Cross validation about synaptic HOT 8 CLOSED

no, we're missing a cross-validation feature, but the Trainer has train and test methods, both take a dataset of the same shape. So you could split the dataset into a training set and a test set, run the train for N iterations and then run the test, that will return the error. Put all that within a while untill the error is smaller than your desired error and it should work.

from synaptic.

Is this something you would like to see implemented as a PR addition?

from synaptic.

Yes! that would be a handy feature that we're are missing right now

from synaptic.

I'm working on it
Currently I am adding a function to the Trainer Object prototype called trainSplit

1. Should this even be a function?
   1. Do you have a better name for the function? (I feel like it may use some help)
   2. I need at least 3 more options, should I add them to the Trainer object or make them separate parameters to my function.
   3. Should I let users to give options to the "inner" trainer I am using or should I use the defaults?
2. Should this be a separate object at all (ie not in the Trainer Object)?
3. I think it could be nice to add a few more options for training the data like KFold,Shuffle and Split etc... How do you see them being incorporated or do you want to keep things simplistic for now?

from synaptic.

mm I what do you think about adding this as another option for the Trainer#train(dataset:Array, options:Object) method? Like crossValidate: { testSize: .3, testError: .01 }, that would take a 30% of the dataset for validation, and the trainer would iterate until the error of the tests go below .01 .. in the end it would be like another break condition for the training, same as error and iterations.

And yea, I think it would be great to add more options for training, but I don't know what do you mean by shuffle tho, we already have a shuffle flag in the options, that shuffle the whole dataset after each iteration, is that what you meant?

from synaptic.

I crated a pull request (#61).
BTW, just as a thought experiment. Maybe we are making our model over-fit to the testSet?

K-means can be the next PR if this gets merged I guess.

I had another Idea to create some kind of plotting tool. When I created my tests it was really useful to plot the inputs and their corresponding outputs that the network made - do you think this is something that is adequate for this repo?

from synaptic.

Hey, I just reviewed and merged your PR. About the plotting tool, I think that would be a great addition, we have support for a very basic DOT lang plotting so far, using Network#toDot(), that returns the DOT lang code and a link to google charts, with a basic plot of your network's layers. Maybe you can use that or expand it if you wish, idk, any contributions are welcome (:

from synaptic.

Fixed the tests and the options bug in #64. I'll look into implementing a plotting tool :)

from synaptic.