Add explanation of tests

chainer · shu65 · Aug 24, 2017 · Aug 10, 2017 · Aug 10, 2017 · Aug 10, 2017
commit 71b25cb319720cde11aaafb933e5a8e1e237381c
diff --git a/tests/links_tests/test_batch_normalization.py b/tests/links_tests/test_batch_normalization.py
@@ -63,6 +63,33 @@ def test_version_check(self):
                 3, self.communicator)
 
     def test_multi_node_bn(self):
+        """Tests correctness of MultiNodeBatchNormalization.
+
+        This test verifies MultiNodeBatchNormalization by comparing
+        the following four configurations.
+        (1) Single worker, normal BatchNormalization
+        (2) Multiple workers, normal BatchNormalization
+        (3) Single worker, MultiNodeBatchNormalization
+        (4) Multiple workers, MultiNodeBatchNormalization
+
+        Single worker: only using the result of worker 0, which uses the whole
+            batch.
+        Multiple workers: Each worker uses the 1/n part of the whole batch,
+            where n is the number of nodes, and gradient is aggregated.
+
+        This test conducts the forward and backward computation once for the
+        deterministic model parameters and an input batch, and checks the
+        gradients of parameters.
+
+        The purpose of MultiNodeBatchNormalization is to make the results of
+        (4) to be exactly same as (1). Therefore, the essential part is to
+        check that the results of (1) and (4) are the same. The results of (3)
+        should also be also same as them. In contrast, the results of (2) is
+        not necessarily always same as them, and we can expect that it is
+        almost always different. Therefore, we also check that the results of
+        (2) is different from them, to see that this test working correctly.
+        """
+
         if chainer.__version__.startswith('1.'):
             raise nose.plugins.skip.SkipTest()