The training week | Grid Search
TLDR
The time is passing quick, almost all the parts of the current project are done this last part of optimising the hyperparameters was done this week.
- Grid search without separate embeddings
- More questions to be added
TLDR ends
At last the Grid search has come to an end, we tested on a total of 13 cases without testing different embedding mechanism. The best BLEU achieved by us was ~85 with an accuracy of 45%. The whole grid chart contained information about configuration of each of the test cases. The charts will be made public after 25th August’19.
Only 2 weeks are left for the GSoC to finally finish, I have tried to work relentlessly on the project from both development and research aspects and will continue to try to be an integral part of the DBpedia community in the comming years.
The next 2 weeks are Buffer periods and the following tasks have been assigned to be completed by the end of the week:
- Adding more questions and finalizing all other aspects of code.
- Creating the final project report with comparison with other models on compositionality namely convs2s.
- Upload train log in the spreadsheet for mentor reference
- Testing on QALD and LC-QuAD
I will finish the task before the end of this week so that the project comes to a finish and as discussed with the mentors we can look into publishing some aspects of it and continue to work further to discover more findings.