wangwei created SINGA-162:
-----------------------------
Summary: Overview of features for V1.x
Key: SINGA-162
URL: https://issues.apache.org/jira/browse/SINGA-162
Project: Singa
Issue Type: Wish
Reporter: wangwei
This ticket gives an overview of the features to be developed for V1.x.
First, we will implement a set of core abstractions,
1. Tensor, which provides basic linear algebra operations (e.g., addition) and
neural net specific operations (e.g., conv). It is a finer abstraction than
Layer in V0.x, and thus could be able to support a wider range of applications.
[Autograd|https://github.com/HIPS/autograd] would also be implemented.
2. Device, which abstract the execution and memory allocation for Tensor using
different hardware/software, including Nvidia GPU (with Cuda/Cudnn) and other
GPUs using OpenCL.
3. Scheduler, which maximizes the parallelism of executions.
4. Memory manager, which manages a memory pool for a device, for garbage
collection, and optimization.
Second, on top of these core abstractions, we will develop a set of modules
specific for neural networks
1. Layer for feature transformation, e.g., conv and pool
2. Model for typical models including feed-forward, RNN and energy models.
3. Updater for updating parameters on single node or in a distributed
environment.
Third, some utility modules would be implemented for IO/Log/Network.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
