README.md

metadata

license: apache-2.0
language:
  - en
base_model:
  - Ultralytics/YOLO11
tags:
  - yolo
  - yolo11
  - nsfw
pipeline_tag: object-detection

EraX-NSFW-V1.0

A Highly Efficient Model for NSFW Detection

• Developed by:

  • Phạm Đình Thục ([email protected])
  • Mr. Nguyễn Anh Nguyên ([email protected])

• License: Apache 2.0

Model Details / Overview

• Model Architecture: YOLO11m
• Task: Object Detection (NSFW Detection)
• Dataset: Private datasets (from Internet).
• Training set: 31890 images.
• Validation set: 11538 images.
• Classes: anus, make_love, nipple, penis, vagina.

Labels

Training Configuration

• Model Weights File: erax_nsfw_yolo11m.pt
• Number of Epochs: 100
• Learning Rate: 0.01
• Batch Size: 208
• Image Size: 640x640
• Training server: 8 x NVIDIA RTX A4000 (16GB GDDR6)
• Training time: ~10 hours

Evaluation Metrics

Below are the key metrics from the model evaluation on the validation set:

• Precision: 0.726
• Recall: 0.68
• mAP50: 0.724
• mAP50-95: 0.434

Format	Status	Size (MB)	metrics/mAP50-95(B)	Inference time (ms/im)	FPS
PyTorch	✅	38.6	0.4332	16.97	58.91
TorchScript	✅	77	0.4153	12.09	82.69
ONNX	✅	76.7	0.4153	103.94	9.62
OpenVINO	❌	-	-	-	-
TensorRT	✅	89.6	0.4155	7	142.92
CoreML	❌	-	-	-	-
TensorFlow SavedModel	✅	192.3	0.4153	40.19	24.88
TensorFlow GraphDef	✅	76.8	0.4153	36.71	27.24
TensorFlow Lite	❌	-	-	-	-
TensorFlow Edge TPU	❌	-	-	-	-
TensorFlow.js	❌	-	-	-	-
PaddlePaddle	✅	153.3	0.4153	1024.24	0.98
NCNN	✅	76.6	0.4153	187.36	5.34

Training Validation Results

Training and Validation Losses

Confusion Matrix

Inference

To use the trained model, follow these steps:

1. Install the necessary packages:

pip install ultralytics supervision

2. Simple Use Case:

from ultralytics import YOLO
from PIL import Image
import supervision as sv

IOU_THRESHOLD        = 0.3
CONFIDENCE_THRESHOLD = 0.2
KERNEL_SIZE          = 20

pretrained_path = "erax_nsfw_yolo11m.pt"
image_path_list = ["img_0.jpg", "img_1.jpg"]

model = YOLO(pretrained_path)
results = model(image_path_list,
                  conf=CONFIDENCE_THRESHOLD,
                  iou=IOU_THRESHOLD
                )


for result in results:
  annotated_image = result.orig_img.copy()
  selected_classes = [0, 2, 3, 4, 5] # hidden make_love class
  detections = sv.Detections.from_ultralytics(result)
  detections = detections[np.isin(detections.class_id, selected_classes)]

  box_annotator = sv.BoxAnnotator()
  blur_annotator = sv.BlurAnnotator(kernel_size=KERNEL_SIZE)
  label_annotator = sv.LabelAnnotator(text_color=sv.Color.BLACK)

  annotated_image = blur_annotator.annotate(
        annotated_image.copy(),
        detections=detections
    )
  annotated_image = box_annotator.annotate(
        annotated_image,
        detections=detections
    )

  annotated_image = label_annotator.annotate(
        annotated_image,
        detections=detections
    )
    
  sv.plot_image(annotated_image, size=(10, 10))

Training

Scripts for training: https://github.com/EraX-JS-Company/EraX-NSFW-V1.0

Examples

1. Example 01:

2. Example 02:

3. Example 03:

4. Example 04:

5. Example 05:

6. Example 06:

Citation

If you find our project useful, we would appreciate it if you could star our repository and cite our work as follows:

@article{EraX-NSFW-V1.0,
  author    = {Phạm Đình Thục and
              Mr. Nguyễn Anh Nguyên and
              Đoàn Thành Khang and
              Mr. Trần Hải Khương and
              Mr. Trương Công Đức and 
              Mr. Phan Nguyễn Tuấn Kha and 
              Phạm Huỳnh Nhật},
  title     = {EraX-NSFW-V1.0: A Highly Efficient Model for NSFW Detection},
  organization={EraX JS Company},
  year={2024},
  url={https://huggingface.co/erax-ai/EraX-NSFW-V1.0}
}