search

lium-lst / nmtpytorch

Sequence-to-Sequence Framework in PyTorch
Other
392 stars 51 forks source link

Unchanged training result #9

Closed sampalomad closed 6 years ago

sampalomad commented 6 years ago

Hi, I trained and test the mnmt Greman model on Multi30K several times, but every training test gave me the same BLEU precise to 0.001, and there is no fluctuation in train_loss after at any given epoch across these training logs (for example, the loss after epoch 28 is the same across all 5 training logs), is it normal?

ozancaglayan commented 6 years ago

Hello,

Are you fixing the seed parameter to something different than 0?

sampalomad commented 6 years ago

Hi,

yes, the seed is set to 1234

ozancaglayan commented 6 years ago

That is the reason. Same seed will produce same results. You can set it 0 for random seeds.