dritchie / adnn Goto Github PK

Modified version of Adam that tracks the relative change in the objective value to better adapt the learning rate:
https://arxiv.org/abs/1611.01505

I really love adnn. It's beautiful and powerful. Thank you.

I am trying to work out some example for adnn.

An simple nnXor.js example run successfully as following.

https://github.com/ccckmit/ai6/tree/master/book/nn/adnn

But for the nnMlp.js example, the network output fail to approximate the training data.

https://github.com/ccckmit/ai6/blob/master/book/nn/adnn/nnMlp.js

Do I use the wrong optimization method ?

I don't know why ?

If there are some simple example for user like me. It will be great !

In WebPPL, I've done something like this:

// Re-export Math constants etc.
Object.getOwnPropertyNames(Math).forEach(function(p) {
  if (!ad.scalar.hasOwnProperty(p)) {
    ad.scalar[p] = Math[p];
  }
});

(Which is similar to what happens in ad.js.)

Would you accept a PR adding something similar to functions.js, or would you prefer to handle it in some other way?

Hi,

how can I save a nn I've trained up to disk, and load into NodeJS again?

thanks in advance.