algernonx / sexyyolo

An implementation of Yolov3 with Tensorflow1.x, which could detect COCO and sexy or porn person simultaneously.

Sexy Yolo

---

Introduction

A hierarchical detection method based on Yolov3, which can work on joint classification and detection dataset, such as COCO and NSFW. So that Yolo could detect coco categories and sexy or porn person simultaneously. (Tensorflow 1.x)

---

Requirements

Python version: >=Python3.6
Packages:

• tensorflow >=1.13 (do not support Tensorflow2.0)
• opencv-python
• numpy
• pyyaml

---

Quick Start

if model weights is ready and basic code is cloned. Please edit configs/XXX.yaml, add your local model weights path to MODEL:MODEL_WEIGHTS:, such as "path/to/yolov3".

common coco:

$ python tools/infer.py test_pic.jpg(or folder) --config-file configs/Yolov3_Coco.yaml --result-file data/demo_result

sexy and coco:

$ python tools/infer.py test_pic.jpg(or folder) --config-file configs/Yolov3_Joint_CocoAndNsfw.yaml --result-file data/demo_result

---

Model weights

Two model weights are provided: common COCO model and Sexy&COCO model.
Common COCO weights is converted from author of Yolov3(pjreddie.com), Sexy&COCO weights are trained on image size of 416 * 416, so mAP is less than the former but close to Yolov3 paper's 416 image size model mAP.

Name	inference time(s/im)	category	AP@IoU=0.50:0.95	AP@IoU=0.50	AP@IoU=0.75	download
COCO	0.046	80	0.357	0.616	0.376	model
SexyCOCO	0.047	83	0.309	0.551	0.308	model

---

Train your own data

Sexy Yolo support train your own detection dataset and also support joint classification and detection dataset.

Train detection dataset

The detection annotation's format must be same as COCO's format. Add training dataset photo folder and annotation file to config file and run:

$ python tools/train.py --config-file configs/Yolov3_Coco.yaml

Train joint classification and detection dataset

As for classification dataset, put all the classification photos into each folder, the folder's name is catogory's name. Make your dataset folder structure as:

• NSFW

• -train

• --neutral

• --sexy

• --porn

• -val

• --neutral

• --sexy

• --porn

Then run:

$ python tools/joint_json_creator.py path/to/NSFW

After dataset is ready, edit Yolov3_Joint_CocoAndNsfw.yaml with new dataset path. Run this start to train:

python tools/train.py --config-file configs/Yolov3_Joint_CocoAndNsfw.yaml

You can change config parameters by editing config file.

---

Dataset

COCO dataset: http://cocodataset.org/
NSFW dataset: https://github.com/alex000kim/nsfw_data_scraper

---

Evaluate model

CocoApi is needed, please install cocoapi first: https://github.com/cocodataset/cocoapi
Run this to evaluate your model:

$ python tools/eval_on_coco.py --config-file configs/Yolov3_Joint_CocoAndNsfw.yaml

---

More Details

• please see the blog Description and discussion

sexyyolo's Issues

nooc_weight 是什么意思

CATEGORY_NUM: (80, 196)

CATEGORY_NUM: (80, 196) # coco categoty
When I give the new category number 196, the training does not work. I could not survive. Can I ask for your help?

测试实例错误

python tools/infer.py data/demo/10.jpg --config-file configs/Yolov3_Coco.yaml --result-folder data/

Use standard file APIs to check for files with this prefix.
load model weights success!
Traceback (most recent call last):
File "tools/infer.py", line 64, in
cfg['INFER']['SECOND_CATE_THRESH'])
KeyError: 'SECOND_CATE_THRESH'

Sex S

Suckme

够Sexy，我喜欢