First release of CoralAPI, an HTTP inference service for Google Coral Edge TPUs. POST an image and a model name; get JSON back.
Tasks
- Classification
- Object detection (SSD and YOLO)
- Semantic segmentation
- Pose estimation
- Image embeddings
- Raw output tensors for any other model
Runtime and deploy
- Current Python, with a matched libedgetpu and tflite that work together
- Build profiles for AVX2 x86, no-AVX2 x86, and Raspberry Pi (armhf)
- Per-request model selection with on-demand download and label auto-fetch
- Async jobs, optional Prometheus metrics, device rebind via
POST /v1/reset - One instance per TPU for multi-TPU hosts
No CPU fallback by design; a running instance always means real acceleration.