在本地将 keras model 加载到 tensorflow.js 中

Question

I want to load keras model which has been converted into tensorflow.js compatible format to perform inference. My code looks like this

import * as tf from '@tensorflow/tfjs';
import "regenerator-runtime/runtime.js";
import 'bootstrap/dist/css/bootstrap.css';

const model_path = '/home/user/Desktop/Github/tfjs_model/tfjs/tfjs_model/model.json'; 

async function loadModel(path){
  console.log("Model loading in progress from ".concat(path));
  const model =  await tf.loadLayersModel(path);
  console.log("Model Loaded Successfully");
  return model;
};

const model = loadModel(model_path);

const input = tf.tensor1d([1], [13]);
console.log(model.predict(input));

I am getting this error Error: Failed to parse model JSON of response from /home/user/Desktop/Github/tfjs_model/tfjs/tfjs_model/model.json'. Please make sure the server is serving valid JSON for this request. Error: Failed to parse model JSON of response from /home/user/Desktop/Github/tfjs_model/tfjs/tfjs_model/model.json'. Please make sure the server is serving valid JSON for this request. I've looked in a lot of placed (Github issues, Stack overflow) but can't seem to find a fix. How can this be resolved? `

Answer 1

I had this problem too.

Using const model = await tf.loadLayersModel('directory/model.json'); generated the same error you had. I tried changing the import to find file model2.json , which didn't exist yet gave the same error (I notice there's an outstanding issue with TensorFlow.js to give a more meaningful error).

Looking at my Parcel setup, I realised that the model.json and .bin files were being copied into the root of the dist folder. So I removed the directory from the import and now it works with:

const model = await tf.loadLayersModel('model.json');

Note that my files are still in <content_root>/directory/ .