batch norm single features before adding

answer_classifier_cached_features/inference.py

@@ -74,6 +74,7 @@ class AnswerInference():
             obj_atr_qa_feat = [None]*self.batch_size
             q_a_feat_list = [None]*self.batch_size
             obj_atr_det_list = [None]*self.batch_size
+            single_feats_list = [None]*self.batch_size
             for j in xrange(self.batch_size):
                 if j==0:
                     reuse_vars = False
@@ -175,6 +176,14 @@ class AnswerInference():
                     1,
                     [0*yes_no_feat_, q_feat, a_feat])
 
+                single_feats_list[j] = tf.concat(
+                    1,
+                    [
+                        self.per_answer_avg_ans_noun_scores[j], 
+                        self.per_answer_avg_ques_noun_scores[j], 
+                        self.per_answer_avg_ans_adj_scores[j],
+                        self.per_answer_avg_ques_adj_scores[j],
+                    ])
 
                 # obj_atr_qa_feat[j] = tf.expand_dims(
                 #     obj_atr_qa_feat[j],
@@ -187,6 +196,10 @@ class AnswerInference():
             self.q_a_feat_packed = tf.concat(
                 0,
                 q_a_feat_list)
+
+            self.single_feats_packed = tf.concat(
+                0,
+                single_feats_list)
             
             self.obj_atr_det_fc_bn = self.fc_bn(
                 self.obj_atr_det_packed,
@@ -198,9 +211,14 @@ class AnswerInference():
                 2500,
                 'q_a_feat_fc_bn')
 
+            self.single_feats_fc_bn = self.fc_bn(
+                self.single_feats_packed,
+                2500,
+                'single_feats_fc_bn')
+
             # (25*18 x 2500)
             self.per_answer_feat = tf.nn.relu(
-                self.obj_atr_det_fc_bn + self.q_a_feat_fc_bn)
+                self.obj_atr_det_fc_bn + self.q_a_feat_fc_bn + self.single_feats_fc_bn)
             
             # (25*18 x 1)
             self.answer_score = self.fc_bn(