Added a normalization term to ctc_beam_search_decoder

tensorflow/core/util/ctc/ctc_beam_search.h

@@ -259,6 +259,16 @@ void CTCBeamSearchDecoder<CTCBeamState, CTCBeamComparer>::Step(
   } else {
     max_coeff = raw_input.maxCoeff();
   }
+
+  // Get normalization term of softmax: log(sum(exp(logit[j]-max_coeff))).
+  float logsumexp = 0.0;
+  for (int j = 0; j < raw_input.size(); ++j) {
+    logsumexp += Eigen::numext::exp(raw_input(j) - max_coeff);
+  }
+  logsumexp = Eigen::numext::log(logsumexp);
+  // Final normalization offset to get correct log probabilities.
+  float norm_offset = max_coeff + logsumexp;
+
   const float label_selection_input_min =
       (label_selection_margin_ >= 0) ? (max_coeff - label_selection_margin_)
                                      : -std::numeric_limits<float>::infinity();
@@ -290,10 +300,10 @@ void CTCBeamSearchDecoder<CTCBeamState, CTCBeamComparer>::Step(
                       beam_scorer_->GetStateExpansionScore(b->state, previous));
       }
       // Plabel(l=abc @ t=6) *= P(c @ 6)
-      b->newp.label += raw_input(b->label) - max_coeff;
+      b->newp.label += raw_input(b->label) - norm_offset;
     }
     // Pblank(l=abc @ t=6) = P(l=abc @ t=5) * P(- @ 6)
-    b->newp.blank = b->oldp.total + raw_input(blank_index_) - max_coeff;
+    b->newp.blank = b->oldp.total + raw_input(blank_index_) - norm_offset;
     // P(l=abc @ t=6) = Plabel(l=abc @ t=6) + Pblank(l=abc @ t=6)
     b->newp.total = LogSumExp(b->newp.blank, b->newp.label);
 
@@ -328,6 +338,8 @@ void CTCBeamSearchDecoder<CTCBeamState, CTCBeamComparer>::Step(
       const float logit = top_k ? top_k_logits[ind] : raw_input(ind);
       // Perform label selection: if input for this label looks very
       // unpromising, never evaluate it with a scorer.
+      // We may compare logits instead of log probabilities, 
+      // since the difference is the same in both cases.
       if (logit < label_selection_input_min) {
         continue;
       }
@@ -341,7 +353,7 @@ void CTCBeamSearchDecoder<CTCBeamState, CTCBeamComparer>::Step(
         //   Plabel(l=abcd @ t=6) = P(l=abc @ t=5) * P(d @ 6)
         beam_scorer_->ExpandState(b->state, b->label, &c.state, c.label);
         float previous = (c.label == b->label) ? b->oldp.blank : b->oldp.total;
-        c.newp.label = logit - max_coeff +
+        c.newp.label = logit - norm_offset +
                        beam_scorer_->GetStateExpansionScore(c.state, previous);
         // P(l=abcd @ t=6) = Plabel(l=abcd @ t=6)
         c.newp.total = c.newp.label;

tensorflow/python/keras/backend_test.py

@@ -1384,8 +1384,8 @@ def test_ctc_decode(self):
           np.array([seq_len_0], dtype=np.int32))
       # batch_size length vector of negative log probabilities
       log_prob_truth = np.array([
-          0.584855,  # output beam 0
-          0.389139  # output beam 1
+          -3.5821197,  # output beam 0
+          -3.777835    # output beam 1
       ], np.float32)[np.newaxis, :]
 
       decode_truth = [np.array([1, 0]), np.array([0, 1, 0])]

tensorflow/python/kernel_tests/ctc_decoder_ops_test.py

@@ -188,23 +188,23 @@ def testCTCDecoderBeamSearch(self):
         ],
         dtype=np.float32)
     # Add arbitrary offset - this is fine
-    input_log_prob_matrix_0 = np.log(input_prob_matrix_0) + 2.0
+    input_prob_matrix_0 = input_prob_matrix_0 + 2.0
 
     # len max_time_steps array of batch_size x depth matrices
     inputs = ([
-        input_log_prob_matrix_0[t, :][np.newaxis, :] for t in range(seq_len_0)
+        input_prob_matrix_0[t, :][np.newaxis, :] for t in range(seq_len_0)
     ]  # Pad to max_time_steps = 8
               + 2 * [np.zeros(
                   (1, depth), dtype=np.float32)])
 
     # batch_size length vector of sequence_lengths
     seq_lens = np.array([seq_len_0], dtype=np.int32)
 
-    # batch_size length vector of negative log probabilities
+    # batch_size length vector of log probabilities
     log_prob_truth = np.array(
         [
-            0.584855,  # output beam 0
-            0.389139  # output beam 1
+            -5.811451,  # output beam 0
+            -6.63339  # output beam 1
         ],
         np.float32)[np.newaxis, :]
 
@@ -215,11 +215,11 @@ def testCTCDecoderBeamSearch(self):
             [[0, 0], [0, 1]], dtype=np.int64), np.array(
                 [1, 0], dtype=np.int64), np.array(
                     [1, 2], dtype=np.int64)),
-        # beam 1, batch 0, three outputs decoded
+        # beam 1, batch 0, one output decoded
         (np.array(
-            [[0, 0], [0, 1], [0, 2]], dtype=np.int64), np.array(
-                [0, 1, 0], dtype=np.int64), np.array(
-                    [1, 3], dtype=np.int64)),
+            [[0, 0]], dtype=np.int64), np.array(
+                [1], dtype=np.int64), np.array(
+                    [1, 1], dtype=np.int64)),
     ]
 
     # Test correct decoding.