The open-source tool Adlik for optimizing and deploying trained deep learning models offers more power and new options in version Eagle 0.5.
Project Adlik codenamed Eagle version 0.5 reached: The open source toolkit is in the incubator of the LF AI & Data Foundation, a subsidiary of the Linux Foundation, where it is currently undergoing its probationary phase. As an optimization framework, it is designed for deep learning models and includes a number of components with which trained models can be transferred to productive operation more quickly. According to its publishers, the new release promises more speed and expanded options for fine tuning. With Adlik Eagle, models can be adapted to the respective application environment.
Provide customized deep learning models
Adlik Eagle’s model compiler provides various optimization methods such as pruning, quantization and structural compression available with which developers can further adjust their deep learning models from common frameworks such as TensorFlow (Keras) or Caffe in order to achieve higher computing power and lower latencies during inference. With version 0.5, the Adlik team for compressing YOLOv5s models introduces support for Quantization and Knowledge Distillation methods. In connection with the OpenVINO Runtime, the inference performance of such DL models can be increased by about two and a half times.
For this purpose, the team also has the model compiler and the inference engine from Adlik Eagle on the OpenVINO version 2022.1.0 updated. The compiler enables developers to transfer models from OneFlow to ONNX format. The inference engine, with which deep learning models can be adapted to the respective deployment environments in the cloud or embedded environment and to the underlying hardware (CPU, GPU, FPGA), also covers the Open, which is designed for scientific calculations, from Adlik Eagle 0.5 source library Torch.
Um To save already optimized deep learning models and keep them ready for further use, Adlik users can access a new repository in the Eagle 0.5 Model Zoo, which now also includes ResNet and YOLOv5 models, among others. An overview of all other innovations in the release can be found in the announcement of the foundation and in the release notes of the open source project on GitHub.
