valueError: need ar least one array to stack about zero-shot-knowledge-graph-relational-learning HOT 4 OPEN

I get the exact same error. The issue is that some of the queries has less than 20 candidates, so it will be skipped, yielding lists of length 0.

The issue arises when query P2629 is sampled. I overrided this by simply changing

yield support_pairs, query_pairs, false_pairs, support_left, support_right, query_left, query_right, false_left, false_right

to

if len(support_pairs) > 0:
     yield support_pairs, query_pairs, false_pairs, support_left, support_right, query_left, query_right, false_left, false_right

in the Extractor_generate function.

However, I would very much like to hear the reply from @Panda0406 on what the original code was, since this bug appears in the settings given. I used the hyperparameters given by @Panda0406 in #5 . I doubt @genggengcss was able to reproduce the results with this, since this is a bug in the code; if a query with less than 20 candidates is sampled, the loop will continue and not yield any data whatsoever, causing the crash.

@Panda0406 , it would be great if you could provide the specific code for the results when you ran on the Wiki-ZS dataset, since the current code is not functional for the Wiki-ZS dataset.

from zero-shot-knowledge-graph-relational-learning.

I can further add that I have found several things hardcoded in the code, specifically for the NELL-dataset. Some examples are row 178, row 181 and row 623 in trainer.py

from zero-shot-knowledge-graph-relational-learning.

Sorry for my late reply. I have double-checked my code and realize that it is truly a bug. I think I did not catch this bug because of this query-randomly-selecting:

Please modify these codes in data_loader.py and try again.

to

for k in task_pool:
if len(rel2candidates[k]) <= 20:
v = 0
else:
v = min(len(rel2candidates[k]), 1000)
t_num.append(v)

Really sorry for this bug.

from zero-shot-knowledge-graph-relational-learning.

Hello @Panda0406 , I realize I sounded a bit harsh here. Sorry for that - I have managed to reproduce your results and everything looks fine :) thank you for a great paper!

from zero-shot-knowledge-graph-relational-learning.