Added swapping model back in
Intermediate commit
Produces poster results for v2
Updates to include epsilon inference
Working status as of 20200114
Still a bug working with some v3 chem v2 cellranger input files where genes seem to get re-ordered? Still slow.
Faster and improved log prob, learning rate schedule
Rho as a latent variable, epsilon set to 1
Approx log_prob = big speedup; better cell prob encoder
Big speedup from an approximate log_prob computation.
Important changes with the encoder and regularizers lead to better performance on difficult test data.
Prevent NaNs late in training
It was noticed that (presumably due to underflow) late in training, alpha can include zero values, which become NaN. Also put tougher constraints on rho params.
Solid on benchmarks
Code to include gene_ids, feature_types, antibodies
Data reading and writing
Write to cellranger format matching input file. Include gene_ids. Include all the things in v3 needed for scanpy to read it using sc.read_10x_h5()
Fix "simple" model errors
Add z to latent encoder, and fixes for "simple" model
Stale import, and a gene_ids MTX fix
Output dataset in same format (v2 or v3) as input
Fix typo
Fix gene_id parse error for CellRanger v2 h5
Extend the range of epsilon, correct an error in lam
Calculation of lambda had an error in its use of epsilon. Epsilon previously had very limited range. Most of the responsibility was on d. Now most of the responsibility is shifted back to epsilon. This improves noise removal especially in the case of swapping.
Update posterior inference to FPR calculation
updates to lambda multiplier handling
global overdispersion, no Dirichlet
Eliminate use of deprecated scheduler epoch param, preclude lambda infinite loop
Remove cooldown from LR scheduler, as it is deprecated
Code cleanup, no substance changes
Default learning rate 1e-4
Minor logging tweak
Tiny change: make a constant
Cosmetic code cleanup
More code cleanup
Move values to consts