence Transformer is a deep learning model built using TensorFlow for recognizing human emotions from continuous video clips.
It uses 478 Mediapipe facial landmarks per frame to capture spatiotemporal patterns of facial movements across time.
The model predicts one of six basic emotions by analyzing both facial geometry and temporal variation within sequences of up to 300 frames.

This model is suitable for real-time video-based emotion detection, affective computing, human-computer interaction, and emotion-aware AI systems.

📊 Dataset

This model was trained on the Optimized 478-Point 3D Facial Landmark Dataset —
a dataset derived from the Video Emotion Dataset, optimized for emotion recognition using Mediapipe’s 3D face mesh landmarks.

Each sample in the dataset includes:

Up to 300 frames per clip
478 facial landmarks per frame
Corresponding emotion label

🧩 Model Architecture

The architecture is based on a Transformer encoder design that processes sequential data of facial landmarks.

Pipeline:

Input normalization using precomputed mean and std (global stats)
Sequence embedding via positional encodings
Transformer encoder blocks to capture temporal and spatial dependencies
Dense layers for emotion classification (6 output neurons with softmax)

Core Components:

Transformer Encoder Layers (Multi-Head Self-Attention)
Layer Normalization and Dropout
Dense classification head

📈 Performance

Metric	Value
Test Accuracy	0.7289
Test Loss	1.1336
Macro F1-Score	0.73
Weighted F1-Score	0.73
Max Clip Length	300 frames
Input Shape	(300, 478)

🧾 Classification Report

Emotion	Precision	Recall	F1-score	Support
Angry	0.75	0.73	0.74	139
Disgust	0.88	0.70	0.78	128
Fear	0.52	0.60	0.55	114
Happy	0.88	0.97	0.92	129
Neutral	0.66	0.79	0.72	101
Sad	0.70	0.58	0.64	134
Overall Accuracy	0.73		Macro Avg: 0.73	745

📊 Visualizations

🔹 Training Accuracy and Loss

🔹 Confusion Matrix

🔹 ROC Curves (Per Class)

📂 Repository Structure

TF-Emotion-Sequence-Transformer/
├── tf_emotion_sequence_transformer_mp478_seq256.h5
├── tf_emotion_sequence_transformer_mp478_seq256_optimized.tflite
├── tf_emotion-sequence-transformer-bilstm-usage.ipynb
├── assets/
│   ├── global_mean.npy
│   ├── global_std.npy
│   ├── label_encoder.pkl
│   └── metadata.json
└── README.md

File Descriptions

File	Description
`tf_emotion_sequence_transformer_mp478_seq256.h5`	Main TensorFlow model trained on 478 landmarks (300 frames max).
`tf_emotion_sequence_transformer_mp478_seq256_optimized.tflite`	Optimized TensorFlow Lite version for deployment (mobile, edge).
`tf_emotion-sequence-transformer-bilstm-usage.ipynb`	Example notebook demonstrating how to use the model for emotion prediction from Mediapipe landmarks.
`assets/global_mean.npy`	Precomputed global mean for normalization.
`assets/global_std.npy`	Precomputed global standard deviation for normalization.
`assets/label_encoder.pkl`	Encoder mapping integer labels to emotion names.
`assets/metadata.json`	Model metadata and configuration details.

🚀 Example Usage

🔸 TensorFlow (.h5) Model

python

import numpy as np
import tensorflow as tf
import joblib
import json

# Load Model
model = tf.keras.models.load_model("tf_emotion_sequence_transformer_mp478_seq256.h5")

# Load assets
mean = np.load("assets/global_mean.npy")
std = np.load("assets/global_std.npy")
label_encoder = joblib.load("assets/label_encoder.pkl")

# Preprocess input
input_seq = np.load("example_input.npy")  # shape: (300, 478)
input_seq = (input_seq - mean) / std
input_seq = np.expand_dims(input_seq, axis=0)

# Predict
pred = model.predict(input_seq)
emotion = label_encoder.inverse_transform([np.argmax(pred)])[0]
print("Predicted Emotion:", emotion)

🔸 TensorFlow Lite (Optimized) Model

python

import numpy as np
import tensorflow as tf
import joblib

# Load TFLite model
interpreter = tf.lite.Interpreter(model_path="tf_emotion_sequence_transformer_mp478_seq256_optimized.tflite")
interpreter.allocate_tensors()

# Get input and output tensors
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Load preprocessing assets
mean = np.load("assets/global_mean.npy")
std = np.load("assets/global_std.npy")
label_encoder = joblib.load("assets/label_encoder.pkl")

# Prepare input
input_seq = np.load("example_input.npy")  # shape: (300, 478)
input_seq = (input_seq - mean) / std
input_seq = np.expand_dims(input_seq, axis=0).astype(np.float32)

# Inference
interpreter.set_tensor(input_details[0]['index'], input_seq)
interpreter.invoke()
pred = interpreter.get_tensor(output_details[0]['index'])

# Decode emotion
emotion = label_encoder.inverse_transform([np.argmax(pred)])[0]
print("Predicted Emotion:", emotion)

🔖 Version Information

Version: v1.0
Date: November 2025
Author: P.S. Abewickrama Singhe
Framework: TensorFlow 2.x
Exported Models: .h5, .tflite
Landmarks per frame: 478
Max frames per clip: 300

🏷️ Tags

tensorflow • emotion-recognition • mediapipe • transformer • sequence-model • facial-landmarks • video-analysis • tflite • human-emotion-ai • affective-computing • computer-vision • deep-learning

📚 Citation

If you use this model in your research, please cite it as:

bibtex

@misc{pasindu_sewmuthu_abewickrama_singhe_2025,
	author       = { Pasindu Sewmuthu Abewickrama Singhe },
	title        = { EmotionFormer-BiLSTM (Revision f329517) },
	year         = 2025,
	url          = { https://huggingface.co/PSewmuthu/EmotionFormer-BiLSTM },
	doi          = { 10.57967/hf/6899 },
	publisher    = { Hugging Face }
}

🪪 License

This model is released under the Apache 2.0 License — free for academic and commercial use with attribution.