Merge pull request #16527 from sneaxiy/feature/check_default_grad_maker

Check default grad op maker revision

paddle/fluid/framework/details/op_registry.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <string>
 #include <tuple>
+#include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
@@ -183,6 +184,10 @@ struct OpInfoFiller<T, kGradOpDescMaker> {
       T maker(fwd_op, no_grad_set, grad_to_var, grad_block);
       return maker();
     };
+
+    info->use_default_grad_op_desc_maker_ =
+        std::is_base_of<DefaultGradOpDescMaker<true>, T>::value ||
+        std::is_base_of<DefaultGradOpDescMaker<false>, T>::value;
   }
 };
 

paddle/fluid/framework/grad_op_desc_maker.h

@@ -147,7 +147,7 @@ class SingleGradOpDescMaker : public GradOpDescMakerBase {
  public:
   using GradOpDescMakerBase::GradOpDescMakerBase;
 
-  std::vector<std::unique_ptr<OpDesc>> operator()() const {
+  std::vector<std::unique_ptr<OpDesc>> operator()() const final {
     std::vector<std::unique_ptr<OpDesc>> retv;
     retv.emplace_back(this->Apply());
     return retv;
@@ -158,14 +158,14 @@ class SingleGradOpDescMaker : public GradOpDescMakerBase {
 };
 
 template <bool DropEmptyIG = true>
-class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
+class DefaultGradOpDescMaker final : public SingleGradOpDescMaker {
  public:
   using SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  virtual std::unique_ptr<OpDesc> Apply() const {
+  std::unique_ptr<OpDesc> Apply() const final {
     auto* grad = new OpDesc();
-    grad->SetType(this->GradOpType());
+    grad->SetType(this->ForwardOpType() + "_grad");
 
     for (auto& input_param : this->InputNames()) {
       grad->SetInput(input_param, this->Input(input_param));
@@ -182,18 +182,12 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
 
     return std::unique_ptr<OpDesc>(grad);
   }
-
-  virtual std::string GradOpType() const {
-    return this->ForwardOpType() + "_grad";
-  }
 };
 
-class EmptyGradOpMaker : public GradOpDescMakerBase {
+class EmptyGradOpMaker final : public GradOpDescMakerBase {
  public:
   using GradOpDescMakerBase::GradOpDescMakerBase;
-  std::vector<std::unique_ptr<OpDesc>> operator()() const override {
-    return {};
-  }
+  std::vector<std::unique_ptr<OpDesc>> operator()() const final { return {}; }
 };
 
 }  // namespace framework

paddle/fluid/framework/op_info.cc

@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/op_info.h"
+#include <set>
+#include <string>
+#include <vector>
 
 namespace paddle {
 namespace framework {
@@ -24,5 +27,17 @@ OpInfoMap& OpInfoMap::Instance() {
   static OpInfoMap g_op_info_map;
   return g_op_info_map;
 }
+
+std::vector<std::string> OpInfoMap::GetUseDefaultGradOpDescMakerOps() const {
+  // Use set to sort op names
+  std::set<std::string> result_ops;
+  for (auto& pair : map_) {
+    if (pair.second.use_default_grad_op_desc_maker_) {
+      result_ops.insert(pair.first);
+    }
+  }
+  return std::vector<std::string>(result_ops.begin(), result_ops.end());
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/op_info.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <map>
 #include <string>
 #include <unordered_map>
+#include <vector>
 
 #include "paddle/fluid/framework/attribute.h"
 #include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
@@ -42,6 +43,10 @@ struct OpInfo {
   InferInplaceOpFN infer_inplace_;
   InferNoNeedBufferVarsFN infer_no_need_buffer_vars_;
 
+  // NOTE(zjl): this flag is added to check whether
+  // the grad maker is the default one.
+  bool use_default_grad_op_desc_maker_{false};
+
   bool HasOpProtoAndChecker() const {
     return proto_ != nullptr && checker_ != nullptr;
   }
@@ -105,6 +110,8 @@ class OpInfoMap {
 
   std::unordered_map<std::string, OpInfo>* mutable_map() { return &map_; }
 
+  std::vector<std::string> GetUseDefaultGradOpDescMakerOps() const;
+
  private:
   OpInfoMap() = default;
   std::unordered_map<std::string, OpInfo> map_;

