From 08a772cb462e02209f4bc5d7910579b120dc7d41 Mon Sep 17 00:00:00 2001
From: mapingshuo <mps2012@yeah.net>
Date: Mon, 9 Mar 2020 18:03:03 +0800
Subject: [PATCH] fix API param bug of recompute.backward() (#22582)

* fix API param bug of recompute.backward(), test=develop
---
 python/paddle/fluid/optimizer.py               | 18 ++++++++----------
 .../fluid/tests/unittests/test_optimizer.py    |  3 +--
 2 files changed, 9 insertions(+), 12 deletions(-)

diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
index 6b42b40403..96488bfc96 100644
--- a/python/paddle/fluid/optimizer.py
+++ b/python/paddle/fluid/optimizer.py
@@ -3850,12 +3850,12 @@ class RecomputeOptimizer(Optimizer):
 
                 sgd = fluid.optimizer.Adam(learning_rate=0.01)
                 sgd = fluid.optimizer.RecomputeOptimizer(sgd)
+                sgd._set_checkpoints([fc_1, pred])
                 params_grads = sgd.backward(
                     cost,
                     startup_program=None,
                     parameter_list=None,
-                    no_grad_set=None,
-                    checkpoints=[fc_1, pred])
+                    no_grad_set=None)
 
                 program = cost.block.program
                 with framework.program_guard(program, None):
@@ -3871,8 +3871,7 @@ class RecomputeOptimizer(Optimizer):
                  startup_program=None,
                  parameter_list=None,
                  no_grad_set=None,
-                 callbacks=None,
-                 checkpoints=None):
+                 callbacks=None):
         """
         call append_backward with checkpoints.
 
@@ -3906,12 +3905,12 @@ class RecomputeOptimizer(Optimizer):
     
                 sgd = fluid.optimizer.Adam(learning_rate=0.01)
                 sgd = fluid.optimizer.RecomputeOptimizer(sgd)
+                sgd._set_checkpoints([fc_1, pred])
                 params_grads = sgd.backward(
                     cost,
                     startup_program=None,
                     parameter_list=None,
-                    no_grad_set=None,
-                    checkpoints=[fc_1, pred])
+                    no_grad_set=None)
                 print("Finished backward")
         """
 
@@ -3958,12 +3957,12 @@ class RecomputeOptimizer(Optimizer):
                 
                 sgd = fluid.optimizer.Adam(learning_rate=0.01)
                 sgd = fluid.optimizer.RecomputeOptimizer(sgd)
+                sgd._set_checkpoints([fc_1, pred])
                 params_grads = sgd.backward(
                     cost,
                     startup_program=None,
                     parameter_list=None,
-                    no_grad_set=None,
-                    checkpoints=[fc_1, pred])
+                    no_grad_set=None)
                 
                 optimize_ops = sgd.apply_optimize(
                     cost, startup_program=None, params_grads=params_grads)
@@ -3993,8 +3992,7 @@ class RecomputeOptimizer(Optimizer):
             loss,
             startup_program=startup_program,
             parameter_list=parameter_list,
-            no_grad_set=no_grad_set,
-            checkpoints=self._checkpoints)
+            no_grad_set=no_grad_set)
 
         if grad_clip:
             # TODO(guru4elephant): should add grad_clip for static graph
diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py
index f97c40b6d9..fefcac6ede 100644
--- a/python/paddle/fluid/tests/unittests/test_optimizer.py
+++ b/python/paddle/fluid/tests/unittests/test_optimizer.py
@@ -791,8 +791,7 @@ class TestRecomputeOptimizer(unittest.TestCase):
             mean_out,
             startup_program=None,
             parameter_list=None,
-            no_grad_set=None,
-            checkpoints=[b1_out])
+            no_grad_set=None)
 
         # apply gradient
         program = mean_out.block.program
-- 
GitLab