Awesome

<a href="https://www.plasticorigins.eu/"><img width="80px" src="https://github.com/surfriderfoundationeurope/The-Plastic-Origins-Project/blob/master/assets/PlasticOrigins_logo.png" width="5%" height="5%" align="left" hspace="0" vspace="0"></a>

Read Latest Documentation - Browse GitHub Code Repository

Welcome to MOT, the current Plastic Origins model for garbage detection on river banks.

MOT stands for Multi-Object Tracking as we detect and then track the different plastic trash instances.

The object detection part is based on tensorpack.

The next subsections are useful to read if you want to train models or perform advanced tasks. However, if you just want to launch a serving container or perform inferences on one of those, directly jump to this file.

Getting Started

Prerequisites

Before you begin, ensure you have met the following requirements:

• You have a <Linux> machine.
• Preferably, you have a GPU on your machine.

Technical stack

• Language: Python , Tensorflow , Docker , Tensorpack
• Framework: Python 3.3+ ,1.6 <= tensorflow < 2.0

Dataset

You can download a training dataset on this link.

Installation

You may run directly the notebook in colab.

For more details on training and inference of the object detection please see the following file which is based on the README of tensorpack.

Classic

To install locally, make sure you have Python 3.3+ and 1.6 <= tensorflow < 2.0

apt install libsm6 libxrender-dev libxext6 libcap-dev ffmpeg
pip3 install --user .

Docker

The following command will build a docker for development and run interactively.

PORT_JUPYTER=22222 PORT_TENSORBOARD=22223 make docker-training

You don't have to specify the ports at the beginning of the command, but do so if you want to assign a specific port to access jupyter notebook and / or tensorboard.

You can add arguments to the docker run command by specifying RUN_ARGS, for example:

RUN_ARGS="-v /srv/data:/srv/data" make docker-training

Do the following command to exec an already running container:

make docker-exec-training

Usage

Internal tools

You can launch a jupyter notebook or a tensorboard server by running the command.

./scripts/run_jupyter.sh

or

./scripts/run_tensorboard.sh /path/to/the/model/folders/to/track

Then, access those servers through the ports you used in the Make command.

Train

See the original tensorpack README for more details about the configurations and weights.

python3 -m mot.object_detection.train --load /path/to/pretrained/weights --config DATA.BASEDIR=/path/to/the/dataset --config TODO=SEE_TENSORPACK_README

The next files are pretrained weights on the dataset introduced previously:

• https://files.heuritech.com/raw_files/surfrider/resnet50_fpn/model-6000.index
• https://files.heuritech.com/raw_files/surfrider/resnet50_fpn/model-6000.data-00000-of-00001

The command used to train this model was:

python3 -m mot.object_detection.train --load /path/to/pretrained_weights/COCO-MaskRCNN-R50FPN2x.npz --logdir /path/to/logdir --config DATA.BASEDIR=/path/to/dataset MODE_MASK=False TRAIN.LR_SCHEDULE=250,500,750

Put those files in a folder, which will be /path/to/your/trained/model in the export section.

Export

First, you need to train an object detection model. Then, you can export this model in SavedModel format:

python3 -m mot.object_detection.predict --load /path/to/your/trained/model --serving /path/to/serving --config DATA.BASEDIR=/path/to/the/dataset SAME_CONFIG=AS_TRAINING

The dataset should be the one downloaded following the instructions above. You can also use a folder with only this file inside if you don't want to download the whole dataset. Also remember to use the same config as the one used for training (using FPN.CASCADE=True for instance).

Serving

Refer to this file.

Test

Developper installation

You need to install the repository in dev:

pip install -e ./

The following libraries are needed to run the tests: pytest, pytest-cov

Use with pyenv

pyenv activate my_amazing_surfrider_project
pip install .

Run the tests

• Within your local environement:

• To run all the tests:

make tests

• To run a specific test:

pytest my_file.py::my_function

• Within a docker environement:

• To run all the tests:

make docker-tests

• To run a specific test:

make up-tests
pytest my_file.py::my_function

Contributing

It's great to have you here! We welcome any help and thank you in advance for your contributions.

• Feel free to report a problem/bug or propose an improvement by creating a new issue. Please document as much as possible the steps to reproduce your problem (even better with screenshots). If you think you discovered a security vulnerability, please contact directly our Maintainers.

• Take a look at the open issues labeled as help wanted, feel free to comment to share your ideas or submit a pull request if you feel that you can fix the issue yourself. Please document any relevant changes.

For more information please read the CONTRIBUTING.md file for developers.

Maintainers

If you experience any problems, please don't hesitate to ping:

• @charlesollion
• @mchagneux

Special thanks to all our Contributors.

License

We’re using the MIT License. For more details, check LICENSE file.

Additional information

Model & training:

• Object detection training

• Improving train, validation and test dataset

• Model improvements

• [ ] Connection with dataset to query dataset

• Tracking model (WIP)

• test dataset for tracking

Inference and deployment:

• Object detection inference notebook

• Inference on video (WIP)

• Connection with input data and inference

• Small webserver and API (in local)

• [ ] Docker build and deployment