fix elbo call in pgpr training loop which causes NxN crash

shgp/classification/demos/demo_pgpr_vs_svgp_uci.py

@@ -5,25 +5,6 @@
 tf.random.set_seed(0)
 
 
-# TODO: Run fertility metric experiment
-def load_fertility():
- # https://archive.ics.uci.edu/ml/datasets/Fertility
- dataset = "../../data/fertility.txt"
-
- data = np.loadtxt(dataset, delimiter=",")
- X = data[:, :-1]
- Y = data[:, -1].reshape(-1, 1)
-
- NUM_INDUCING = 100 # quicker with 100 than with 9
- BERN_ITERS = 500 # best with 100: -38.982949 (with 9: -38.991571)
- PGPR_ITERS = (5, 25, 5) # best with 100: -39.354674
- GREEDY_THRESHOLD = 1e-6 # (early stops at 9): -39.354684
-
- return X, Y, NUM_INDUCING, BERN_ITERS, PGPR_ITERS, GREEDY_THRESHOLD
-
-
-# TODO: Try fix memory errors (maybe I can run PGPR, but not svgp?)
-# This is why it won't fit in memory???
 def load_electricity():
  # https://datahub.io/machine-learning/electricity
  dataset = "../../data/electricity.csv"

shgp/utilities/train_pgpr.py

@@ -118,7 +118,7 @@ def _train_sparse_reinit_pgpr(model, inner_iters, opt_iters, ci_iters, M, reinit
  For example: greedy_variance() or h_greedy_variance().
  """
  opt = gpflow.optimizers.Scipy()
- prev_elbo, elbos = model.elbo(), []
+ prev_elbo, elbos = -float("inf"), []
  outer_iters = reinit_metadata.outer_iters
 
  return_metrics = X_test is not None