fix k_means large N subsampling

shgp/classification/experiments/metrics_experiment.py

@@ -19,6 +19,8 @@
 likelihood and fixed inducing points initialised with kmeans. 
 These are the 'best' case scenarios of best method used for a thorough evaluation of PGPR vs SVGP. 
 In particular we evaluate the performance of very sparse models to see whether HGV is beneficial.
+
+Please note that many of the experiments were run on a GPU so are not entirely reproducible on CPU.
 """
 
 
@@ -92,4 +94,4 @@ def run_iteration(metadata, X, Y, X_test, Y_test):
 
 
 if __name__ == '__main__':
- run_metrics_experiment(TwonormMetricsMetaDataset())
+ run_metrics_experiment(MagicMetricsMetaDataset())

shgp/data/metadata_metrics.py

@@ -26,9 +26,6 @@ class MetricsMetaDataset:
  ci_iters: int
 
 
-# TODO: Fertility experiment?
-
-
 """
 SVGP Distribution: (kmeans++, no grad-optim, with unconstrained/default)
 ELBO - max: -89.519647, min: -103.250001, median: -98.954888, mean: -98.629271, std: 3.787107.
@@ -198,34 +195,30 @@ def __init__(self):
 
 
 """
-SVGP Distribution: (kmeans++, no grad-optim, with unconstrained/default, M=300)
+SVGP Distribution: (kmeans++, no grad-optim, with unconstrained/default, M=500)
 
-PGPR Distribution: (hetero greedy var, no grad-optim, with unconstrained/default, M=300)
+PGPR Distribution: (hetero greedy var, no grad-optim, with unconstrained/default, M=500)
 """
 
 
-# TODO: Doesn't fit in memory -> crashes
-# TODO: Try fix memory errors (maybe I can run PGPR, but not svgp?)
-# This is why it won't fit in memory???
-# This experiment was run on a GPU so is not reproducable on CPU.
+# TODO: Run experiment
+# TODO: M=300?
 class MagicMetricsMetaDataset(MagicDataset, MetricsMetaDataset):
  def __init__(self):
  MagicDataset.__init__(self)
- MetricsMetaDataset.__init__(self, 10, 200, 500, 20, 500, 20)
+ MetricsMetaDataset.__init__(self, 10, 500, 500, 10, 250, 10)
 
 
 """
-SVGP Distribution: (kmeans++, no grad-optim, with unconstrained/default, M=300)
+SVGP Distribution: (kmeans++, no grad-optim, with unconstrained/default, M=500)
 
-PGPR Distribution: (hetero greedy var, no grad-optim, with unconstrained/default, M=300)
+PGPR Distribution: (hetero greedy var, no grad-optim, with unconstrained/default, M=500)
 """
 
 
-# TODO: Doesn't fit in memory -> crashes
-# TODO: Try fix memory errors (maybe I can run PGPR, but not svgp?)
-# This is why it won't fit in memory???
-# This experiment was run on a GPU so is not reproducable on CPU.
+# TODO: Run experiment
+# TODO: M=300?
 class ElectricityMetricsMetaDataset(ElectricityDataset, MetricsMetaDataset):
  def __init__(self):
  ElectricityDataset.__init__(self)
- MetricsMetaDataset.__init__(self, 10, 300, 500, 10, 250, 10)
+ MetricsMetaDataset.__init__(self, 10, 500, 500, 10, 250, 10)

shgp/inducing/initialisation_methods.py

@@ -43,7 +43,7 @@ def k_means(training_inputs: np.ndarray, M: int):
 
  # If N is large, take a uniform subset
  if N > 20000:
- training_inputs, _ = uniform_subsample(training_inputs, 20000)
+ training_inputs = uniform_subsample(training_inputs, 20000)
 
  # Scipy k-means++
  centroids, _ = scipy.cluster.vq.kmeans(training_inputs, M)