.github/workflows/conda-build-pytest.yml

@@ -19,11 +19,11 @@ jobs:
       max-parallel: 5
 
     steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@v3
       with:
         submodules: recursive
     - name: Set up Python 3.8
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: '3.8'
 
@@ -40,7 +40,7 @@ jobs:
       shell: bash -l {0}
       run: |
         conda install git pytest conda-build conda-verify anaconda-client conda-forge::grayskull conda-forge::conda-souschef conda-forge::flit conda-forge::coveralls conda-forge::conda-souschef sgbaird::pqdm sgbaird::elmd
-        pip install ElM2D==0.4.1
+        {{ PYTHON }} -m pip install ElM2D==0.4.1
         
     - name: Miniconda build
       shell: bash -l {0}
@@ -60,4 +60,4 @@ jobs:
       shell: bash -l {0}
       run: |
         export NUMBA_ENABLE_CUDASIM="1" # i.e. disable GPU
-        python -m pytest
+        {{ PYTHON }} -m pytest

.gitignore

@@ -176,3 +176,6 @@ figures/lc_data/UnnamedModel_lc.csv
 examples/supercon.csv
 examples/super_con_mat_discover_pypi.ipynb
 examples/super_con_mat_discover.py
+matbench_kvrh_comp_similarity.csv
+matbench_kvrh_grid_similarity.csv
+matbench_kvrh_total_similarity.csv

examples/adaptive_design_compare.py

@@ -122,9 +122,9 @@
 x = list(range(n_iter))
 y = np.zeros((rows, cols, n_repeats + 4, n_iter))
 formula = rows * [cols * [(n_repeats + 4) * [None]]]
-for (col, experiment) in enumerate(experiments):
-    for (row, y_name) in enumerate(y_names):
-        for (page, sub_experiment) in enumerate(experiment):
+for col, experiment in enumerate(experiments):
+    for row, y_name in enumerate(y_names):
+        for page, sub_experiment in enumerate(experiment):
             y[row, col, page] = sub_experiment[y_name].values.tolist()
             formula[row][col][page] = sub_experiment["formula"].values.tolist()
 

examples/bare_bones.py

@@ -129,6 +129,7 @@ def my_mvn(mu_x, mu_y, r):
 
 val_k_neigh_avg = k_neigh_avg_targ[val_ids]
 
+
 # %% 7. Weighted scores
 def weighted_score(pred, proxy, pred_weight=1.0, proxy_weight=1.0):
     """Calculate weighted discovery score using the predicted target and proxy."""

examples/crabnet_performance.py

@@ -103,10 +103,10 @@
 x = list(range(n_iter))
 y = np.zeros((rows, cols, n_repeats, n_iter))
 formula = rows * [cols * [n_repeats * [None]]]
-for (col, experiment) in enumerate(experiments):
-    for (row, y_name) in enumerate(y_names):
+for col, experiment in enumerate(experiments):
+    for row, y_name in enumerate(y_names):
         # loop through DataFrame rows: https://stackoverflow.com/a/11617194/13697228
-        for (page, sub_experiment) in enumerate(experiment):
+        for page, sub_experiment in enumerate(experiment):
             y[row, col, page] = sub_experiment[y_name].values.tolist()
             formula[row][col][page] = sub_experiment["formula"].values.tolist()
 

