Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into doc_fix

python/paddle/fluid/__init__.py

@@ -44,8 +44,8 @@ import transpiler
 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 @@ class TestDistTranspiler(unittest.TestCase):
         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 @@ class TestDistTranspiler(unittest.TestCase):
         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("fc_w@GRAD.trainer_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_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 @@ class TestSplitVar(unittest.TestCase):
                 # 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:
@@ -49,7 +49,7 @@ class TestSplitVar(unittest.TestCase):
             [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],
@@ -57,7 +57,7 @@ class TestSplitVar(unittest.TestCase):
         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"
 ]

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -39,6 +39,7 @@ Steps to transpile pserver:
 from __future__ import print_function
 
 import math
+import numpy as np
 
 from ps_dispatcher import RoundRobin, HashName, PSDispatcher
 from .. import core, framework
@@ -70,7 +71,7 @@ def same_or_split_var(p_name, var_name):
     return p_name == var_name or p_name.startswith(var_name + ".block")
 
 
-def split_variable(var_list, service_count, min_block_size=8192):
+def slice_variable(var_list, slice_count, min_block_size=8192):
     """
     We may need to split dense tensor to one or more blocks and put
     them equally onto parameter server. One block is a sub-tensor
@@ -78,25 +79,25 @@ def split_variable(var_list, service_count, min_block_size=8192):
 
     We need to have a minimal block size so that the calculations in
     the parameter server side can gain better performance. By default
-    minimum block size 8K elements (maybe 16bit or 32bit or 64bit). 
+    minimum block size 8K elements (maybe 16bit or 32bit or 64bit).
 
     Args:
         var_list (list): List of variables.
-        service_count (int): Numel of pserver services. A pserver may have two
-            or more listening ports.
+        slice_count (int): Numel of count that variables will be sliced, which
+            could be the pserver services' count.
         min_block_size (int): Minimum splitted block size.
     Returns:
-        blocks (list[(varname, block_id, current_block_size)]): A list 
+        blocks (list[(varname, block_id, current_block_size)]): A list
             of VarBlocks. Each VarBlock specifies a shard of the var.
     """
     blocks = []
     for var in var_list:
-        split_count = service_count
+        split_count = slice_count
         var_numel = reduce(lambda x, y: x * y, var.shape)
         max_pserver_count = int(math.floor(var_numel / float(min_block_size)))
         if max_pserver_count == 0:
             max_pserver_count = 1
-        if max_pserver_count < service_count:
+        if max_pserver_count < slice_count:
             split_count = max_pserver_count
         block_size = int(math.ceil(var_numel / float(split_count)))
 
@@ -177,7 +178,7 @@ class DistributeTranspiler:
                     for index in range(len(self.pserver_endpoints))
                 ]
 
-    def _init_splited_vars(self, split_method):
+    def _init_splited_vars(self, slice_var_up):
         # update these mappings for further transpile:
         # 1. param_var_mapping: param var name -> [splited params vars]
         # 2. grad_var_mapping: grad var name -> [splited grads vars]
@@ -196,9 +197,19 @@ class DistributeTranspiler:
         self._update_dist_lookup_table_vars(param_list, grad_list,
                                             self.params_grads)
 
-        grad_blocks = split_variable(grad_list, len(self.pserver_endpoints))
-        param_blocks = split_variable(param_list, len(self.pserver_endpoints))
+        if slice_var_up:
+            # when we slice var up into blocks, we will slice the var according to
+            # pserver services' count. A pserver may have two or more listening ports.
+            grad_blocks = slice_variable(grad_list, len(self.pserver_endpoints))
+            param_blocks = slice_variable(param_list,
+                                          len(self.pserver_endpoints))
+        else:
+            # when we do NOT slice var up into blocks, we will always slice params
+            # grads into one block.
+            grad_blocks = slice_variable(grad_list, 1)
+            param_blocks = slice_variable(param_list, 1)
         assert (len(grad_blocks) == len(param_blocks))
+
         # origin_varname -> [splited_var]
         self.param_var_mapping = self._create_vars_from_blocklist(
             self.origin_program, param_blocks)
@@ -229,6 +240,7 @@ class DistributeTranspiler:
                   program=None,
                   pservers="127.0.0.1:6174",
                   trainers=1,
+                  slice_var_up=True,
                   split_method=RoundRobin,
                   sync_mode=True):
         """
@@ -262,13 +274,27 @@ class DistributeTranspiler:
         self.has_distributed_lookup_table = self._has_distributed_lookup_table()
 
         # split and create vars, then put splited vars in dicts for later use.
-        self._init_splited_vars(split_method)
+        self._init_splited_vars(slice_var_up)
 
         # step 3.1: insert send op to send gradient vars to parameter servers
         ps_dispatcher.reset()
         send_vars = []
-        for orig_varname, splited_vars in self.grad_var_mapping.items():
+
+        # in general cases, the number of pservers is times of 2, and this
+        # will lead to uneven distribution among weights and bias:
+        #       fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1
+        #       fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2
+        # shuffle the map will avoid the uneven distribution above
+        grad_var_mapping_items = self.grad_var_mapping.items()
+        if not slice_var_up:
+            np.random.shuffle(grad_var_mapping_items)
+
+        for orig_varname, splited_vars in grad_var_mapping_items:
             eplist = ps_dispatcher.dispatch(splited_vars)
+
+            if not slice_var_up:
+                assert (len(splited_vars) == 1)
+
             if len(splited_vars) == 1:
                 orig_varname = splited_vars[0].name
                 index = find_op_by_output_arg(program.global_block(),
@@ -316,6 +342,7 @@ class DistributeTranspiler:
         for i, ep in enumerate(eplist):
             self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
             self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
+
         # step4: Concat the parameters splits together after recv.
         for varname, splited_var in self.param_var_mapping.iteritems():
             eps = []
@@ -788,8 +815,8 @@ class DistributeTranspiler:
             program (ProgramDesc): ProgramDesc which gradients blong.
             block_list (list[(varname, block_id, block_size)]): List of gradient blocks.
             add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True.
-        Returns: 
-            var_mapping (dict(varname->[new_varname_variable])):A dict mapping 
+        Returns:
+            var_mapping (dict(varname->[new_varname_variable])):A dict mapping
                 from original var name to each var split.
         """