yixuan
partial_fit method?
Hi Yixuan. Thanks for starting this.
What do you think about partial_fit fit method (similar to scikit-learn API)? Since we are fitting neural network with SGD, it is natural to extend it to out-of-core computation. While fit will be "default" method, partial_fit can be more low level method - i.e. manually decrease learning rate between epochs/iterations, make some checks, etc.
And in general I'm thinking to create unified API for several types of R models (mostly inspired by scikit-learn). Mb you can be interested and can share your experience (we can discuss by email).
I see. I'll take a look then.
Just to mention that tinydnn is more like an experimental package, since its performance is to some degrees lower than other finely tuned DNN frameworks such as MXNet. I'm not sure whether tinydnn will finally be published as a "formal" package or not.
But on other side for small networks it should be enough (especially for not convolutional networks). And simplicity of installation is a big advantage.
Not sure I understand what do you mean. Better in which sense? Gluon is python library, isn't it?
вт, 10 июл. 2018 г., 13:03 HarryZhu [email protected]:
it seems to migrate gluon to R is a better solution than this package.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/yixuan/tinydnn/issues/2#issuecomment-403752970, or mute the thread https://github.com/notifications/unsubscribe-auth/AE4u3eYNzv5-I36ReyaL_cnjPhuJbkeRks5uFG3hgaJpZM4MMR3T .