Model with one latent factor less than specified

[flcello]

Hi,

It seems to me like the model is built with one factor less than specified in set_model_options():

from mofapy2.run.entry_point import entry_point

ent = entry_point()

ent.set_data_options(
    scale_groups = False, 
    scale_views = False
    )

ent.set_data_matrix([[data]], likelihoods = ["gaussian"])

ent.set_model_options(
    factors = 3, 
    spikeslab_weights = False, 
    ard_factors = False,
    ard_weights = False
    )

ent.set_train_options(
    iter = 1000, 
    convergence_mode = "slow", 
    startELBO = 1, 
    freqELBO = 1, 
    dropR2 = 0.000, 
    gpu_mode = False, 
    verbose = False, 
    seed = 1
    )

ent.build()
ent.run()

This should result in a model with 3 factors, but from the first iteration on I only get 2 factors:

######################################
## Training the model with seed 1 ##
######################################

ELBO before training: -160245.23 

Iteration 1: time=0.01, ELBO=-38881.59, deltaELBO=121363.642 (75.73619616%), Factors=2
Iteration 2: time=0.03, ELBO=-35202.00, deltaELBO=3679.586 (2.29622203%), Factors=2
Iteration 3: time=0.02, ELBO=-34646.27, deltaELBO=555.731 (0.34680037%), Factors=2
...

[gtca]

Hey, thanks for the report!

If you drop the dropR2 parameter, it should work as expected. It should be set to None (as default) in order not to drop any factors. I think that would solve this issue.

To expand a bit on this topic, in order to set a threshold and not to drop a factor at the first iteration, there's a start_drop parameter to the training options that can be set to 2: ent.set_train_options(..., start_drop=2). The reason behind this is that the factors are dropped in the beginning of each iteration. Together these things don't make much sense though... What do you think, @rargelaguet?

[flcello]

perfect, thanks for clarifying!