Fixbug Beam Search

lvapeab · Jan 11, 2017 · aed24df · aed24df
1 parent b7ca708
commit aed24df
Show file tree

Hide file tree

Showing 2 changed files with 64 additions and 103 deletions.
diff --git a/keras_wrapper/beam_search_ensemble.py b/keras_wrapper/beam_search_ensemble.py
@@ -25,7 +25,7 @@ def __init__(self, models, dataset, params_prediction, verbose=0):
             params_prediction.get('optimized_search') is not None else False
         self.verbose = verbose
         if self.verbose > 0:
-            print "Using optimized_search=", self.optimized_search
+            print "Using optimized_search =", self.optimized_search
 
     # PREDICTION FUNCTIONS: Functions for making prediction on input samples
 
@@ -126,7 +126,6 @@ def beam_search(self, X, params, null_sym=2):
                                                                prev_outs=prev_outs)
             else:
                 probs = self.predict_cond(self.models, X, state_below, params, ii)
-
             # total score for every sample is sum of -log of word prb
             cand_scores = np.array(hyp_scores)[:, None] - np.log(probs)
             cand_flat = cand_scores.flatten()
@@ -141,11 +140,13 @@ def beam_search(self, X, params, null_sym=2):
 
             # Form a beam for the next iteration
             new_hyp_samples = []
+            new_trans_indices = []
             new_hyp_scores = np.zeros(k-dead_k).astype('float32')
             if params['pos_unk']:
                 new_hyp_alphas = []
             for idx, [ti, wi] in enumerate(zip(trans_indices, word_indices)):
                 new_hyp_samples.append(hyp_samples[ti]+[wi])
+                new_trans_indices.append(ti)
                 new_hyp_scores[idx] = copy.copy(costs[idx])
                 if params['pos_unk']:
                     new_hyp_alphas.append(hyp_alphas[ti]+[alphas[ti]])
@@ -164,7 +165,7 @@ def beam_search(self, X, params, null_sym=2):
                         sample_alphas.append(new_hyp_alphas[idx])
                     dead_k += 1
                 else:
-                    indices_alive.append(idx)
+                    indices_alive.append(new_trans_indices[idx])
                     new_live_k += 1
                     hyp_samples.append(new_hyp_samples[idx])
                     hyp_scores.append(new_hyp_scores[idx])