[GitHub] [incubator-mxnet] anko-intel commented on a change in pull request #18644: Fix BatchNorm backward synchronization

[GitBox <gi...@apache.org>]

anko-intel commented on a change in pull request #18644:
URL: https://github.com/apache/incubator-mxnet/pull/18644#discussion_r448157104



##########
File path: tests/python/unittest/test_gluon.py
##########
@@ -665,6 +665,34 @@ def transpose(shape):
         assert (layer(x).shape==ceil_out_shape)
 
 
+@with_seed()
+@pytest.mark.parametrize('cudnn_off', [True, False])
+@pytest.mark.parametrize('variable', ['running_var', 'running_mean'])
+def test_batchnorm_backward_synchronization(cudnn_off, variable):
+    """
+    Tests if synchronization of BatchNorm running variables is done correctly.
+    If not, the test sometimes fails - depending on the timing.
+    """
+    ctx = mx.cpu() if cudnn_off else mx.gpu()

Review comment:
       Thanks for the tip @wkcn,  
   Is there any command line option or environment value to enforce particular context  when I run pytest locally?

##########
File path: tests/python/unittest/test_gluon.py
##########
@@ -665,6 +665,34 @@ def transpose(shape):
         assert (layer(x).shape==ceil_out_shape)
 
 
+@with_seed()
+@pytest.mark.parametrize('cudnn_off', [True, False])
+@pytest.mark.parametrize('variable', ['running_var', 'running_mean'])
+def test_batchnorm_backward_synchronization(cudnn_off, variable):
+    """
+    Tests if synchronization of BatchNorm running variables is done correctly.
+    If not, the test sometimes fails - depending on the timing.
+    """
+    ctx = mx.cpu() if cudnn_off else mx.gpu()
+    read_op = 'layer.' + variable + '.data().asnumpy()'
+
+    for _ in range(20):
+        layer = nn.BatchNorm()
+        layer.initialize(ctx=ctx)
+        for _ in range(3):
+            data = mx.nd.random.normal(loc=10, scale=2, shape=(1, 3, 10, 10), ctx=ctx)
+            with mx.autograd.record():
+                out = layer(data)
+            out.backward()
+
+        # check if each read give the same value
+        var1 = eval(read_op)

Review comment:
       Thanks!




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org