Merge pull request #8652 from jacquesqiao/fix-parallel-do-with-non-pa…

…rameter parallel-do should not merge the gradient of parameter that stop_gradient=True
PaddlePaddle · Feb 28, 2018 · be9a2b2 · be9a2b2
2 parents f9c1b6f + c4996fb
commit be9a2b2
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 0 deletions.
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
@@ -220,7 +220,10 @@ def _callback_lookup_(op):
     :return: callback function
     """
     if op.type == 'parallel_do' and op.attr('use_nccl'):
+        all_vars = op.block.vars
         param_names = set(op.input('parameters'))
+        param_names = filter(lambda name: all_vars[name].stop_gradient is False,
+                             param_names)
         param_grad_names = [n + "@GRAD" for n in param_names]
 
         class ParallelDoCallBack(object):

diff --git a/python/paddle/fluid/tests/unittests/test_parallel_op.py b/python/paddle/fluid/tests/unittests/test_parallel_op.py
@@ -170,6 +170,7 @@ def __network__():
         x = fluid.layers.data(shape=[784], dtype='float32', name='img')
         x = yield x
         hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
+        hidden = fluid.layers.batch_norm(input=hidden)
         loss = fluid.layers.mean(hidden)
         yield loss