[GitHub] SumNeuron opened a new issue #8806: Not all LSTM layout options are specified

[GitBox <gi...@apache.org>]

SumNeuron opened a new issue #8806: Not all LSTM layout options are specified
URL: https://github.com/apache/incubator-mxnet/issues/8806
 
 
   In the docs for the [LSTM layer](https://mxnet.incubator.apache.org/api/python/gluon.html?highlight=lstm#mxnet.gluon.rnn.LSTM) it states that one can specify the layout with the following options:
   
   - T = `Sequence Length`
   - N = `Batch Size`
   - C = `Feature Dimensions`
   
   and the default is `TNC`.
   
   The other accepted format is `NTC` (batch x seq x feat). 
   
   This is an issue for two reasons:
   
   1.  it is not stated in the auto-generated docs. Mostly this is because of the hard-coded assert statement:
   
   ```
   assert layout == 'TNC' or layout == 'NTC', \
                "Invalid layout %s; must be one of ['TNC' or 'NTC']"%layout
   ```
   
   which is why this is not being picked up by the auto-doc generation. However this should be there.
   
   2. because there are `3!` ways for a user to order the characters 'T', 'N', and 'C':
   
   - TNC (seq x batch x feat)
   - TCN (seq x batch x feat)
   - NTC (batch x seq x feat)
   - NCT (batch x feat x seq)
   - CNT (feat x batch x seq)
   - CTN (feat x seq x batch)
   
   While some of these  formats are more likely to be used than others, I would at least expect `NCT` to be implemented, as it parallels 1D time series (batch, channels, values over time).
   
   In short, I would say for maximum utility that all implementations should be implemented, however I would like to place priority on NTC.
   
   Also maybe add a drop-down element to the docs to show assertions (so viable options are known)
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services