theta linkage and init beta

[elibol]

For discussion of implementation

[elibol]

The init beta part is clear.
I've renamed doc_id to doc_indexes (this makes it a bit easier to follow, given the old name of patient_id is doc_id).

[finaledoshivelez]

Makes sense to me.

[elibol]

@finaledoshivelez modified assignment to psi. Results:

[finaledoshivelez]

Great to see that it's doing better with the initialization! Let's chat
about this with Matt and Scott tomorrow (maybe leave a longer run going to
see whether LDA ever catches up if given another 50 or so iterations --
hopefully not)

On Tue, Jul 28, 2015 at 10:52 AM, Melih Elibol [email protected]
wrote:

In pgmult/lda.py
#1 (comment):

@@ -391,9 +413,20 @@ def init(self, data, timestamps, K, alpha_theta):
def initialize_parameters(self):
self.sigmasq_states = 0.1 # TODO make this learned, init from hypers

•    mean_psi = compute_uniform_mean_psi(self.V)[0][None,:,None]
  

- self.psi = np.tile(mean_psi, (self.T, 1, self.K))

•    # finale: added this to the initialization (must already be
  

•    # correct size, do that in the top level file or as another
  

•    # function in the top level file), not here
  

•    #
  

•    # Melih: pi_to_psi expects rows to be normalized
  

•    # and only takes arrays of size <= 2.
  

•    # Need confirmation on whether double transpose below
  

•    # results in desired behavior.
  

•    if self.param_set['use_init_beta']:
  

•        self.psi = np.array(map(lambda x: pi_to_psi(x.T).T, self.param_set['init_beta']))
  

@finaledoshivelez https://github.com/finaledoshivelez modified
assignment to psi. Results:
[image: dtm_from_lda]
https://cloud.githubusercontent.com/assets/1884921/8934613/804e7dd6-3516-11e5-98cf-ad91933d8467.png

—
Reply to this email directly or view it on GitHub
https://github.com/dtak/pgmult/pull/1/files#r35656341.