Support for storing models from N last epochs and decoding with those models.

[curly1]

User can define number of models to store from n_mdls_store last epochs, controlled with n_mdls_store option in a config file. After training, user can decide to decode with the model stored prior to the final iteration, similarly to the iter option in Kaldi decoding scripts. This can be done with ep_to_decode option. Epoch indexing starts from 0, so to match e.g. the res.res file, here indexing also starts from 0. This means that if the training is for 5 epochs, and we'll set ep_to_decode to 4, the final model will be used. If the ep_to_decode option is left empty, the final model will be used. A separate decoding directory will be created for each ep_to_decode.
Maybe changing epoch indexing to start from 1 throughout the code would be less confusing?

[mravanelli]

Hi Joanna, thank you for your pull request. I think the modification you propose is interesting and can give users more flexibility. This modification requires a change in the config files, but I would like to keep them frozen until the Interspeech deadline (many people are working with the current config file, and it' better to avoid any incompatibility with past and new config files at least before the deadline). I can review your pull request now and merge it after that! Thank you very much for your contribution! Mirco

…

On Mon, 4 Mar 2019 at 04:37, Joanna Rownicka ***@***.***> wrote: User can define number of models to store from n_mdls_store last epochs, controlled with n_mdls_store option in a config file. After training, user can decide to decode with the model stored prior to the final iteration, similarly to the iter option in Kaldi decoding scripts. This can be done with ep_to_decode option. Epoch indexing starts from 0, so to match e.g. the res.res file, here indexing also starts from 0. This means that if the training is for 5 epochs, and we'll set ep_to_decode to 4, the final model will be used. If the ep_to_decode option is left empty, the final model will be used. A separate decoding directory will be created for each ep_to_decode. Maybe changing epoch indexing to start from 1 throughout the code would be less confusing? ------------------------------ You can view, comment on, or merge this pull request online at: #61 Commit Summary - Support added for storing a specified number of models from N last epochs and decoding with those models. - Update to latest pytorch-kaldi File Changes - *M* cfg/TIMIT_baselines/TIMIT_MLP_fmllr.cfg <https://github.com/mravanelli/pytorch-kaldi/pull/61/files#diff-0> (7) - *M* run_exp.py <https://github.com/mravanelli/pytorch-kaldi/pull/61/files#diff-1> (74) - *M* utils.py <https://github.com/mravanelli/pytorch-kaldi/pull/61/files#diff-2> (132) Patch Links: - https://github.com/mravanelli/pytorch-kaldi/pull/61.patch - https://github.com/mravanelli/pytorch-kaldi/pull/61.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#61>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQGs1gpnJd9lQgqcL8Cctig6FrxWH5__ks5vTOlogaJpZM4bb7R_> .

[curly1]

OK, sure, thank you.

[Mistobaan]

Ping for this Pr. Is the Interspeech over?

[mravanelli]

Yes, but we are planning to significantly expand this project in the next months with several other functionalities (including some of this pull request). Best, Mirco

…

On Tue, 13 Aug 2019 at 16:50, Fabrizio Milo ***@***.***> wrote: Ping for this Pr. Is the Interspeech over? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#61?email_source=notifications&email_token=AEA2ZVW3LZNHAKPY6BQ4YBLQEMNAHA5CNFSM4G3PWR72YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4G6FIY#issuecomment-521003683>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEA2ZVRD3S33HDC2PLR7TZ3QEMNAHANCNFSM4G3PWR7Q> .

[TParcollet]

HI, we are sorry for the delay. Lot of stuffs have changed since then and few conflicts exist. If you are ok and have time, could you make a new PR with the current PyTorch-Kaldi version for this feature? Please do not make this field mandatory on the config file so we don't have to refractor every configuration file.