expt: Testing bottom to top approach

code/lib/lmdk_sel.py

@@ -269,6 +269,41 @@ def get_opts_from_bottom(seq, lmdks):
   return opts
 
 
+def get_opts_from_bottom_h(seq, lmdks):
+  # Create histogram
+  hist, h = get_hist(seq, lmdks)
+  # Keep track of points
+  hist_cur = np.array([h]*len(hist))
+  while np.sum(hist_cur) > len(seq):
+    hist_cur[-1] -= 1
+  # The options to be returned
+  hist_opts = []
+  # Keep removing points until the minimum is reached
+  while np.sum(hist_cur) > np.sum(hist):
+    # Track the minimum (best) evaluation
+    diff_min = float('inf')
+    # The candidate option
+    hist_cand = np.copy(hist_cur)
+    # Check every possibility
+    for i, h_i in enumerate(hist_cur):
+      # Can we remove one more point?
+      if h_i - 1 >= hist[i]:
+        hist_tmp = np.copy(hist_cur)
+        hist_tmp[i] -= 1
+        # Find difference from original
+        # diff_cur = get_norm(hist, hist_tmp)  # Euclidean
+        diff_cur = get_emd(hist, hist_tmp)  # Wasserstein
+        # Remember if it is the best that you've seen
+        if diff_cur < diff_min:
+          diff_min = diff_cur
+          hist_cand = np.copy(hist_tmp)
+    # Update current histogram
+    hist_cur = np.copy(hist_cand)
+    # Add current best to options
+    hist_opts.append(hist_cand)
+  # Return options
+  return hist_opts
+
 def get_non_opts_from_bottom(seq, lmdks):
   # Evaluate the original
   eval_orig = eval_seq(get_rel_dists(seq, [], lmdks))