Tfjs: Manually setting weights for model

[ArthDh]

TensorFlow.js version

0.11.6

Browser version

Describe the problem or feature request

I'm trying to manually set weights for a model(genetic algorithm) but cannot find the required functionality, I have tried model.setWeights() but it doesn't work for me. I don't want to save the entire model and load those weights each time as that would be computationally expensive.

Error

Uncaught TypeError: Cannot read property 'length' of undefined
at [email protected]:1
at e.tidy ([email protected]:1)
at t.setWeights ([email protected]:1)
at t.setWeights ([email protected]:1)

Any help towards using model.setWeights() or an alternative function will be appreciated! Thanks!

[caisq]

@ArthDh

You can access a model's layer by using model.layers. You can set a layer's weights with layer.setWeights(). Therefore you can use code like the following to set the weights of a single layer:
model.layers[2].setWeights(...). You still can't set individual weights. But at least this helps you narrow down to a smaller set of weights. Does this help with your use case?

[ArthDh]

@caisq
Thanks a ton! That worked!

[smith-kyle]

It appears that Model now has both instance methods setWeights and getWeights.

Is this the equivalent of getting each layer's weights individually and setting them individually?

[ArthDh]

Yes, for model.setWeights() weights Should be a list of Tensors with shapes and types matching that of model.getWeights().

[smith-kyle]

Ah ok thanks for your help. Perhaps this should be added to the documentation

[smith-kyle]

I'm getting the error Input Tensor ID not referenced when running predict on a model that has had its weights set.

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

const model = tf.sequential();
model.add(tf.layers.dense({ units: 1, inputShape: [1] }));

// Prepare the model for training: Specify the loss and the optimizer.
model.compile({ loss: "meanSquaredError", optimizer: "sgd" });

// Generate some synthetic data for training.
const xs = tf.tensor2d([1, 2, 3, 4], [4, 1]);
const ys = tf.tensor2d([1, 3, 5, 7], [4, 1]);

// Train the model using the data.
model.fit(xs, ys, { epochs: 10 }).then(() => {
  // Use the model to do inference on a data point the model hasn't seen before:
  // Open the browser devtools to see the output
  //   model.setWeights(model.getWeights());
  for (let i = 0; i < model.layers.length; i++) {
    model.layers[i].setWeights(model.layers[i].getWeights());
  }
  model.predict(tf.tensor2d([5], [1, 1])).print();
});

Error: Input Tensor ID not referenced
    at NodeJSKernelBackend.executeSingleOutput (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-node/dist/nodejs_kernel_backend.js:96:43)
    at NodeJSKernelBackend.matMul (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-node/dist/nodejs_kernel_backend.js:150:21)
    at environment_1.ENV.engine.runKernel.$a (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/ops/matmul.js:46:83)
    at /Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/engine.js:115:26
    at Engine.scopedRun (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/engine.js:95:23)
    at Engine.runKernel (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/engine.js:113:14)
    at matMul_ (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/ops/matmul.js:46:37)
    at Object.matMul (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-core/dist/ops/operation.js:23:29)
    at Object.dot (/Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-layers/dist/backend/tfjs_backend.js:215:24)
    at /Users/kyle/git-projects/hashback/tfjs-playground/node_modules/@tensorflow/tfjs-layers/dist/layers/core.js:142:28

Same error is thrown when using

model.setWeights(model.getWeights())

before calling predict

[bileschi]

Verified that this is reproducible in web:
tfjs: "0.12.4"
tfjs-converter: "0.5.5"
tfjs-core: "0.12.8"
tfjs-layers: "0.7.2"

[bileschi]

This only happens when setting the weights after fitting a model. Setting the weights before fitting the model works fine.

[bileschi]

This should be fixed with tfjs-layers commit #285
https://github.com/tensorflow/tfjs-layers/commit/934134f677f4464ab9287004a2f6b4b4edb9a039