A model with empty weights doesn't load

[thekevinscott]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow.js): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Mac
• TensorFlow.js installed from (npm or script link): npm
• TensorFlow.js version (use command below): 4.8.0

Describe the current behavior
A model.json with an empty array of weights fails to be loaded:

$ tf.loadLayersModel('/path/to/my/model');
        var key = Object.keys(weights)[0].split('/');
                                         ^
TypeError: Cannot read properties of undefined (reading 'split')

Why would I want a model with empty weights?

I maintain a tool, UpscalerJS, for doing image upscaling in Javascript. It has a number of integration tests that require a functioning Tensorflow.js model, and it's helpful to keep that model as simple possible (a single upsampling layer) to assert behavior without significant test latency.

At some point tests were working, but a version upgrade somewhere along the line broke things. I haven't yet been able to isolate the version things changed.

Describe the expected behavior

I would expect a model with an empty weights array to load.

Standalone code to reproduce the issue

https://codesandbox.io/p/sandbox/jolly-firefly-2g63jk

Or, run the following code in Node:

const tf = require("@tensorflow/tfjs-node");
const fs = require("fs");

(async () => {
  const modelPath = `/${__dirname}/model`;
  const modelPathWithFile = `file://${modelPath}`;
  if (!fs.existsSync(modelPath)) {
    const model = tf.sequential();
    const scale = 2;
    model.add(
      tf.layers.upSampling2d({
        size: [scale, scale],
        dataFormat: "channelsLast",
        inputShape: [null, null, 3],
      })
    );

    model.compile({ loss: "meanSquaredError", optimizer: "sgd" });

    await model.save(modelPathWithFile);
  }

  await tf.loadLayersModel(`${modelPathWithFile}/model.json`);
})();

[gaikwadrahul8]

Hi, @thekevinscott

Thank you for bringing this issue to our attention and I tried to replicate the same issue from my end and I'm also getting the same error message which you've mentioned above.

If I'm not wrong model with empty weights won't load in TFJs because TFJs uses the weights to initialize the model layers. If the weights are empty, the layers won't be initialized and the model won't be able to load or run

Here is output from above code snippet for reference :

gaikwadrahul-macbookpro:test-7865 gaikwadrahul$ node index.js
/Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:21666
        var key = Object.keys(weights)[0].split('/');
                                         ^

TypeError: Cannot read properties of undefined (reading 'split')
    at Container.loadWeights (/Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:21666:42)
    at /Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:25621:27
    at step (/Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:159:27)
    at Object.next (/Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:108:53)
    at fulfilled (/Users/gaikwadrahul/Desktop/TFJS/test-7865/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:89:28)

Node.js v18.15.0
gaikwadrahul-macbookpro:test-7865 gaikwadrahul$ 

I tried to load model.json file in https://netron.app/ and I got below output :

[thekevinscott]

Thanks for the feedback, @gaikwadrahul8

If I create the model and run a prediction through it, it works:

const t = tf.ones([1, 2,2,3]);
const prediction = model.predict(t)
// prediction has shape [1,4,4,3]

So it looks like a model can have empty weights if it’s created on the fly, but it can’t be loaded from a file.

[gaikwadrahul8]

@thekevinscott

You're welcome and you understood correctly and I have edited my above comment where I loaded model.json file in https://netron.app/ to see model topology.

If you don't need further assistance on this issue, Could you please confirm if this issue is resolved for you ? Please feel free to close the issue if it is resolved ? Thank you!

[thekevinscott]

The point I was trying to make above is that creating a model on the fly with empty weights works and successfully runs on an input.

Saving that model and reloading it fails.

I would expect the behavior to be identical for both cases. Either both work, or both fail. Or is there a reason that loaded models exhibit different behavior?

[thekevinscott]

It appears the change was introduced in version 4.9.x. Version 4.8.x loads a model with empty weights fine.

Here is a code sandbox pinned to 4.8.x demonstrating the previous working behavior:

https://codesandbox.io/p/sandbox/admiring-snow-5hvhkl?file=%2Findex.js%3A1%2C1

The model loads and is able to run inference against a given tensor.