Documentation

Installation for User

Open anaconda powershell, activate an environment with anaconda, navigate into the trainer repo and execute the following to install trainer using pip, including its dependencies:

pip install ai-trainer

For Online Learning you have to install PyTorch:

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

AI-Trainer helps with building a data generator and it relies on imgaug for it:

conda install imgaug -c conda-forge

Getting started with training models

Trainer currently supports annotating images and videos. First, create a dataset using

trainer init-ds
cd YOUR_DATASET

Getting started with using trainer in python

For using the annotated data, you can use trainer as a python package. After activating the environment containing the trainer and its dependencies, feel free to inspect some of the tutorials in ./tutorials/.

Development Setup

Execute the user installation, but instead of using pip install ai-trainer, clone the repo locally.

git clone https://github.com/Telcrome/ai-trainer

Both vsc and pycharm are used for development with their configurations provided in .vscode and .idea

Recommended environments

For development we recommend to install the conda environment into a subfolder of the repo. This allows for easier experimentation and the IDE expects it this way.

conda env create --prefix ./envs -f environment.yml
conda activate .\envs\.

Now install a deep learning backend. PyTorch provides well-working conda install commands.

For Tensorflow with GPU:

conda install cudatoolkit=10.0 cudnn=7.6.0=cuda10.0_0
pip install tensorflow-gpu

Testing Development for pip and cli tools

Installing the folder directly using pip does not work due to the large amount of files inside the local development folder, especially because in the local development setup the environment is expected to be a subfolder of the repo.

pip install -e .

Uploading to PyPi by hand

python setup.py sdist bdist_wheel
twine upload dist/* # The asterisk is important

Using Docker

Docker and the provided DOCKERFILE support is currently experimental as it proved to slow down the annotation GUI too much. When the transition to a web GUI is completed docker will be supported again.

Contribution

Docs

Currently, Read the Docs is used for CI of the docs. Before submitting changes, test the make command in the environment:

conda env create -f environment.yml
conda activate trainer_env
make html

If this throws warnings or errors, Read the Docs won`t publish them.

Tutorials inside the repo

• Do not use jupyter notebooks
• Should be testable without preparing data by hand where possible.