AI on mobile: Powering your field force – Part 4 of 5


TensorFlow Lite deployment in browser using WebAssembly

Dear readers, I am a software developer, and this is my very first blog post. I would like to thank Gunter Albrecht for inspiring me to research on the topic of WebAssembly. There have been several attempts to run inferencing for ML models on the web browser using TensorFlow.js, ONNX runtime web etc. We wanted to explore the possibility of using WebAssembly backend for this purpose as it is lighter with smaller binary size and faster than javascript.

Abstract

We have seen in the Part 3 of this blog post series from Sumin Lee, “Building deep learning model within SAP BTP” who has explained how we trained the model and converted to tflite and stored the model. In this blog post you will learn how to load a TensorFlow Lite Machine Learning model in C++ and run inference on the web browser with the help of WebAssembly backend. Here we take the example of passing the meter reading image as an input to the TensorFlow Lite model and decode the output predictions to the string matching the digits of the meter reading captured.

Overview%20of%20WebAssembly%20Binary%20creation

Overview of WebAssembly Binary creation

Overview

In this blog post we will look into the following learning journey:

  • Creating WebAssembly binary
    • System Information
    • Pre-requisites
    • Build libraries for the application
    • C++ application code for the meter reading scenario
  • Test WebAssembly binary

Creating WebAssembly Binary

A simple “Hello World” example is provided in this emscripten tutorial for generating the WebAssembly binary and the related JavaScript and html glue code from the C++ application code:

https://emscripten.org/docs/getting_started/Tutorial.html

In the following sections I will describe some of the pre-requisites to compile a C++ application code to WebAssembly binary for our meter reading scenario.

System Information

I am using a MacBook Pro with Apple M1 Max chip. Here I am using a Linux VDI to compile the TensorFlow Lite library with CMake.

Pre-requisites

The following need to be installed on the Linux VDI:

  1. Git
  2. Emscripten SDK
  3. CMake (version 3.20.2 used here) (Optional at this point)
  4. TensorFlow – Clone the TensorFlow git repository using the following command (Optional at this point):
    • git clone https://github.com/tensorflow/tensorflow.git tensorflow_src
  5. Python (to test the application) – here I have used version 2.7

Build libraries for the application

I will now describe the steps to build the library files:

  1. Follow steps 1-3 from the following TensorFlow Lite documentation:
  2. Open a bash terminal and make sure to set the EMSDK environment variables as explained in the Emscripten tutorial https://emscripten.org/docs/getting_started/downloads.html.
  3. Run the build commands with the following compiler flags. In the below code snippet replace the path to the respective “emcmake”, “cmake”, “tensorflow lite source directory” according to your local settings as depicted in the picture below from my local setup:
/path/to/emcmake /path/to/cmake_binary /path/to/tensorflow_lite/source -DCMAKE_CXX_FLAGS="-lpthread" -DTFLITE_ENABLE_MMAP=OFF -DTFLITE_ENABLE_NNAPI=OFF -DTFLITE_ENABLE_RUY=ON -DTFLITE_ENABLE_XNNPACK=ON

CMake%20build%20Step

CMake build Step

This build initially failed for me as Emscripten is not one of the CMAKE_SYSTEM_NAME in couple of library paths generated duringthe build. This is the failure I observed:CMake%20build%20error

CMake build error

  • Perform the following changes to make the build successful. Here I have updated the CMakeLists.txt file inside the “cpuinfo” directory as follows:
  • File%20Diff

File Diff

Also similarly add the missing CMAKE_SYSTEM_NAME for “Emscripten” to the “xnnpack” directory’s CMakelists.txt file and add all the source files to the same CMakeListst.txt inside the “xnnpack” directory to avoid getting any “undefined reference errors” during the build. Repeat the step – 3 CMake build with flags of this section to re-build.

Next run the following command to build the library files:

/path/to/emcmake /path/to/cmake_binary --build .


Once the build is successful you will get the following message:Build%20Successful

Build Successful

C++ application code for the meter reading scenario

We will create a CMakeLists.txt file for the project like one below:

cmake_minimum_required(VERSION 2.4) project(meter_reading) include_directories("/path/to/flatbuffers/include")
include_directories("path/to/current/directory/of/CMakeLists.txt")
link_directories("/path/to/libraries/")
link_directories("/path/to/libraries/pthreadpool")
link_directories("./path/to/_deps/ruy-build/ruy")
link_directories("/path/to/_deps/fft2d-build")
link_directories("/path/to/_deps/flatbuffers-build")
link_directories("/path/to/_deps/xnnpack-build")
link_directories("/path/to/_deps/farmhash-build")
link_directories("/path/to/_deps/clog-build")
link_directories("/path/to/_deps/cpuinfo-build")
add_executable(meter_reading demo.cpp)
set(CMAKE_EXECUTABLE_SUFFIX ".html")
set_target_properties(meter_reading PROPERTIES LINK_FLAGS "-s LLD_REPORT_UNDEFINED -s WASM=1 -s NO_EXIT_RUNTIME=1 -s EXPORTED_FUNCTIONS='['_evaluate', '_malloc', '_free']' -s 'EXPORTED_RUNTIME_METHODS=['ccall', 'printErr']' -s TOTAL_MEMORY=65536000 -s INITIAL_MEMORY=65601536 -s ALLOW_MEMORY_GROWTH=1 -s ABORTING_MALLOC=0") target_link_libraries(meter_reading -lrt -Wl --shared-memory --no-check-features
-ldl
-lpthread
-lflatbuffers
-lpthreadpool
-lruy_kernel_avx2_fma
-lruy_prepacked_cache
-lruy_kernel_arm
-lruy_have_built_path_for_avx512
-lruy_have_built_path_for_avx2_fma
-lruy_frontend
-lruy_pack_avx512
-lruy_pack_arm -lruy_block_map -lruy_blocking_counter -lruy_denormal -lruy_kernel_avx512
-lruy_context -lruy_context_get_ctx -lruy_wait -lruy_allocator -lruy_system_aligned_alloc -lruy_tune -lruy_have_built_path_for_avx
-lruy_apply_multiplier -lruy_ctx -lruy_pack_avx2_fma -lruy_trmul -lruy_pack_avx -lruy_prepare_packed_matrices -lruy_kernel_avx
-lruy_thread_pool -lruy_cpuinfo
-lclog
-lXNNPACK
-lcpuinfo
-lfft2d_fftsg2d
-lfft2d_fftsg
-lfarmhash
-ltensorflow-lite
)

