[
https://issues.apache.org/jira/browse/SINGA-193?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15367221#comment-15367221
]
ASF subversion and git services commented on SINGA-193:
-------------------------------------------------------
Commit 790b7b4cdec305356ce5e18ccf153832d7d25c13 in incubator-singa's branch
refs/heads/dev from [~flytosky]
[ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=790b7b4 ]
SINGA-193 Add Python layers
Add python classes for all layers implemented in src/model/layer/.
Update factory for layer registration, new layers can be registered using the
help function in layer.h; Add gtest for CreateLayer.
Update cmake file to include the .cc file for each layer as python_cxx
which are necessary to get all registered layers by the swig code.
Swig can automatically convert python list into vector for function args, and
convert vector to python tuple.
Update CreateLayer to return shared_ptr<Layer>; Add a function to list all
registered layers.
Add py version of cifar10 alexnet model in alexnet.py; update download_data.py
for the py version data.
Add the FeedForwardNet python class
Update py layer constructor to make name as a required field of each layer's
constructor.
Create symbolic link (in build/python) to python source files, not copy
Replace Tensor* to Tensor (for cpp functions) to make it easier for creating
python functions.
Otherwise, py Tensor needs to carry both cpp Tensor* and cpp Tensor,
which is ambiguous and hard to maintain.
Add py loss, metric
Fix bugs from default arg for W_specs and b_specs; cannot set the default value
for args of type dict.
Otherwise the same args would be used by later functions.
running foodlg training using vgg model
Add net::save() and net::load() for params
Add predict.py
Add random shuffle
clean code and pass tests
> Add Python layers
> -----------------
>
> Key: SINGA-193
> URL: https://issues.apache.org/jira/browse/SINGA-193
> Project: Singa
> Issue Type: New Feature
> Reporter: wangwei
> Assignee: wangwei
>
> This ticket creates the python layers for V1, which differ to the python
> layers of v0,3 in terms of the layer construction.
> The python layers should provide intuitive APIs for uses to configure the
> layer. Keras API would be a reference.
> We need to consider the param specs in the layer creation which is missing in
> Keras APIs.
> The python layer would be uniform for cudnn, opencl and cpp. A backend (e.g.,
> 'cudnn', 'cuda') field would be used to create the C++ layer, e.g.,
> cudnnPooling or openclPooling.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
