knorth55 / 73b2_kitchen_edgetpu_object_detection

This repo is merged to knorth55/coral_usb_ros

Coral TPU training scripts for 73B2 Kitchen dataset

Preparation

Build Docker image

git clone https://github.com/knorth55/73b2_kitchen_edgetpu_object_detection.git
cd 73b2_kitchen_edgetpu_object_detection
mkdir learn/
docker build docker/ --tag 73b2-kitchen-object-detection

Download and prepare 73b2 kitchen dataset

bash run.bash

Inside docker, run commands below

cd /tensorflow/models/research/73b2_kitchen_learn
gdown https://drive.google.com/uc?id=1iBSxX7I0nFDJfYNpFEb1caSQ0nl4EVUa
tar zxvf kitchen_dataset.tgz

Download and prepare your own dataset

First you need VOC format annotation. For the annotation, please read here.

Inside docker, plear download and rename, setup your dataset directory like below.

73b2_kitchen_edgetpu_object_detection/learn/kitchen_dataset  # or 73b2_kitchen_learn in docker container
|-- train  # train dataset
|   |-- JPEGImages
|   |-- SegmentationClass
|   |-- SegmentationClassPNG
|   |-- SegmentationClassVisualization
|   |-- SegmentationObject
|   |-- SegmentationObjectPNG
|   |-- SegmentationObjectVisualization
|   `-- class_names.txt
`-- test   # test dataset
    |-- JPEGImages
    |-- SegmentationClass
    |-- SegmentationClassPNG
    |-- SegmentationClassVisualization
    |-- SegmentationObject
    |-- SegmentationObjectPNG
    |-- SegmentationObjectVisualization
    `-- class_names.txt

Training

Fine tuning in docker container

bash run.bash

Inside docker

cd /tensorflow/models/research/73b2_kitchen_scripts
# prepare dataset
./prepare_checkpoint_and_dataset.sh --train_whole_model false --network_type mobilenet_v2_ssd
# retraining on GPU 0
CUDA_VISIBLE_DEVICES=0 ./retrain_detection_model.sh --num_training_steps 500 --num_eval_steps 100 
# change to edgetpu model
./convert_checkpoint_to_edgetpu_tflite.sh --checkpoint_num 500

Whole retraining in docker container

bash run.bash

Inside docker

cd /tensorflow/models/research/73b2_kitchen_scripts
# prepare dataset
./prepare_checkpoint_and_dataset.sh --train_whole_model true --network_type mobilenet_v2_ssd
# retraining on GPU 0
CUDA_VISIBLE_DEVICES=0 ./retrain_detection_model.sh --num_training_steps 50000 --num_eval_steps 2000 
# change to edgetpu model
./convert_checkpoint_to_edgetpu_tflite.sh --checkpoint_num 50000

Run Tensorboard for visualization (Optional)

docker exec -it 73b2-kitchen-edgetpu /bin/bash

Inside docker

tensorboard --logdir=./73b2_kitchen_learn/train

You can see Tensorboard in localhost:6006.

Compile

Compile the model to Edge TPU

Compile and get output_tflite_graph_edgetpu.tflite model file!

bash run.bash

Inside docker

cd /tensorflow/models/research/73b2_kitchen_learn/models
edgetpu_compiler output_tflite_graph.tflite