Released ailia SDK 1.2.13
We are pleased to introduce version 1.2.13 of ailia SDK, a cross-platform framework to perform fast AI inference on GPU or CPU. You can find more information about ailia SDK on the official website.
Acceleration through enhanced TensorCore support on NVIDIA platforms
Enhanced TensorCore support on the NVIDIA platform provides significant speedups in the FP16 environment; Jetson NX can infer yolox 1.54 times faster than ailia SDK 1.2.12.
In ailia-models, available backends (execution environments) can be enumerated with the env_list option.
$ python3 yolox.py --env_list
By default the FP32 backend is used so please manually select the FP16 backend from the enumerated list; for Jetson, backend number 2 corresponds to FP16.
$ python3 yolox.py -e 2
Support for Jetson Orin
ailia 1.2.13 support Jetpack 5.0.2 and cuDNN 8.3 and are now compatible with Jetson Orin, enabling fast inference using TensorCore.
Metal and Vulkan support for optimized specific activation functions
The composite activations added for CPU and CUDA in ailia SDK 1.2.12 have been added for Metal and Vulkan. In addition, we have extended the GPU execution range of ResizeNearest and Concat. macOS (M1 Max) can now infer yolox 1.48 times faster than ailia SDK 1.2.12.
Reduction of memory copying
Improved performance of the ailiaSetBlobData API for models with multiple inputs and outputs; Python essentially uses the ailiaSetBlobData API, which improves overall performance when used from Python.
ONNX opset 16 support
Bernoulli, CastLike, GreaterOrEqual, GridSample, LessOrEqual are implemented, and ONNX opset 16 is supported.
Acceleration of DiffusionModel by speeding up Einsum
Accelerated Einsum for faster inference in DiffusionModel. ailia-models DiffusionModel are now faster using GPU.
End of support for C++ AMP
Microsoft has deprecated C++ AMP thus we have decided to discontinue the support of C++ AMP. Vulkan Backend can be used instead.
Enhanced Unity documentation
The documentation for each API of Unity has been enhanced. You can check it from the following URL.
Introduction of new models
- Crestereo, MobileStereoNet : Depth Estimation from Stereo Images
- Glip : Object detection from arbitrary text
- Dehamer : Fog removal model
- Detic : GridSampler integrated version of object detection model
The opset16 version is available with the following command: GridSample can be run inside the ailia SDK, thus avoiding the problem of possible conflicts with Pytorch’s cuDNN version.
$ python3 detic.py --opset16
Evaluation version of ailia SDK
ailia SDK 1.2.13 evaluation version can be downloaded at the link below.
ailia SDK is a self-contained cross-platform high speed inference SDK for AI developed by ax Inc.
ax Inc. provides a wide range of services from consulting and model creation, to the development of AI-based applications and SDKs. Feel free to contact us for any inquiry.