incubator-cvs mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Apache Wiki <wikidi...@apache.org>
Subject [Incubator Wiki] Update of "SingaProposal" by ThejasNair
Date Wed, 28 Jan 2015 18:54:57 GMT
Dear Wiki user,

You have subscribed to a wiki page or wiki category on "Incubator Wiki" for change notification.

The "SingaProposal" page has been changed by ThejasNair:
https://wiki.apache.org/incubator/SingaProposal?action=diff&rev1=4&rev2=5

Comment:
remove H20 reference, formatting

  == Initial Goals ==
  We have developed a system for SINGA running on a commodity computer
  cluster. The initial goals include,
+  * improving the system in terms of scalability and efficiency, e.g., using Infiniband for
network communication and multi-threading for one node computation. We would consider extending
SINGA to GPU clusters later.
- * improving the system in terms of scalability and efficiency, e.g., using
- Infiniband for network communication and multi-threading for one node computation.
- We would consider extending SINGA to GPU clusters later.
- * benchmarking with larger datasets (hundreds of millions of training instances)
+  * benchmarking with larger datasets (hundreds of millions of training instances) and models
(billions of parameters).
- and models (billions of parameters).
- * adding more built-in deep learning models. Users can train the built-in models
+  * adding more built-in deep learning models. Users can train the built-in models on their
datasets directly.
- on their datasets directly.
  
  
  == Current Status ==
@@ -150, +146 @@

  Furthermore, the need for scalability for such a platform is obvious.
  
  === Relationship with Other Apache Products ===
- Apache H2O implemented two simple deep learning models, namely the Multi-Layer
- Perceptron and Deep Auto-encoders. There are two significant differences between
- H2O and SINGA. First, H2O adopts the Map-Reduce framework which runs a set of
- computing nodes in parallel againsts of the training set. Model parameters
- trained by all computing nodes are averaged as the final model parameters. This
- training algorithm is different from the distributed training algorithm used by
- DistBelief, Adam and SINGA, which frequently synchronizes the parameters trained
- from different nodes. SINGA adopts the parameter server framework to support a wide
- range of distributed training algorithms and parallelization methods (e.g., data
- parallelism, model parallelism and hybrid parallelism. H2O only support data
- parallelism) . Second, in H2O, users are restricted to use the two built-in models.
- In SINGA, we provide simple programming model to let users implement their own
- deep learning models. A new deep learning model can be implemented by customizing
- the base Layer class for each layer involved in the model. It is similar to
- writing Hadoop programs where users only need to override the base Mapper and
- Reducer. We also provide built-in models for users to use directly. 
  
  == Documentation ==
  The project is hosted at http://www.comp.nus.edu.sg/~dbsystem/project/singa.html.

---------------------------------------------------------------------
To unsubscribe, e-mail: cvs-unsubscribe@incubator.apache.org
For additional commands, e-mail: cvs-help@incubator.apache.org


Mime
View raw message