Access Tensorflow.js Layers Properties Inside Model

Question

I am using a pretrained (in Python) Keras model by loading it into JavaScript with tensorflow.js and using the tensorflow.js library. This model has a GaussianNoise input layer, and a GaussianDropout layer, with respective properties stddev and rate.

After loading the model, say:

let model;
(async function () {
    model = await tf.loadLayersModel("TensorFlowModels/dnn_fscav.json");
})();

In the same path I keep the weights in a .bin file. Now, my problem is that I need to change the stddev of the GaussianNoise layer and the rate of the GaussianDropout after loading the model, ideally directly from the model. I have looked into the model tensorflow object and its layers, but I could not find these properties.

I also know I can directly change these properties from the JSON file, but that means reading the JSON, modify it, and then read it with tensorflow.. which is not ideal. I have also looked at the tensorflow.js API reference, but found no explanation on how to do this.

Does anyone know a cleaner way to change layers properties from a loaded model?

OMS OMS · Accepted Answer · 2021-01-03T11:43:54

In case this helps anyone, after looking closer to the Tensorflow model object, I realised it has a circular structure. That meaning, some object properties call other object properties, which at the same time call the previous object properties. I could find the layer properties in the following object properties:

model.layers[layer_number].outboundNodes[0].outboundLayer.property = new_value

In my case, property is stddev for GaussianNoise layer of the neural network and rate for the GaussianDropout layer. By changing that property to a new value and saving the model, I could confirm that the properties of the layers are changed. I guess you can also access the properties from any of the other circular features in the object, as they all point to the same value.

Access Tensorflow.js Layers Properties Inside Model

1 Answers