ailia Tech BLOG
Released ailia SDK 1.2.9
We are pleased to introduce version 1.2.9 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.
Optimization for Arm NEON
The inference performance in Arm environments such as Android and Mac M1 has been greatly improved using inline NEON assembly and various optimizations based on Arm big.LITTLE architecture.
Unification of the Android inference backend to Vulkan
Google has announced the end of support for RenderScript from Android 12 and recommends to migrate to Vulkan. In line with this, we have ended support for RenderScript in the ailia SDK and migrated the inference backend to Vulkan.
Multiple input/output support for JNI
The support for multiple input/output has been added to JNI in order to run more complex models on Android platform.
Enhancements of memory saving mode in Python
We added a memory-saving mode to the Classifier and Detector APIs in Python, similar to the Predict API, where memory_mode can be given as an argument when creating an instance.
YOLOX support via the Detector API
The Detector API now supports YOLOX, a highly accurate object detection model, and existing apps can be converted to YOLOX by simply rewriting the arguments of the Detector API. Since the pre-processing and post-processing is now done in C++, it runs faster than doing the pre-processing and post-processing in Python. In addition, BGR data can now be given as argument instead of BGRA so that unnecessary color format conversions can be avoided.
#yolov4
detector = ailia.Detector( MODEL_PATH, WEIGHT_PATH, len(COCO_CATEGORY), format=ailia.NETWORK_IMAGE_FORMAT_RGB, channel=ailia.NETWORK_IMAGE_CHANNEL_FIRST, range=ailia.NETWORK_IMAGE_RANGE_U_FP32, algorithm=ailia.DETECTOR_ALGORITHM_YOLOV4, env_id=args.env_id)#yolox
detector = ailia.Detector( MODEL_PATH, WEIGHT_PATH, len(COCO_CATEGORY), format=ailia.NETWORK_IMAGE_FORMAT_BGR, channel=ailia.NETWORK_IMAGE_CHANNEL_FIRST, range=ailia.NETWORK_IMAGE_RANGE_U_INT8, algorithm=ailia.DETECTOR_ALGORITHM_YOLOX, env_id=args.env_id)
Python 3.10 support
collections.abc.Sequence alias is no longer supported in Python 3.10 causing issues with the run API. If you encounter the following errors with ailia SDK 1.2.8, please update to ailia SDK 1.2.9.
File "Python310\lib\site-packages\ailia\wrapper.py", line 407, in run
elif isinstance(input, collections.Sequence) :
AttributeError: module 'collections' has no attribute 'Sequence'
Support of new layers
The support of the SoftSign layer and opset=12 for some layers was added. We are planning to officially support opset=12 to 14 in ailia SDK 1.2.10. For example the model HitNet recently added to ailia MODELS uses opset=12.
Automatic management of evaluation license file
When installing the evaluation version of ailia SDK with pip, the license file is now automatically copied. The license file will also be automatically copied when you build your app in Unity.
Addition of new models
The following models will be newly supported in ailia SDK 1.2.9.
SiamMOT : Siamese Multi-Object Tracking
Source: https://vimeo.com/60139361
BackGroundMattingV2 : Background removal model
Source: https://github.com/PeterL1n/BackgroundMattingV2#video--image-examples
HitNet : Depth estimation model from stereo camera
Source: https://vision.middlebury.edu/stereo/data/scenes2003/
LapDepth : Monocular depth estimation model
Source: https://github.com/tjqansthd/LapDepth-release/blob/master/example/kitti_demo.jpg
RestyleEncoder : “Toonification” of face images
Source: https://github.com/yuval-alaluf/restyle-encoder/tree/main/notebooks/images
SAM : Age transformation for face images
Source: https://github.com/yuval-alaluf/SAM/blob/master/notebooks/images/866.jpg
ailia Inc. has developed ailia SDK, which enables cross-platform, GPU-based rapid inference.
ailia 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.