Don't be a Turtle Project

Author: Jaewook Kang

About

Project objectives

This Don’t be a Turtle Project makes all of IT people have right posture and feel good while they are working! We investigate a mobile machine learning based methodology providing feedbacks with respect to your neck posture. For this purpose, we monitor neck, detecting whether you are maintaining good working posture. If you are working in an overhanging posture, you will be alerted to maintain a good posture.

Note that the above model in .gif was trained only by 865 custom dataset.

Release Benchmarks

• Pose Estimation Accuracy (PCKh): TBU

Version	Framework	Device	size (KB)
0.5.0	Android Pie + Tflite	Google Pixel2	749 KB
0.5.0	iOS 11.4.1 + CoreML	iPhoneX	811 KB

Repository Components

.
├── images              # some images for documentation
├── dataset/coco_form   # Unzip the dontbeturtle dataset at ./dataset/coco_form
├── note                # Some notes under Google Camps
├── sh_scripts          # A collection of shell scripts for easy operations
├── release             # dontbe turtle tflite and mlmodel here
└── tfmodules           # A collection of TF python files

Mobile Apps

• Android repo
• iOS repo

Frameworks

Technical Stacks

• Tensorflow (+ Tf Slim) >= 1.9
• Tf plot == 0.2.0.dev0
• opencv-python >= 3.4.2
• pycocotools == 2.0.0
• Cython == 0.28.4
• tensorpack == 0.8.0
• tfcoreml == 0.2.0

Repository Installation

git clone https://github.com/motlabs/dont-be-turtle
# cd dont-be-turtle/
git init
git submodule init
git submodule update

pip install -r requirement.txt
./sh_scripts/install_tensorflow_gpu.sh

How to Run Training

export MODEL_BUCKET=./tfmodules/export/model/       # set path for exporting ckpt and tfsummary
export DATA_BUCKET=./dataset/coco_form/dontbeturtle # set path for placing dataset
export SOURCE=./tfmodules/trainer_gpu.py            # set path for tensorflow trainer

python ${SOURCE}\
  --data_dir=${DATA_BUCKET}\
  --model_dir=${MODEL_BUCKET}\
  --is_ckpt_init=False\
  --ckptinit_dir=None

• You have an option to use ./sh_scripts/run_train_gpu.sh with some customization

How to Get .tflite and .mlmodel

Note that you need to configure ./tfmodule/model/model_config_released.py before executing the below command.

python gen_tflite_coreml.py  --is-summary=False --import-ckpt-dir=<ckpt path directory>
# Args:
#  1) is-summary==True : collect tf summary for model graph
#     is-summary==False: None
#  2) --import-ckpt-dir: global path directory .ckpt stored
#
# An example:
# python gen_tflite_coreml.py  --is-summary=False --import-ckpt-dir=/Users/jwkangmacpro2/SourceCodes/dont-be-turtle/tfmodules/export/model/run-20180815075050/
#

Donbeturtle Dataset v1.0

You need to create ./dataset/coco_form/ and place the data set

• Donbeturtle dataset v1.0 (trainset only, 865 images) download

• Keypoint annotator repos

  • For iOS Mobile
  • For OSX

Baseline Papers

• MobileNet v2
• Stacked Hourglass

Related Materials

• Jaewook Kang, " Don't be turtle project beyond Google Camp," GDGDevFest 2018 Pangyo, 2018 Nov
• Jaewook Kang, "From NIN to Inception V3," Modulabs Machine Learning of Things (MoT) Lab 2018 Mar
• Jaewwok Kang, "Machine Learning on Your Hands: Introduction to Tensorflow Lite Preview," Tensorflow dev Exteneded X Modulabs, 2018 Apr
• Jaewook Kang, "Mobile Vision Learning: Model Compression and Efficient Convolution perspective," ETRI, 2018 June 12th
• Jaewook Kang, "Let's use Cloud TPU", Aug 2018

Project Contributors

• Dontbeturtle v0.5
  • Jaewook Kang (PI)
  • Doyoung Gwak
  • Jeongah Shin
  • YongGeunLee
  • Joonho Lee
  • DongSeok Yang

Acknowledgement

• This project was supported by Google Deep Learning Camp Jeju 2018.