Tensorflow Object Detection Api Training In Docker

The purpose of these images is to allow you to quickly train an object detection model by transfer learning onto a model from the Tensorflow Object Detection Model Zoo (https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md). Big thanks to people who publish models like these.

How to use these images

It's assumed that you are starting with unlabeled images and will want to use labelImage by tzutalin.

Clone the repo to get the workspace scaffolded (git clone https://github.com/MadLadLabs/TensorflowObjectDetectionApiTrainingInDocker.git) and cd into the workspace directory in the project
Put your unlabeled images in the workspace/images/raw directory
Run labelImg (if you have it locally or use the snipped below to use it from a container) and label your images.

xhost +

docker run -ti --rm \
    -e DISPLAY=$DISPLAY \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    -v $(pwd)/images/raw:/images/ \
    spiridonovpolytechnic/labelimg:latest

Move your images from the workspace/images/raw directory to the workspace/images/train and workspace/images/test directories. Typically, you might have 85% training data and 15% test data or thereabouts.
Update the workspace/config.yml file. The labels array should match the names of the objects you have labeled in the images. You can swap out the pretrained model for another from the object detection model zoo.
Run the image with the one of the training tags (latest or latest-gpu if you would like to use GPU to train)

docker run \
    --volume $(pwd):/tf-workspace \
    --gpus all -it \
    spiridonovpolytechnic/tensorflow-object-detection-training:latest-gpu

Once the training finishes, you can run another image to test inferrence using your webcam with one of the images tagged with latest-infer-webcam or latest-infer-webcam-gpu. ALternative, you can use latest-infer-file-to-screen or latest-infer-file-to-screen-gpu tags to test inferrence against an input video file.

docker run -e DISPLAY=$DISPLAY \
    --device=/dev/video0:/dev/video0 \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    --volume $(pwd):/tf-workspace \
    --gpus all \
    spiridonovpolytechnic/tensorflow-object-detection-training:latest-infer-webcam-gpu

(when using image with inferrence on file input, set the input property in config.yml)

docker run -e DISPLAY=$DISPLAY \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    --volume $(pwd):/tf-workspace \
    --volume /path/to/directory_with_input_video:/input/ \
    --gpus all \
    spiridonovpolytechnic/tensorflow-object-detection-training:latest-infer-file-to-screen-gpu

Build and run container locally for testing via

docker build -t tfod_test -f Dockerfile.gpu .
docker run --volume $(pwd):/tf-workspace --gpus all -it tfod_test bash

Object detection tested within the container using

cd ~/tf/models/research/
python object_detection/builders/model_builder_tf2_test.py

Credit

The containers are originally based on code from a tutorial project by nicknochnack (https://github.com/nicknochnack/RealTimeObjectDetection). I hit a couple of snags when trying to set up my environment, got frustrated while trying to get all the right dependencies and decided to containerize the process.

abrahamperz/TensorflowObjectDetectionApiTrainingInDocker

Tensorflow Object Detection Api Training In Docker

How to use these images

Credit