Rik Koncel-Kedziorski
Rik Koncel-Kedziorski
I don't know how to include position encoding in a graph because there is no ordering to the nodes.
This is a standard training technique called "teacher forcing" where the model is trained with the gold sequence prefix. You'll notice that in the inference code (beam_generate function lines 157...
as I understand it, in teacher forcing we observe loss (roughly) as P(y_i) - P(\hat{y}_i | y_0 ... y_{i-1}) , where the y are gold tokens and the \hat{y} are...
This is an error in the figure. the left side should also have v1->v2, v3->v4. Sorry!
Hi Dheeraj, this code is built with an older torchtext, probably around v0.4 or earlier. Please modify as needed!
I think if you change line 27 of generator to be a more appropriate filename it will work. I will clean up this part of the code ASAP. Thanks!
sorry this code is so messy. I think the point of these two lines is to get the final hidden state from the top layer of the encoder BiLSTM and...
the dataset class in lastDataset.py will help you
Are you running on Windows/DOS OS? This error might be a problem with the meteor eval script and windows. I am unsure how to debug... maybe change it to self.meteor_p.write(score_line...
i think the -save flag is expecting a checkpoint file