Log (2019 01 10) - Gandhie/AICS-Project GitHub Wiki
What did I do?
I had previously managed to fix the issues I was having with creating the 11D vectors (although I forgot to write a log post about fixing the issue).
Once the vectors were done, I started working on getting the cosine similarity and standard deviation from these vectors. While talking with Mehdi through email we came to the conclusion that cosine similarity and standard deviation was not necessarily the right solution here, and instead we are going with cosine distance and getting the mean of cosine distances for each preposition's vectors.
When writing and running code to get the cosine distances for each preposition vector, I ran into an issue of getting disconnected from the Jupyter Notebook that I'm running remotely too often to be able to run the cell. The looping through all vectors and comparing all vectors to all other vectors was taking a long time to run, of course, and I would get disconnected before it was even close to done.
For a few days, I was looking into solving this. I managed to figure out that it is my NIC on my motherboard of my desktop computer that is giving up and then tried a USB 3.0 Ethernet adapter, only to find out that the USB 3.0 ports were also not functioning as they should. As such, the only way I will solve this problem is a complete change of motherboard - which I will not be doing quite yet.
I instead decided to save the vectors to a file, which I moved to my local desktop and I am now running the cosine calculations locally and saving the results to a file that I can then load back into the notebook. As of right now I have not added this code to the repo, since it is currently running and I have yet to see if the final results are entirely as they should be.
What is left to do?
Once I have been able to load cosine distance scores for the vectors of each preposition back into the notebook, I will get the mean and save this info (i.e. make a table of it for the report). When that is done, the WHERE part of this project should also be done.
I will then move on to the WHAT part, the materials for which Mehdi is uploading to the mltgpu server for me. He will also be instructing me on how to use this data. I will then get measures for all prepositions regarding the visual contents of the bounding boxes of targets and landmarks and save these results as well.
As of right now, I have not added any additional prepositions other than [over, under, above, below]. I was initially intending to, but at this point it will end up being a question of time. Running the cosine distance for WHERE is taking a good while, so there might not be time to rerun everything for more prepositions in the end.
Lastly, once I have the results - I will go on to write the project report.