fix save_inferece_model bug (#15365)

paddle/fluid/framework/ir/CMakeLists.txt

@@ -65,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference)
 pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
+pass_library(identity_scale_op_clean_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will

paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc

@@ -0,0 +1,80 @@
+// Copyright (c) 2019 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.
+
+#include "paddle/fluid/framework/ir/identity_scale_op_clean_pass.h"
+#include <string>
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  FusePassBase::Init("identity_scale_op_clean", graph.get());
+
+  // pre_op -> scale_in -> scale_op -> scale_out
+  // ->
+  // pre_op -> scale_out
+  GraphPatternDetector detector;
+  auto pre_op = detector.mutable_pattern()->NewNode("pre_op")->assert_is_op();
+  auto scale_in = detector.mutable_pattern()
+                      ->NewNode("scale_in")
+                      ->assert_is_op_input("scale")
+                      ->AsIntermediate();
+  auto scale_op = detector.mutable_pattern()
+                      ->NewNode("scale_fuse")
+                      ->assert_is_op("scale")
+                      ->assert_op_attr<float>("scale", 1.)
+                      ->assert_op_attr<float>("bias", 0.);
+  auto scale_out = detector.mutable_pattern()
+                       ->NewNode("scale_out")
+                       ->assert_is_op_output("scale");
+
+  pre_op->LinksTo({scale_in});
+  scale_op->LinksFrom({scale_in}).LinksTo({scale_out});
+
+  GraphPatternDetector::handle_t handler = [&](
+      const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
+    Node* scale_op_var = subgraph.at(scale_op);
+    Node* scale_in_var = subgraph.at(scale_in);
+    Node* scale_out_var = subgraph.at(scale_out);
+    Node* pre_op_var = subgraph.at(pre_op);
+    // Link pre_op directly to scale_out
+    const std::string scale_in_name = scale_in_var->Name();
+    const std::string scale_out_name = scale_out_var->Name();
+    // Remove links in graph
+    GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var});
+    // Modify proto message
+    auto* pre_op_desc = pre_op_var->Op();
+    for (auto& parameter : *pre_op_desc->Proto()->mutable_outputs()) {
+      auto* arguments = parameter.mutable_arguments();
+      auto it = std::find(arguments->begin(), arguments->end(), scale_in_name);
+      PADDLE_ENFORCE(it != arguments->end());
+      *it = scale_out_name;
+    }
+
+    IR_NODE_LINK_TO(pre_op_var, scale_out_var);
+  };
+
+  detector(graph.get(), handler);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(identity_scale_op_clean_pass,
+              paddle::framework::ir::IdentityScaleOpCleanPass);

paddle/fluid/framework/ir/identity_scale_op_clean_pass.h

@@ -0,0 +1,33 @@
+// Copyright (c) 2019 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.
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class IdentityScaleOpCleanPass : public FusePassBase {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+ private:
+  virtual ~IdentityScaleOpCleanPass() = default;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -117,6 +117,7 @@ class CpuPassStrategy : public PassStrategy {
         "conv_bn_fuse_pass",             //
         "conv_eltwiseadd_bn_fuse_pass",  //
         "is_test_pass",                  //
+        "identity_scale_op_clean_pass",  //
     });
     use_gpu_ = false;
   }
@@ -155,6 +156,7 @@ class GpuPassStrategy : public PassStrategy {
   GpuPassStrategy() : PassStrategy({}) {
     passes_.assign({
       "infer_clean_graph_pass",                        //
+          "identity_scale_op_clean_pass",              //
           "conv_affine_channel_fuse_pass",             //
           "conv_eltwiseadd_affine_channel_fuse_pass",  //
           "conv_bn_fuse_pass",                         //

python/paddle/fluid/io.py

@@ -21,9 +21,10 @@ import shutil
 import six
 from functools import reduce
 
+from paddle.fluid import layers
 from paddle.fluid.executor import Executor
 from paddle.fluid.evaluator import Evaluator
-from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable
+from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable, program_guard
 from . import core
 
 __all__ = [
@@ -931,6 +932,17 @@ def save_inference_model(dirname,
     if main_program is None:
         main_program = default_main_program()
 
+    # fix the bug that the activation op's output as target will be pruned.
+    # will affect the inference performance.
+    # TODO(Superjomn) add an IR pass to remove 1-scale op.
+    with program_guard(main_program):
+        uniq_target_vars = []
+        for var in target_vars:
+            if isinstance(var, Variable):
+                var1 = layers.scale(var, 1.)
+            uniq_target_vars.append(var1)
+        target_vars = uniq_target_vars
+
     # when a pserver and a trainer running on the same machine, mkdir may conflict
     try:
         os.makedirs(dirname)

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

@@ -82,7 +82,8 @@ class TestBook(unittest.TestCase):
 
         self.assertEqual(feed_var_names, ["x", "y"])
         self.assertEqual(len(fetch_vars), 1)
-        self.assertEqual(str(fetch_vars[0]), str(avg_cost))
+        print("fetch %s" % str(fetch_vars[0]))
+        self.assertTrue("scale" in str(fetch_vars[0]))
         self.assertEqual(expected, actual)