Checkpoints and frozen model in fine tuning

[JRMeyer]

>>> halake.kumar22
[September 6, 2019, 11:36am]

I am going through the
documentation
to retrain existing deepspeech-0.5.1-models with my own data, where only
checkpoints are used to retrain the model. Here I downloaded the
checkpoints directory and the code looks as below.

> python3 -u DeepSpeech.py slash
> --train_files '/home/dev_ds/deepspeech_dir/corpus/corpus-train.csv' slash
> --dev_files '/home/dev_ds/deepspeech_dir/corpus/corpus-dev.csv' slash
> --test_files '/home/dev_ds/deepspeech_dir/corpus/corpus-test.csv' slash
> --alphabet_config_path '/home/dev_ds/deepspeech_dir/deepspeech-0.5.1-models/alphabet.txt' slash
> --lm_binary_path '/home/dev_ds/deepspeech_dir/my-model/lm.binary' slash
> --lm_trie_path '/home/dev_ds/deepspeech_dir/my-model/trie' slash
> --checkpoint_dir /home/dev_ds/deepspeech_dir/deepspeech-0.5.1-checkpoint/ slash
> --train_batch_size 2 slash
> --learning_rate 0.000001 slash
> --export_dir '/home/dev_ds/deepspeech_dir/my-model/'

I have two questions

1. Should I use checkpoints (--checkpoint_dir) as well as frozen
graph ( --initialize_from_frozen_model) from existing model to
adapt the model to get better
accuracy? slash [1 slash ]

2. What is the difference between --checkpoint_dir and
--source_model_checkpoint_dir? slash [2 slash ]

[This is an archived TTS discussion thread from discourse.mozilla.org/t/checkpoints-and-frozen-model-in-fine-tuning]

[JRMeyer]

>>> reuben
[September 6, 2019, 11:50am]

The documentation you linked to is for master, if you're using v0.5.1
you should look at the documentation for v0.5.1:
https://github.com/mozilla/DeepSpeech/tree/v0.5.1



> Should I use checkpoints ( --checkpoint_dir ) as well as frozen
> graph ( --initialize_from_frozen_model ) from existing model to
> adapt the model to get better
> accuracy? slash [1

--initialize_from_frozen_model no longer exists in v0.5.1.



> What is the difference between --checkpoint_dir and
> --source_model_checkpoint_dir
> ? slash [2 slash ]

--source_model_checkpoint_dir does not exist in v0.5.1. I think that's
a flag from the transfer learning branches? So it's only useful if
you're doing transfer to a different alphabet, and requires checking out
that branch.

[Archived Post]