Skip to content

System requirements

This page covers the checkmaite Python library. For the batch container, see Batch container.

Software

checkmaite requires Python 3.10, 3.11, 3.12, or 3.13 and PyTorch 2.2.0 or later. The installer pulls in PyTorch and every other dependency. See Setup.

Hardware

Resource Minimum Recommended
CPU 2 cores 4 or more cores
Memory 8 GB 16 GB or more
Disk 10 GB free 25 GB or more free
GPU None NVIDIA GPU with 4 GB or more memory
  • Memory grows with the dataset. DataEval capabilities keep per-image results for the whole dataset in memory, so size memory to your largest dataset. Each tutorial in this documentation peaks below 3 GB, and the end-to-end object detection workflow completes on 2 CPU cores with 8 GB.
  • Disk holds the Python environment, downloaded model weights, and your data, caches, and results. On Linux the environment is about 6 GB, because PyPI's PyTorch includes the CUDA libraries; on macOS it is about 2 GB. The models used in the tutorials add about 250 MB.
  • A GPU is optional. Every capability runs on the CPU. When no device is given, checkmaite uses CUDA if it is available, then Apple's MPS, then the CPU. The CUDA versions and GPUs supported depend on the PyTorch build you install.

Supported architectures

Verified on Linux x86_64 and aarch64, and on macOS arm64 (Apple silicon). macOS arm64 requires brew install libomp; see Setup. Other platforms are not tested.

Internet access

  • Installation needs PyPI (or a mirror), or conda-forge for conda installs. checkmaite-plugins installs from gitlab.jatic.net.
  • First use of a pretrained model downloads its weights once: torchvision weights from download.pytorch.org into TORCH_HOME (default ~/.cache/torch), and the VisDrone model from data.kitware.com into ~/.cache/modelmaite. To work offline, copy these caches from a connected machine.
  • Optional features use the network only when you configure them: datasets or an analytics store on S3 or Google Cloud Storage, and Ray clusters or Ray Serve endpoints given by address.

Everything else runs offline. A local Ray runtime also sends anonymous usage statistics unless you set RAY_USAGE_STATS_ENABLED=0.