Streaming Image Recognition by WebAssembly

This project demonstrates how to process a video stream in real-time using WebAssembly and apply a pre-trained food classification model to each frame of the video in order to determine if food is present in that frame, all by integrating WasmEdge into YoMo serverless.

Open-source projects that we used:

• Serverless stream processing framework YoMo
• Integrate with WasmEdge to introduce WebAssembly, interop TensorFlow Lite model
• A deep learning model found on TensorFlow Hub; make sure to download TFLite (aiy/vision/classifier/food_V1), which is authored by Google

Values:

• Low-latency: Deep learning in computer vision can be processed on far edge data centers
• Security: WasmEdge isolates code in sandbox
• Low Overhead: As compared to popular container software such as Docker, WebAssembly provides a higher level of abstraction to launch on-the-fly
• Edge Computing: By introducing high-performance, lightweight Wasm VMs, better for resource-constrained edge devices

Steps to run

1. Clone This Repository

$ git clone https://github.com/yomorun/yomo-wasmedge-tensorflow.git

2. Install YoMo CLI

$ go install github.com/yomorun/cli/yomo@latest
$ yomo version
YoMo CLI version: v0.0.5

Or, you can download the pre-built binary tarball yomo-v0.0.5-x86_64-linux.tgz.

Details about YoMo CLI installation can be found here.

3. Install WasmEdge Dependencies

Install WasmEdge

$ wget https://github.com/WasmEdge/WasmEdge/releases/download/0.8.0/WasmEdge-0.8.0-manylinux2014_x86_64.tar.gz
$ tar -xzf WasmEdge-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo cp WasmEdge-0.8.0-Linux/include/wasmedge.h /usr/local/include
$ sudo cp WasmEdge-0.8.0-Linux/lib64/libwasmedge_c.so /usr/local/lib
$ sudo ldconfig

Or, you can build from the source.

Install WasmEdge-tensorflow

Install tensorflow dependencies for manylinux2014 platform

$ wget https://github.com/second-state/WasmEdge-tensorflow-deps/releases/download/0.8.0/WasmEdge-tensorflow-deps-TF-0.8.0-manylinux2014_x86_64.tar.gz
$ wget https://github.com/second-state/WasmEdge-tensorflow-deps/releases/download/0.8.0/WasmEdge-tensorflow-deps-TFLite-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo tar -C /usr/local/lib -xzf WasmEdge-tensorflow-deps-TF-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo tar -C /usr/local/lib -xzf WasmEdge-tensorflow-deps-TFLite-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo ln -sf libtensorflow.so.2.4.0 /usr/local/lib/libtensorflow.so.2
$ sudo ln -sf libtensorflow.so.2 /usr/local/lib/libtensorflow.so
$ sudo ln -sf libtensorflow_framework.so.2.4.0 /usr/local/lib/libtensorflow_framework.so.2
$ sudo ln -sf libtensorflow_framework.so.2 /usr/local/lib/libtensorflow_framework.so
$ sudo ldconfig

Install WasmEdge-tensorflow:

$ wget https://github.com/second-state/WasmEdge-tensorflow/releases/download/0.8.0/WasmEdge-tensorflow-0.8.0-manylinux2014_x86_64.tar.gz
$ wget https://github.com/second-state/WasmEdge-tensorflow/releases/download/0.8.0/WasmEdge-tensorflowlite-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo tar -C /usr/local/ -xzf WasmEdge-tensorflow-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo tar -C /usr/local/ -xzf WasmEdge-tensorflowlite-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo ldconfig

Install WasmEdge-image:

$ wget https://github.com/second-state/WasmEdge-image/releases/download/0.8.0/WasmEdge-image-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo tar -C /usr/local/ -xzf WasmEdge-image-0.8.0-manylinux2014_x86_64.tar.gz
$ sudo ldconfig

If you have any questions about installation, please refer to the official documentation. Currently, this project works on Linux machines only.

Install video and image processing dependencies

$ sudo apt-get update
$ sudo apt-get install -y ffmpeg libjpeg-dev libpng-dev

4. Write your Streaming Serverless function

Write app.go to integrate WasmEdge-tensorflow:

Get WasmEdge-go:

$ cd flow
$ go get -u github.com/second-state/WasmEdge-go/wasmedge

Download pre-trained TensorflowLitee model: lite-model_aiy_vision_classifier_food_V1_1.tflite, store to rust_mobilenet_foods/src:

$ wget 'https://storage.googleapis.com/tfhub-lite-models/google/lite-model/aiy/vision/classifier/food_V1/1.tflite' -O ./rust_mobilenet_food/src/lite-model_aiy_vision_classifier_food_V1_1.tflite

Compile to wasm file:

Install rustc and cargo

$ curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
$ export PATH=$PATH:$HOME/.cargo/bin
$ rustc --version

Set default rust version to 1.50.0: $ rustup default 1.50.0

Install rustwasmc

$ curl https://raw.githubusercontent.com/second-state/rustwasmc/master/installer/init.sh -sSf | sh
$ cd rust_mobilenet_food
$ rustwasmc build
# The output WASM will be `pkg/rust_mobilenet_food_lib_bg.wasm`.

Copy pkg/rust_mobilenet_food_lib_bg.wasm to flow directory:

$ cp pkg/rust_mobilenet_food_lib_bg.wasm ../.

5. Run YoMo Orchestrator Server

  $ yomo serve -c ./zipper/workflow.yaml

6. Run Streaming Serverless function

$ cd flow
$ go run --tags "tensorflow image" app.go

7. Demonstrate video stream

Download this demo vide: hot-dog.mp4, store to source directory, then run：

$ wget -P source 'https://github.com/yomorun/yomo-wasmedge-tensorflow/releases/download/v0.1.0/hot-dog.mp4'
$ go run ./source/main.go ./source/hot-dog.mp4

8. Result