practical translate. Help needed

[remotejob]

Can you add real translate function for 'transformer' as for 'padded_seq2seq.py'

def translate_sentence(sentence): # tokenize the sentence tokenized = tokenize_de(sentence) tokenized = ['<sos>'] + [t.lower() for t in tokenized] + ['<eos>']

Thanks.

[whoszus]

I have the same issue with @remotejob, that I need a function to implement the translation and to get accuracy.
Thanks.

[graviraja]

Hi, It means that in the case of training, the decoder of transformer takes the encoded src sentence, and the target sentence words as inputs in order to predict the sentence. Where as in the case of evaluation/inference, decoder takes the encoded src sentence, and the predicted word as input. In simple terms, For training we will use the target sentence as input to decoder, in order to predict the target sentence and calculate the loss on how much different is the predicted from target sentence. In evaluation, we will use the predicted word at time step t as the input to time step t+1. We use target sentence to calculate the loss. Hope this helps. Thanks, Ravi.

…

On Sat, Aug 24, 2019 at 7:54 PM TK ***@***.***> wrote: I read a sentence on your blog that you said that 'We will send the trg sentence and the encoded src sentence to the decoder, to get the target sentence.' What's it means? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ABZUYF3YHCAR5PEEVUHPDG3QGFAA5A5CNFSM4IFBUVX2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5CA7BQ#issuecomment-524554118>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABZUYF5YLXO5OYF37FRB2C3QGFAA5ANCNFSM4IFBUVXQ> .

[whoszus]

Great, Thx.

[whoszus]

And I found another implementation of this paper, hope this helps @remotejob
https://github.com/jadore801120/attention-is-all-you-need-pytorch