From f95ee9c09fa8459516df47bf3e72f54ab19afa66 Mon Sep 17 00:00:00 2001
From: Wu Yi <typhoonzero1986@gmail.com>
Date: Wed, 12 Dec 2018 20:32:15 +0800
Subject: [PATCH] fix nccl dist test acc (#14867)

* fix nccl dist test acc test=develop

* fix test=develop
---
 .../fluid/tests/unittests/dist_mnist.py       |  2 +-
 .../fluid/tests/unittests/test_dist_base.py   | 21 ++++++++++++-------
 .../fluid/tests/unittests/test_dist_mnist.py  |  2 +-
 3 files changed, 15 insertions(+), 10 deletions(-)

diff --git a/python/paddle/fluid/tests/unittests/dist_mnist.py b/python/paddle/fluid/tests/unittests/dist_mnist.py
index 1cda2711f7..1c45a10a9d 100644
--- a/python/paddle/fluid/tests/unittests/dist_mnist.py
+++ b/python/paddle/fluid/tests/unittests/dist_mnist.py
@@ -93,7 +93,7 @@ class TestDistMnist2x2(TestDistRunnerBase):
         # TODO(typhoonzero): fix distributed adam optimizer
         # opt = fluid.optimizer.AdamOptimizer(
         #     learning_rate=0.001, beta1=0.9, beta2=0.999)
-        opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
+        opt = fluid.optimizer.Momentum(learning_rate=self.lr, momentum=0.9)
 
         # Reader
         train_reader = paddle.batch(
diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py
index 26fa20291b..cedb3383ed 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_base.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_base.py
@@ -32,7 +32,7 @@ DEFAULT_BATCH_SIZE = 2
 
 
 class TestDistRunnerBase(object):
-    def get_model(self, batch_size=DEFAULT_BATCH_SIZE):
+    def get_model(self, batch_size=DEFAULT_BATCH_SIZE, lr=0.1):
         raise NotImplementedError(
             "get_model should be implemented by child classes.")
 
@@ -56,6 +56,7 @@ class TestDistRunnerBase(object):
         return t
 
     def run_pserver(self, args):
+        self.lr = args.lr
         self.get_model(batch_size=args.batch_size)
         # NOTE: pserver should not call memory optimize
         t = self.get_transpiler(args.trainer_id,
@@ -71,6 +72,7 @@ class TestDistRunnerBase(object):
         exe.run(pserver_prog)
 
     def run_trainer(self, args):
+        self.lr = args.lr
         test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
             self.get_model(batch_size=args.batch_size)
 
@@ -189,6 +191,7 @@ def runtime_main(test_class):
     parser.add_argument(
         '--use_reader_alloc', action='store_true', required=False)
     parser.add_argument('--batch_size', required=False, type=int, default=2)
+    parser.add_argument('--lr', required=False, type=float, default=0.001)
     parser.add_argument(
         '--batch_merge_repeat', required=False, type=int, default=1)
 
@@ -234,6 +237,7 @@ class TestDistBase(unittest.TestCase):
         self._dc_asgd = False  # must use with async mode
         self._use_reader_alloc = True
         self._nccl2_mode = False
+        self._lr = 0.001
         self._setup_config()
         self._after_setup_config()
 
@@ -284,7 +288,8 @@ class TestDistBase(unittest.TestCase):
                    batch_size=DEFAULT_BATCH_SIZE,
                    batch_merge_repeat=1):
 
-        cmd = "%s %s --role trainer" % (self._python_interp, model)
+        cmd = "%s %s --role trainer --lr %f" % (self._python_interp, model,
+                                                self._lr)
         if batch_size != DEFAULT_BATCH_SIZE:
             cmd += " --batch_size %d" % batch_size
         if batch_merge_repeat > 1:
@@ -330,13 +335,13 @@ class TestDistBase(unittest.TestCase):
 
         ps0_ep, ps1_ep = self._ps_endpoints.split(",")
 
-        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver"
+        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver --lr %f"
         tr0_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   0, ps0_ep, self._trainers)
+                   0, ps0_ep, self._trainers, self._lr)
         tr1_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   1, ps1_ep, self._trainers)
+                   1, ps1_ep, self._trainers, self._lr)
 
         if self._sync_mode:
             tr0_cmd += " --sync_mode"
@@ -425,13 +430,13 @@ class TestDistBase(unittest.TestCase):
         worker_endpoints = self._ps_endpoints.split(",")
         w0_ep, w1_ep = worker_endpoints
 
-        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method nccl2"
+        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method nccl2 --lr %f"
         tr0_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   0, w0_ep)
+                   0, w0_ep, self._lr / 2)
         tr1_cmd = tr_cmd % \
                   (self._python_interp, model, self._ps_endpoints,
-                   1, w1_ep)
+                   1, w1_ep, self._lr / 2)
 
         if self._mem_opt:
             tr0_cmd += " --mem_opt"
diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist.py b/python/paddle/fluid/tests/unittests/test_dist_mnist.py
index 630bed198f..49a2ca40e3 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_mnist.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_mnist.py
@@ -36,7 +36,7 @@ class TestDistMnistNCCL2(TestDistBase):
     def test_dist_train(self):
         import paddle.fluid as fluid
         if fluid.core.is_compiled_with_cuda():
-            self.check_with_place("dist_mnist.py", delta=1)
+            self.check_with_place("dist_mnist.py", delta=1e-5)
 
 
 class TestDistMnist2x2Lars(TestDistBase):
-- 
GitLab