Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Char-rnn aurkezpena

Char-rnn aurkezpena

2a2707abeffc7d8abb8487969c78eaf6?s=128

Manex Agirrezabal

March 14, 2016
Tweet

Transcript

  1. Poesiaren metrika DL bidez Manex Agirrezabal https://github.com/manexagirrezabal/char-rnn/

  2. Proba ezberdinak TensorFlow: Sequence-to-sequence models https://www.tensorflow.org/versions/master/tutorials/seq2seq/index.html Torch: char-rnn (Andrew Karpathy)

    https://github.com/karpathy/char-rnn/
  3. Char-rnn http://karpathy.github.io/2015/05/21/rnn-effectiveness/ Karaktere mailako hizkuntz-ereduak sortzeko balio du. Sarrera gisa

    testu hutsa.
  4. Char-rnn Gure beharretarako moldatu behar: to swell the gourd and

    plump the ha zel shells - ' - ' - ' - ' - ' wo man much missed how you call to me call to me ' - - ' - - ' - - ' - -
  5. Char-rnn Dataset-a testu soil gisa: To_= swell_+ the_= gourd_+ and_=

    plump_+ the_= ha_+ zel_= shells_+ To_= swell_+ the_= gourd_+ and_= plump_+ the_= hazel_+= shells_+ Wo_+ man_= much_= missed_+ how_= you_= call_+ to_= me_= call_+ to_= me_= Woman_+= much_= missed_+ how_= you_= call_+ to_= me_= call_+ to_= me_=
  6. Char-rnn (training) $ th train.lua Parametroak: Model: [RNN, LSTM edo

    GRU] rnn_size: LSTMaren (zelda) barruko tamaina num_layers: LSTMaren kapa kopurua seq_length: sekuentzian ikasteko karaktere kopurua
  7. Char-rnn (prediction) $ th sample(mod).lua Parametroak: Model: eredu entrenatua Primetext:

    sarrera testua (_ karakterearekin amaituta)
  8. Char-rnn (prediction) Python programa bat (callSampleMod.py) aurreko programari deitzeko pausuz

    pausu: $ th sampleMod.lua ­model M1 ­primetext “to_” = $ th sampleMod.lua ­model M1 ­primetext “to_= swell_” + $ th sampleMod.lua ­model M1 ­primetext “to_= swell_+ the_” = ...
  9. Char-rnn (prediction) Arazoa: Hasieran, informazio gutxi duenez, batzuetan hanka sartzen

    (+ propagatzen) du predikzioan. Adibidez, “to_” sarrerarekin Horrentzako soluzioa, predikzioa bi aldetara egitea.
  10. Char-rnn (FW) Parametroak optimizatu nahi ditugu (seq_length, batch_size, rnn_size, ...)

    Embedding-ak erabili nahi ditugu, baina gure hipotesia da ez dutela asko lagunduko.