Schaaaafrichter

The successor of the original Schaafrichter. Now with Advanced AI-Technolgies!

Installation

You can build the project with or without support for a CUDA capable GPU. Steps only applicable to one of them will be denoted (GPU Support/CPU). Also you can either install this in your own system, or use a docker image instead.

Install on System

1. Make sure to install Python 3 on your device
   • Windows: Get it here
   • Mac: Get it here or use your favourite package manager
   • Linux: Use your favourite package manager i.e. pacman -S python or apt install python3
2. GPU Support:
   • install CUDA
   • install cudnn
3. Create a virtualenvironment
   • you can do so with python3 -m venv --system-site-packages <path to virtualenv>
   • If you are using Linux, we recommend that you install virtualenvwrapper and organize all virtual environments with this tool, its quite neat.
   • INFO: Make sure to include the global site-packages that contain Opencv (make sure to use --system-site-packages)!
4. Load the virtual environment
5. Clone the repository
6. For Ubuntu: Install header files for alsa: apt install libasound2-dev
7. Install all necessary libraries:
   • GPU Support: pip install -r requirements.txt
   • CPU: pip install -r requirements_cpu.txt

Install with Docker

1. Install Docker
   • Windows: Get it here
   • Mac: Get it here
   • Linux: User your favourite package manager i.e. pacman -S docker, or use this guide for Ubuntu.
2. GPU Support: In case your device has a CUDA capable GPU, you should do the following:
   • install CUDA
   • install cudnn
   • install nvidia-docker (Ubuntu, Arch Like OS)
3. Build the Docker Image:
   • CPU:

   docker build -t sheep --build-arg FROM_IMAGE=ubuntu:16.04 --build-arg CPU_ONLY=true .
   

   • GPU Support:

   docker build -t sheep .
   

   If your host system uses CUDA with a version earlier than 9.1, specify the corresponding docker image to match the configuration of your machine (see this list for available options). For example, for CUDA 8 and CUDNN 6 use the following instead:

   docker build -t sheep --build-arg FROM_IMAGE=nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04 .
   

Training

If you want to train the Schaaafrichter you will first need a dataset. You can either create or download a dataset for training the Schaaafrichter.

Once you have all necessary data, you can start the training. The training scipt is train.py. We will now quickly go over all possible command-line arguments you can/have to use:

• dataset path to the json file containing your training dataset
• test_dataset path to the json file containing the validation dataset
• --dataset-root specify a dataset root dir that might be different from the directory of the dataset file locations, which is used as default.
• --model available choices are ssd300 and ssd512, you can choose which kind of model you want to train. Default is ssd512.
• --batchsize the batch size to use for training. Default is 32.
• --gpu which gpu to use (e.g. 0 means your first gpu). You can also give more than one gpu id. The model will then be trained in data parallel fashion. Default is -1, which means run on CPU.
• --out specifies the output directory for the trained model and log. Default is result
• --resume specify a trainer_snapshot and continue training.
• --lr set the default learning rate for the optimizer. Default is 1e-3.

Once you started the training, grab a coffee/tea and enjoy the rest of your day.

Inference

Once you got a trained model, you can do inference and have fun!

Using Docker

Execute this on your host, to allow docker to connect to your X server (needs to be done after every system restart):

xhost +local:docker

Run the container and get a command-line (replace nvidia-docker with docker if using only CPU):

nvidia-docker run \
    --rm \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    -e DISPLAY=$DISPLAY \
    --device /dev/video0:/dev/video0 \
    --device /dev/snd \
    -it \
    --volume /absolute/path/to/repository:/app \
    sheep

Note: The --volume option overwrites the content in the docker image and should be used for developing instead of rebuilding the image when changing code. In this option /absolute/path/to/repository should be the absolute path to the root directory of the repository. You can also use: --volume "$( readlink . -f )":/app, which inserts the absolute path to the current directory, but that does not work on Windows.

Known errors

If you receive the following error, you need to execute xhost +local:docker before executing the docker run command (see comment below this answer):

No protocol specified
Failed to connect to Mir: Failed to connect to server socket: No such file or directory
Unable to init server: Could not connect: Connection refused

(sheeper:1): Gtk-WARNING **: cannot open display: :1

If you get the following errors add --env QT_X11_NO_MITSHM=1 to your docker run command (source):

X Error: BadAccess (attempt to access private resource denied) 10
  Extension:    130 (MIT-SHM)
  Minor opcode: 1 (X_ShmAttach)
  Resource id:  0x4200003
X Error: BadShmSeg (invalid shared segment parameter) 128
  Extension:    130 (MIT-SHM)
  Minor opcode: 3 (X_ShmPutImage)
  Resource id:  0x420000a

Running the script

Run something like:

python3 live_sheeping.py data/models/trained_model data/models/log

You can also run on a gpu with --gpu <gpu_id>.

To generate predictions for static images instead (you can add --gpu <gpu_id> again, and --help for other options):

python image_sheeping.py data/models/trained_model data/models/log -j data/generated/test_info.json