AI-Serving

[Attention] This repo's git ignored all *.c file

[Attention] In some system, protobuf will fallback to pure python implementation, which is much slower

• export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=cpp can help to solve this problem
• Please verify that: run.py:L6 sets to a correct protocbuf implementation

[Attention] Currently only verified on python3 environment

Introduction

• AIServing OnBoard is designed to:

  • provide an unified Deep Learning (DL) detection interface
  • a delegate agent of AIServing Cloud, under a distributed deployment

• Currently supported DL framework:

  • Tensorflow, [ ] Tensorflow Serving, [ ] Tensorflow-Lite
  • PyTorch
  • RKNN
  • MXNet

Components

• src stores all source files
• build_release.py and setup.py is used to build binary
• CHANGELOG is used to keep fixing, changing, adding features log
• Dockerfile is used to build containerize image
• ld.conf is used to link dynamic library inside container building

Installation

• Requirements (OutDated)
  • Depends on the backend you are using, install all
  • Ex1. Tensorflow on Jetson GPU

apt-get -y install libhdf5-serial-dev hdf5-tools libhdf5-dev zlib1g-dev zip libjpeg8-dev
pip3 install -U numpy grpcio absl-py py-cpuinfo psutil portpicker six mock requests gast h5py astor termcolor protobuf keras-applications keras-preprocessing wrapt google-pasta
pip3 install tensorflow-estimator==1.13.0 tensorboard==1.13.0
pip3 install --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v42 tensorflow-gpu==1.13.1+nv19.3

• Build for Release (This helps to protect source code)

  • python build_release.py (in some environment, it needs python3 build_release.py)
  • after above, all release binaries are stored in release-pack
  • ATTENTION: We noticed that generated binaries are highly related to the version of python that executed building command above, hence always use the same version of python to build or execute

• Deployment

  • Copy the whole release-pack folder to the path you want

• Deployment by Docker

  • TODO

Concept

• AIServing Model (JSM)

  • JSM is a folder that contains all information

• AIServing has the following concepts

  • For each inference, all data will pass through workflow shown as blow
    • Image -> Pre-DataProcess -> DL Backend -> Post-DataProcess
    • Here, Pre/Post-DataProcess are implemented by the model provider
  • Pre-DataProcess
    • def pre_dataprocess(infer_data):
    • infer_data comes from HTTP Restful request (/api/v1alpha/detect)
    • expect to return dictionary object which includes a data frame needed by DL framework
  • Post-DataProcess
    • def post_dataprocess(pre_p, prediction, classes):
    • pre_p is Pre-DataProcess result, prediction is DL inference result
    • classes is an array contains all classes supported by this DL model
    • expect to return dictionary object which includes a data frame that you want to return to actual users

Usage

• Config settings.py

  • set storage: path of a directory which stores all JSM bundle
  • set preheat: path of an image which is used to preheat nerual network
  • set redis.host: redis server host
  • set redis.port: redis server port, by default it is "6379"
  • other configurations need to check the specific list of each backend

• Prepare JSM

  • Copy JSM bundle to the storage path you set in settings.py
  • See more details in Serving Bundle section

• Run (w/ options) (Not Available after gRPC API checked-in)

  • format: python3 run.py [options]
  • --port, run on a specific port, by default is 8080
  • --debug, enable debug mode, print much more debug messages
  • --profile, enable profile mode, used to profile functions

• Call gRPC API

  • Please check .proto files stored in src/serving/interface

• Call HTTP Restful API (Deprecated)

  • API version (prefix): /api/v1alpha
  • POST /api/v1alpha/detect: block, use to inference an image

{
  path: "/absolute/path/of/image"
}

* `Response`

{
  status: "success" or "failed"     # status of inference
  result:                           # inference result, depens on models
  message: "inference message"
}

• POST /api/v1alpha/switch: non-block, use to load or switch a model

{
  bid:     "backend id"             # backend id, a new backend will be created if not given
  btype:   "backend type"           # backend type, must be given if a new backend need to create
  model:   "model name"             # model's folder name under collection_path
  mode:    "model type"             # indicate which type of model is going to load
  device:  "device name"            # assign work load to a specific device if available
}

* `Response`

{
  status: "success"                 # this API will always returns success
}

• GET /api/v1alpha/switch: block, used to get loading status
  • Response

{
  model:  "model name"             # the loading model's name
  status: "status"                 # includes: cleaning, loading, preheating, loaded or failed
  error:  "error message"          # error message, if failed to load model
}

Serving Bundle

• AIServing will automatically serving a model once this model is constrcuted with a conventional structure

• Tensorflow (Frozen)

  • model_core
  • pre_dataprocess.py
  • post_dataprocess.py
  • distros.json