Where communities thrive

Join over 1.5M+ people
Join over 100K+ communities
Free without limits
Create your own community

Explore more communities

ThoughtWorksInc/DeepLearning.scala

A DSL to create very complicated neural networks

People

Repo info

Activity

Sep 15 00:39

scala-steward closed #273
Sep 15 00:39

scala-steward commented #273
Sep 15 00:39

scala-steward opened #278
Sep 10 01:31

scala-steward closed #276
Sep 10 01:31

scala-steward commented #276
Sep 10 01:31

scala-steward opened #277
Jul 26 17:43

scala-steward closed #275
Jul 26 17:43

scala-steward commented #275
Jul 26 17:43

scala-steward opened #276
Jul 19 08:04

scala-steward opened #275
Jun 16 19:03

scala-steward closed #268
Jun 16 19:03

scala-steward commented #268
Jun 16 19:03

scala-steward opened #274
May 28 13:14

scala-steward closed #267
May 28 13:14

scala-steward commented #267
May 28 13:14

scala-steward opened #273
May 17 00:56

scala-steward closed #271
May 17 00:56

scala-steward commented #271
May 17 00:56

scala-steward opened #272
May 09 16:44

scala-steward closed #270

SemanticBeeng

@SemanticBeeng

I do not know how automatic differentiation works (or is different) on multiple variables as opposed to on one - was hoping you would know and know how the DL4S engines supports this

Will dig further then

Yang, Bo

@Atry

I don't know what is DL4S

SemanticBeeng

@SemanticBeeng

sorry DeepLearning.scala

"Madness: a package for Multivariate Automatic Differentiation" : https://cran.r-project.org/web/packages/madness/vignettes/introducing_madness.pdf

an example of impl explicitly built for multiple variables

anyway, will dig

Yang, Bo

@Atry

https://static.javadoc.io/com.thoughtworks.deeplearning/deeplearning_2.11/2.0.1/com/thoughtworks/deeplearning/DeepLearning.html

Common differentiable types that supports DeepLearning are:
Float, FloatWeight or FloatLayer
Double, DoubleWeight or DoubleLayer
INDArray, INDArrayWeight or INDArrayLayer

SemanticBeeng

@SemanticBeeng

Aha, thanks

SemanticBeeng

@SemanticBeeng

So the question really is how/if nd4j supports multivariate/vector calculus

Yang, Bo

@Atry

nd4j is not differentiable. You may create plugins of new differentiable layers from raw nd4j operators. See CNN plugin for example.

SemanticBeeng

@SemanticBeeng

the gist of my ignorance is about the differences between how automatic differentiation works with a single variable vs a vector of variables

was hoping to find a code example here

but will have to digg deeper to learn. thanks

Yang, Bo

@Atry

Actually no difference here.

SemanticBeeng

@SemanticBeeng

thank you, good to know.

Koen Dejonghe

@koen-dejonghe

Hi, I created Scorch, a deep learning framework in Scala inspired by Pytorch. It has automatic differentiation built in and follows an imperative coding style. Scorch lives here: https://github.com/botkop/scorch. I value your feedback. Thank you.

Yang, Bo

@Atry

@koen-dejonghe It seems too complicated to me

Koen Dejonghe

@koen-dejonghe

@Atry Thank you for the feedback. Could you give me an indication or an example of what exactly you think is too complex? Much appreciated.

Yang, Bo

@Atry

@koen-dejonghe The Module/subModules concepts are implementation detail. A deep learning framework should be able to create computational graph from arbitrary differentiable expression, instead of forcing useres explicitly creating them.

We actually has concepts similar to Module/subModules, in pre-1.0 version of DeepLearning.scala.

Koen Dejonghe

@koen-dejonghe

@Atry You don't need a Module to create a computation graph. All you need are Variables and Functions, and you can back prop in them. See https://github.com/botkop/scorch/blob/master/src/test/scala/scorch/autograd/AutoGradSpec.scala for examples.

Modules are just a convenience class for collecting parameters and wrapping Functions.

Yang, Bo

@Atry

Ah, I see

The pytorch-like API looks more concise than Deeplearning4j. Great work!

Koen Dejonghe

@koen-dejonghe

Thanks.

DL4J, like TF, has a symbolic programming style, while pyTorch/scorch have an imperative coding style.

Which I think is much more intuitive

mghildiy

@mghildiy

I cloned Compute.scala

I try to import in Intellij, but got issues

https://stackoverflow.com/questions/49595341/intellij-issue-while-importing-sbt-project

Scala plugin version is:2017.2.13

Yang, Bo

@Atry

Try to upgrade?

mghildiy

@mghildiy

Upgrade what?

I just upgraded IntelliJ

Yang, Bo

@Atry

Good

mghildiy

@mghildiy

and Scala plugin too is latest one, as shown by Intellij

but still not able to import the project

Yang, Bo

@Atry

The latest Intellij should be 2018.x

mghildiy

@mghildiy

Mine is:2018.1

Yang, Bo

@Atry

You can reopen https://youtrack.jetbrains.com/issue/SCL-12901 if it still does not work in 2018.1

mghildiy

@mghildiy

Anthony Platanios

@eaplatanios

I don't know if this project is still alive, but if it is, would you guys be interested to add support for TensorFlow Scala?

doofin

@doofin

what kind of reverse mode ad does this project use?Is it implemented with tape based reverse ad?

Yang, Bo

@Atry

Similar to that but not actual linear tape

Checkout https://deeplearning.thoughtworks.school/assets/paper.pdf @doofin

doofin

@doofin

thanks!

doofin

@doofin

Hi,are there up to date sbt tutorial ? https://github.com/izhangzhihao/deeplearning-tutorial is too old