Skip to content

Backends

Vladimir Mandic edited this page Nov 18, 2022 · 12 revisions

Backends

For use in Browsers, Human library includes pre-bundled CPU, WASM and WebGL backends
as well as optimized HumanGL backend, there is no need to load them externally

For use in NodeJS, Human library supports tfjs-node and tfjs-node-gpu implementations of tensorflow backend
which have to be loaded before Human library

Experimental support is added for WebGPU backend in Browsers and WASM backend in NodeJS


Backend Performance Compare

This is not real-world expected performance, this is worst-case with optimizations disabled to highlight differences between each backend

Backend Environment Execution Variation Warmup Average Note
CPU NodeJS CPU - 78,750 ms 75,000 ms Reference only
CPU Browser CPU - 92,250 ms 91,000 ms Reference only
Tensorflow NodeJS CPU - 2,635 ms 2,110 ms Recommended
Tensorflow NodeJS GPU - 4,650 ms 1,810 ms Advanced level
WASM NodeJS CPU SIMD 3,730 ms 3,500 ms Experimental
WASM Browser CPU 25,630 ms 21,200 ms
WASM Browser CPU SIMD 4,325 ms 3,670 ms Recommended when running without GPU
WebGL Browser GPU 22,050 ms 1,590 ms
HumanGL Browser GPU 6,170 ms 1,600 ms Recommended when running with GPU
WebGPU Browser GPU GLSL 21,350 ms 1,150 ms Experimental
WebGPU Browser GPU WGSL N/A N/A Not available
TFLite Browser CPU N/A N/A Not available
TFLite Browser CPU SIMD N/A N/A Not available

Notes

  • Human configuration: All models enabled
  • Human configuration: Results caching, interpolation and frame change detection disabled
  • CPU backend is running interpreted JS code and is included for reference only
  • HumanGL backend is the optimized version of WebGL backend for faster warmup
  • WebAssembly SIMD is enabled by default in Chome 93 or higher, otherwise it has to be manually enabled
  • WebGPU GLSL implementation is experimental and is only available in Chrome Canary
  • WebGPU WGSL implementation will be added in future versions
  • TFLite support will be added in future versions

Recommendations

  • For Browser, if you have a GPU, recommended backend is HumanGL
    which reduces warmup time compared to WebGL while maintaining fast execution time
  • For Browser, if you don't have a GPU, recommended backend is WASM
    For optimal WASM performance, SIMD support should be enabled in browser
  • For NodeJS, recommended backend is tfjs-node
    Note that while tfjs-node-gpu achieves slightly higher performance than tfjs-node, it's usage is limited to specific nVidia CUDA enabled environments and does not achieve much higher performance since bottlneck is upload&download of textures from GPU, not execution itself




Tensorflow Backend

When used in NodeJS, load either tfjs-node or tfjs-node-gpu modules before loading Human library
and set Human config.backend to tensorflow

If using tfjs-node-gpu module and there is no available GPU acceleratrion, backend will print warning messages
during execution and will fall back to CPU execution model instead

GPU acceleration in NodeJS is currently supported for nVidia GPU cards

For requirements and how to unable GPU acceleration using tfjs-node-gpu in
native Linux, Windows or Windows WSL2 environments,
see appropriate docs on nVidia CUDA


WebGL Backend

WebGL backend usage is fastest on systems with modern GPU
although startup speed is a bit slower

HumanGL Backend

Human library includes custom version of WebGL backend with specific optimizations
to enable enhanced workflows and set specific WebGL options to increase warmup speed
and overal performance of the WebGL backend

HumanGL is a default backend for Browser implementations


CPU Backend

CPU backend usage is not recommended for performance reasons
as TensorFlow implementation in JavaScript is by far the slowest


WASM Backend

WASM backend has good resuts under specific conditions:

  • Enable WASM SIMD support in your browser
    (CPU parallel processing instructions)
  • Reduce size of input image (e.g. webcam)
    as WASM does not have optimized image extraction methods

Loading WASM Files

WASM backend requires access to wasm files which are part of @tensorflow/tfjs-backend-wasm package and included in
node_modules/@tensorflow/tfjs-backend-wasm/dist/*.wasm (different WASM file is loaded depending on detected capabilities)

Also, to be able to load WASM files your web server, web server has to set appropriate HTTP headers:

Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp

Or configure Human load WASM files directly from a CDN:

wasmPath = 'https://cdn.jsdelivr.net/npm/@tensorflow/[email protected]/dist/';

Note that version of WASM binaries must match version of TFJS used by Human library


How to enable WASM SIMD support

WebAssembly SIMD is enabled by default in Chome 93 or higher, otherwise it has to be manually enabled

Chrome:

  • Navigate to chrome://flags
  • WebAssembly SIMD support set to Enabled

Edge:

  • Navigate to chrome://flags
  • WebAssembly SIMD support set to Enabled

Firefox:

  • Navigate to about:config
  • javascript.options.wasm_simd set to True

WebGPU Backend

In the long-term, WebGPU backend will superseed existing WebGL backend - it's intention is to support same workflows and features while having lower overhead while communicating with GPU

Currently, WebGPU support is only present in Canary version of Google Chrome
And must be explicitly enabled via --enable-unsafe-gpu flag

Due to it's experimental nature WebGPU module is not bundled with Human and instead it has to be manually loaded
After WebGPU has been loaded, simply set Human config.backend to WebGPU and it will perform additional checks and initializations as needed

Note that @tensorflow/tfjs-backend-webgpu is currently in alpha phase of development and released packages are often too old
Instead, it is recommended to build @tensorflow/tfjs-backend-webgpu from sources directly from main branch


Known Issues

See <TODO.md>


Clone this wiki locally