vowpalwabbit

Failed to run vowpal wabbit contextual bandit in daemon mode


I am trying to stream my data to vw in --daemon mode. Right now I'm using just a toy example. For some reasons I failed to use any of strategies that VW has (epsilon, bag, first). I always get an error. Any quess what is wrong there?

My data:

9:0:0.3 | SmartPhone 1 16
7:0:0.3 | SmartPhone 1 21

How I launch the training:

vw --cb 10 --epsilon 0.1 -d train.vw --save_resume -f model.vw

It gives me the error (the same error for --bag 10, --first 2, cb_explore 10--bag 10 btw):

PS C:\Users\Default.PC\Desktop\vw> vw -d train.vw --cb 10 --epsilon 0.1 -f model.vw
final_regressor = model.vw
Num weight bits = 18
learning rate = 0.5
initial_t = 0
power_t = 0.5
Error: unrecognised option '--epsilon'

However, if I don't use --epsilon the code works.

vw --cb 10 -d train.vw --save_resume -f model.vw

And I can use it for prediction:

predict: | SmartPhone 1 19 - result is 3

Solution

  • Not sure which vw version was used, but the command with --epsilon appears to work fine with vw 8.9.2:

    PS C:\Users\xxx\Desktop> vw --version
    8.9.2 (git commit: 884420267)
    
    PS C:\Users\xxx\Desktop> vw -d train.vw --cb 10 --epsilon 0.1 -f model.vw
    final_regressor = model.vw
    Num weight bits = 18
    learning rate = 0.5
    initial_t = 0
    power_t = 0.5
    using no cache
    Reading datafile = train.vw
    num sources = 1
    Enabled reductions: gd, scorer, csoaa, cb
    average  since         example        example  current  current  current
    loss     last          counter         weight    label  predict features
    0.000000 0.000000            1            1.0    known        1        4
    0.000000 0.000000            2            2.0    known        1        4
    
    finished run
    number of examples = 2
    weighted example sum = 2.000000
    weighted label sum = 0.000000
    average loss = 0.000000
    total feature number = 8