I a training a model in batches and am therefore saving its weights into JSON to store/send.

I need to now load those back into tensors - is there a proper way to do this?

tensor.data().then(d => JSON.stringify(d));

// returns
{"0":0.000016666666851961054,"1":-0.00019999999494757503,"2":-0.000183333337190561}

I can iterate over this an convert back to an array manually - but feel there maybe something in the API which would do this cleaner?

I a training a model in batches and am therefore saving its weights into JSON to store/send.

I need to now load those back into tensors - is there a proper way to do this?

tensor.data().then(d => JSON.stringify(d));

// returns
{"0":0.000016666666851961054,"1":-0.00019999999494757503,"2":-0.000183333337190561}

I can iterate over this an convert back to an array manually - but feel there maybe something in the API which would do this cleaner?

• javascript
• tensorflow
• tensorflow.js

Share Improve this question Follow edited Aug 9, 2019 at 10:45 edkeveked 18.4k1010 gold badges5959 silver badges9595 bronze badges asked Jul 25, 2019 at 17:44 dendogdendog 3,36855 gold badges3333 silver badges7070 bronze badges

Add a ment  | 

3 Answers 3

Sorted by: Reset to default 3

There is no need to stringify the result of data(). To save a tensor and restore it later, two things are needed, the data shape and the data flattened array.

s = tensor.shape 
// get the tensor from backend 

saved = {data: await s.data, shape: shape}
retrievedTensor = tf.tensor(saved.data, saved.shape)

The two pieces of information are given when using array or arraySync - the typedarray generated has the same structure as the tensor

saved = await tensor.array()
retrievedTensor = tf.tensor(saved)

This below can solve the issue, because you can export the Weights 'showWeights' in text format to save it in the database, text file ou browser storage for example and after you can apply in your model again with 'setWeightsFromString'.

showWeights() {

    tf.tidy(() => {

        const weights = this.model.getWeights();
        let pesos = '';
        let shapes = '';

        for (let i = 0; i < weights.length; i++) {

            let tensor = weights[i];
            let shape = weights[i].shape;
            let values = tensor.dataSync().slice();

            if (pesos) pesos += ';';
            if (shapes) shapes += ';';
                       
            pesos += values;
            shapes += shape;

        }

        console.log(pesos);  // sValues for setWeightsFromString
        console.log(shapes); // sShapes for setWeightsFromString
        
    });

}

setWeightsFromString(sValues,sShapes) {   
    
   tf.tidy(() => {

        const aValues = sValues.split(';');
        const aShapes = sShapes.split(';');
        const loadedWeights = [];

        for (let i = 0 ; i < aValues.length ; i++) {
            
            const anValues = aValues[i].split(',').map((e) => {return Number(e)});
            const newValues = new Float32Array(anValues);
            const newShapes = aShapes[i].split(',').map((e) => {return Number(e)});

            loadedWeights[i] = tf.tensor(newValues, newShapes);

        }

        this.model.setWeights(loadedWeights);
        
    });
}

This is my code to do this operation.

import { 
    tensor,
    tensor2d,
} from '@tensorflow/tfjs-node'

import { readFile, writeFile } from 'node:fs'

const path2File = './SAVED-TENSOR/obj.json'

//------------------------------------- 2D ------------------------------------
const a = [
    [
        0.9969421029090881,
        9.39412784576416,
        95.00736999511719
    ]
]

const inputs2dT = tensor2d(a)
    
console.log(`@TENSOR  >> `, inputs2dT.dataSync())
// @TENSOR  >>  Float32Array(3) [
//     0.9969421029090881,
//     9.39412784576416,
//     95.00736999511719
// ]

const aa = await inputs2dT.array()
console.log(aa)
// [ [ 0.9969421029090881, 9.39412784576416, 95.00736999511719 ] ]

const aaObj = {
    "tensor": aa
}

writeFile(
    path2File,
    JSON.stringify(aaObj),
    (err) => {
        if (err) throw err

        console.log('@DATA >> Written!')
    }
)

readFile(path2File, (err, rawData) => {
    if (err) throw err
    const obj = JSON.parse(rawData)
    console.log('@DATA >> ', obj.tensor)

    const t = tensor(obj.tensor)
    if (t.constructor.name === 'Tensor') {
        t.print()
    } else {
        console.log('@UNDEFINED >> Tensor')
    }
})