sort bins by occupation

ms2deepscore/train_new_model/inchikey_pair_selection.py

@@ -39,16 +39,24 @@ def select_compound_pairs_wrapper(
         settings.same_prob_bins,
         settings.include_diagonal)
 
-    aimed_nr_of_pairs_per_bin = determine_aimed_nr_of_pairs_per_bin(available_pairs_per_bin_matrix,
-                                                                    settings,
-                                                                    nr_of_inchikeys=len(inchikeys14_unique))
-
-    pair_frequency_matrixes = balanced_selection_of_pairs_per_bin(available_pairs_per_bin_matrix,
-                                                                  settings.max_pair_resampling,
-                                                                  aimed_nr_of_pairs_per_bin)
-
-    selected_pairs_per_bin = convert_to_selected_pairs_list(pair_frequency_matrixes, available_pairs_per_bin_matrix,
-                                                            available_scores_per_bin_matrix, inchikeys14_unique)
+    aimed_nr_of_pairs_per_bin, bin_priorities = determine_aimed_nr_of_pairs_per_bin(
+        available_pairs_per_bin_matrix,
+        settings,
+        nr_of_inchikeys=len(inchikeys14_unique)
+        )
+
+    pair_frequency_matrixes = balanced_selection_of_pairs_per_bin(
+        available_pairs_per_bin_matrix,
+        settings.max_pair_resampling,
+        aimed_nr_of_pairs_per_bin
+        )
+
+    selected_pairs_per_bin = convert_to_selected_pairs_list(
+        pair_frequency_matrixes,
+        available_pairs_per_bin_matrix,
+        available_scores_per_bin_matrix,
+        inchikeys14_unique
+        )
     return [pair for pairs in selected_pairs_per_bin for pair in pairs]
 
 
@@ -143,12 +151,19 @@ def compute_jaccard_similarity_per_bin(
 
 def determine_aimed_nr_of_pairs_per_bin(available_pairs_per_bin_matrix, settings, nr_of_inchikeys):
     """Determines the aimed_nr_of_pairs_per_bin.
-    If the settings given are higher than the highest possible number of pairs it is lowered to that"""
+
+    If the settings given are higher than the highest possible number of pairs it is lowered to that.
+    """
 
     # Select the nr_of_pairs_per_bin to use
     nr_of_available_pairs_per_bin = get_nr_of_available_pairs_in_bin(available_pairs_per_bin_matrix)
     lowest_max_number_of_pairs = min(nr_of_available_pairs_per_bin) * settings.max_pair_resampling
     print(f"The available nr of pairs per bin are: {nr_of_available_pairs_per_bin}")
+
+    # Set bin priority from lowest to highest no. of available pairs
+    bin_priority = np.argsort(nr_of_available_pairs_per_bin)
+    print(f"Bin priorities will be orderd accordingly: {[settings.same_prob_bins[i] for i in bin_priority]}")
+
     aimed_nr_of_pairs_per_bin = settings.average_pairs_per_bin * nr_of_inchikeys
     if lowest_max_number_of_pairs < aimed_nr_of_pairs_per_bin:
         print(f"Warning: The average_pairs_per_bin: {settings.average_pairs_per_bin} cannot be reached, "
@@ -158,13 +173,14 @@ def determine_aimed_nr_of_pairs_per_bin(available_pairs_per_bin_matrix, settings
               f"Instead the lowest number of available pairs in a bin times the resampling is used, "
               f"which is: {lowest_max_number_of_pairs}")
         aimed_nr_of_pairs_per_bin = lowest_max_number_of_pairs
-    return aimed_nr_of_pairs_per_bin
+    return aimed_nr_of_pairs_per_bin, bin_priority
 
 
 def balanced_selection_of_pairs_per_bin(
         available_pairs_per_bin_matrix: np.ndarray,
         max_pair_resampling: int,
-        nr_of_pairs_per_bin: int
+        nr_of_pairs_per_bin: int,
+        bin_priority: np.ndarray = None,
         ) -> np.ndarray:
     """From the available_pairs_per_bin_matrix a balanced selection is made to have a balanced distribution.
 
@@ -190,11 +206,16 @@ def balanced_selection_of_pairs_per_bin(
         Resampling means that the exact same inchikey pair is added multiple times to the list of pairs.
     nr_of_pairs_per_bin:
         The number of pairs that should be sampled for each tanimoto bin.
+    bin_priority:
+        Bins will be processed in the order given in bin_priority. Default is set to None in which case no change
+        to the order will be done.
     """
+    if bin_priority is None:
+        bin_priority = np.arange(0, available_pairs_per_bin_matrix.shape[0])
 
     inchikey_count = np.zeros(available_pairs_per_bin_matrix.shape[1])
     pair_frequency_matrixes = []
-    for pairs_in_bin in available_pairs_per_bin_matrix:
+    for pairs_in_bin in available_pairs_per_bin_matrix[bin_priority]:
         pair_frequencies, inchikey_count = select_balanced_pairs(pairs_in_bin,
                                                                  inchikey_count,
                                                                  nr_of_pairs_per_bin,