Could not create model for v0.6.0

DeepSpeech
#1

I have built an arabic language model before, with the v0.3.0. I could train, infer … etc.

Now I have moved to v0.6.0, with a fresh env, restarted the training, and exported a model. When I try to use the new native client with command:

./native_client/deepspeech --model model_0.6.pbmm --lm lm.binary --trie lm.trie --audio wav/002255_test.wav

I get this error that I cannot escape:

TensorFlow: v1.14.0-21-ge77504a
DeepSpeech: v0.6.0-0-g6d43e21
2019-12-19 16:28:21.312176: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Unable to fetch graph version: Invalid argument: Session was not created with a graph before Run()!
Could not create model.

When I googled the error, it seems like older incompatible files exist somewhere. But the above log message shows the correct version for TF and DeepSpeech. I have also used the new ./native/client/generate_trie to update the lm.trie file.

Seems like I only have lm.binary left from the older format, does it need an update? How can I update it?

#2

Re-regenate it, look at data/lm/ for examples.

#3

@tarekeldeeb Take a look at our repo https://github.com/Tarteel-io/deepspeech for a generated lm.binary and associated files :slight_smile:

I’ve been running into issues myself (see post here). Let me know if you’re able to get up and running using the generated binary and trie from our repo.

#4

Thanks @lissyx for your reply.
I regenerated the lm.binary using

/kenlm_orig/build/bin/build_binary -a 255 -q 8 trie myLM.arpa lm.binary

But ended up with the same error. The problem is now solved anyway.

When I use the model.pb (not the converted pbmm) everything works as expected. That seems to be an incompatibility from 0.3.

Regards

#5

Thanks Anas, binary LM cannot be just shared. It tried it anyway (with my graph) but did not work as expected.

As my issue is now fixed, I have shared my pb. You can try inferring.

#6

That’s excellent to hear!
We’re able to train as well. Will keep you posted.

We’re you able to run on 0.6.0 tho or 0.5.1?