Unable to load Trained Model (LSTMTimeStep) from JSON File

[albertp89]

What is wrong?

I trained a model and saved to a file using
JSON.stringify(net.toJSON());

I've confirmed the object is properly read in and loaded.
When trying to run net.run(someData) it gives this error
UnhandledPromiseRejectionWarning: TypeError: Cannot read property 'runInput' of undefined

Maybe I'm doing something wrong...

Where does it happen?

When trying to load a trained model from a JSON file of type LSTMTimeStep

How do we replicate the issue?

Train a model using LSTMT, JSONify it and try to load it back in.
Can see console output running the following JSFiddle
JSFiddle: http://jsfiddle.net/usrbjkt5/18

How important is this (1-5)?

3

Expected behavior (i.e. solution)

Model is properly loaded and able to be run.

Other Comments

[mubaidr]

Thanks for providing the jsfiddle, we are looking into this issue.

[robertleeplummerjr]

The model loads, but the prediction isn't long enough:

var net = new brain.recurrent.LSTMTimeStep();

net.train([
  [1, 2, 3]
]);

var output = net.run([1, 2]);  // 3
console.log(output);
var model = net.toJSON();

net = new brain.recurrent.LSTMTimeStep();
net.fromJSON(JSON.parse(JSON.stringify(model)));
console.log(net.run([1, 2], 3));

Works fine ^. This 3 is the length of the input + the output length. I think this is a bug that needs solved.

[mrorigo]

I was stuck on this, then found the keepNetworkIntact: true option for learn()

[robertleeplummerjr]

keepNetworkIntact is now removed, and this should just work as of now.