Merge pull request #11054 from velconia/update_simple_distranspiler

Merge simple_dist_transpiler.py into distribute_transpiler.py

Author: Yancey

python/paddle/fluid/__init__.py

@@ -44,8 +44,8 @@
 from param_attr import ParamAttr, WeightNormParamAttr
 from data_feeder import DataFeeder
 from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace
-from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, \
-    InferenceTranspiler, memory_optimize, release_memory
+from transpiler import DistributeTranspiler, InferenceTranspiler, \
+    memory_optimize, release_memory
 from concurrency import (Go, make_channel, channel_send, channel_recv,
                          channel_close, Select)
 from lod_tensor import create_lod_tensor, create_random_int_lodtensor

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

@@ -12,40 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
 import paddle.fluid as fluid
-import paddle.fluid.core as core
-import paddle.fluid.layers as layers
 from paddle.fluid.transpiler.distribute_transpiler import delete_ops
-import numpy
+
+from transpiler_test import TranspilerTest
 
 
-class TestDistTranspiler(unittest.TestCase):
+class TestDistTranspiler(TranspilerTest):
     def setUp(self):
-        self.trainer_id = 0
-        self.trainers = 2
-        self.pservers = 2
-        self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
         self.current_pserver_ep = "127.0.0.1:6174"
 
-    def net_conf(self):
-        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
-
-        y_predict = fluid.layers.fc(input=x,
-                                    size=1000,
-                                    act=None,
-                                    param_attr=fluid.ParamAttr(name='fc_w'))
-
-        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
-
-        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
-        avg_cost = fluid.layers.mean(cost)
-        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
-
-        optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
-        return optimize_ops, params_grads
-
     def test_transpiler(self):
         trainer = self.get_trainer()
         pserver, startup = self.get_pserver(self.current_pserver_ep)
@@ -70,14 +46,6 @@ def test_transpiler(self):
         fc_w_var = startup.global_block().var("fc_w.block1")
         self.assertEqual(fc_w_var.shape, (500, 1000))
 
-    def get_main_program(self):
-        main = fluid.Program()
-
-        with fluid.program_guard(main):
-            self.net_conf()
-
-        return main
-
     def get_expect_trainer_ops(self):
         trainer = fluid.Program()
 
@@ -92,25 +60,6 @@ def get_expect_trainer_ops(self):
         ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars")
         return ops
 
-    def get_trainer(self):
-        return self._transpiler_instance().get_trainer_program()
-
-    def get_pserver(self, ep):
-        t = self._transpiler_instance()
-        pserver = t.get_pserver_program(ep)
-        startup = t.get_startup_program(ep, pserver)
-        return pserver, startup
-
-    def _transpiler_instance(self):
-        main = self.get_main_program()
-        t = fluid.DistributeTranspiler()
-        t.transpile(
-            self.trainer_id,
-            program=main,
-            pservers=self.pserver_eps,
-            trainers=self.trainers)
-        return t
-
 
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_simple_dist_transpiler.py

@@ -0,0 +1,80 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.fluid.transpiler.distribute_transpiler import delete_ops
+
+from transpiler_test import TranspilerTest
+
+
+class TestSimpleDistTranspiler(TranspilerTest):
+    def setUp(self):
+        self.current_pserver_ep = "127.0.0.1:6175"
+
+    def test_simple_transpiler(self):
+        np.random.seed(1)
+
+        trainer = self.get_trainer()
+        pserver, startup = self.get_pserver(self.current_pserver_ep)
+        self.assertEqual([op.type for op in trainer.global_block().ops],
+                         self.get_expect_trainer_ops())
+
+        self.assertEqual(len(pserver.blocks), 2)
+        # block0: listen_and_serv
+        self.assertEqual([op.type for op in pserver.blocks[0].ops],
+                         ["listen_and_serv"])
+        # block1: optimize pass
+        self.assertEqual([op.type for op in pserver.blocks[1].ops],
+                         ["sum", "scale", "sgd"])
+
+        # confirm startup program
+        self.assertEqual([op.type for op in startup.global_block().ops],
+                         ["fill_constant", "uniform_random", "uniform_random"])
+
+        # the variable #fc_w will NOT be splited
+        fc_w_var = startup.global_block().var("fc_w@GRAD")
+        self.assertEqual(fc_w_var.shape, (1000, 1000))
+
+        fc_w_var = startup.global_block().var("[email protected]_0")
+        self.assertEqual(fc_w_var.shape, (1000, 1000))
+
+    def get_expect_trainer_ops(self):
+        trainer = fluid.Program()
+
+        with fluid.program_guard(trainer):
+            optimize_ops, params_grads = self.net_conf()
+
+        delete_ops(trainer.global_block(), optimize_ops)
+        ops = [op.type for op in trainer.global_block().ops] + [
+            "send_vars", "send_barrier", "recv", "recv", "fetch_barrier"
+        ]
+        ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars")
+        return ops
+
+    def _transpiler_instance(self):
+        main = self.get_main_program()
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            self.trainer_id,
+            program=main,
+            pservers=self.pserver_eps,
+            trainers=self.trainers,
+            slice_var_up=False)
+        return t
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_split_var.py renamed to python/paddle/fluid/tests/unittests/test_slice_var.py

