search

macournoyer / neuralconvo

Neural conversational model in Torch
776 stars 347 forks source link

Getting <unknown> as result #53

Closed Vincent717 closed 8 years ago

Vincent717 commented 8 years ago

Hi all,

I have train a model for some days, with 50000 datasize, everything is fine, until I try to eval it. And I got:

you> Hi neuralconvo> <unknown>. you> How are you? neuralconvo> <unknown>. you> What is your name? neuralconvo> <unknown>.

Every response of it is <unknown>. This is wired and I cannot find the reason, because there is no error occur so far. So do you have any thoughts on this?

Thanks Vincent

macournoyer commented 8 years ago

Any word that has been trimmed from the vocabulary (using --maxVocabSize) is replaced w/ <unknown>.

You probably chose a --maxVocabSize that is too big.

Vincent717 commented 8 years ago

I didn't set --maxVocabSize when training actually, but good news is, I have retry with a small model with 100 data and it can reply again.. Can I assume that there may be some problem with the vocab.t7 in my previous try? And thank you anyway!

Btw, one more quick question, what is the meaning of "Epoch"? Since like the traing is different in different epoch, do they use diffrent data?

macournoyer commented 8 years ago

Ah yes! You have to remove data/*.t7 if you change training params.

Epoch is one training iteration over the whole dataset.

Vincent717 commented 8 years ago

That may be the problem!

I see! Thanks