TANGO 2023 11 Release

This release includes several enhancements and updates on following functions:

• Project Manager:project_manager,
• Base Model Selector:bms,
• AutoNN :autonn_yoloe (for detection task),autonn_resnet (newly added for classification task)
• Code Generation:code_gen (formerly known asdeploy_codegen),
• Target Deployment:cloud_deploy,kube_deploy,ondevice_deploy
  • formerly singledeploy_target container did the whole deployment task on different targets, now we have individual deployment containers for each target

To improve the efficiency of theautonn container, which typically requires several days for training even on high-performance GPUs, we have implemented a manual workflow step-forward functionality within theproject manager. This functionality has been tested and validated through two stages:

• bms +autonn_yoloe orautonn_resnet stage: This stage utilizes the dataset and target configurations within the project manager to select the appropriate base model. The selected base model undergoes fine-tuning, generating a trained model and associated codes for the subsequent stage.
• code_gen +*_deploy stage: Building upon the trained model and generated codes from the previous stage, this stage prepares executable neural network codes for deployment on the specified target, as configured in theproject_manager container. Please note that the current release includes the addition of K8s and cloud (ex. Google Cloud Platform) target deployment-related code, although it is still under going.

Notes on the current release:

• We have tested K8s target deployment-related code, which has been developed and built.
• The project manager in this release usesVue.js based front-end.

BMS and AutoNN

This release includes thebms (Base Model Selector) container, developed by ETRI, which serves as a simple test for BMS member container role. Thebms container selects the base model from the Yolo v7 or Resnet and suitable batch size for training based on target type information such as ondevice (PC, Andorid Device, Embedded Board), K8s or cloud, specified within the project configuration step of theproject_manager. The selected base model is utilized in the AutoNN containers for fine-tuning with the dataset, also specifed within the project configuration step. Additionally, the AutoNNautonn_yoloe andautonn_resnet containers (implemented inautonn\YoloE andautonn\ResNet folders) hvae been included for testing the AutoNN member container role within the TANGO project workflow (pipeline).

Code Generation and Deployment on Targets

We have made changes to the source structure related to deployment codes compared to the previous release. The updates are as follows:

• The codes responsible for generating the executable neural network code have been moved to thedeploy_codegen/optimize_codegen folder.
• The codes for deploying the executable neural network have been relocated to thedeploy_target folder.

Thedeploy_target folder now includes sub-folders based on the deployment target:

• cloud: codes for deployment to cloud environments done incloud_deploy container
• k8s: codes for deployment to Kubernetes done inkube_deploy container
• ondevice: codes for deployment to on-device platforms such as PC, android phones, or embedded devices, fullfilled inondevice_deploy container
  • for PC or embedded devices,ondvice_deploy container generates python codes that can be called by python interpreter.
  • for android phone,ondvice_deploy container generates anAPK file.

Notes on deployment:

• The codes in thedeploy_target/k8s folder can be used to build a Docker image, but integration testing is still ongoing. Therefore, in this release, K8s deployment is not functioning properly.
• In this release,code_gen container supportsTensorRT,Apache TVM,PyTorch andACL. Support forRKNN . Support forrknn is still in progress.

To Do:

TBD