tfjs

Custom build of Tensorflow/JS for Browsers as pure ESNext

This solution builds TensorFlow/JS for Browsers, including individual modules as well as a unified bundle

Included Backends: CPU, WebGL, WebGPU, WASM

Node packages (tfjs-node and tfjs-node-gpu) are not covered

Notes

• Does not perform any modifications to tfjs sources
  Instead it relies on custom import resolver during build process
• Does not use build hacks such as symlinks or virtual packages or manually set sideeffects
• Masively smaller:
  • original tf.es2017.js is 3.9MB without wasm and webgpu backends which are another 2MB on top
  • new tfjs.esm.js is 2.5MB with wasm and webgpu included
    and that is non-minfied and with tree shaking disabled
• Entire build is about 10x faster
  Full build takes <2min on a notebook (of which >90% is compiling WASM code)
  Rebuild takes <10sec
• Changing WASM build compiler parameters improves performance slightly
• Changing backend imports to link to source files instead of symlinked packages solves a lot of dependency issues
• Changing imports from files generated in dist to actual sources makes everything tree shakable

<hr>

Install

1. Internal dependencies: npm install
2. External dependencies: git, bazel, python Git: assumed to be already installed if you’re here :)
   Bazel: requires bazel 4.x https://docs.bazel.build/versions/4.2.1/install-ubuntu.html
   Python: requires python 3.x aliased as python, not compatible with python 2.x

Steps

Full Build

Downloads sources and performs full build of all TFJS modules

npm run build executes steps:

1. scripts/download.sh
   on first run it clones @tensorflow/tfjs from git into ./src
   on subsequent runs it rolls forward cloned copy of @tensorflow/tfjs
2. scripts/prepare.sh
   creates versioned .tfjs-*.ts from build templates
   and .js imports from .json files
3. scripts/build-wasm.sh
   build wasm binaries and wasm .js exports using bazel
4. node scripts/build-tfjs.js
   build targets in /dist using esbuild with custom resolver plugins and TypeScript compiler settings

Fast Build

Available once full build has been completed

npm run build executes steps:

1. node scripts/build-tfjs.js
   build targets in /dist using esbuild with custom resolver plugins and TypeScript compiler settings

TypeDefs

Default typedefs in types/index.d.ts are re-exports of official @tensorflow/* modules Buildinging typedefs from sources is an experimental feature

npm run types executes steps:

1. compile all modules using TSC to types/lib
2. build rollup using API-Extractor to types/tfjs-core.d.ts and types/tfjs.d.ts
3. patch rollup

Clean

Remove currently downloaded version of TFJS and reset all cached info

npm run clean

Notes

• /dist: all JS modules, MAP files and WASM binaries
• /dist/tfjs.esm.js is default TFJS bundle that includes all available backends
• /types/index.d.ts re-exports TypeScript type definitions from @tensorflor/tfjs as a single bundle
• npm run fast can be used after completed initial build as it executes last build step only

<hr>

This site is open source. Improve this page.