Refactored loss_modules (#653)

ludwig-ai · Mar 19, 2020 · 1b3105b · 1b3105b
1 parent 5ffe45c
commit 1b3105b
Showing 1 changed file with 47 additions and 79 deletions.
diff --git a/ludwig/models/modules/loss_modules.py b/ludwig/models/modules/loss_modules.py
@@ -121,46 +121,8 @@ def cross_entropy_sequence_loss(logits, targets, sequence_length):
 def sampled_softmax_cross_entropy(output_placeholder, feature_hidden, logits,
                                   vector_labels, class_weights,
                                   class_biases, loss, num_classes):
-    output_exp = tf.cast(tf.expand_dims(output_placeholder, -1), tf.int64)
-    if loss['sampler'] == 'fixed_unigram':
-        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes,
-            unigrams=loss['class_counts'],
-            distortion=loss['distortion']
-        )
-    elif loss['sampler'] == 'uniform':
-        sampled_values = tf.nn.uniform_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes
-        )
-    elif loss['sampler'] == 'log_uniform':
-        sampled_values = tf.nn.log_uniform_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes
-        )
-    elif loss['sampler'] == 'learned_unigram':
-        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes,
-            unigrams=loss['class_counts'],
-            distortion=loss['distortion']
-        )
-    else:
-        raise ValueError('Unsupported sampler {}'.format(loss['sampler']))
-
+    output_exp = tf.cast(tf.expand_dims(output_placeholder, -1), tf.int64)    
+    sampled_values = obtained_sampled_values(num_classes, output_exp, loss)
     train_loss = tf.nn.sampled_softmax_loss(weights=tf.transpose(class_weights),
                                             biases=class_biases,
                                             labels=output_exp,
@@ -198,45 +160,7 @@ def sequence_sampled_softmax_cross_entropy(targets, targets_sequence_length,
     # unpadded_targets = targets[:, :batch_max_seq_length]
     # output_exp = tf.cast(tf.reshape(unpadded_targets, [-1, 1]), tf.int64)
     output_exp = tf.cast(tf.reshape(targets, [-1, 1]), tf.int64)
-
-    if loss['sampler'] == 'fixed_unigram':
-        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes,
-            unigrams=loss['class_counts'],
-            distortion=loss['distortion']
-        )
-    elif loss['sampler'] == 'uniform':
-        sampled_values = tf.nn.uniform_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes
-        )
-    elif loss['sampler'] == 'log_uniform':
-        sampled_values = tf.nn.log_uniform_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes
-        )
-    elif loss['sampler'] == 'learned_unigram':
-        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
-            true_classes=output_exp,
-            num_true=1,
-            num_sampled=loss['negative_samples'],
-            unique=loss['unique'],
-            range_max=num_classes,
-            unigrams=loss['class_counts'],
-            distortion=loss['distortion']
-        )
-    else:
-        raise ValueError('Unsupported sampler {}'.format(loss['sampler']))
+    sampled_values = obtained_sampled_values(num_classes, output_exp, loss)
 
     def _sampled_loss(labels, logits):
         labels = tf.cast(labels, tf.int64)
@@ -324,3 +248,47 @@ def loss_multilabel(logits, vector_labels, loss):
                         'sum': tf.contrib.layers.sum_regularizer,
                         'None': lambda x: None,
                         None: lambda x: None}
+
+
+
+def obtained_sampled_values(num_classes, output_exp, loss):
+    """returns sampled_values using the chosen sampler"""
+    if loss['sampler'] == 'fixed_unigram':
+        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
+            true_classes=output_exp,
+            num_true=1,
+            num_sampled=loss['negative_samples'],
+            unique=loss['unique'],
+            range_max=num_classes,
+            unigrams=loss['class_counts'],
+            distortion=loss['distortion']
+        )
+    elif loss['sampler'] == 'uniform':
+        sampled_values = tf.nn.uniform_candidate_sampler(
+            true_classes=output_exp,
+            num_true=1,
+            num_sampled=loss['negative_samples'],
+            unique=loss['unique'],
+            range_max=num_classes
+        )
+    elif loss['sampler'] == 'log_uniform':
+        sampled_values = tf.nn.log_uniform_candidate_sampler(
+            true_classes=output_exp,
+            num_true=1,
+            num_sampled=loss['negative_samples'],
+            unique=loss['unique'],
+            range_max=num_classes
+        )
+    elif loss['sampler'] == 'learned_unigram':
+        sampled_values = tf.nn.fixed_unigram_candidate_sampler(
+            true_classes=output_exp,
+            num_true=1,
+            num_sampled=loss['negative_samples'],
+            unique=loss['unique'],
+            range_max=num_classes,
+            unigrams=loss['class_counts'],
+            distortion=loss['distortion']
+        )
+    else:
+        raise ValueError('Unsupported sampler {}'.format(loss['sampler']))
+    return sampled_values