Releases: roboflow/inference
v0.9.0
Summary
This release includes:
- The new
inference-cli
to make starting the inference server easy and automated - A new
inference-client
to as a helpful utility when interacting with theinference
HTTP API - Updates and added features to the Device Manager (enterprise feature)
- Unified model APIs so that all Roboflow models adhere to a consistent processing pipeline
- Bug fixes, maintenance
Breaking Changes:
- Some model APIs have been updated (see instance segmentation and classification)
v0.8.9
Summary
This release includes a new env var DISABLE_INFERENCE_CACHE
. When set to true
, internal inference caching will be disabled. Also, logging has been updated to be less verbose by default. To increase verbosity, set LOG_LEVEL=DEBUG
.
v0.8.8
Summary
Contains a fix in imread/imdecode logic. Also moves logic out of version.py
to fix github actions.
v0.8.7
Summary
- Abandons Pillow in favor of OpenCV for faster end to end processing
- Fixes a bug with new device management logic
- Upgrades version checking logic
- Adds env var to fix Jetson 5.1.1 images
v0.8.6
Summary
This release includes logic to detect and log if there is a newer release available. It also contains a new enterprise device manager.
v0.8.5
Summary
Contains bug fixes for configurations that use the LICENSE_SERVER
setting.
v0.8.4
Summary:
- Image loading is now multi-threaded for batch requests. This should increase total FPS, especially for batch requests that include large images.
- The regression test Github action now runs on a Github actions runner.
- The extras require has been fixed for the various distribution packages
v0.8.2
Summary
Updated the Jetson images so that the default execution provider is CUDA. TensorRT is now an optional configuration via the environment variable ONNXRUNTIME_EXECUTION_PROVIDERS=TensorrtExecutionProvider
. The images are also renamed to:
- roboflow/roboflow-inference-server-jetson-4.5.0
- roboflow/roboflow-inference-server-jetson-4.6.1
- roboflow/roboflow-inference-server-jetson-5.1.1
v0.8.1
Summary
- Optional Byte Track for UDP interface
- Updated SAM and CLIP requirements and added README quickstarts
- Bug fix for single channel numpy strings
Breaking Changes
- The output for UDP JSON messages has updated the key
class_name
toclass
to match HTTP responses.
v0.8.0
Summary
This release includes a bit of an overhaul for the model APIs. As this repository started as an internal tool for hosting inference logic, the APIs were tailored for an HTTP interface. With this release, we have made using inference
within your python code much smoother and easier. We also updated imports to be less verbose. See the README and docs for new usage. Additionally, a new interface is provided for consuming video streams, and then broadcasting the results over UDP. This interface is tuned for low latency and is ideal for use cases that need to the most up to date information as possible from a video stream. See https://blog.roboflow.com/udp-inference/ for more details.
Breaking Changes
The main change was creating new definitions for model infer()
functions that now take many keyword arguments instead of a single request
argument. To continue inferring using request objects, a new method infer_from_request()
is provided.