Load Tensorflowjs from json object not json

Question

I am trying to load a Tensorflowjs model using the model.json which is an in memory browser side object.

https://js.tensorflow.org/api/latest/#loadLayersModel

One approach may be to return the json from a dummy fetch method.

fetchFunc (Function) A function used to override the window.fetch function.

Alternatively, it is possible to create a custom IOHandler, but there is very little documentation on this.

An tf.io.IOHandler object that loads model artifacts with its load method.

Does anyone know how to achieve this using the tensorflow load methods?

var modelJson = "{...ModelAndWeightsConfig}";

//Do something here to load it.

var newModel =  tf.loadLayersModel("/model_0/model.json", {
                    onProgress: onProgressCallback}).then(model =>{});

Regards,

Thomas Dondorf · Accepted Answer · 2019-09-22 14:46:39Z

5

Yes, you can write your own IOHandler to load the model. Check out the definition of an IOHandler here. You have to implement the load function that returns a Promise<ModelArtifacts>.

That means, to load a model saved by the file IOHandler, you can check out the source code and reimplement the load function yourself.

Code Sample

Here is example to get you started. The load() part is mostly copied from the loadJSONModel function from the file IOHandler. Basically, the JSON string is passed as an argument and then used when the load function is called by Tensorflow.js.

export class JSONHandler implements tfc.io.IOHandler {
  constructor(jsonString) {
     this.jsonString = jsonString;
  }
  async load() {
    const modelJSON = JSON.parse(jsonString);
    const modelArtifacts: tfc.io.ModelArtifacts = {
      modelTopology: modelJSON.modelTopology,
      format: modelJSON.format,
      generatedBy: modelJSON.generatedBy,
      convertedBy: modelJSON.convertedBy
    };
    if (modelJSON.weightsManifest != null) {
      // load weights (if they exist)
    }
    if (modelJSON.trainingConfig != null) {
      modelArtifacts.trainingConfig = modelJSON.trainingConfig;
    }
    if (modelJSON.userDefinedMetadata != null) {
      modelArtifacts.userDefinedMetadata = modelJSON.userDefinedMetadata;
    }
    return modelArtifacts;
  }
}

To use the model, you can then create an instance of it and pass it to the load function:

const modelJson = '{ ... }';
const handler = new JSONHandler(modelJson);
const model = await tf.loadLayersModel(handler);

edited Sep 22, 2019 at 14:46

answered Sep 22, 2019 at 14:23

Thomas Dondorf

25.4k6 gold badges97 silver badges113 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Jama Mohamed Over a year ago

hi, i seem to get the following error, Uncaught (in promise) TypeError: specs is not iterable at Module.decodeWeights (io_utils.js:105), is there a way to resolve this

WebSight · Accepted Answer · 2019-09-22 13:16:48Z

0

Fetch the models

    var fetchPromise = function(url,p1,p2,) {
        return new Promise(function(resolve, reject) {
            fetch(url)
                    .then(response => {
                        resolve(response);
                    }).catch(err =>{
                        reject();
                    });
        });
    };

answered Sep 22, 2019 at 13:16

WebSight

6802 gold badges14 silver badges32 bronze badges

Collectives™ on Stack Overflow

Load Tensorflowjs from json object not json

2 Answers 2

1 Comment

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

1 Comment

Comments

Your Answer

Sign up or log in

Post as a guest

Related