flashlight / wav2letter

Facebook AI Research's Automatic Speech Recognition Toolkit
https://github.com/facebookresearch/wav2letter/wiki
Other
6.39k stars 1.01k forks source link

CUDA driver version is insufficient for CUDA runtime version #24

Closed saisrinivas047 closed 6 years ago

saisrinivas047 commented 6 years ago

HI when I am running the below command 'luajit /wav2letter/test.lua /experiments/hello_librispeech/001_model_dev-clean.bin -progress -show -test dev-clean -save'

I am getting the following error 'CUDA driver version is insufficient for CUDA runtime version'. can someone help me solve this.

Ps- i am trying to run the pretrained model in gcloud gpu instance

saisrinivas047 commented 6 years ago

@VitaliyLi

VitaliyLi commented 6 years ago

What CUDA version are you using? You need CUDA 8.0.

saisrinivas047 commented 6 years ago

this is what I get after running nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2016 NVIDIA Corporation Built on Tue_Jan_10_13:22:03_CST_2017 Cuda compilation tools, release 8.0, V8.0.61

VitaliyLi commented 6 years ago

Not sure what exactly is wrong. This solution might work.

saisrinivas047 commented 6 years ago

sure will try that, thanks for the fast reply.

saisrinivas047 commented 6 years ago

@VitaliyLi Can you please briefly tell the process to run the pre-trained model in cpu. Does running the model in cpu require Cuda. ?

saisrinivas047 commented 6 years ago

how much time does running a pretrained model takes approximately. for me it is showing around 6days to complete

VitaliyLi commented 6 years ago

For training - on 4 GPUs 1 epoch takes 1 day. If you are running 1 GPU only, 6 days is reasonable. You can make epochs shorter with flag -itersz.

saisrinivas047 commented 6 years ago

there is no need for training for pretrained model right. so for running the pre-trained model also will ir take same time to complete ?

saisrinivas047 commented 6 years ago

when I am trying to transcribe speech using "luajit ~/wav2letter/test.lua ~/librispeech-glu-highdropout-cpu.bin -progress -show -test dev-clean -save -datadir ~/librispeech-proc/ -dictdir ~/librispeech-proc/ -gfsai" command it is showling

[Sentence WER: 000.00%, dataset WER: 003.70%] [.................... 2/2703 ..................] ETA: 6D13h | Step: 3m29s

will it really take 6 days to complete ?

VitaliyLi commented 6 years ago

6 days indeed is too long for 3k sequences. Do you use Intel MKL?

srinivas047 commented 6 years ago

I did a mistake actually, I want to use pretrained model to transcribe few audio files but instead I am transcribing librispeech data. Can you tell me the audio format to transcribe single audio file because I see that the librispeech audio is splitted into different parts should I also split my audio file. Can you specify the steps for data preparation to transcribe sample audio files