patryklaskowski / Convert_Darknet_YOLO_to_TensorFlow

Darknet YOLO architecture implemented in Tensorflow and Tensorflow Lite.

Convert_Darknet_YOLO_to_TensorFlow

Darknet YOLO architectures implemented in Tensorflow and Tensorflow Lite.

---

Before You start:

• In the first place You need to have Darknet YOLOv3 or v4 weights to work with. Weights might be either custom trained or pre-trained on benchmark COCO dataset. To download pre-trained yolov4.weights click here.
• Except weights, .names file is required for model to have class labels reference. For benchmark COCO dataset, file coco.names is already available here.

Start

1. Prepare environment

git clone https://github.com/patryklaskowski/Convert_Darknet_YOLO_to_TensorFlow.git && \
cd Convert_Darknet_YOLO_to_TensorFlow && \
python3.7 -m venv env && \
source env/bin/activate && \
python3.7 -m pip install -U pip && \
python3.7 -m pip install -r requirements.txt

2. Put .weights file in ./data/ folder.

Darknet YOLOv4 weights to download.

yolov4.weights (COCO dataset)	yolov4_licence_plate.weights
Download	Download

My .weights file is here: ./data/yolov4_licence_plate.weights

3. Prepare .names file respectively to .weights.

.names file contains all class labels for specific YOLO weights where each line represents one class name.

NOTE

.names file for domain yolov4.weights is already prepared on path ./data/classes/coco.names. coco.names has 80 rows -> each one corresponds to single label.

coco.names	licence_plate.names
Show on path	Download

My .names file is here: ./data/classes/licence_plate.names

4. Adjust config.py file.

File is here: ./core/config.py. Edit only __C.YOLO.CLASSES value to be path that points prepared .names file.

NOTE

By default __C.YOLO.CLASSES points to ./data/classes/coco.names file.
Therefore if you use domain coco.names there is no need to change.

According to my .names file: __C.YOLO.CLASSES = "./data/classes/licence_plate.names"

---

Summarize

So far we have:

• .weights on path ./data/yolov4_licence_plate.weights
• .names on path ./data/classes/licence_plate.names
• adjusted __C.YOLO.CLASSES param inside 'config.py' on path ./core/config.py

You have environment prepared to perform conversion.

---

5. Convert Darknet weights into Tensorflow.

a. Into regular Tensorflow .pb model

save_model.py does the job.
Required flags:

• --weights : path to weights ./data/yolov4_licence_plate.weights
• --output : where to save output ./checkpoints/license_plate-416
• --input_size : size of YOYLO input data 416 (px)
• --model : one of ['yolov3', yolov3] yolov4

python3.7 save_model.py --weights ./data/yolov4_licence_plate.weights --output ./checkpoints/license_plate-416 --input_size 416 --model yolov4

This creates new folder ./checkpoints/license_plate-416 that stores saved_model.pb - actual Tensorflow model.

b. Into Tensorflow Lite .tflite model

This option is lightweight. This solution trade off speed over accuracy.
Great for edge devices such as mobile phones, raspberry pi and others.

python3.7 save_model.py --weights ./data/yolov4_licence_plate.weights --output ./checkpoints/license_plate-416 --input_size 416 --model yolov4 --framework tflite
python3.7 convert_tflite.py --weights ./checkpoints/license_plate-416 --output ./checkpoints/yolov4_license_plate-416.tflite

Difference makes --framework tflite flag.

This creates new light weight Tensorflow object ./checkpoints/yolov4_license_plate-416.tflite - actual Tensorflow Lite model.

---

Model (one of [.pb, .tflite]) has been successfully converted.

Now ready to run.

---

a. Run regular Tensorflow .pb model

Detect image

python3.7 detect.py --weights ./checkpoints/license_plate-416 --size 416 --model yolov4 --images ./data/images/license_plate.jpg

NOTE

To run multiple image detection, change flag --images using following pattern --images './path/to/image1.jpg, ./path/to/image2.jpg, ./another/path/image.jpg'

Detect video

python3.7 detect_video.py --weights ./checkpoints/license_plate-416 --size 416 --model yolov4 --video ./data/video/road.mp4 --output ./detections/results.avi

NOTE

To run predictions from webcam set --video flag argument to 0 as follows --video 0.

b. Run lightweight Tensorflow Lite .tflite model

Detect image

python3.7 detect.py --weights ./checkpoints/yolov4_license_plate-416.tflite --size 416 --model yolov4 --images ./data/images/license_plate.jpg --framework tflite

Detect video

python3.7 detect_video.py --weights ./checkpoints/yolov4_license_plate-416.tflite --size 416 --model yolov4 --video ./data/video/road.mp4 --output ./detections/results.avi --framework tflite

---

Command Line Args Reference

save_model.py:
  --weights: path to weights file
    (default: './data/yolov4.weights')
  --output: path to output
    (default: './checkpoints/yolov4-416')
  --[no]tiny: yolov4 or yolov4-tiny
    (default: 'False')
  --input_size: define input size of export model
    (default: 416)
  --framework: what framework to use (tf, trt, tflite)
    (default: tf)
  --model: yolov3 or yolov4
    (default: yolov4)

detect.py:
  --images: path to input images as a string with images separated by ","
    (default: './data/images/kite.jpg')
  --output: path to output folder
    (default: './detections/')
  --[no]tiny: yolov4 or yolov4-tiny
    (default: 'False')
  --weights: path to weights file
    (default: './checkpoints/yolov4-416')
  --framework: what framework to use (tf, trt, tflite)
    (default: tf)
  --model: yolov3 or yolov4
    (default: yolov4)
  --size: resize images to
    (default: 416)
  --iou: iou threshold
    (default: 0.45)
  --score: confidence threshold
    (default: 0.25)

detect_video.py:
  --video: path to input video (use 0 for webcam)
    (default: './data/video/video.mp4')
  --output: path to output video (remember to set right codec for given format. e.g. XVID for .avi)
    (default: None)
  --output_format: codec used in VideoWriter when saving video to file
    (default: 'XVID)
  --[no]tiny: yolov4 or yolov4-tiny
    (default: 'false')
  --weights: path to weights file
    (default: './checkpoints/yolov4-416')
  --framework: what framework to use (tf, trt, tflite)
    (default: tf)
  --model: yolov3 or yolov4
    (default: yolov4)
  --size: resize images to
    (default: 416)
  --iou: iou threshold
    (default: 0.45)
  --score: confidence threshold
    (default: 0.25)

---

References:

• hunglc007