Hi, I have a question on the code

to_skip['lhs'][(rhs, rel + n_relations, ts)].add(lhs) # reciprocals
to_skip['rhs'][(lhs, rel, ts)].add(rhs)

It means that for each triple (s, p, o, t) in the dataset, add another inverse triple (s, p_reciprocal, o, t) in the corresponding dataset. Therefore, predicting (o, p_reciprocal, ?, t) is equivalent to predicting (?, p, o, t)?

in datasets.py 'while id_timeline < len(self.events) and self.events[id_timeline][0] <= eval_events[id_event][3]:' should it be 'eval_events[id_event][0]' instead of 'eval_events[id_event][3]'?

Hi, very interesting work!
Would you please share the hyper-params use in the 10x rank experiments (Table 6 in the paper)?

I tried to rerun the model on the dataset wikidata and the command used to do so is 'python tkbc/learner.py --dataset wikidata --model TNTComplEx --no_time_emb --emb_reg 1e-2 --time_reg 1'. The trained model was evaluated on valid data set in every five steps/epochs and i found that the performance on valid datasets was not improving after the first evaluation. I was wondering whether it is because of the hyperparameters. Could you please provide the hyperparameter on the dataset or even better the trained model?

Your paper has reported the results of a tenfold increase of parameters, but how to reproduce it？ (I follow your readme.md to run the code, but can only get the results of the basic mode.)