This is a typical way to create a CMake project with the required include directories and link directories added. The library files generated from the above section could be copied to a local directory, here “assets” and linked as shown above to the “meter_reading” project.

In addition, the link flags are added using “set_target_properties” to link the application to a webassembly binary file along with JavaScript and html glue code as shown above.

The export function “_evaluate” is the one that is exposed from the C++ application to the front end via the webassembly binary.

Now let us talk about the content of the demo.cpp file which contains the application code.

  1. The “evaluate” method will contain the main logic of the application where we read the input parameters sent by the front end, load the ML model, and run prediction on the input image and return the output string decoded.
  2. Include files:
    1. Include the relevant files for the application, here I have included the system headers taking reference from the TensorFlow lite minimal example:https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/minimal/minimal.cc
    2. Include the emscripten system header to export the “evaluate” method.
  3. Write a decoding function to map the output predictions to the dictionary as per the model output.

Build this application code to generate the webassembly binary and the glue code as below:

Create the “build_test” directory where the source files are present and run the following commands.

/path/to/emcmake /path/to/cmake_binary ..

Find below my local setup command for reference:
Application%20CMake%20build

Application CMake build

Then execute the following in the same directory:

/path/to/emcmake /path/to/cmake_binary --build .

The build directory will contain the following output:

File%20Structure

File Structure

Test WebAssembly binary

I have updated the “meter_reading.html” file generated above to read an input image of the meter reading and fetch the ML model from the smart meter service endpoint URL and pass these parameters as input to the “evaluate” method.

I have written the following methods in JavaScript:

    1. Get the Meter Reading:

      Here 350*50 is the input image resolution with 3 channels which is accepted by the model and 11 is each output tensor array’s number of columns.

      // get the meter reading function getMeterReading() { //Read Image var canvas = document.getElementById("myCanvas"); var ctx = canvas.getContext("2d"); var img = document.getElementById("test-image"); ctx.drawImage(img, 0, 0, canvas.width, canvas.height); //Read pixel data let pixData = tf.browser.fromPixels(ctx.getImageData(0, 0, canvas.width, canvas.height)) // Convert it into float let pixDataFloat = pixData.div(255.0); // Transpose Image data to 350 * 50 let transposedPixDataFloat = tf.transpose(pixDataFloat, perm=[1, 0, 2]); // Expand dimension to 1 * 350 * 50 let dimsImgPixDataFloat = tf.expandDims(transposedPixDataFloat, 0); // Convert pixel data array to pointer const arrayPointer = arrayToPtr(dimsImgPixDataFloat.dataSync(), Module, 4); const modelPointer = modelToPtr(modelData, Module, 1); // Import function from Emscripten generated file var result = Module.ccall( 'evaluate', // name of exported C function 'string', // return type ['number', 'number', 'number' ,'number', 'number', 'number', 'number'], // argument types [arrayPointer, 50, 350, 3, 11, modelPointer, modelLen] // arguments ); console.log("Result from Meter Reader: ", result); Module._free(arrayPointer); }
    2. Array to Pointer Conversion methods for input image and model:
      // convert input array to a pointer function arrayToPtr(array, Module, bytesPerPixel) { // Get data byte size, allocate memory on Emscripten heap, and get pointer, for 4 bytes per pixel var ptr = Module._malloc(array.length * bytesPerPixel) // Copy data to Emscripten heap Module.HEAPF32.set(array, ptr / bytesPerPixel) return ptr } function modelToPtr(array, Module, bytesPerPixel) { // Get data byte size, allocate memory on Emscripten heap, and get pointer, for 4 bytes per pixel var ptr = Module._malloc(array.length * bytesPerPixel) // Copy data to Emscripten heap Module.HEAPU8.set(array, ptr / bytesPerPixel) return ptr }​
    3. I call this getMeterReading function from the html file:
      <button id="btn-capture" type="button" class="button" onclick="getMeterReading()">Get Meter Reading</button>
      <canvas id="myCanvas" width="350px" height="50px"style="border:1px solid #d3d3d3;">
      <image src="https://localhost:8000/00025332.png" id="test-image"/>​
    4. Finally using a python http server, we can test the end-to-end implementation:
      python -m SimpleHTTPServer​

      This starts the http server on the localhost at port 8000 in my case.HTTP%20Server

      HTTP Server

    5. We can see the output when we open the meter_reading.html file on the http://localhost:8000 as I have logged in the console upon pressing the “Get Meter Reading” button.Console%20Output

Console Output

I hope by following this blog post you can build your very first application running inference for TensorFlow Lite model on the web browser using WebAssembly. In the next part of this blog post series we will learn from Naman Gupta about the front end application.

Please feel free to Like, Comment and ask any questions you have about this blog post.