Released ailia SDK 1.2.12
We are pleased to introduce version 1.2.12 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.
Optimisation of specific activation functions
We developed optimized implementations for activation functions such as SiLU used for YOLOX and YOLOv5, or Mish used in YOLOv4 for cuDNN and CPU (SIMD).
In order to reduce memory transfers between CPU and GPU, the cuDNN implementation was extended to support Resize (Nearest) and Transpose in CUDA.
These optimizations result in 33% speedup for YOLOX tiny, 18% for YOLOX, 25% for YOLOv5, and 15% for YOLOv4 on Jetson NX using cuDNN. On Intel CPUs using SIMD, YOLOX tiny is 20% faster, 19% for YOLOX, 23% for YOLOv5, and 18% for YOLOv4.
Faster model loading by sorting graphs in execution order
We have added the ability to sort nodes in execution order when loading graphs to speed up model loading. This speeds up graph exploration and makes loading of huge models such as Detic about 30% faster.
Support for new operators
The support for operators Compress and Det was added, along with the possibility to have input/output of more than 5 dimensions for the CumSum layer.
Support for ONNX with FP16 weights
ONNX models using FP16 weights can now be loaded in ailia SDK 1.2.12.
Introduction of new models
- SberSwap: Real-time face replacement model
- SwinIR: Transformer-based super-resolution model
- DabDetr: Transformer-based object detection model
Introduction of new samples
Samples of ailia.audio were added for Unity and C++ as experimental feature which can be used to classify audio.
Evaluation version of ailia SDK
ailia SDK 1.2.12 evaluation version can be downloaded at the link below.
ax Inc. has developed ailia SDK, which enables cross-platform, GPU-based rapid inference.
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.
