Load trainability status of layers when importing them from JSON configuration

[mkaze]

Similar to #107, the trainability status of a layer is not taken into account when importing it from JSON configuration, e.g. if a layer is freezed in the saved JSON configuration it might be loaded as trainable (depending on the default value of isTrainable property for that layer).

To fix this, either each of the layer creator functions in ModelLoader.kt should be modified to set the isTrainable property of the created layer with value of config.trainable; or alternatively, this could be done in a single place like convertToLayer function for all layers. Further, one should take into account this issue when implementing a fix.

[zaleslaw]

I agree that it could be implemented now, but in reality, we need to add more check/IllegalStateExceptions if the layer implementing the NoGradients layer has an isTrainable == true status.

The idea is the following: layers with NoGradient marker interface should have this status as false.

[dosier]

I can do this one

[dosier]

I agree that it could be implemented now, but in reality, we need to add more check/IllegalStateExceptions if the layer implementing the NoGradients layer has an isTrainable == true status.

The idea is the following: layers with NoGradient marker interface should have this status as false.

Should these checks u mention for NoGradients layers be applied only to these, or to other layers as well. Because some layers do set isTrainable to false but do not implement NoGradients. So should these other layers implement NoGradients, or should we make another interface that every non trainable layers extends?

[dosier]

Yo @zaleslaw or @mkaze what do ya think about something like this:
dosier@8e41418

I don't really like how the default value is true but can't nicely change it to something like:

    public var isTrainable: Boolean = this is NoGradients

without leaking this in the constructor.

I see three options:

• keep it like in the commit I made
• leak this in constructor of Layer class
• add isTrainable as constructor param

But maybe one of your has a better idea (or anyone else!!)