ailia Tech BLOG
Released ailia SDK 1.2.7
We are pleased to introduce version 1.2.7 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.
ailia SDK 1.2.7 is a release that focuses on speeding up the inference process.
Performance increase on CPU
ConvTransposeND has benefited a substantial speed up. In particular, the performance of the audio processing system has greatly improved. The CPU performance on Mac M1 has also increased overall.
Performance increase on GPU
In Eltwise, a dedicated process for Tensor Broadcast has been added. The speedup is even more significant with EfficientNet. Also, GPU support for Padding has been enhanced. In Vulkan, MatMul and Gemm kernels have also gained in performance.
Performance increase of Yolov4
The Detector API now supports YOLOv4. In the past, running YOLOv4 required pre-processing (scale conversion, channel order swap) and post-processing (NMS) in Python, and with the low CPU performance of Jetson for example, the load of pre-processing and post-processing was higher than the inference time on the GPU.
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = np.transpose(img, [2, 0, 1])
img = img.astype(np.float32) / 255
img = np.expand_dims(img, 0)
（Example of heavy preprocessing）
The Detector API now performs these pre-processing and post-processing on the C++ side of the ailia SDK, which greatly improves performance. With this new flow, the above processing can be calculated in a single operation without going through buffers.
Addition of new layers
EyeLike、RandomNormal、RandomNormalLike、RandomUniform、RandomUniformLikeReduceLogSum、ReduceLogSumExp、Size are now supported. Also, MatMul now supports 5D input.
Improvements of the compression tools
A new min_tensor_size option has been added to give the possibility to exclude tensors with a small number of elements from quantization.
Addition of new supported models
3d-photo-inpainting : Video generation from still images
Source：https://github.com/vt-vl-lab/3d-photo-inpainting/blob/master/image/moon.jpg
Source：https://github.com/wtjiang98/PSGAN
GAST : Highly accurate 3D skeletal estimation
Source：https://github.com/fabro66/GAST-Net-3DPoseEstimation/blob/master/data/video/baseball.mp4
FLAVR : Video frame interpolation
Source：https://github.com/tarun005/FLAVR
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.