Add visualization of 3D QDax repertoires

HISTORY.md

@@ -17,6 +17,7 @@
 - Add `rasterized` arg for heatmaps (#359)
 - Support 1D cvt_archive_heatmap ({pr}`362`)
 - Add 3D plots for CVTArchive ({pr}`371`)
+- Add visualization of 3D QDax repertoires ({pr}`372`)
 
 #### Documentation
 

ribs/visualize/__init__.py

@@ -20,21 +20,24 @@
     ribs.visualize.grid_archive_heatmap
     ribs.visualize.parallel_axes_plot
     ribs.visualize.sliding_boundaries_archive_heatmap
+    ribs.visualize.qdax_repertoire_3d_plot
     ribs.visualize.qdax_repertoire_heatmap
 """
 from ribs.visualize._cvt_archive_3d_plot import cvt_archive_3d_plot
 from ribs.visualize._cvt_archive_heatmap import cvt_archive_heatmap
 from ribs.visualize._grid_archive_heatmap import grid_archive_heatmap
 from ribs.visualize._parallel_axes_plot import parallel_axes_plot
-from ribs.visualize._qdax_repertoire_heatmap import qdax_repertoire_heatmap
 from ribs.visualize._sliding_boundaries_archive_heatmap import \
     sliding_boundaries_archive_heatmap
+from ribs.visualize._visualize_qdax import (qdax_repertoire_3d_plot,
+                                            qdax_repertoire_heatmap)
 
 __all__ = [
     "cvt_archive_3d_plot",
     "cvt_archive_heatmap",
     "grid_archive_heatmap",
     "parallel_axes_plot",
     "sliding_boundaries_archive_heatmap",
+    "qdax_repertoire_3d_plot",
     "qdax_repertoire_heatmap",
 ]

ribs/visualize/_qdax_repertoire_heatmap.py → ribs/visualize/_visualize_qdax.py

@@ -1,10 +1,34 @@
-"""Provides qdax_repertoire_heatmap."""
+"""Provides visualization functions for QDax repertoires."""
 import numpy as np
 
 from ribs.archives import CVTArchive
+from ribs.visualize._cvt_archive_3d_plot import cvt_archive_3d_plot
 from ribs.visualize._cvt_archive_heatmap import cvt_archive_heatmap
 
 
+def _as_cvt_archive(repertoire, ranges):
+    """Converts a QDax repertoire into a CVTArchive."""
+
+    # Construct a CVTArchive. We set solution_dim to 0 since we are only
+    # plotting and do not need to have the solutions available.
+    cvt_archive = CVTArchive(
+        solution_dim=0,
+        cells=repertoire.centroids.shape[0],
+        ranges=ranges,
+        custom_centroids=repertoire.centroids,
+    )
+
+    # Add everything to the CVTArchive.
+    occupied = repertoire.fitnesses != -np.inf
+    cvt_archive.add(
+        np.empty((occupied.sum(), 0)),
+        repertoire.fitnesses[occupied],
+        repertoire.descriptors[occupied],
+    )
+
+    return cvt_archive
+
+
 def qdax_repertoire_heatmap(
     repertoire,
     ranges,
@@ -33,22 +57,36 @@ def qdax_repertoire_heatmap(
     """
     # pylint: enable = line-too-long
 
-    # Construct a CVTArchive. We set solution_dim to 0 since we are only
-    # plotting and do not need to have the solutions available.
-    cvt_archive = CVTArchive(
-        solution_dim=0,
-        cells=repertoire.centroids.shape[0],
-        ranges=ranges,
-        custom_centroids=repertoire.centroids,
-    )
+    cvt_archive_heatmap(_as_cvt_archive(repertoire, ranges), *args, **kwargs)
 
-    # Add everything to the CVTArchive.
-    occupied = repertoire.fitnesses != -np.inf
-    cvt_archive.add(
-        np.empty((occupied.sum(), 0)),
-        repertoire.fitnesses[occupied],
-        repertoire.descriptors[occupied],
-    )
 
