[jira] [Commented] (SINGA-324) Extend RNN layer to accept variant seq length across batches
[ https://issues.apache.org/jira/browse/SINGA-324?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16068193#comment-16068193 ] ASF subversion and git services commented on SINGA-324: --- Commit 2ce7229ad688f54b7fb88fe46c23e771dfb87365 in incubator-singa's branch refs/heads/master from [~foxconnwang...@163.com] [ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=2ce7229 ] SINGA-324 Extend RNN layer to accept variant seq length across batches The cudnn rnn layer is updated to handle mini-batches with different seq lengths. The internal data structures are re-allocated if max_length_ < seq_length_ (the longest sample of the current mini-batch). > Extend RNN layer to accept variant seq length across batches > > > Key: SINGA-324 > URL: https://issues.apache.org/jira/browse/SINGA-324 > Project: Singa > Issue Type: Improvement >Reporter: wangwei > > The current RNN layer can accept samples with different length within one > mini-batch, however, it assumes that the longest samples of each mini-batch > is the same, i.e. seq_length is fixed. > To make it more flexible, this ticket will extend the RNN layer to accept > mini-batches with different longest samples. > In particular, we will set two variables, max_length_ and seq_length_, where > the first one is for the max seq length over all mini-batches, and the > seq_length_ is the effective seq length for the current mini-batch, which is > updated based on the input data for every iteration (mini-batch). The > max_length_ is updated when max_length_ < seq_length_. -- This message was sent by Atlassian JIRA (v6.4.14#64029)
[jira] [Commented] (SINGA-324) Extend RNN layer to accept variant seq length across batches
[ https://issues.apache.org/jira/browse/SINGA-324?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16068194#comment-16068194 ] ASF subversion and git services commented on SINGA-324: --- Commit 9eea5b53f25319c0649138fc3059253f7c2efa38 in incubator-singa's branch refs/heads/master from [~zhaojing] [ https://git-wip-us.apache.org/repos/asf?p=incubator-singa.git;h=9eea5b5 ] Merge branch 'SINGA-324' > Extend RNN layer to accept variant seq length across batches > > > Key: SINGA-324 > URL: https://issues.apache.org/jira/browse/SINGA-324 > Project: Singa > Issue Type: Improvement >Reporter: wangwei > > The current RNN layer can accept samples with different length within one > mini-batch, however, it assumes that the longest samples of each mini-batch > is the same, i.e. seq_length is fixed. > To make it more flexible, this ticket will extend the RNN layer to accept > mini-batches with different longest samples. > In particular, we will set two variables, max_length_ and seq_length_, where > the first one is for the max seq length over all mini-batches, and the > seq_length_ is the effective seq length for the current mini-batch, which is > updated based on the input data for every iteration (mini-batch). The > max_length_ is updated when max_length_ < seq_length_. -- This message was sent by Atlassian JIRA (v6.4.14#64029)
[jira] [Created] (SINGA-326) Add Inception V4 for ImageNet classification
wangwei created SINGA-326: - Summary: Add Inception V4 for ImageNet classification Key: SINGA-326 URL: https://issues.apache.org/jira/browse/SINGA-326 Project: Singa Issue Type: New Feature Reporter: wangwei Assignee: wangwei The [Inception V4 |http://arxiv.org/abs/1602.07261] model has much better performance than GoogleNet. We convert the parameters pre-trained from Tensorflow into pickle dictionary and load it into the net created using SINGA FeedForwardNet. -- This message was sent by Atlassian JIRA (v6.4.14#64029)