M&M ELBO¶

Using DSC/finemapping.ipynb benchmark, I evaluate for situation with different effects the ELBO updates, under current implementation, which I believe is still incorrect (the way single effect posterior is computed is problematic).

J = 1000¶

cd ~/GIT/github/mvarbvs/DSC
dsc mnm.dsc --target debug_mnm_2

INFO: Load command line DSC sequence: debug_mnm_2
INTERNAL WARNING: "liter_data" has multiple output files, but only meta-file signature is tracked. Rigorous support of multiple output files is not yet implemented in current version of DSC.
INFO: DSC script exported to benchmark.html
INFO: Constructing DSC from mnm.dsc ...
INFO: Building execution graph & running DSC ...
DSC: 100%|████████████████████████████████████████| 6/6 [00:02<00:00,  2.45it/s]
INFO: Building DSC database ...
INFO: DSC complete!
INFO: Elapsed time 5.815 seconds.

fitted_again = readRDS('benchmark/fit_mnm_debug/liter_data_1_original_Y_1_get_sumstats_1_init_mnm_1_fit_mnm_debug_1.rds')$fitted
plot(sapply(1:20, function(i) fitted_again[[i]]$elbo), ylab = 'ELBO', xlab = 'n_iter')

J = 2000¶

cd ~/GIT/github/mvarbvs/DSC
dsc mnm.dsc --target debug_mnm_1

INFO: Load command line DSC sequence: debug_mnm_1
INTERNAL WARNING: "lite_data" has multiple output files, but only meta-file signature is tracked. Rigorous support of multiple output files is not yet implemented in current version of DSC.
INFO: DSC script exported to benchmark.html
INFO: Constructing DSC from mnm.dsc ...
INFO: Building execution graph & running DSC ...
DSC: 100%|████████████████████████████████████████| 6/6 [02:05<00:00, 26.21s/it]
INFO: Building DSC database ...
INFO: DSC complete!
INFO: Elapsed time 129.249 seconds.

fitted_again = readRDS('benchmark/fit_mnm_debug/lite_data_1_original_Y_1_get_sumstats_1_init_mnm_1_fit_mnm_debug_1.rds')$fitted
plot(sapply(1:20, function(i) fitted_again[[i]]$elbo), ylab = 'ELBO', xlab = 'n_iter')