開始使用 LiteRT.js

npm install @litertjs/core

import {loadLiteRt} from '@litertjs/core';

// They are located in node_modules/@litertjs/core/wasm/
// Serve them statically on your server.
await loadLiteRt(`your/path/to/wasm/`);

import {CompileOptions, loadAndCompile, loadLiteRt, setWebGpuDevice} from '@litertjs/core';
import {runWithTfjsTensors} from '@litertjs/tfjs-interop';

// TensorFlow.js imports
import * as tf from '@tensorflow/tfjs';
import '@tensorflow/tfjs-backend-webgpu'; // Only WebGPU is supported
import {WebGPUBackend} from '@tensorflow/tfjs-backend-webgpu';

async function main() {
  // Initialize TensorFlow.js WebGPU backend
  await tf.setBackend('webgpu');

  // Initialize LiteRT.js's Wasm files
  await loadLiteRt('your/path/to/wasm/');

  // Make LiteRt use the same GPU device as TFJS (for tensor conversion)
  const backend = tf.backend() as WebGPUBackend;
  setWebGpuDevice(backend.device);
  // ...
}

main();

const model = await loadAndCompile('path_to_model.tflite', {
  accelerator: 'webgpu', // or 'wasm'
});

// Wrap in a tf.tidy call to automatically clean up intermediate TensorFlow.js tensors.
// (Note: tidy only supports synchronous functions).
const top5 = tf.tidy(() => {
  // Get RGB data values from an image element and convert it to range [0, 1).
  const image = tf.browser.fromPixels(dogs, 3).div(255);

  // These preprocessing steps come from https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py#L315
  // The mean and standard deviation for the image normalization come from https://github.com/pytorch/vision/blob/main/torchvision/transforms/_presets.py#L38
  const imageData = image.resizeBilinear([224, 224])
    .sub([0.485, 0.456, 0.406])
    .div([0.229, 0.224, 0.225])
    .reshape([1, 224, 224, 3])
    .transpose([0, 3, 1, 2]);

  // You can pass inputs as a single tensor, an array, or a JS Object
  // where keys are the tensor names in the TFLite model.
  // When passing an Object, the output is also an Object.
  // Here, we're passing a single tensor, so the output is an array.
  const probabilities = runWithTfjsTensors(model, imageData)[0];

  // Get the top five classes.
  return tf.topk(probabilities, 5);
});

const values = await top5.values.data();
const indices = await top5.indices.data();
top5.values.dispose(); // Clean up the tfjs tensors.
top5.indices.dispose();

// Print the top five classes.
const classes = ... // Class names are loaded from a JSON file in the demo.
for (let i = 0; i < 5; ++i) {
  const text = `${classes[indices[i]]}: ${values[i]}`;
  console.log(text);
}

// Imports, initialization, and model loading...
// Create fake inputs for the model
const fakeInputs = model.getInputDetails().map(
    ({shape, dtype}) => tf.ones(shape, dtype));

// Run the model
const outputs = runWithTfjsTensors(model, fakeInputs);
console.log(outputs);