From 724d49da1844b8b4bd552c6381e5a9cd87d5abaf Mon Sep 17 00:00:00 2001
From: 0x45f <23097963+0x45f@users.noreply.github.com>
Date: Mon, 17 Jan 2022 10:34:08 +0800
Subject: [PATCH] [Dy2St]close enable_inplace PASS for PE and open
 test_mnist_pure_fp16.py for windows (#38752)

* close enable_inplace PASS for PE, and test dy2st pure fp16 training stability

* add some comment

* enlarge atol
---
 .../unittests/dygraph_to_static/test_mnist_pure_fp16.py   | 8 ++++++--
 .../unittests/dygraph_to_static/test_resnet_pure_fp16.py  | 6 +++++-
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist_pure_fp16.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist_pure_fp16.py
index 029e3e9a53..62878f5cfc 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist_pure_fp16.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist_pure_fp16.py
@@ -33,7 +33,7 @@ class TestPureFP16(TestMNIST):
         return self.train(to_static=False)
 
     def test_mnist_to_static(self):
-        if paddle.fluid.is_compiled_with_cuda() and os.name != 'nt':
+        if paddle.fluid.is_compiled_with_cuda():
             dygraph_loss = self.train_dygraph()
             static_loss = self.train_static()
             # NOTE: In pure fp16 training, loss is not stable, so we enlarge atol here.
@@ -52,7 +52,11 @@ class TestPureFP16(TestMNIST):
 
         if to_static:
             print("Successfully to apply @to_static.")
-            mnist = paddle.jit.to_static(mnist)
+            build_strategy = paddle.static.BuildStrategy()
+            # Why set `build_strategy.enable_inplace = False` here?
+            # Because we find that this PASS strategy of PE makes dy2st training loss unstable.
+            build_strategy.enable_inplace = False
+            mnist = paddle.jit.to_static(mnist, build_strategy=build_strategy)
 
         optimizer = paddle.optimizer.Adam(
             learning_rate=0.001, parameters=mnist.parameters())
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet_pure_fp16.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet_pure_fp16.py
index 6620703ab7..cf5c2b7311 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet_pure_fp16.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet_pure_fp16.py
@@ -106,7 +106,11 @@ def train(to_static, build_strategy=None):
 class TestResnet(unittest.TestCase):
     def train(self, to_static):
         program_translator.enable(to_static)
-        return train(to_static)
+        build_strategy = paddle.static.BuildStrategy()
+        # Why set `build_strategy.enable_inplace = False` here?
+        # Because we find that this PASS strategy of PE makes dy2st training loss unstable.
+        build_strategy.enable_inplace = False
+        return train(to_static, build_strategy)
 
     def test_resnet(self):
         if fluid.is_compiled_with_cuda():
-- 
GitLab