Topology mismatch with 0.20.0.alpha releases

[j3nsykes]

Not wanting to open duplicate issues but I have narrowed this issue
to be related to the last 3 alpha releases of ML5.

It is not present with ml5@latest/dist/ml5.min.js

However, it seems when saving and loading the files with the neural net function in the alpha releases the tensorflow topology is expecting a different structure ?

I have tried this with nextgen handpose and also a simple mouse X,Y regression. Both cause this error as soon as the version is 0.20.0.alpha 1,2 or 3.

[shiffman]

Thank you for reporting this @j3nsykes! Can you share one of your examples where you are getting this error. I'd like to take a look. I have been using the ml5.neuralNetwork() class for my new chapters in The Nature of Code, but I never actually tested the saving and loading model features!

it's absolutely fine to open duplicate issues if they are present in this new repo! I realize we've caused some extra confusion and additional labor in moving to a new repo, but it helped us re-engage with the project after a dormant period! Hopefully we'll be able to bring everything together smoothly into a new release soon!!!

In case you are curious to take a look: https://natureofcode.com/neural-networks/ https://natureofcode.com/neuroevolution/)

[j3nsykes]

Hey Dan,
Sure its happening here when I'm adapting one of Golan's hand tracking regression examples to be saved and loaded but also in a simple mouse regression example too (so not related to handpose.js)

The mouse example I've been using to test which version of ML5 it breaks with.

[shiffman]

Ok, so I've dug into this error a bit more! The code still works if you are loading the model files directly with a single filename rather than an object:

This works:

  classifier.load("model.json", modelLoaded);

However, this does not work:

  const modelDetails = {
    model: "model.json",
    metadata: "model_meta.json",
    weights: "model.weights.bin",
  };
  classifier.load(modelDetails, modelLoaded);

However, only the second option is compatible with the web editor where the files are stored separately and the code needs to extrapolate the hosted path (?). It does that via something along the lines of:

 this.model = await tf.loadLayersModel(tf.io.browserFiles([file1, file21]));

Rather than the more simple:

   this.model = await tf.loadLayersModel(file1);

So I think something changed in a new version of tf.js with tf.io.browserFiles(). @ziyuan-linn offered in our meeting this morning to look into this further! Hopefully this helps you get started with a place to look!

[j3nsykes]

Ah amazing thank you. I actually plan to not use the web editor so can probably progress with a single filename option for now.
Thanks for digging in to this!

[ziyuan-linn]

Thank you @j3nsykes for reporting the issue and thank you @shiffman for pointing me in the right direction! The issue should be fixed now with the new PR.

The original code passes a twice-stringified JSON file to TensorFlow.js, and it only gets parsed once by tfjs, hence the modelTopology field missing error. I have removed the JSON.stringify call on ml5's side to fix the issue.

I didn't look into it, but my best guess is that a previous version of Axios automatically parsed JSON results and no longer does so in the current version. Which is why we have the redundant JSON.stringify.

[j3nsykes]

thank you @ziyuan-linn and @shiffman !