Is there a way to access weights and biases in tfjs layers api?

[Nico Boh]

Hello,

is there a way to access weights and biases of a layer in tfjs with the layers api? To be more precisely I want to get and also set them to use a generic algorithm on my neural net.

If not, is it possible with just tfjs core?

[Shanqing Cai]

The answer is yes.

Suppose you have defined your model as follows:

const model = tf.sequential();

model.add(tf.layers.dense({units: 2, inputShape: [10]}));

model.add(tf.layers.dense({units: 4}));

model.compile({loss: 'categoricalCrossentropy', optimizer: 'sgd'});

To access the weights (i.e., kernel and bias) of the first dense layer, for example:

// kernel:

model.layers[0].getWeights()[0].print()

// bias:

model.layers[0].getWeights()[1].print()

Likewise for the second dense layer.

Also note that this workflow is the same as Python Keras.

[Shanqing Cai]

Sorry - overlooked the second part of your question. To set weights, use the setWeights() method. For example:

const model = tf.sequential();

model.add(tf.layers.dense({units: 2, inputShape: [10]}));

model.add(tf.layers.dense({units: 4}));

model.compile({loss: 'categoricalCrossentropy', optimizer: 'sgd'});

// To set the weights (kernels and bias) of the first dense layer to all ones:

model.layers[0].setWeights([tf.ones([10, 2]), tf.ones([2])]);

[Nico Boh]

Ahh thank you so much <3

[Tiago Vasconcelos]

Can this method be used to save the model? I'd like to train the model and then save for re use without having to do training every time...

[Nikhil Thorat]

See this issue which is high-priority: https://github.com/tensorflow/tfjs/issues/13

--
You received this message because you are subscribed to the Google Groups "TensorFlow.js Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfjs+unsubscribe@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tfjs/.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfjs/6be5e49e-a964-4326-b23f-22957714dd2f%40tensorflow.org.

[tensorflow.r...@gmail.com]

This is great! I have spent one day by trying with the method get_weight() before posting the question in the other thread where you have replied to refer to this post.

now, How to know that such method already exists? I referred the API documentation and this method does not seem to be there in the documentation. Am I missing something very trivial

Thank you!

[Jeremy Ellis]

Any hints on how to manipulate getWeights so that it can be imported into setWeights. 

[Jeremy Ellis]

      model.layers[0].setWeights(model.layers[0].getWeights()) 

The above code works, but not after I have put getWeights into a textarea.  JSON.stringify  JSON.parse,  split() don't seem to help. Any suggestions for getting getWeights back from a string into a usable form for setWeights?

[Jeremy Ellis]

Looks like this code does work to convert getWeights back from a textarea string into something setWeights can work with.

      model.layers[0].setWeights(model.layers[0].getWeights(document.getElementById('myArea01').value.split('\r\n|\n'))) 

 

[Jeremy Ellis]

Sorry for posting so many times, but I am really impressed that this bit of code works to load your entire model from a textarea not just a layer at a time. 

      model.setWeights(model.getWeights(document.getElementById('myArea01').value.split('\r\n|\n'))) 

[Jeremy Ellis]

oops. The above 2 posts are totally wrong. Still stuck on how to convert getWeights to a string and then back into setWeights.

[Shanqing Cai]

The work on saving models including their weights from the browser is underway and being tracked by

https://github.com/tensorflow/tfjs/issues/13

Depending on your schedule, you might want to wait for the features to land instead of writing your

own saving logic.

[Jeremy Ellis]

Got it working at:

https://hpssjellis.github.io/beginner-tensorflowjs-examples-in-javascript/beginner-examples/tfjs11-load-weights.html

Needed to use a  JSON weights object