How to export pretrained Nvidia's tao model .tlt to .riva?

AI & Data Science Deep Learning (Training & Inference) Riva
1

Hey all,

I am working on Nvidia’s pretrained TAO model. Only .tlt file is given at NGC. I am trying to convert this .tlt to .riva using commands below (reference link)-

tao speech_to_text_citrinet export -e <experiment_spec>
-m <model_checkpoint>
-r <results_dir>
-k <encryption_key>
export_format=RIVA

This needs <experiment_spec> as mandatory but I dont know about it at all.

TAO Toolkit
Riva
Please help me to get this conversion done.

Thanks.

Hardware - GPU A100 DGX
TLT Version - latest one 1.4

2

Hi @user115721

Thanks for your interest in Riva

If we want to convert the pretrained model directly (i.e without having to do any training or finetuning), we can directly download the riva model from ngc, Please find the link below

ngc registry model download-version 
"nvidia/tao/speechtotext_zh_cn_citrinet:deployable_v2.0"

Once Downloaded you can go for deployement (link provided at end of this thread)

In case, if you are training or finetuning the model with your own dataset, we may require the experiment_spec, it can be downloaded from below command (Reference)

tao speech_to_text_citrinet download_specs -o <target_path> \
                                   -r <results_path>

After downloading you will modify according to your training dataset (Reference)

Please find the below doc link which would help in case if you are doing training
https://docs.nvidia.com/deeplearning/riva/user-guide/docs/tutorials/asr-python-advanced-finetune-am-citrinet-tao-finetuning.html?highlight=speech_to_text_citrinet

Please find the below doc which would help in deployement
https://docs.nvidia.com/deeplearning/riva/user-guide/docs/tutorials/asr-python-advanced-finetune-am-citrinet-tao-deployment.html?highlight=speech_to_text_citrinet

Please let us know if it helps, if you need further information please feel free to post in this thread

Thanks