DrGM/DrGM-ConvNeXt-V2L-Facial-Emotion-Recognition

DrGM-ConvNeXt-V2-Large-FER

Model Description

This is a State-of-the-Art (SOTA) Facial Emotion Recognition (FER) model based on the ConvNeXt V2 Large architecture. It has been fine-tuned to recognize 7 distinct facial emotions with high accuracy.

Model Architecture

• Base Model: facebook/convnextv2-large-22k-224
• Parameters: ~198M
• Fine-tuning: Optimized with BF16 mixed precision, Label Smoothing (0.1), and RandAugment on an A100 GPU.

⚠️ License & Usage

This model is released under the CC-BY-NC-4.0 license.

• Personal Use: ✅ Allowed. You can use this for personal projects, research, and education.
• Commercial Use: ❌ Forbidden without prior permission.
• Commissions: If you wish to use this model for commercial applications or commissions, please contact the author for licensing.

Training History

The model was trained for 15 epochs on an A100 GPU. Below is the detailed progression of loss and metrics:

Epoch	Training Loss	Validation Loss	Accuracy	F1 (Weighted)
1	1.0126	0.9488	76.31%	0.7629
2	0.8600	0.8394	81.91%	0.8177
3	0.7161	0.7932	85.02%	0.8495
4	0.6340	0.7552	87.52%	0.8748
5	0.5956	0.7405	88.34%	0.8829
6	0.5568	0.7247	88.94%	0.8893
7	0.5259	0.7251	89.03%	0.8902
8	0.5149	0.7208	89.16%	0.8913
9	0.5071	0.7172	89.66%	0.8964
10	0.4984	0.7156	89.66%	0.8963
11	0.4933	0.7101	89.88%	0.8989
12	0.4857	0.7071	89.92%	0.8991
13	0.4803	0.7038	90.25%	0.9025
14	0.4718	0.7031	90.43%	0.9042
15	0.4730	0.7013	90.40%	0.9039

Final Training Metrics

• Total Training Time: ~49 minutes (2933.82 seconds)
• Global Steps: 11,805
• Final Training Loss: 0.5977
• Throughput: 257.37 samples/second

Performance

The model achieves exceptional performance on the Facial Emotion Expressions dataset.

Final Evaluation Results (Test Set)

After training, the model was evaluated on the unseen test set:

Metric	Value
Accuracy	90.43%
F1 Score (Weighted)	0.9042
Validation Loss	0.7031
Inference Time (Batch)	23.63s (Total)
Throughput	532.68 samples/sec

(Note: These metrics are from the held-out test split, confirming the model generalizes well and is not just memorizing data.)

Classification Report (Full Dataset Evaluation)

Class	Precision	Recall	F1-Score	Support
Angry	0.97	0.97	0.97	8989
Disgust	1.00	1.00	1.00	8989
Fear	0.97	0.96	0.97	8989
Happy	0.98	0.98	0.98	8989
Neutral	0.96	0.97	0.97	8989
Sad	0.96	0.96	0.96	8989
Surprise	0.99	0.99	0.99	8989
Accuracy			0.98	62923

(Note: Full dataset evaluation includes both training and validation samples, indicating high model capacity and learning)

Confusion Matrix

📊 Advanced Model Statistics

Global Accuracy Metrics:

• Top-1 Accuracy: 97.78%
• Top-2 Accuracy: 99.05% (Correct emotion is in the top 2 predictions)
• Top-3 Accuracy: 99.41%

Per-Emotion Performance Breakdown

Emotion	Accuracy	Avg Confidence	Samples
angry	97.49%	89.99%	8989
disgust	100.00%	91.34%	8989
fear	96.07%	89.66%	8989
happy	98.16%	90.72%	8989
neutral	97.35%	90.39%	8989
sad	96.27%	89.55%	8989
surprise	99.13%	90.81%	8989

Inference Speed Benchmark

Tested on an NVIDIA A100 GPU with a batch size of 1 (simulating real-time usage):

• Average Latency: 20.75 ms per image
• Frame Rate: 48.19 FPS

This performance indicates the model may be suitable for real-time video processing applications.

Usage

from transformers import AutoImageProcessor, AutoModelForImageClassification
import torch
from PIL import Image

# Load Model
repo_name = "DrGM/DrGM-ConvNeXt-V2L-Facial-Emotion-Recognition" 
processor = AutoImageProcessor.from_pretrained(repo_name)
model = AutoModelForImageClassification.from_pretrained(repo_name)

# Predict
image = Image.open("path/to/your/image.jpg").convert("RGB")
inputs = processor(image, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
    predicted_label = logits.argmax(-1).item()
    print(model.config.id2label[predicted_label])