@@ -14,14 +14,14 @@
 
 import math
 import unittest
-from paddle.fluid.transpiler.distribute_transpiler import split_variable
+from paddle.fluid.transpiler.distribute_transpiler import slice_variable
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import random
 
 
-class TestSplitVar(unittest.TestCase):
-    def check_split_output(self, shapes, expected_sizes, min_size):
+class TestSliceVar(unittest.TestCase):
+    def check_slice_output(self, shapes, expected_sizes, min_size):
         var_list = []
         program = fluid.Program()
         for shape in shapes:
@@ -31,7 +31,7 @@ def check_split_output(self, shapes, expected_sizes, min_size):
                 # dtype=core.VarDesc.VarType.LOD_TENSOR,
                 shape=shape)
             var_list.append(var)
-        blocks = split_variable(var_list, 10, min_size)
+        blocks = slice_variable(var_list, 10, min_size)
         all_sizes = []
         for s in expected_sizes:
             for s2 in s:
@@ -49,15 +49,15 @@ def test_1k(self):
             [1150, 1150, 1150, 1150, 1150, 1150, 1100]
         ]
 
-        self.check_split_output(shapes, expected_sizes, 1024)
+        self.check_slice_output(shapes, expected_sizes, 1024)
 
     def test_check_output_8k(self):
         shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10],
                   [6, 33, 33, 33]]
         expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000],
                           [35937, 35937, 35937, 35937, 35937, 35937]]
 
-        self.check_split_output(shapes, expected_sizes, 8192)
+        self.check_slice_output(shapes, expected_sizes, 8192)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/unittests/transpiler_test.py

@@ -0,0 +1,73 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import paddle.fluid.layers as layers
+
+
+class TranspilerTest(unittest.TestCase):
+    @classmethod
+    def setUpClass(self):
+        self.trainer_id = 0
+        self.trainers = 2
+        self.pservers = 2
+        self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
+
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'))
+
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
+
+        optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+        return optimize_ops, params_grads
+
+    def get_main_program(self):
+        main = fluid.Program()
+
+        with fluid.program_guard(main):
+            self.net_conf()
+
+        return main
+
+    def get_trainer(self):
+        return self._transpiler_instance().get_trainer_program()
+
+    def get_pserver(self, ep):
+        t = self._transpiler_instance()
+        pserver = t.get_pserver_program(ep)
+        startup = t.get_startup_program(ep, pserver)
+        return pserver, startup
+
+    def _transpiler_instance(self):
+        main = self.get_main_program()
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            self.trainer_id,
+            program=main,
+            pservers=self.pserver_eps,
+            trainers=self.trainers)
+        return t

python/paddle/fluid/transpiler/__init__.py

@@ -15,10 +15,9 @@
 from distribute_transpiler import DistributeTranspiler
 from inference_transpiler import InferenceTranspiler
 from memory_optimization_transpiler import memory_optimize, release_memory
-from distribute_transpiler_simple import SimpleDistributeTranspiler
 from ps_dispatcher import HashName, RoundRobin
 
 __all__ = [
-    "DistributeTranspiler", "InferenceTranspiler", "SimpleDistributeTranspiler",
-    "memory_optimize", "release_memory", "HashName", "RoundRobin"
+    "DistributeTranspiler", "InferenceTranspiler", "memory_optimize",
+    "release_memory", "HashName", "RoundRobin"
 ]