[help wanted] run server streaming

[alidabaghi123]

hello, everyone. I have been installing sherpa and when I wanted to run streaming server.py. encounter with this error,

![Uploading image.png…]()

[csukuangfj]

Could you re-upload the image? It is not readable.

[alidabaghi123]

ok. thanks for concern

[csukuangfj]

Please recheck that your have downloaded the pre-trained models correctly (using git lfs install). That is, you can use

ls -lh

to show the file size of the .pt file.

[alidabaghi123]

I have checked icefall files and found pt model.

[alidabaghi123]

do i downloaded from another repository???

[csukuangfj]

you can see that the .pt file is only 134 bytes

Please use

cd exp
git lfs pull --include "cpu-jit-epoch-30-avg-10-torch-1.6.0.pt"

(I think you have missed the command git lfs install as described in the documentation)

[alidabaghi123]

thank you, i am downloading.

[alidabaghi123]

hello again, I downloaded the model and encounter this error

[alidabaghi123]

can you help me, again

[csukuangfj]

Are you using the latest version of sherpa? If not, please use the latest version of sherpa.

[alidabaghi123]

I installed this version\

can you help me, again

[csukuangfj]

I installed this version

can you help me, again

We have been in 1.0 for a long time. It is not clear which commit you are using.

How did you install sherpa?

What is the output of

sherpa-version

The line numbers in your error logs don't match the latest file https://github.com/k2-fsa/sherpa/blob/master/sherpa/bin/pruned_stateless_emformer_rnnt2/streaming_server.py

[alidabaghi123]

[csukuangfj]

How did you install sherpa?

[csukuangfj]

I suggest that you uninstall sherpa and reinstall it using the latest master.

[alidabaghi123]

with pip

[csukuangfj]

The git sha1 and git date are empty. It is not possible to tell which commit you are using.

---

Please install the latest master of sherpa.

[alidabaghi123]

i used this page for install sherpa https://k2-fsa.github.io/sherpa/python/installation/index.html

[csukuangfj]

Please select this one

git clone https://github.com/k2-fsa/sherpa
cd sherpa

# Install the dependencies
pip install -r ./requirements.txt

# Install the C++ extension.
# Use one of the following methods:
#
# (1)
python3 setup.py install --verbose

[alidabaghi123]

thank you. i am trying installl again

[csukuangfj]

Remember to uninstall sherpa first

pip uninstall k2-sherpa

[alidabaghi123]

I am installing. With 96 mycomputer locked

[alidabaghi123]

I am waiting long time even now

[csukuangfj]

How large is your RAM and how many CPUs do you have?

[alidabaghi123]

[alidabaghi123]

i forced restart my pc

[csukuangfj]

Could you change https://github.com/k2-fsa/sherpa/blob/b933ac22b6f1d1871d10d52bc2ee60ea0b115f3b/cmake/cmake_extension.py#L104 to

-j1

before running python3 setup.py install?

[alidabaghi123]

thank you for replying again. i am installing now

[alidabaghi123]

is that true???

[csukuangfj]

Did you use wget or git clone to download the source code? It's strange that git sha1 and git date is empty for sherpa.

Can you re-run your streaming_server.py and see if it still causes errors?

[alidabaghi123]

hello agian. i have downloded sherpa from master and succesfully installed web server. i am very thrilling in speed of decoding.
do i read how to improving speed of decoding??? can i replace my model????

[alidabaghi123]

thank you very lot for helping me

[alidabaghi123]

can i train this models with mexican language???

[csukuangfj]

do i read how to improving speed of decoding???

Do you have a problem with the speed of decoding? Does it run very slowly?

---

can i replace my model????

What is the type of your model? Transducer, ctc, or attention based encoder-decoder, or any other kinds of models?

---

can i train this models with mexican language???

Yes, as long as you have the data and some GPUs. The training code is at http://github.com/k2-fsa/icefall This repo is for speech recognition with pre-trained models from icefall.

[alidabaghi123]

do i read how to improving speed of decoding???

Do you have a problem with the speed of decoding? Does it run very slowly?

can i replace my model????

What is the type of your model? Transducer, ctc, or attention based encoder-decoder, or any other kinds of models?

can i train this models with mexican language???

Yes, as long as you have the data and some GPUs. The training code is at http://github.com/k2-fsa/icefall This repo is for speech recognition with pre-trained models from icefall.

no i mean, very fast.

[alidabaghi123]

I have trained conformer with nemo toolkit.

[csukuangfj]

I have trained conformer with nemo toolkit.

What kind of loss are you using?

CTC loss ?

[alidabaghi123]

yes

[csukuangfj]

yes

Then it is supported by sherpa.

Are you able to export your model via torchscript, either with torch.jit.script() or torch.jit.trace()?

If so, are you able to make it public? (If so, we can add it to sherpa. Otherwise, we can tell you how to modify the code and you have to make the changes by yourself).

[alidabaghi123]

I have pretraind conformer bpe model and use ctc loss. i use pyctc for decoding.

[alidabaghi123]

my model havent high quality now. and i am trying improve model accuracy. then i public model

[csukuangfj]

If you are going to make the model public, is it possible to provide a model for testing so that I can add it to sherpa. (The model does not need to be of high recognition accuracy).

You only need to provide a torchscript model and a token list (mapping token to its integer ID).

[alidabaghi123]

I have trained this model from this link https://github.com/NVIDIA/NeMo/tree/b8cc5030d4cb95a7e3ce7a8c22ef9d736c5c0bfc/nemo/collections/asr/models

[alidabaghi123]

i have trained squeez former

[csukuangfj]

https://github.com/NVIDIA/NeMo/tree/b8cc5030d4cb95a7e3ce7a8c22ef9d736c5c0bfc/nemo/collections/asr/models

Thanks!

Are you able to provide a torchscript model?

[csukuangfj]

You may find https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/core/export.html useful.

[alidabaghi123]

thank you. i will be trying load model in torch script.

[alidabaghi123]

i may be trying to train your model with Mexican language????

[csukuangfj]

i may be trying to train your model with Mexican language????

Yes, that is also possible. How large is your dataset (i.e., how many hours of your training data)?

---

If you already have a pre-trained model, it may be easier to export it via torchscript and use it with sherpa. It may take some time to train a model.

please have a look at http://github.com/k2-fsa/icefall if you want to train your model with icefall. Feel free to create any issues if you encounter problems.

[alidabaghi123]

can i help you to how preproccesing data and train it???