Installing DeepSpeech GPU version - do i need to install Tensorflow-GPU also?

Archives DeepSpeech
#1

or does the DeepSpeech install add the correct version of tensorflow-gpu? (i have looked in the manual [https://deepspeech.readthedocs.io/en/latest/USING.html] and i cannot see anything obvious that tells me this)

i have just tried installing both DS CPU and DS GPU and there is no difference in inference speeds, hence my question. there may be some other reason why DS-GPU is not invoking my installed gpu. in fact i have two gpus, a P104 and GT710. the latter is only there to give monitor output, but i wonder if it is interfering with DS or Tensorflow and stopping it/them using the P104?

#2

You need to document whether you are referring to training or inference from the lib. In both case, tensorflow is a dependency (python wheel dep in the training case, static linkage in the lib case). So you don’t have to install anything manually.

#3

Please share exact repro steps, we can’t know what you did by “installed both” …

#4

inference = from the audio samples (not training).

#5

DS CPU = sudo pip3 install deepspeech
DS GPU = sudo pip3 install deepspeech-gpu

then i ran:

deepspeech --model deepspeech-0.9.3-models.pbmm --scorer deepspeech-0.9.3-models.scorer --audio my_audio_file.wav

the inference time was .5 secs in both cases, though i did notice that in the case of the CPU the first time i ran it, it was more 1.5 secs inference time. running the GPU version a second time made no difference (in fact it was a tiny fraction longer).

is this the kind of repro info you mean?

#6

How long is the audio ? Have you verified CUDA deps?

Please don’t, this is messing with everything we document. Especially, you are not leveraging your virtualenv and you might have a mix of CPU and GPU libs somewhere …

#7

You should also mention the length of this file ; based on the processing times you shared, it’s a few seconds at most, CUDA has some setup costs and it will not go much much faster than CPU on a small dataset.

#8

Inference took 0.486s for 1.975s audio file.
and the audio file is the sample “experience proves this” (from the DS samples)

i believe that i verified the CUDA deps when i installed them, but i can check again.

re

Please don’t, this is messing with everything we document. Especially, you are not leveraging your virtualenv and you might have a mix of CPU and GPU libs somewhere

i don’t fully understand. i am using sudo because i was told to do so. or do you mean that i should not try to install two different versions of DS on the same machine, even in virtual environments? if i should not do this, i don’t know how i can compare inferences with and without GPU. maybe there is another and better way of doing it.

#9

thanks for the info re short audio files and setup costs. i will try something more like twenty seconds.

#10

Who told you ? I explicitely stated you should not.

I never said that, please. it’s perfectly fine to setup two virtualenv side by side …