How to genrate .pb model from saved checkpoint

DeepSpeech
#1

Hello Everyone,
I am new to this community,
I trained Deepspeech model on Tamil Dataset from common voice.
Want to know that how to genrate an infrence ready model from latest checkpoint saved.
This is the command i use for training

./DeepSpeech.py --train_files ./data/ta/clips/train.csv --dev_files ./data/ta/clips/dev.csv --test_files ./data/ta/clips/test.csv --checkpoint_dir ./data/checkpoint --export_dir ./data/export_model --n_hidden 2048 --learning_rate 0.001 --automatic_mixed_precision=True --epochs 9

And these are the saved checkpoints dir.

image
image728×200 23.6 KB

Please Help!

#2

What’s in the export_dir?

1 Like
#3

Thanks for the rply @baconator
export_dir is empty. I deleted model earlier. After that it is empty.
Question: If i run training command and give epoch=1 then it could genrate the model in export_dir?

#4

Check the forum, but I remember that you can simply run a test with a checkpoint and it will generate a model in the export dir. No need to train.

#5

So you simply leave out the train and dev flags

#7

I run the test from checkpoints and it create the model in export dir, but the seems to failed as res string is empty. But when i train, loss is like max to 2 or 3 , What would be the reason for this empty string.

image

Please help!

#8

Could it be that you didn’t build a Tamil language model and scorer? Search the forum, I remember Tamil questions some time ago.

1 Like
#9

I trained this model on the common voice tamil dataset as listed

https://deepspeech.readthedocs.io/en/v0.7.0/TRAINING.html

here.

#10

Those instructions are focused on the importer side of things, and they’re written for English. For training on non-English data you need to create your own external scorer or you need to disable it entirely (pass an empty scorer path: --scorer_path '').

2 Likes
#11

Thanks @reuben For reply.
It solve the problem, and results are looking fine.

I just add --scorer_path to a empty string.

image

#12

There is one more query,
Do i have to train another model for other language or I can use this model to use both the language?

And I am quite at the beginning, So any one can suggest a book so i can clear my concepts.