examples/matbench_log_kvrh_gridrdf.py

@@ -0,0 +1,133 @@
+"""
+An example of how to compute GRID distances directly without writing intermediate files.
+This example uses manually created PyMatGen Structure objects, but could be modified to 
+compute Structures from another source.
+NOTE: If the number of Structures gets too large, this could have memory implications.
+
+Modified from source by @sgbaird: https://github.com/CumbyLab/gridrdf/blob/master/examples/direct_GRID_calculation_without_files.py
+"""
+
+__author__ = "James Cumby"
+__email__ = "james.cumby@ed.ac.uk"
+
+import gridrdf
+import os
+from matbench.bench import MatbenchBenchmark
+import pandas as pd
+
+data_source_loc = os.path.join("examples")
+
+dummy = False
+
+mb = MatbenchBenchmark(autoload=True, subset=["matbench_log_kvrh"])
+task = list(mb.tasks)[0]
+fold = 0
+train_inputs, train_outputs = task.get_train_and_val_data(fold)
+test_inputs, test_outputs = task.get_test_data(fold, include_target=True)
+if dummy:
+    train_inputs = train_inputs.head(10)
+    train_outputs = train_outputs.head(10)
+    test_inputs = test_inputs.head(5)
+    test_outputs = test_outputs.head(5)
+
+structures = pd.concat((train_inputs, test_inputs), axis=0).tolist()
+
+
+# Variables to define cutoffs etc
+maximum_grid_distance = 20
+bin_size = 0.1
+number_of_shells = 100
+
+
+# Empty list to hold GRID arrays
+grid_representations = []
+
+
+# Calculate all GRIDS iteratively
+for structure in structures:
+    prim_cell_list = list(range(len(structure)))
+    rdf_atoms = gridrdf.extendRDF.get_rdf_and_atoms(
+        structure=structure,
+        prim_cell_list=prim_cell_list,
+        max_dist=maximum_grid_distance,
+    )
+
+    GRID = gridrdf.extendRDF.rdf_kde(
+        rdf_atoms=rdf_atoms, max_dist=maximum_grid_distance, bin_size=bin_size
+    )
+
+    assert (
+        GRID.shape[0] >= number_of_shells
+    ), f"Distance cutoff should be increased so that there are at least {number_of_shells} GRID shells for {structure} (only {GRID.shape[0]} computed)."
+
+    grid_representations.append(GRID[:number_of_shells])
+
+# Calculate EMD similarity
+# Currently, rdf_similarity_matrix requires a list of dicts, each with a 'task_id' for each structure
+# NOTE - this is likely to change in a future release
+structure_ids = [
+    {"task_id": i, "structure": structures[i]} for i in range(len(grid_representations))
+]
+
+# Calculate EMD similarity between grids
+grid_similarity = gridrdf.earth_mover_distance.rdf_similarity_matrix(
+    structure_ids, grid_representations, method="emd"
+)
+
+grid_similarity = grid_similarity.add(grid_similarity.T, fill_value=0)
+
+# First, convert composition to vector encoding
+elem_vectors, elem_symbols = gridrdf.composition.composition_one_hot(
+    structure_ids, method="percentage"
+)
+
+# Now computed EMD similarity based on "distances" between species contained in `similarity_matrix.csv`)
+# This is essentially Pettifor distance, but with non-integer steps defined by data-mining of probabilities.
+comp_similarity = gridrdf.earth_mover_distance.composition_similarity_matrix(
+    elem_vectors,
+    elem_similarity_file=os.path.join(data_source_loc, "similarity_matrix.csv"),
+)
+
+comp_similarity = comp_similarity.add(comp_similarity.T, fill_value=0)
+
+total_similarity = 10 * grid_similarity + comp_similarity
+
+
+# print("\nStructural Similarity:")
+# print(grid_similarity)
+
+# print("\nComposition Similarity:")
+# print(comp_similarity)
+
+# print("\nTotal similarity (= 10*GRID + Composition):")
+# print(total_similarity)
+
+grid_similarity.to_csv("matbench_kvrh_grid_similarity.csv", index=False, header=False)
+comp_similarity.to_csv("matbench_kvrh_comp_similarity.csv", index=False, header=False)
+total_similarity.to_csv("matbench_kvrh_total_similarity.csv", index=False, header=False)
+
+1 + 1
+
+# %% Code Graveyard
+
+# from pymatgen.core.lattice import Lattice
+# from pymatgen.core.structure import Structure
+
+# # Set up dummy pymatgen Structures with different cells/compositions
+# dummy_structures = [
+#     Structure(
+#         Lattice.from_parameters(a=3.84, b=3.84, c=3.84, alpha=120, beta=90, gamma=60),
+#         ["Si", "Si"],
+#         [[0, 0, 0], [0.75, 0.5, 0.75]],
+#     ),
+#     Structure(
+#         Lattice.from_parameters(a=3.84, b=3.84, c=3.84, alpha=120, beta=90, gamma=60),
+#         ["Ni", "Ni"],
+#         [[0, 0, 0], [0.75, 0.5, 0.75]],
+#     ),
+#     Structure(
+#         Lattice.from_parameters(a=3.94, b=3.94, c=3.94, alpha=120, beta=90, gamma=60),
+#         ["Si", "Si"],
+#         [[0, 0, 0], [0.75, 0.5, 0.75]],
+#     ),
+# ]

examples/not_precomputed.py

@@ -198,6 +198,7 @@ def my_mvn(mu_x, mu_y, r):
 
 val_k_neigh_avg = k_neigh_avg_targ[val_ids]
 
+
 # %% 7. Weighted scores
 def weighted_score(pred, proxy, pred_weight=1.0, proxy_weight=1.0):
     """Calculate weighted discovery score using the predicted target and proxy."""