Pytorch Implement of YOLO

Introduction

The full name of YOLO is You Only Look Once. It is a popular model with high speed and accuracy used for Object Detection. You can learn more in the official website.

Dataset

VOC2012

Requirements

• Python>=3.5
• Pytorch>=1.4
• OpenCV
• moviepy
• scipy
• PIL

Detect

You can run:

python detect.py
    --model_load_path=models/yolov3.pth
    --class_path=data/coco-ch.names
    --input_path=data/dog.jpg
    --output_path=data/dog_pred.jpg
    --device_ids=0

Now enjoy!

All usages and optional arguments:

 usage: detect.py [-h] [--model_load_path MODEL_LOAD_PATH]
                 [--class_path CLASS_PATH] [--color_path COLOR_PATH]
                 [--anchor_path ANCHOR_PATH] [--input_path INPUT_PATH]
                 [--output_path OUTPUT_PATH] [--not_show]
                 [--score_threshold SCORE_THRESHOLD]
                 [--iou_threshold IOU_THRESHOLD] [--device_ids DEVICE_IDS]
                 [--num_processes NUM_PROCESSES]

Object detection.

optional arguments:
  -h, --help            show this help message and exit
  --model_load_path MODEL_LOAD_PATH
                        Input path to models.
  --class_path CLASS_PATH
                        Path to a file to store names and colors of the
                        classes.
  --color_path COLOR_PATH
                        Path to a file which stores colors.
  --anchor_path ANCHOR_PATH
                        Input path to anchors.
  --input_path INPUT_PATH
                        Path to the file used for detection. If zero, camera
                        on your computer will be used.
  --output_path OUTPUT_PATH
                        Path to the output image or video. If Empty, the
                        predicted image will not be saved.
  --not_show            Whether not to show predictions.
  --score_threshold SCORE_THRESHOLD
                        Threshold of score(IOU * P(Object)).
  --iou_threshold IOU_THRESHOLD
                        Threshold of IOU used for calculation of NMS.
  --device_ids DEVICE_IDS
                        Device ids. Should be seperated by commas. -1 means
                        cpu.
  --num_processes NUM_PROCESSES
                        number of processes.

Run with flask

python run_flask.py

Transform .weights to .pth

We firstly use .weights and .cfg files to generate and save a Tensorflow model. The table below shows how to do this.

Model	repo	outputs
yolov1, yolov1-tiny	https://github.com/thtrieu/darkflow	a .pb file and a .meta file
yolov3	https://github.com/jinyu121/DW2TF	3 .ckpt files and a file named checkpoint

Then we use these files to generate a Pytorch model by running pb2pth.py.

Thanks

darkflow

DW2TF