Reproduce examples with new GeneVector class structure.

[jordantgh]

You have obviously updated the GeneVector class substantially since writing e.g. the PBMC example. Do you have any ideas how to reproduce the original behaviour? I have tried a few different combinations of parameters and get things like this:

It's not as it should be, of course.

#7 is obviously related but the suggested changes do not make a difference (although I do not have cuda configured for GPU in this env, it doesn't matter).

[jordantgh]

@nceglia

I ran a naive grid search overnight to see if I could observe anything remotely as good as what was acheived before the GeneVector refactor. It is quite primitive, but I don't want to put too much effort at this early stage. The parameter values I tried were:

parameter_grid = {
    'emb_dimension': [100, 200],
    'batch_size': [100000, 1000000, 10000000],
    'gain': [0.5, 1, 2],
    'init_ortho': [True, False],
    'epochs': [1000, 5000, 10000],
    'alpha': [0.01, 0.1],
    'beta': [0.01, 0.1]
}

You can see the results in the attached .csv. I have not done any GO/GSEA analysis but it is obvious that none of these retrieve anything like the enrichment of ISGs expected and observed previously. I can confirm however that I was able to reproduce your earlier results running the code as is on c5c7b0e

grid_search_results.csv

[nceglia]

Hi! Thanks for bringing this up. I am going to try to reproduce the original results with the new structures in the next few days. Sorry for any inconveniences!

[jordantgh]

Good luck. Sorry that I can't be of more help. I did try just hardcoding scaling factors into the mi score, that did not work at all.

[nceglia]

You've helped a ton just letting me know! I should've tested this tho, im sorry for that. To give you some background, we were trying to make it more sensitive doing CD8 and CD4 T cells phenotypes, when only running on T cells. It seemed to improve significantly in that task, hopefully we can adjust the parameters to work for both cases.