Added PerDimWeightedAverage component for chain models in SWBD, works…

[vijayaditya]

… as good as jesus (tdnn_4f) model

[vijayaditya]

I have created a new config generator for TDNNs similar to those available for LSTM. This will not affect any existing TDDNN/Jesus scripts.

[danpovey]

Why do you need a separate train_tdnn.sh script for this? Can't you use the same pattern as for the jesus-configs, to create the configs?

[danpovey]

It's the duplication of code that I'm not crazy about, and the name of the
script train_tdnn_b.sh. Why couldn't you just slightly extend the existing
train_tdnn.sh script?
BTW, did the low-pass filtering help?
Dan

On Wed, Jan 20, 2016 at 9:41 PM, Vijayaditya Peddinti <
notifications@github.com> wrote:

I have created a new config generator for TDNNs similar to those available
for LSTM. This will not affect any existing TDDNN/Jesus scripts.

—
Reply to this email directly or view it on GitHub
#461 (comment).

[vijayaditya]

the only thing different in the new train_tdnn.sh script is the make_configs.py which is being called. This new steps/nnet3/tdnn/make_configs.py script makes use of components defined in steps/nnet3/components.py

[vijayaditya]

Low-pass filtering didn't help but I would like to keep it as it was just overtraining a lot and I hope it will improve after l2 regularization. Will start this experiment tomorrow.

[danpovey]

OK.
Better to just make changes to train_tdnn.sh than to duplicate the script.
Dan

On Wed, Jan 20, 2016 at 9:47 PM, Vijayaditya Peddinti <
notifications@github.com> wrote:

Low-pass filtering didn't help but I would like to keep it as it was just
overtraining a lot and I hope it will improve after l2 regularization. Will
start this experiment tomorrow.

—
Reply to this email directly or view it on GitHub
#461 (comment).

[vijayaditya]

OK, will do.

[vijayaditya]

just realized I am using a tracking branch for pushing this commit. so I will close this and recreate a pull request.