React Native 与 TypeScript tfjs-models/universal-sentence-encoder 模型加载错误

Question

我正在尝试通过关注these instructions 在 React Native 应用程序中使用 tfjs-models/universal-sentence-encoder。但是，当我尝试加载模型时出现以下错误：

ERROR: TypeError: undefined is not an object (evaluating '_universalSentenceEncoder.default.load'

代码：

import React, { useEffect, useState } from 'react';
require('@tensorflow/tfjs');
const use = require('@tensorflow-models/universal-sentence-encoder');

export default function App() {

  useEffect(() => {
    console.log("App is starting...")
    
    const init = async () => {
      // initialize state variables 
      // console.log("App is initializing services...")
      
      // Load the model.
      try {
        use.load().then((model: any) => {
          // Embed an array of sentences.
          const sentences = [
            'Hello.',
            'How are you?'
          ];
          model.embed(sentences).then((embeddings: any) => {
            // `embeddings` is a 2D tensor consisting of the 512-dimensional embeddings for each sentence.
            // So in this example `embeddings` has the shape [2, 512].
            embeddings.print(true /* verbose */);
          });
        });
      }
      catch (err) {
        console.log(`ERROR: ${err}`);
      }
    };
  }, []);

软件包版本：

• react-native@0.63.3
• @tensorflow-models/universal-sentence-encoder@1.3.2
• @tensorflow/tfjs@3.6.0
• @tensorflow/tfjs-react-native@0.5.0

Answer 1

我认为这与您导入通用句子编码器的方式有关，请尝试以下方法：

import React, { useEffect, useState } from 'react';
import * as use from '@tensorflow-models/universal-sentence-encoder';

export default function App() {

  useEffect(() => {
    console.log("App is starting...")
    
    const init = async () => {
      // initialize state variables 
      // console.log("App is initializing services...")
      
      // Load the model.
      try {
        use.load().then((model: any) => {
          // Embed an array of sentences.
          const sentences = [
            'Hello.',
            'How are you?'
          ];
          model.embed(sentences).then((embeddings: any) => {
            // `embeddings` is a 2D tensor consisting of the 512-dimensional embeddings for each sentence.
            // So in this example `embeddings` has the shape [2, 512].
            embeddings.print(true /* verbose */);
          });
        });
      }
      catch (err) {
        console.log(`ERROR: ${err}`);
      }
    };
  }, []);

Answer 2

yudhiesh 提供的答案修复了引用的错误，但产生了另一个错误。需要对代码进行更多更改，并且需要将 @tensorflow/tfjs 从 3.3.0 降级到 3.0.0（紧随 this post）。

这是更新后的代码：

import React, { useEffect, useState } from 'react';
import "@tensorflow/tfjs-react-native";
import * as tf from '@tensorflow/tfjs';
import * as use from '@tensorflow-models/universal-sentence-encoder';

export default function App() {

  useEffect(() => {
    console.log("App is starting...")
    
    const init = async () => {
      // initialize state variables 
      // console.log("App is initializing services...")
      
      await tf.ready();
      
      // Load the model.
      try {
        use.load().then((model: any) => {
          // Embed an array of sentences.
          const sentences = [
            'Hello.',
            'How are you?'
          ];
          model.embed(sentences).then((embeddings: any) => {
            // `embeddings` is a 2D tensor consisting of the 512-dimensional embeddings for each sentence.
            // So in this example `embeddings` has the shape [2, 512].
            embeddings.print(true /* verbose */);
          });
        });
      }
      catch (err) {
        console.log(`ERROR: ${err}`);
      }
    };
  }, []);

注意@tensorflow/tfjs-react-native 的导入，@tensorflow/tfjs、@tensorflow-models/universal-sentence-encoder 导入方式的变化，以及添加调用await tf.ready() 以确保在尝试加载使用模型之前初始化 tensorflow。

此代码在 iOS 模拟器上运行时会按预期生成嵌入，但在我的物理 iPhone 上运行时会生成一个包含 NaNs 的嵌入向量。这可能是一个单独的问题，因此请考虑解决此问题。