paddle/fluid/op_use_default_grad_op_maker.spec

+abs
+acos
+asin
+atan
+attention_lstm
+bilinear_tensor_product
+brelu
+conv_shift
+cos
+cos_sim
+dequantize
+elementwise_div
+elementwise_max
+elementwise_min
+elu
+fc
+flatten
+fsp
+fused_embedding_fc_lstm
+fused_embedding_seq_pool
+fusion_gru
+fusion_lstm
+fusion_repeated_fc_relu
+fusion_seqconv_eltadd_relu
+fusion_seqexpand_concat_fc
+fusion_seqpool_concat
+fusion_squared_mat_sub
+gelu
+gru
+hard_shrink
+hierarchical_sigmoid
+hinge_loss
+huber_loss
+leaky_relu
+log
+logsigmoid
+lookup_table
+lrn
+lstm_unit
+lstmp
+max_pool2d_with_index
+max_pool3d_with_index
+maxout
+modified_huber_loss
+nce
+norm
+pool2d
+pool3d
+pow
+prelu
+quantize
+rank_loss
+reduce_max
+reduce_mean
+reduce_min
+reduce_prod
+reduce_sum
+requantize
+reshape
+rnn_memory_helper
+round
+row_conv
+sequence_concat
+sequence_conv
+sequence_expand
+sequence_expand_as
+sequence_pad
+sequence_scatter
+sequence_slice
+sequence_softmax
+sequence_unpad
+sigmoid_cross_entropy_with_logits
+sin
+softplus
+softshrink
+softsign
+space_to_depth
+spp
+square
+squared_l2_distance
+squared_l2_norm
+squeeze
+stanh
+swish
+tanh_shrink
+teacher_student_sigmoid_loss
+temporal_shift
+tensor_array_to_tensor
+thresholded_relu
+transpose
+tree_conv
+unpool
+unsqueeze
+warpctc

paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc

@@ -107,17 +107,6 @@ And the output will change the LoD information with input Ids.
   }
 };
 
-class FusedEmbeddingSeqPoolOpGradDescMaker
-    : public framework::DefaultGradOpDescMaker<true> {
-  using ::paddle::framework::DefaultGradOpDescMaker<
-      true>::DefaultGradOpDescMaker;
-
- protected:
-  virtual std::string GradOpType() const {
-    return "fused_embedding_seq_pool_grad";
-  }
-};
-
 class FusedEmbeddingSeqPoolOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -160,7 +149,7 @@ class FusedEmbeddingSeqPoolOpGradVarTypeInference
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(fused_embedding_seq_pool, ops::FusedEmbeddingSeqPoolOp,
-                  ops::FusedEmbeddingSeqPoolOpGradDescMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>,
                   ops::FusedEmbeddingSeqPoolOpMaker);
 REGISTER_OPERATOR(fused_embedding_seq_pool_grad,
                   ops::FusedEmbeddingSeqPoolOpGrad,

paddle/fluid/operators/lookup_table_op.cc

@@ -119,15 +119,6 @@ or not. And the output only shares the LoD information with input Ids.
   }
 };
 
-class LookupTableOpGradDescMaker
-    : public framework::DefaultGradOpDescMaker<true> {
-  using ::paddle::framework::DefaultGradOpDescMaker<
-      true>::DefaultGradOpDescMaker;
-
- protected:
-  virtual std::string GradOpType() const { return "lookup_table_grad"; }
-};
-
 class LookupTableOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -169,7 +160,8 @@ class LookupTableOpGradVarTypeInference : public framework::VarTypeInference {
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(lookup_table, ops::LookupTableOp,
-                  ops::LookupTableOpGradDescMaker, ops::LookupTableOpMaker);
+                  paddle::framework::DefaultGradOpDescMaker<true>,
+                  ops::LookupTableOpMaker);
 REGISTER_OPERATOR(lookup_table_grad, ops::LookupTableOpGrad,
                   ops::LookupTableOpGradVarTypeInference);
 

paddle/fluid/operators/nce_op.cc

@@ -187,14 +187,6 @@ By default this operator uses a uniform distribution for sampling.
   }
 };
 
-class NCEOpGradDescMaker : public framework::DefaultGradOpDescMaker<true> {
-  using ::paddle::framework::DefaultGradOpDescMaker<
-      true>::DefaultGradOpDescMaker;
-
- protected:
-  virtual std::string GradOpType() const { return "nce_grad"; }
-};
-
 class NCEOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -259,7 +251,9 @@ class NCEOpGradVarTypeInference : public framework::VarTypeInference {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(nce, ops::NCEOp, ops::NCEOpGradDescMaker, ops::NCEOpMaker);
+REGISTER_OPERATOR(nce, ops::NCEOp,
+                  paddle::framework::DefaultGradOpDescMaker<true>,
+                  ops::NCEOpMaker);
 REGISTER_OPERATOR(nce_grad, ops::NCEOpGrad, ops::NCEOpGradVarTypeInference);
 REGISTER_OP_CPU_KERNEL(nce, ops::NCEKernel<paddle::platform::CPUPlace, float>,
                        ops::NCEKernel<paddle::platform::CPUPlace, double>);

paddle/fluid/pybind/pybind.cc

@@ -29,6 +29,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/parallel_executor.h"
 #include "paddle/fluid/framework/prune.h"
@@ -156,6 +157,9 @@ PYBIND11_MODULE(core, m) {
         return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
       });
 
