Adds a test for Vector output feature

ye-man · Jul 29, 2019 · d460062 · d460062
1 parent c8ecab0
commit d460062
Show file tree

Hide file tree

Showing 3 changed files with 35 additions and 17 deletions.
diff --git a/ludwig/features/base_feature.py b/ludwig/features/base_feature.py
@@ -307,6 +307,8 @@ def concat_dependencies_and_build_output(
         final_hidden[self.name] = (feature_hidden, feature_hidden_size)
 
         # ================ Outputs ================
+        kwargs['is_training'] = is_training
+        kwargs['dropout_rate'] = dropout_rate
         train_mean_loss, eval_loss, output_tensors = self.build_output(
             feature_hidden,
             feature_hidden_size,

diff --git a/ludwig/features/vector_feature.py b/ludwig/features/vector_feature.py
@@ -176,6 +176,8 @@ def __init__(self, feature):
 
         _ = self.overwrite_defaults(feature)
 
+        self.decoder = 'fc_stack'
+        feature['fc_size'] = self.vector_size
         self.decoder_obj = self.get_vector_decoder(feature)
 
     def get_vector_decoder(self, decoder_parameters):
@@ -194,6 +196,7 @@ def _get_output_placeholder(self):
     def _get_measures(self, targets, predictions):
 
         with tf.variable_scope('measures_{}'.format(self.name)):
+            import pdb; pdb.set_trace()
             error_val = get_error(
                 targets,
                 predictions,
@@ -208,18 +211,17 @@ def _get_measures(self, targets, predictions):
                 get_squared_error(targets, predictions, self.name), axis=1
             )
 
-            r2_val = tf.reduce_sum(
-                get_r2(targets, predictions, self.name), axis=1
-            )
+            # TODO - not sure if this is correct
+            r2_val = tf.reduce_sum(get_r2(targets, predictions, self.name))
 
         return error_val, squared_error_val, absolute_error_val, r2_val
 
     def vector_loss(self, error_val, squared_error_val):
         with tf.variable_scope('loss_{}'.format(self.name)):
             if self.loss['type'] == MEAN_SQUARED_ERROR:
-                train_loss = tf.reduce_sum(squared_error_val, axis=1)
+                train_loss = tf.reduce_sum(squared_error_val)
             elif self.loss['type'] == MEAN_ABSOLUTE_ERROR:
-                train_loss = tf.reduce_sum(error_val, axis=1)
+                train_loss = tf.reduce_sum(error_val)
             else:
                 train_mean_loss = None
                 train_loss = None
@@ -246,8 +248,9 @@ def build_output(
             self.decoder_obj,
             hidden,
             hidden_size,
+            kwargs['is_training'],
+            kwargs['dropout_rate'],
             regularizer=regularizer,
-            kwarg=kwargs
         )
         return train_mean_loss, eval_loss, output_tensors
 
@@ -257,8 +260,9 @@ def build_vector_output(
             decoder,
             hidden,
             hidden_size,
+            is_training,
+            dropout_rate,
             regularizer=None,
-            **kwargs
     ):
         feature_name = self.name
         output_tensors = {}
@@ -268,11 +272,12 @@ def build_vector_output(
 
         # ================ Predictions ================
         predictions, predictions_size = self.vector_predictions(
-            targets,
             decoder,
             hidden,
             hidden_size,
-            regularizer=regularizer
+            is_training,
+            dropout_rate,
+            regularizer=regularizer,
         )
 
         output_tensors[PREDICTIONS + '_' + feature_name] = predictions
@@ -317,21 +322,22 @@ def build_vector_output(
 
     def vector_predictions(
             self,
-            targets,
             decoder,
             hidden,
             hidden_size,
+            is_training,
+            dropout_rate,
             regularizer=None,
-            is_timeseries=False
     ):
         with tf.variable_scope('predictions_{}'.format(self.name)):
             output, output_size = decoder(
-                dict(self.__dict__),
-                targets,
+                # dict(self.__dict__),
+                # targets,
                 hidden,
                 hidden_size,
                 regularizer,
-                is_timeseries=is_timeseries
+                dropout_rate,
+                is_training
             )
 
         return output, output_size
@@ -424,8 +430,9 @@ def populate_defaults(output_feature):
 
         set_default_value(output_feature, LOSS, {})
         set_default_value(output_feature[LOSS], 'type', MEAN_SQUARED_ERROR)
-        set_default_value(output_feature, 'reduce_input', SUM)
-        set_default_value(output_feature, 'reduce_dependencies', SUM)
+        set_default_value(output_feature[LOSS], 'weight', 1)
+        set_default_value(output_feature, 'reduce_input', None)
+        set_default_value(output_feature, 'reduce_dependencies', None)
         set_default_value(output_feature, 'decoder', 'fc_stack')
         set_default_value(output_feature, 'dependencies', [])
 

diff --git a/tests/integration_tests/test_experiment.py b/tests/integration_tests/test_experiment.py
@@ -549,7 +549,7 @@ def test_experiment_h3(csv_filename):
         run_experiment(input_features, output_features, data_csv=rel_path)
 
 
-def test_experiment_vector_feature(csv_filename):
+def test_experiment_vector_feature_1(csv_filename):
     input_features = [vector_feature()]
     output_features = [binary_feature()]
     # Generate test data
@@ -558,6 +558,15 @@ def test_experiment_vector_feature(csv_filename):
     run_experiment(input_features, output_features, data_csv=rel_path)
 
 
+def test_experiment_vector_feature_2(csv_filename):
+    input_features = [vector_feature()]
+    output_features = [vector_feature()]
+    # Generate test data
+    rel_path = generate_data(input_features, output_features, csv_filename)
+
+    run_experiment(input_features, output_features, data_csv=rel_path)
+
+
 if __name__ == '__main__':
     """
     To run tests individually, run: