ailia Tech BLOG
Released ailia SDK 1.2.8
We are pleased to introduce version 1.2.8 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.
Addition of new layers
This release adds support for opset=11, as well as operatorsConcatFromSequence, SequenceEmpty, SequenceConstruct, SequenceInsert, SequenceErase, SequenceAt and SplitToSequence which are sometimes used in Pytorch. HardSigmoid, If, MaxUnpool, and OneHot are also newly supported.
Support memory saving mode for Android (JNI)
ailia’s memory saving mode can now be used on Android (JNI) platform. You can specify one of the memory modes available obtained via AiliaModel.getMemoryMode to the AiliaModel instance as shown below.
EnumSet<AiliaMemoryMode> memory_mode = AiliaModel.getMemoryMode(true, true, false, true);
ailia = new AiliaModel(envId, Ailia.MULTITHREAD_AUTO, memory_mode,
loadRawFile(R.raw.u2netp_opset11_proto), loadRawFile(R.raw.u2netp_opset11_weight));
For more information about memory saving mode, please refer to the tutorial below.
Propagation of tensor types
All the inference processing was done using floats up to ailia SDK 1.2.7. Starting from ailia SDK 1.2.8, each tensor type in ONNX is now propagated, so it is possible to infer models that perform operations with integer precision and dynamically calculate tensor sizes. It mainly improves the compatibility of models exported from TensorFlow.
Introducing ailia.audio
As an experimental feature, we added ailia.audio, a library for audio pre and post-processing.
Speech recognition and generation requires pre-processing of the input waveform of the AI model, such as MelSpectrum transformation and resampling. torch.audio and librosa are usually used for these pre-processing tasks. However, these speech processing libraries do not work on edge devices such as Jetson and RaspberryPi, therefore models for speech recognition and generation do not work.
To address this issue, ailia.audio provides the equivalent of torch.audio and librosa for all platforms, enabling consistent audio pre and post-processing.
To use ailia.audio, use the --ailia_audio option for speech recognition models available in ailia MODELS.
python3 pytorch-dc-tts.py --ailia_audio
New supported models
Thanks to the addition of new supported layers and the implementation of type propagation, the following models are supported.
M-LSD : Wireframe Extraction Model
ailia-models/line_segment_detection/mlsd at master · ailia-ai/ailia-models(Image from…github.com
3D Object Detection Pytorch
Source: https://github.com/google-research-datasets/Objectron/blob/master/notebooks/Download%20Data.ipynb
MoveNet : Pose estimation for videos with intense motion
ailia-models/pose_estimation/movenet at master · ailia-ai/ailia-models(Image from…github.com
MMFashion VirtualTryOn : Virtual try-on of fashion items
Pytorch ENet : Image Segmentation model
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.