How to get the gaussian parameter(mean,cov,weights) ?

[liuziyang365]

if i have data = batch1+batch2, are there 2 ways to run the online-clustering ?
1 way: set nBatch = 2
bnpy.run(
data, 'DPMixtureModel', 'Gauss', 'memoVB',
output_path='/home/pinga/lzy/bnpy-test/tmp/3/',
nLap=100, nTask=1, nBatch=2,
sF=0.1, ECovMat='eye',
initname=dict_2['task_output_path'],
moves='birth,merge,shuffle',
m_startLap=5, b_startLap=0, b_Kfresh=4)

2 way ：set batch2_initname = batch1_out_put_path
gamma = 5
sF = 5
K = 1
model_1, dict_1 = bnpy.run(
batch2, 'DPMixtureModel', 'Gauss', 'memoVB',
output_path='/home/pinga/lzy/bnpy-test/tmp/2/',
nLap=100, nTask=1, nBatch=1,
sF=sF,K=K,gamma0=gamma ECovMat='eye',
initname='randexamples',
moves='birth,merge,shuffle',
m_startLap=5, b_startLap=0, b_Kfresh=4)
model_2, dict_2 = bnpy.run(
batch2, 'DPMixtureModel', 'Gauss', 'memoVB',
output_path='/home/pinga/lzy/bnpy-test/tmp/2/',
nLap=100, nTask=1, nBatch=1,
sF=sF,K=K,gamma0=gamma ECovMat='eye',
initname=dict_1['task_output_path'],
moves='birth,merge,shuffle',
m_startLap=5, b_startLap=0, b_Kfresh=4)

Are these two ways the same?
How to get the gaussian parameters (mean,cov,weight) after clustering each batch ?
How to get the weighted log probabilities for each sample?

[michaelchughes]

Are these two ways the same? No

No. Those two approaches will deliver very different models (I'm assuming you intended the 2nd approach's first call to bnpy.run to refer to batch1, and the second call to batch2. As written your second example never uses batch1).

The first approach will fit a model that tries to jointly explain batch 1 and batch 2. It will run for 100 laps or epochs (100 passes thru both batch1 and batch2).

The second approach will be different. It will first fit a good mixture model to batch1 only, using 100 laps/epochs. Then, it will use this model as initialization (but only an initialization) when fitting to batch2, again running the batch2-specific training for 100 laps/epochs.

If batch2 looks very different than batch1, the results will be very different.

• first approach will try to find clusters that fit both batches.
• second approach, because it has birth/merge moves, will only fit batch2 in the end. Any preliminary clusters that were useful for batch1 would probably be discarded/edited by merge moves or standard update steps unless they provide benefit for batch2.