-    # Plot the archive.
-    cvt_archive_heatmap(cvt_archive, *args, **kwargs)
+def qdax_repertoire_3d_plot(
+    repertoire,
+    ranges,
+    *args,
+    **kwargs,
+):
+    # pylint: disable = line-too-long
+    """Plots a QDax MapElitesRepertoire with 3D measure space.
+
+    Internally, this function converts a
+    :class:`~qdax.core.containers.mapelites_repertoire.MapElitesRepertoire` into
+    a :class:`~ribs.archives.CVTArchive` and plots it with
+    :meth:`cvt_archive_3d_plot`.
+
+    Args:
+        repertoire (qdax.core.containers.mapelites_repertoire.MapElitesRepertoire):
+            A MAP-Elites repertoire output by an algorithm in QDax.
+        ranges (array-like of (float, float)): Upper and lower bound of each
+            dimension of the measure space, e.g. ``[(-1, 1), (-2, 2), (-3, 3)]``
+            indicates the first dimension should have bounds :math:`[-1,1]`
+            (inclusive), the second dimension should have bounds :math:`[-2,2]`,
+            and the third dimension should have bounds :math:`[-3,3]`
+            (inclusive). This is needed since the MapElitesRepertoire does not
+            store measure space bounds.
+        *args: Positional arguments to pass to :meth:`cvt_archive_3d_plot`.
+        **kwargs: Keyword arguments to pass to :meth:`cvt_archive_3d_plot`.
+    """
+    # pylint: enable = line-too-long
+
+    cvt_archive_3d_plot(_as_cvt_archive(repertoire, ranges), *args, **kwargs)

tests/visualize_qdax/visualize_qdax_test.py

@@ -11,7 +11,7 @@
 from qdax.core.containers.mapelites_repertoire import (MapElitesRepertoire,
                                                        compute_cvt_centroids)
 
-from ribs.visualize import qdax_repertoire_heatmap
+from ribs.visualize import qdax_repertoire_3d_plot, qdax_repertoire_heatmap
 
 
 @pytest.fixture(autouse=True)
@@ -26,10 +26,11 @@ def clean_matplotlib():
     plt.close("all")
 
 
-@image_comparison(baseline_images=["qdax_repertoire_heatmap"],
-                  remove_text=False,
-                  extensions=["png"],
-                  tol=0.1)  # See CVT_IMAGE_TOLERANCE in visualize_test.py
+@image_comparison(
+    baseline_images=["qdax_repertoire_heatmap"],
+    remove_text=False,
+    extensions=["png"],
+    tol=0.1)  # See CVT_IMAGE_TOLERANCE in cvt_archive_heatmap_test.py
 def test_qdax_repertoire_heatmap():
     plt.figure(figsize=(8, 6))
 
@@ -59,3 +60,44 @@ def test_qdax_repertoire_heatmap():
 
     # Plot heatmap.
     qdax_repertoire_heatmap(repertoire, ranges=[(-1, 1), (-1, 1)])
+
+
+@image_comparison(
+    baseline_images=["qdax_repertoire_3d_plot"],
+    remove_text=False,
+    extensions=["png"],
+    tol=0.1)  # See CVT_IMAGE_TOLERANCE in cvt_archive_3d_plot_test.py
+def test_qdax_repertoire_3d_plot():
+    plt.figure(figsize=(8, 6))
+
+    random_key = jax.random.PRNGKey(42)
+
+    # Compute the CVT centroids.
+    random_key, subkey = jax.random.split(random_key)
+    centroids, _ = compute_cvt_centroids(
+        num_descriptors=3,
+        num_init_cvt_samples=1000,
+        num_centroids=500,
+        minval=-1,
+        maxval=1,
+        random_key=subkey,
+    )
+
+    # Create initial population.
+    random_key, *subkeys = jax.random.split(random_key, 4)
+    x = jax.random.uniform(subkeys[0], (10000,), minval=-1.0, maxval=1.0)
+    y = jax.random.uniform(subkeys[1], (10000,), minval=-1.0, maxval=1.0)
+    z = jax.random.uniform(subkeys[2], (10000,), minval=-1.0, maxval=1.0)
+    init_pop = jnp.stack((x, y, z), axis=1)
+
+    # Create repertoire with the initial population inserted.
+    repertoire = MapElitesRepertoire.init(
+        genotypes=init_pop,
+        # Negative sphere function.
+        fitnesses=-jnp.sum(jnp.square(init_pop), axis=1),
+        descriptors=init_pop,
+        centroids=centroids,
+    )
+
+    # Plot heatmap.
+    qdax_repertoire_3d_plot(repertoire, ranges=[(-1, 1), (-1, 1), (-1, 1)])