Using GPU to train a french deepspeech

Deep Speech
#1

Hey again :blush:

Please How can I use the GPU to train the deepspeech model on my own data.

I have nvidia Geforce GTX 1070

Thx

#2

Have you read the documentation ? https://github.com/mozilla/DeepSpeech/blob/master/README.md#training-your-own-model

Have you read the links I gave you ? https://github.com/Common-Voice/commonvoice-fr/issues https://github.com/Common-Voice/commonvoice-fr/blob/master/DeepSpeech/CONTRIBUTING.md

#3

yeah I read the documentation and I followed every step in it :

I can run deepspeech but it is so slow and take a lot of time

Ok I will see those links

#4

Please provide more context, than just “it is so slow”. We can’t help you if you keep on hiding everything.

Have you installed tensorflow-gpu as documented?

#5

yeah I installed tensorflow-gpu = 1.13.1and all the libraries mentioned in the requirements file
and when I run the training, it took about 1 hour with 1 epoch

#6

It also depends on your dataset, as well as on some training parameters. But one hour per epoch, on a single GTX 1070 with default n_hidden=2048 seems okay.

1 Like
#7

try nvidia-smi while deepspeech is running to see whether you are using the gpu, if instead htop show alls cpus running, you have no gpu support

#8

There’s also nvtop available.

#9

capture
capture.png907×387 53.4 KB

What do you think?

#10

@yasine.nifa
Il faut s’assurer que tu disposes d’une bonne carte GPU + PC.
L’idĂ©al est une bonne station de travail.
Il est prĂ©fĂ©rable aussi d’entraĂźner sur des langues moins rĂ©pandues car les autres sont dĂ©jĂ  prises en charge.

#11

I do think that images are not a good thing, it’s not searchable etc. I also think you still not have followed the docs and installed tensorflow-gpu as we document. So training being slower than it should.