+  m.def("_get_use_default_grad_op_desc_maker_ops",
+        [] { return OpInfoMap::Instance().GetUseDefaultGradOpDescMakerOps(); });
+
   // NOTE(zjl): ctest would load environment variables at the beginning even
   // though we have not `import paddle.fluid as fluid`. So we add this API
   // to enable eager deletion mode in unittest.

paddle/scripts/paddle_build.sh

@@ -425,6 +425,13 @@ function assert_api_not_changed() {
     sed -i '/.*ComposeNotAligned.*/d' new.spec
 
     python ${PADDLE_ROOT}/tools/diff_api.py ${PADDLE_ROOT}/paddle/fluid/API.spec new.spec
+
+    # Currently, we only check in PR_CI python 2.7
+    if [ "$SYSTEM" != "Darwin" ]; then
+      if [ "$1" == "" ] || [ "$1" == "cp27-cp27m" ] || [ "$1" == "cp27-cp27mu" ]; then
+        python ${PADDLE_ROOT}/tools/diff_use_default_grad_op_maker.py ${PADDLE_ROOT}/paddle/fluid/op_use_default_grad_op_maker.spec
+      fi
+    fi
     deactivate
 }
 
@@ -434,9 +441,12 @@ function assert_api_spec_approvals() {
     fi
 
     API_FILES=("paddle/fluid/API.spec"
+               "paddle/fluid/op_use_default_grad_op_maker.spec"
                "python/paddle/fluid/parallel_executor.py"
                "paddle/fluid/framework/operator.h"
                "paddle/fluid/framework/tensor.h"
+               "paddle/fluid/framework/details/op_registry.h"
+               "paddle/fluid/framework/grad_op_desc_maker.h"
                "paddle/fluid/framework/lod_tensor.h"
                "paddle/fluid/framework/selected_rows.h"
                "paddle/fluid/framework/op_desc.h"

tools/diff_use_default_grad_op_maker.py

+# 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.
+
+import os
+
+os.environ['CUDA_VISIBLE_DEVICES'] = ''
+
+import paddle.fluid as fluid
+import sys
+
+
+def get_op_diff(filename):
+    ops_created_by_py_func = set(
+        fluid.core._get_use_default_grad_op_desc_maker_ops())
+
+    with open(filename, 'r') as f:
+        ops_read_from_file = set([line.strip() for line in f.readlines()])
+
+    diff_ops = []
+
+    for op in ops_read_from_file:
+        if op not in ops_created_by_py_func:
+            diff_ops.append(op)
+        else:
+            ops_created_by_py_func.remove(op)
+
+    err_msg = []
+    diff_ops = list(diff_ops)
+    if len(diff_ops) > 0:
+        err_msg.append('Added grad op with DefaultGradOpDescMaker: ' + str(
+            diff_ops))
+
+    ops_created_by_py_func = list(ops_created_by_py_func)
+    if len(ops_created_by_py_func) > 0:
+        err_msg.append('Remove grad op with DefaultGradOpDescMaker: ' + str(
+            ops_created_by_py_func))
+
+    return err_msg
+
+
+if len(sys.argv) != 2:
+    print('Usage: python diff_use_default_grad_op_maker.py [filepath]')
+    sys.exit(1)
+
+file_path = str(sys.argv[1])
+err_msg = get_op_diff(file_path)
+
+if len(err_msg) > 0:
+    _, filename = os.path.split(file_path)
+    print('File `{}` is wrong compared to your PR revision!'.format(filename))
+    print(
+        'Please use `python generate_op_use_grad_op_desc_maker_spec.py [filepath]` to generate new `{}` file'.
+        format(filename))
+    print('Error message is: ' + '; '.join(err_msg))
+    sys.exit(1)

tools/generate_op_use_grad_op_desc_maker_spec.py

+# 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.
+
+import os
+
+os.environ['CUDA_VISIBLE_DEVICES'] = ''
+
+import paddle.fluid as fluid
+import sys
+
+if len(sys.argv) != 2:
+    print('Usage: python generate_op_use_grad_op_desc_maker_spec.py [filepath]')
+    sys.exit(1)
+
+with open(sys.argv[1], 'w') as f:
+    ops = fluid.core._get_use_default_grad_op_desc_maker_ops()
+    for op in ops:
+        f.write(op + '\n')