Error while deserialising alphabet

Hi,I used pre-trained model of deepspeech v0.6.1 and used code on the transfer-learning2 branch for transfer learning. After training,at the time of evaluation I am geting error like:

  File "", line 183, in <module>
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/platform/", line 125, in run
  File "", line 173, in main
    samples = evaluate(FLAGS.test_files.split(','), create_model, try_loading)
  File "", line 49, in evaluate
  File "/usr/local/lib/python3.5/dist-packages/ds_ctcdecoder/", line 31, in __init__
    raise ValueError('Error when deserializing alphabet.')
ValueError: Error when deserializing alphabet.

The specifications of the training requirements are:
tensorflow == 1.13.1
ds_ctcdecoder == 0.7.0-alpha.2
numpy == 1.18.1
os- ubuntu 16.04 LTS
Can someone please help me with what shall I do?

Please install matching deepspeech version.

You also have the wrong version of tensorflow, and are mixing incompatible versions of the code (v0.6.1 and the transfer-learning2 branch).

Apologies,I have DeepSpeech version 0.5.0-alpha.6 and I am unable to install ds-ctcdecoder version 0.5.0a6…getting this error:

Collecting ds-ctcdecoder==0.5.0a6
ERROR: HTTP error 404 while getting
ERROR: Could not install requirement ds-ctcdecoder==0.5.0a6 from because of error 404 Client Error: Not Found for url:
ERROR: Could not install requirement ds-ctcdecoder==0.5.0a6 from because of HTTP error 404 Client Error: Not Found for url: for URL

Can you please help with me it?

Unfortunately, TaskCluster has been migrated to a new instance and those artifacts are not available anymore. This is very old as well, so I would recommend switching to newer version anyway.

Oh I see thank you very much…but as I wanted to do transfer-learning any suggestions on what should I do? Since deepspeech version of trasnfer-learning2 branch is 0.5.0-alpha.6 it will require ds-ctcdecoder version 0.5.0a6. What can I do in this case?

You could stay with the 0.6 branch and do some fine tuning for the released model. Everything is there.

At the same time install the 0.7 version in a different place and start experimenting with the newly available transfer learning:

You will then be able to compare both approaches once the new model is available.

As all of Europe is staying at home, I guess there is not much else to do but getting the next release out :slight_smile:

1 Like