Object Detection Using Tensorflow on the Raspberry Pi

Script for object detection from training new model on dataset to exporting quantized graph

Step 1. Setup

Using docker registry

This is the fastest way to use the repo

# For cpu
docker pull docker.nanonets.com/pi_training
# For gpu
docker pull docker.nanonets.com/pi_training:gpu

OR

Building locally

Docker build script

Should run this script from repository root

# For cpu
docker build -t pi_training -f docker/Dockerfile.training .
docker image tag pi_training docker.nanonets.com/pi_training

# For gpu
docker build -t pi_training:gpu -f docker/Dockerfile.training.gpu .
docker image tag pi_training:gpu docker.nanonets.com/pi_training:gpu

---

Step 2. Preparing dataset

Dataset for object detection consists of images of objects you want to detect and annotations which are xml files with coordinates of objects inside images in Pascal VOC format. If you have collected images, you can use tool like LabelImg to create dataset.

Copy dataset with images folder containing all training images and annotations folder containing all respective annotations inside data folder in repo which will be mounted by docker as volume

Step 3. Starting training

Tensorboard will be started at port 8000 and run in background You can specify -h parameter to get help for docker script

If you have a GPU instance, you need to install nvidia-docker

# For cpu
sudo docker run -p 8000:8000 -v `pwd`/data:/data docker.nanonets.com/pi_training -m train -a ssd_mobilenet_v1_coco -e ssd_mobilenet_v1_coco_0 -p '{"batch_size":8,"learning_rate":0.003}'
# For gpu
sudo nvidia-docker run -p 8000:8000 -v `pwd`/data:/data docker.nanonets.com/pi_training:gpu -m train -a ssd_mobilenet_v1_coco -e ssd_mobilenet_v1_coco_0 -p '{"batch_size":8,"learning_rate":0.003}'

Usage

The docker instance on startup runs a script run.sh which takes the following parameters:

run.sh [-m mode] [-a architecture] [-h help] [-e experiment_id] [-c checkpoint] [-p hyperparameters]

-h          display this help and exit
-m          mode: should be either `train` or `export`
-p          key value pairs of hyperparameters as json string
-e          experiment id. Used as path inside data folder to run current experiment
-c          applicable when mode is export, used to specify checkpoint to use for export

List of Models (that can be passed to -a):

1. ssd_mobilenet_v1_coco
2. ssd_inception_v2_coco
3. faster_rcnn_inception_v2_coco
4. faster_rcnn_resnet50_coco
5. rfcn_resnet101_coco
6. faster_rcnn_resnet101_coco
7. faster_rcnn_inception_resnet_v2_atrous_coco
8. faster_rcnn_nas

Possible hyperparameters to override from -p command in json

Name	Type
learning_rate	float
batch_size	int
train_steps	int
eval_steps	int

---

Step 4. Exporting trained model

This command would export trained model in quantized graph that can be used for prediction. You need to specify one of the trained checkpoints from experiment directory that you want to use for prediction with -c command as follows:

# For cpu
sudo docker run -v `pwd`/data:/data docker.nanonets.com/pi_training -m export -a ssd_mobilenet_v1_coco -e ssd_mobilenet_v1_coco_0 -c /data/0/model.ckpt-8998

# For gpu
sudo nvidia-docker run -v `pwd`/data:/data docker.nanonets.com/pi_training:gpu -m export -a ssd_mobilenet_v1_coco -e ssd_mobilenet_v1_coco_0 -c /data/0/model.ckpt-8998

Once your done training the model and have exported it you can move this onto a client device like the Raspberry Pi. For details of how to use on the Raspberry Pi click see https://github.com/NanoNets/TF-OD-Pi-Test