October 29, 2020

854 words 5 mins read

pythonlessons/TensorFlow-2.x-YOLOv3

pythonlessons/TensorFlow-2.x-YOLOv3

YOLOv3 implementation in TensorFlow 2.3.1

repo name pythonlessons/TensorFlow-2.x-YOLOv3
repo link https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3
homepage https://pylessons.com/
language Jupyter Notebook
size (curr.) 72448 kB
stars (curr.) 174
created 2020-04-21
license MIT License

TensorFlow-2.x-YOLOv3 and YOLOv4 tutorials

YOLOv3 and YOLOv4 implementation in TensorFlow 2.x, with support for training, transfer training, object tracking mAP and so on… Code was tested with following specs:

  • i7-7700k CPU and Nvidia 1080TI GPU
  • OS Ubuntu 18.04
  • CUDA 10.1
  • cuDNN v7.6.5
  • TensorRT-6.0.1.5
  • Tensorflow-GPU 2.3.1
  • Code was tested on Ubuntu and Windows 10 (TensorRT not supported officially)

Installation

First, clone or download this GitHub repository. Install requirements and download pretrained weights:

pip install -r ./requirements.txt

# yolov3
wget -P model_data https://pjreddie.com/media/files/yolov3.weights

# yolov3-tiny
wget -P model_data https://pjreddie.com/media/files/yolov3-tiny.weights

# yolov4
wget -P model_data https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

# yolov4-tiny
wget -P model_data https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights

Quick start

Start with using pretrained weights to test predictions on both image and video:

python detection_demo.py

Quick training for custom mnist dataset

mnist folder contains mnist images, create training data:

python mnist/make_data.py

./yolov3/configs.py file is already configured for mnist training.

Now, you can train it and then evaluate your model

python train.py
tensorboard --logdir=log

Track training progress in Tensorboard and go to http://localhost:6006/:

Test detection with detect_mnist.py script:

python detect_mnist.py

Results:

Custom YOLOv3 & YOLOv4 object detection training

Custom training required to prepare dataset first, how to prepare dataset and train custom model you can read in following link: https://pylessons.com/YOLOv3-TF2-custrom-train/ More about YOLOv4 training you can read on this link. I didn’t have time to implement all YOLOv4 Bag-Of-Freebies to improve the training process… Maybe later I’ll find time to do that, but now I leave it as it is. I recommended to use Alex’s Darknet to train your custom model, if you need maximum performance, otherwise, you can use my implementation.

Google Colab Custom Yolo v3 training

To learn more about Google Colab Free gpu training, visit my text version tutorial

Yolo v3 Tiny train and detection

To get detailed instructions how to use Yolov3-Tiny, follow my text version tutorial YOLOv3-Tiny support. Short instructions:

  • Get YOLOv3-Tiny weights: wget -P model_data https://pjreddie.com/media/files/yolov3-tiny.weights
  • From yolov3/configs.py change TRAIN_YOLO_TINY from False to True
  • Run detection_demo.py script.

Yolo v3 Object tracking

To learn more about Object tracking with Deep SORT, visit Following link. Quick test:

  • Clone this repository;
  • Make sure object detection works for you;
  • Run object_tracking.py script

YOLOv3 vs YOLOv4 comparison on 1080TI:

YOLO FPS on COCO 2017 Dataset:

Detection 320x320 416x416 512x512
YoloV3 FPS 24.38 20.94 18.57
YoloV4 FPS 22.15 18.69 16.50

TensorRT FPS on COCO 2017 Dataset:

Detection 320x320 416x416 512x512 608x608
YoloV4 FP32 FPS 31.23 27.30 22.63 18.17
YoloV4 FP16 FPS 30.33 25.44 21.94 17.99
YoloV4 INT8 FPS 85.18 62.02 47.50 37.32
YoloV3 INT8 FPS 84.65 52.72 38.22 28.75

mAP on COCO 2017 Dataset:

Detection 320x320 416x416 512x512
YoloV3 mAP50 49.85 55.31 57.48
YoloV4 mAP50 48.58 56.92 61.71

TensorRT mAP on COCO 2017 Dataset:

Detection 320x320 416x416 512x512 608x608
YoloV4 FP32 mAP50 48.58 56.92 61.71 63.92
YoloV4 FP16 mAP50 48.57 56.92 61.69 63.92
YoloV4 INT8 mAP50 40.61 48.36 52.84 54.53
YoloV3 INT8 mAP50 44.19 48.64 50.10 50.69

Converting YOLO to TensorRT

I will give two examples, both will be for YOLOv4 model,quantize_mode=INT8 and model input size will be 608. Detailed tutorial is on this link.

Default weights from COCO dataset:

  • Download weights from links above;
  • In configs.py script choose your YOLO_TYPE;
  • In configs.py script set YOLO_INPUT_SIZE = 608;
  • In configs.py script set YOLO_FRAMEWORK = "trt";
  • From main directory in terminal type python tools/Convert_to_pb.py;
  • From main directory in terminal type python tools/Convert_to_TRT.py;
  • In configs.py script set YOLO_CUSTOM_WEIGHTS = f'checkpoints/{YOLO_TYPE}-trt-{YOLO_TRT_QUANTIZE_MODE}–{YOLO_INPUT_SIZE}';
  • Now you can run detection_demo.py, best to test with detect_video function.

Custom trained YOLO weights:

  • Download weights from links above;
  • In configs.py script choose your YOLO_TYPE;
  • In configs.py script set YOLO_INPUT_SIZE = 608;
  • Train custom YOLO model with instructions above;
  • In configs.py script set YOLO_CUSTOM_WEIGHTS = f"{YOLO_TYPE}_custom";
  • In configs.py script make sure that TRAIN_CLASSES is with your custom classes text file;
  • From main directory in terminal type python tools/Convert_to_pb.py;
  • From main directory in terminal type python tools/Convert_to_TRT.py;
  • In configs.py script set YOLO_FRAMEWORK = "trt";
  • In configs.py script set YOLO_CUSTOM_WEIGHTS = f'checkpoints/{YOLO_TYPE}-trt-{YOLO_TRT_QUANTIZE_MODE}–{YOLO_INPUT_SIZE}';
  • Now you can run detection_custom.py, to test custom trained and converted TensorRT model.

What is done:

To be continued… (not anytime soon)

  • Converting to TensorFlow Lite
  • YOLO on Android (Leaving it for future, will need to convert everythin to java… not ready for this)
  • Generating anchors
  • YOLACT: Real-time Instance Segmentation
  • Model pruning (Pruning is a technique in deep learning that aids in the development of smaller and more efficient neural networks. It’s a model optimization technique that involves eliminating unnecessary values in the weight tensor.)
comments powered by Disqus