ax Inc. machine learning training platform.
It currently supports the training of object detection models YOLOX and YOLOv4.
Trainings can be done on custom datasets as well as already available datasets from Open Images Dataset.
ailia TRAINER also offers a feature of auto-annotation using Detic for automatically generating annotations.
All the features are described in great details in the blog posts mentionned in the following section.
We published a blog post below as an introduction to the concept of model training for users who are not familiar with the fundamental concepts.
The second blog post below is a manual on how to setup and how to use the training platform.
ailia TRAINER — Getting Started
ailia TRAINER is compatible with Linux and Windows platforms.
Environment variables section.
The following prerequisites are required to run ailia Trainer:
- Docker, also tested with Docker Desktop on Windows platform
- NVIDIA GPU drivers for training on GPU. Training will be performed on CPU otherwise, if supported depending on the model type.
git lfs pull
- Pull the container images
./docker-pull.sh- Start the containers
docker compose create
docker compose startThen open your browser and go to:
http://localhost:19998to access the ailia Trainer YOLOv4http://localhost:19999to access the ailia Trainer YOLOX
By using the command below you can watch all the logs coming from the containers
Several environment variables can be commented/uncommented from the docker-compose.yaml file to enforce some specific configuration changes.
FORCE_CPU: force the use of cuda/cpu deviceUSE_WSL2: If you are using docker on Windows wsl2, be sure the environment variableUSE_WSL2is set to1as a workaround to avoid network error while downloading dataset from internet.
The weights are included in the repository and stored using Git LFS, but for reference the sources can be found below.