add find_unused_parameters param (#668)

* add find_unused_parameters param

* fix default val

tools/train.py

@@ -70,7 +70,9 @@ def main(args):
         config, parameter_list=net.parameters())
 
     if config["use_data_parallel"]:
-        net = paddle.DataParallel(net)
+        find_unused_parameters = config.get("find_unused_parameters", False)
+        net = paddle.DataParallel(
+            net, find_unused_parameters=find_unused_parameters)
 
     # load model from checkpoint or pretrained model
     init_model(config, net, optimizer)