make pgpr vs svgp comparisons non-sparse

GiovanniPasserello · May 28, 2021 · ef9cff5 · ef9cff5
1 parent 23b3b36
commit ef9cff5
Show file tree

Hide file tree

Showing 3 changed files with 16 additions and 21 deletions.
diff --git a/shgp/classification/demos/demo_pgpr_vs_svgp_1D.py b/shgp/classification/demos/demo_pgpr_vs_svgp_1D.py
@@ -4,7 +4,6 @@
 import tensorflow as tf
 
 from shgp.data.dataset import PlatformDataset
-from shgp.inducing.initialisation_methods import uniform_subsample
 from shgp.likelihoods.pg_bernoulli import PolyaGammaBernoulli
 from shgp.utilities.general import invlink
 from shgp.utilities.train_pgpr import train_pgpr
@@ -18,24 +17,20 @@ def model_comparison():
  # Model Optimisation #
  ######################
 
- num_inducing = 10
-
  # PGPR
  pgpr, pgpr_elbo = train_pgpr(
  X, Y,
- 10, 1000, 10,
- kernel_type=gpflow.kernels.Matern52,
- M=num_inducing,
- init_method=uniform_subsample
+ 20, 2000, 20,
+ kernel_type=gpflow.kernels.SquaredExponential
  )
  print("pgpr trained")
 
  # TODO: This comparison of Bernoulli vs PG is worth showing in `Evaluation'.
  # SVGP (choose Bernoulli or PG likelihood for comparison)
- # likelihood = gpflow.likelihoods.Bernoulli(invlink=tf.sigmoid)
- likelihood = PolyaGammaBernoulli()
+ likelihood = gpflow.likelihoods.Bernoulli(invlink=tf.sigmoid)
+ #likelihood = PolyaGammaBernoulli()
  svgp = gpflow.models.SVGP(
- kernel=gpflow.kernels.Matern52(),
+ kernel=gpflow.kernels.SquaredExponential(),
  likelihood=likelihood,
  inducing_variable=pgpr.inducing_variable.Z
  )
@@ -72,7 +67,10 @@ def model_comparison():
  # Meta
  plt.title('PGPR vs SVGP - Platform Dataset')
  plt.xlim((-2, 2))
+ plt.xlabel('x')
  plt.ylim((-0.5, 1.5))
+ plt.yticks([0, 0.5, 1])
+ plt.ylabel('p( y=1 | x )')
 
  # Display
  plt.legend(loc='upper right')

diff --git a/shgp/classification/demos/demo_pgpr_vs_svgp_2D.py b/shgp/classification/demos/demo_pgpr_vs_svgp_2D.py
@@ -4,7 +4,6 @@
 import tensorflow as tf
 
 from shgp.data.dataset import BananaDataset
-from shgp.inducing.initialisation_methods import uniform_subsample
 from shgp.likelihoods.pg_bernoulli import PolyaGammaBernoulli
 from shgp.utilities.general import invlink
 from shgp.utilities.train_pgpr import train_pgpr
@@ -18,22 +17,18 @@ def model_comparison():
  # Model Optimisation #
  ######################
 
- num_inducing = 20
-
  # PGPR
  pgpr, pgpr_elbo = train_pgpr(
  X, Y,
- 20, 250, 10,
- kernel_type=gpflow.kernels.SquaredExponential,
- M=num_inducing,
- init_method=uniform_subsample
+ 20, 2000, 20,
+ kernel_type=gpflow.kernels.SquaredExponential
  )
  print("pgpr trained")
 
  # TODO: This comparison of Bernoulli vs PG is worth showing in `Evaluation'.
  # SVGP (choose Bernoulli or PG likelihood for comparison)
- # likelihood = gpflow.likelihoods.Bernoulli(invlink=tf.sigmoid)
- likelihood = PolyaGammaBernoulli()
+ likelihood = gpflow.likelihoods.Bernoulli(invlink=tf.sigmoid)
+ # likelihood = PolyaGammaBernoulli()
  svgp = gpflow.models.SVGP(
  kernel=gpflow.kernels.SquaredExponential(),
  likelihood=likelihood,
@@ -87,6 +82,8 @@ def model_comparison():
  c1.collections[0].set_label('PGPR ({:.2f})'.format(pgpr_elbo))
  c2.collections[0].set_label('SVGP ({:.2f})'.format(svgp_elbo))
 
+ plt.xlabel('x1')
+ plt.ylabel('x2', rotation=1)
  plt.title('PGPR vs SVGP - Banana Dataset')
  plt.legend(loc='upper left')
  plt.show()

diff --git a/shgp/utilities/train_pgpr_svgp.py b/shgp/utilities/train_pgpr_svgp.py
@@ -115,7 +115,7 @@ def _try_train_pgpr_svgp(X, Y, M, opt_iters, kernel_type, init_method, reinit_me
 # return result(svgp, svgp.elbo((X, Y)), X_test, Y_test)
 
 
-# TODO: Interleaved training (good, but slow and *very* non-monotonic unreliable)
+# Interleaved training (good, but slow and *very* non-monotonic unreliable)
 # def _train_sparse_pgpr_svgp(X, Y, M, opt_iters, kernel_type, reinit_method, reinit_metadata, optimise_Z, X_test, Y_test):
 # """
 # Train a sparse PGPR-SVGP model with a given reinitialisation method.
@@ -197,7 +197,7 @@ def _try_train_pgpr_svgp(X, Y, M, opt_iters, kernel_type, init_method, reinit_me
 # return pgpr_svgp, np.max(elbos) # return the highest ELBO
 
 
-# TODO: With greedy variance reinit (no PGPR)
+# With greedy variance reinit (no PGPR)
 def _train_sparse_pgpr_svgp(X, Y, M, opt_iters, kernel_type, reinit_method, reinit_metadata, optimise_Z, X_test, Y_test):
  """
  Train a sparse PGPR-SVGP model with a given reinitialisation method.