model.save() | ValueError: Cannot find any save handlers for URL

System information

• OS Platform and Distribution: Linux Debian 5.15.5-2kali2
• TensorFlow.js installed from: npm
• TensorFlow.js version: ^3.19.0
• Node version: v16.13.1
• NPM version: 8.1.2

Describe the current behavior I am trying to save the trained sequential model to the file system for later use. I have the same problem as the issues described in #4280 and #723. I have tried to any possible solution suggested in those posts but could not solve it. I am running an Electron application, the training code is called within the main process so it’s backed by NodeJS.

// Traning code
// ...

// Save model code
try {
    let saveResults = await model.save('file://./myModel');
    console.log(saveResults);
}catch ( e ) {
    console.log(e);
}

This is the error log

ValueError: Cannot find any save handlers for URL 'file://./myModel'
    at Sequential.save (webpack:///./node_modules/@tensorflow/tfjs-layers/dist/engine/training.js?:1561:23)

Describe the expected behavior According to Save and load models, the sequential tensorflow model files - [model].json & [model].weights.bin, should be saved in the file system.

[vladmandic]

I really hope to get tfjs-node working

i really really wouldn’t go down that path

but if you insist, check how package.json build.extraFiles works for electronjs
in short, you can use it to package node_modules/@tensorflow/tfjs-node which will include already built binary bindings for your platform. but if you need electron app built for anything but same platform as your dev system, no chance.

i’ve already suggested two much cleaner ideas

• use indexdb as model store and transfer to and from indexdb using electron ipc to nodejs main process.
• implement your i/o handler for tfjs so .save() and .load() methods will actually use electron ipc to your nodejs main process

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No