l2 on word vectors and put back relu and bn in obj atr network

constants_vision_gpu_1.py

@@ -5,7 +5,7 @@ def mkdir_if_not_exists(dir_name):
     if not os.path.exists(dir_name):
         os.mkdir(dir_name)
         
-experiment_name = 'obj_atr_through_ans_mil'
+experiment_name = 'obj_atr_through_ans_mil_wordvecl2'
 #experiment_name = 'object_attribute_classifier_large_images'
 # Global output directory (all subexperiments will be saved here)
 global_output_dir = '/data/tanmay/GenVQA_Exp_Results'
@@ -121,7 +121,7 @@ region_fine_tune_from_iter = 3000
 region_fine_tune_from = region_model + '-' + str(region_fine_tune_from_iter)
 
 # Object Attribute Classifier Evaluation Params
-region_eval_on = 'val' # One of {'val','test','train'}
+region_eval_on = 'train_held_out' # One of {'train_subset','train_held_out','test'}
 region_model_to_eval = region_model + '-' + '77500'
 
 region_attribute_scores_dirname = os.path.join(
@@ -214,7 +214,7 @@ answer_fine_tune_from = answer_model + '-' + str(answer_fine_tune_from_iter)
 
 # Answer eval params
 answer_eval_on = 'val'
-answer_model_to_eval = answer_model + '-69500'
+answer_model_to_eval = answer_model + '-55000'
 
 answer_eval_data_json = os.path.join(
     answer_output_dir,

object_attribute_classifier_cached_features/eval.py

@@ -291,8 +291,8 @@ if __name__=='__main__':
     initializer = create_initializer(
         graph, 
         sess, 
-        constants.pretrained_model)
-        #constants.answer_model_to_eval)
+        #constants.pretrained_model)
+        constants.answer_model_to_eval)
         #constants.region_model_to_eval)
 
     print 'Creating feed dict creator...'

object_attribute_classifier_cached_features/inference.py

@@ -78,10 +78,10 @@ class ObjectAttributeInference():
                     out_dim,
                     'fc',
                     func = None)
-                # fc2_out = layers.batch_norm(
-                #     fc2_out,
-                #     tf.constant(self.training))
-                # fc2_out = tf.nn.relu(fc2_out)
+                fc2_out = layers.batch_norm(
+                    fc2_out,
+                    tf.constant(self.training))
+                fc2_out = tf.nn.relu(fc2_out)
 
         return fc2_out
 
@@ -108,10 +108,10 @@ class ObjectAttributeInference():
                     out_dim,
                     'fc',
                     func = None)
-                # fc2_out = layers.batch_norm(
-                #     fc2_out,
-                #     tf.constant(self.training))
-                # fc2_out = tf.nn.relu(fc2_out)
+                fc2_out = layers.batch_norm(
+                    fc2_out,
+                    tf.constant(self.training))
+                fc2_out = tf.nn.relu(fc2_out)
         
         return fc2_out
     

word2vec/word_vector_management.py

@@ -43,6 +43,7 @@ class word_vector_manager():
                 'word_vec_fc2',
                 func = None)
 
+            tf.add_to_collection('to_regularize',word_vecs_fc2)
         return word_vecs_fc2
         
     def init_word_vector_tensor(self):