You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@singa.apache.org by "Lee Chonho (JIRA)" <ji...@apache.org> on 2015/09/04 03:58:46 UTC
[jira] [Commented] (SINGA-65) Add an example of writing
user-defined layers
[ https://issues.apache.org/jira/browse/SINGA-65?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14730181#comment-14730181 ]
Lee Chonho commented on SINGA-65:
---------------------------------
Document should be revised accordingly.
For such example containing user-defined layers (and proto)
make sure to check the followings
- For layer type: built-in layer type is LayerType (integer), but user-defined layer type is string; and in job.conf, it is specified by user_type.
e.g.,
type: kMnist <---- kMnist is defined in job.proto
user_type: "kHidden" <---- "kHidden" is registered in main.cc
- To specify user-defined proto in job.conf, you need square-brackets
e.g.,
[singa.hidden_conf] <---- refer to myproto.proto
how to run
bin/singa-run.sh -exec examples/mlp/mlp.bin -conf examples/mlp/job.conf (or deep.conf)
> Add an example of writing user-defined layers
> ---------------------------------------------
>
> Key: SINGA-65
> URL: https://issues.apache.org/jira/browse/SINGA-65
> Project: Singa
> Issue Type: New Feature
> Reporter: wangwei
>
> SINGA's code has been modularized to support user extension.
> This ticket creates a full example of writing training using user defined layers.
> This example has the same neural net as the deep MLP example for digits recognition over the MNIST dataset. Instead of using the built-in InnerProductLayer and STanhLayer, we create a HiddenLayer as user-defined layer which applies inner-product and scaled Tanh transformations together.
> Makefile.example is provided to compile and link user code.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)