Skip to content

Latest commit

 

History

History
47 lines (34 loc) · 2.72 KB

IS2016-Experiments.md

File metadata and controls

47 lines (34 loc) · 2.72 KB

Experimenting with the VM

You can do the following experiments in the VM:

  • After logging in, you should be greeted with a terminal window that looks like this:
  Florians-MBP:aws-sandbox metze$ ssh -i ~/.ssh/ec2fmetze.pem [email protected]
  Last login: Wed Sep  7 16:24:45 2016 from florians-mbp.wv.cc.cmu.edu

         __|  __|_  )
         _|  (     /   Amazon Linux AMI
        ___|\___|___|

  https://aws.amazon.com/amazon-linux-ami/2016.03-release-notes/
  [ec2-user@ip-10-41-184-80 ~]$

  • Let's try a command (and you want to do this command the first time you login to the VM):

    sudo chmod 777 /media/ephemeral0

  • Now let's change into our experiment directory

    cd babel-exps/egs/asr/s5c/201-haitian-flp

  • And let's run our experiments:

    • ln -s spkList.dev spkList.dev2h (a small bug fix)
    • ./run-1-main.sh >& log/my-run.log
    • ./train-7-gpu.sh >& log/my-train.log
    • ./test-7-v1_x3.sh >& log/my-test.log

  • The three commands will do the following:

    • run-1-main.sh will run the pre-processing of the data. It will expect the Haitian Creole data in /media/s3fs/IARPA-babel201b-v0.2b.build/ and (re-)create the data and exp folders in the current directory, which will (mostly) contain the extracted features (lMEL), along with the language model and dictionaries - everything that is required for the training
    • train-7-gpu.sh will run the actual DNN training. It will need a few hours to run (or a day and a half), and you can see the progress in the exp/train_l7_c100_n60_h0.7_v7v/log directory. You should reach an accuracy on the validation set of around 64%.
    • test-7-v1_x3.sh will test the acoustic model. It will take a model, and put the decoding result (including WER, etc.) in exp/train_l7_c100_n60_h0.7_v7v/decode_dev10h_v1
    • The results will be in a file exp/train_l7_c100_n60_h0.7_v7v/decode_dev10h_v1/score_p1.0_10/dev10h.ctm.filt.sys. You'll get a word error rate of around 49% as in:
  | Sum/Avg |21530  95893 | 55.9   30.2   13.9    4.9   49.0   32.6 | -0.812 |
  • There are many more things that you can do, like training multi-lingual systems or porting features to new languages or domains. We will add more documentation about this here and to the Eesen repository as things stabilize. Get in touch! For a first experiment, turn off -shuffle true in the training script. Your system should now train faster, because you are no longer shuffling all the utterances together. Surprisingly, there is no loss in accuracy (or almost none).

Back to the main README.

Contact

Contact Florian Metze (https://www.cs.cmu.edu/directory/fmetze) or Eric Riebling (https://www.cs.cmu.edu/directory/er1k).