A bug report

[jgkimi]

Hi, I'm trying to reproduce the results in the paper using this source code.

I've found some bugs to make errors, so I'm reporting them.

First, loading the glove pretrained file cannot be done.
This is an error caused by different shapes between loaded entries(400,000) and declared one(400,001).
Maybe, one is for UNK, which should be positioned at the last(indexed as [400,000]), and would better be randomly initialized as a vector and added to the embedding list in the "read_GloVe" function.

Second, an error occurs when running on the "process_batch" function, which is caused by different dimensions of a vector generated from "one_hot" and expected one.
By the "one_hot" function, vectors of attribute type and mix type have 22 dim and 27 dim, respectively, but the comment above the code said "intent dim is 26, 5 is act, 21 is uttribute".

What are the exact ones? I've checked the codes run well when I replace them to 21 and 26, respectively.

Finally, there's no "row_h". I want to be sure that each model is assigned to the corresponding model of the paper correctly.

• Additionally, I found all the directory names are wrong. They must be shifted (e.g. row_a-> row_b, ..., row_g -> row_h).

Thank you for consideration!