+1. It would be better to use implement it as a new layer that combines the conv and lstm operations here https://github.com/apache/incubator-singa/blob/master/python/singa/autograd.py
On Sun, May 5, 2019 at 3:29 AM Faouzi Amrouche (JIRA) <j...@apache.org> wrote: > Faouzi Amrouche created SINGA-452: > ------------------------------------- > > Summary: Adding the ConvLSTM layer implementation > Key: SINGA-452 > URL: https://issues.apache.org/jira/browse/SINGA-452 > Project: Singa > Issue Type: Improvement > Reporter: Faouzi Amrouche > > > ConvLSTM is an improved version of the LSTM layer. It replaces the > traditional matrix multiplications by convolutions. In this new > implementation, the input transformations and recurrent transformations are > both convolutional. > > This layer has become really important recently and is already supported > in many existing frameworks (Keras, Tensorflow...etc). > > > > -- > This message was sent by Atlassian JIRA > (v7.6.3#76005) >
