Ref tensorflow#57: Moving to use class weight on logits before cross …

…entropy. The math should work as -weight[class]*x[class] + log( sum ( exp weighted x))

skflow/ops/losses_ops.py

@@ -46,11 +46,11 @@ def softmax_classifier(tensor_in, labels, weights, biases, class_weight=None, na
     """
     with tf.op_scope([tensor_in, labels], name, "softmax_classifier"):
         logits = tf.nn.xw_plus_b(tensor_in, weights, biases)
+        if class_weight:
+            logits = tf.mul(logits, class_weight)
         xent = tf.nn.softmax_cross_entropy_with_logits(logits,
                                                        labels,
                                                        name="xent_raw")
-        if class_weight:
-            xent = tf.mul(xent, class_weight)
         loss = tf.reduce_mean(xent, name="xent")
         predictions = tf.nn.softmax(logits, name=name)
         return predictions, loss

skflow/ops/tests/test_ops.py

@@ -20,6 +20,19 @@
 
 class OpsTest(tf.test.TestCase):
 
+    def test_softmax_classifier(self):
+        with self.test_session() as session:
+            features = tf.placeholder(tf.float32, [None, 3])
+            labels = tf.placeholder(tf.float32, [None, 2])
+            weights = tf.constant([[0.1, 0.1], [0.1, 0.1], [0.1, 0.1]])
+            biases = tf.constant([0.2, 0.3])
+            class_weight = tf.constant([0.1, 0.9])
+            prediction, loss = ops.softmax_classifier(features, labels, weights, biases, class_weight)
+            self.assertEqual(prediction.get_shape()[1], 2)
+            self.assertEqual(loss.get_shape(), [])
+            value = session.run(loss, {features: [[0.2, 0.3, 0.2]], labels: [[0, 1]]})
+            self.assertAllClose(value, 0.55180627)
+
     def test_embedding_lookup(self):
         d_embed = 5
         n_embed = 10