first commit (#5286)

paddle/operators/rnn_memory_helper_op.cc

@@ -0,0 +1,154 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+class RNNMemoryHelperOp : public framework::OperatorBase {
+ public:
+  RNNMemoryHelperOp(const std::string &type,
+                    const framework::VariableNameMap &inputs,
+                    const framework::VariableNameMap &outputs,
+                    const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto mem_var_name = Input("X");
+    auto *mem_var = scope.FindVar(mem_var_name);
+    PADDLE_ENFORCE(mem_var != nullptr,
+                   "Cannot find mem_var in scope, mem_var_name is %s",
+                   mem_var_name);
+
+    auto out_name = this->Output("Out");
+    auto *out_var = scope.FindVar(out_name);
+    PADDLE_ENFORCE(out_var != nullptr,
+                   "Cannot find out_var in scope, out_var_name is %s",
+                   out_name);
+
+    auto *out_tensor = out_var->GetMutable<framework::LoDTensor>();
+    auto &mem_tensor = mem_var->Get<framework::LoDTensor>();
+    out_tensor->ShareDataWith(mem_tensor);
+    out_tensor->set_lod(mem_tensor.lod());
+  }
+};
+
+class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "");
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class RNNMemoryHelperOpInfoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  RNNMemoryHelperOpInfoMaker(framework::OpProto *proto,
+                             framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "");
+    AddOutput("Out", "");
+    AddAttr<int>("data_type",
+                 "(int, default 5 (FP32)) "
+                 "Output data type")
+        .SetDefault(framework::DataType::FP32);
+    AddComment("");
+  }
+};
+
+class RNNMemoryHelperGradOp : public framework::OperatorBase {
+ public:
+  RNNMemoryHelperGradOp(const std::string &type,
+                        const framework::VariableNameMap &inputs,
+                        const framework::VariableNameMap &outputs,
+                        const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto out_grad_var_name = Input(framework::GradVarName("Out"));
+    auto *out_grad_var = scope.FindVar(out_grad_var_name);
+
+    auto in_grad_var_name = Output(framework::GradVarName("X"));
+    auto *in_grad_var = scope.FindVar(in_grad_var_name);
+    PADDLE_ENFORCE(in_grad_var != nullptr,
+                   "Cannot find in_grad_var in scope, name is %s",
+                   in_grad_var_name);
+
+    if (out_grad_var == nullptr) {
+      VLOG(5) << "Using fill constant 0 as starting gradient";
+      auto in_var_name = Input("X");
+      auto *in_var = scope.FindVar(in_var_name);
+      auto &in_var_tensor = in_var->Get<framework::LoDTensor>();
+
+      framework::AttributeMap attrs;
+      attrs["data_type"] = framework::ToDataType(in_var_tensor.type());
+      attrs["shape"] = framework::vectorize2int(in_var_tensor.dims());
+      attrs["value"] = 0.0f;
+
+      auto zero_op = framework::OpRegistry::CreateOp(
+          "fill_constant", {}, {{"Out", {in_grad_var_name}}}, attrs);
+      zero_op->Run(scope, dev_ctx);
+    } else {
+      auto &out_grad_tensor = out_grad_var->Get<framework::LoDTensor>();
+      auto *in_grad_tensor = in_grad_var->GetMutable<framework::LoDTensor>();
+      in_grad_tensor->ShareDataWith(out_grad_tensor);
+      in_grad_tensor->set_lod(out_grad_tensor.lod());
+    }
+  }
+};
+
+class RNNMemoryHelperGradOpInfoMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  RNNMemoryHelperGradOpInfoMaker(framework::OpProto *proto,
+                                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(framework::GradVarName("Out"), "");
+    AddInput("X", "");
+    AddInput("Out", "");
+    AddOutput(framework::GradVarName("X"), "");
+    AddAttr<int>("data_type",
+                 "(int, default 5 (FP32)) "
+                 "Output data type")
+        .SetDefault(framework::DataType::FP32);
+    AddComment("");
+  }
+};
+
+class RNNMemoryHelperGradOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    auto x_grad_name = framework::GradVarName("X");
+    auto out_grad_name = framework::GradVarName("Out");
+    PADDLE_ENFORCE(ctx->HasInput(out_grad_name), "");
+    PADDLE_ENFORCE(ctx->HasOutput(x_grad_name), "");
+    ctx->SetOutputDim(x_grad_name, ctx->GetInputDim(out_grad_name));
+    ctx->ShareLoD(out_grad_name, /*->*/ x_grad_name);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OPERATOR(rnn_memory_helper, paddle::operators::RNNMemoryHelperOp,
+                  paddle::operators::RNNMemoryHelperOpInfoMaker,
+                  paddle::operators::RNNMemoryHelperOpShapeInference,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(rnn_memory_helper_grad,
+                  paddle::operators::RNNMemoryHelperGradOp,
+                  paddle::operators::RNNMemoryHelperGradOpInfoMaker,
+                  paddle::operators::RNNMemoryHelperGradOpShapeInference);

python/paddle/v2/framework/framework.py

@@ -264,7 +264,9 @@ class Operator(object):
                     self.desc.set_attr(attr_name, attrs[attr_name])
 
         self.desc.check_attrs()
-        no_kernel_op_set = {'feed', 'fetch', 'save', 'load'}
+        no_kernel_op_set = {
+            'feed', 'fetch', 'save', 'load', 'rnn_memory_helper_grad'
+        }
         if type not in no_kernel_op_set:
             self.desc.infer_var_type(self.block.desc)
             self.desc.infer_shape(self.block.desc)

python/paddle/v2/framework/tests/test_rnn_memory_helper_op.py

@@ -0,0 +1,130 @@
+import unittest
+
+from paddle.v2.framework.framework import Program
+from paddle.v2.framework.executor import Executor
+from paddle.v2.framework.backward import append_backward_ops
+import numpy as np
+import paddle.v2.framework.core as core
+
+
+def create_tensor(np_data, place):
+    tensor = core.LoDTensor()
+    tensor.set(np_data, place)
+    return tensor
+
+
+class RNNMemoryHelperOpTest(unittest.TestCase):
+    def setUp(self):
+        self.program = Program()
+        self.place = core.CPUPlace()
+
+        self.X = self.program.global_block().create_var(
+            name='X', shape=[2, 3], dtype='float32')
+        self.Out = self.program.global_block().create_var(
+            name='Out', shape=[2, 3], dtype='float32')
+        self.program.global_block().append_op(
+            type='rnn_memory_helper',
+            inputs={"X": self.X},
+            outputs={"Out": self.Out},
+            attrs={})
+
+    def test_forward(self):
+        x_np = np.random.normal(size=(2, 3)).astype("float32")
+        self.feed_map = {'X': create_tensor(x_np, self.place)}
+        self.fetch_list = [self.Out]
+        exe = Executor(self.place)
+        out = exe.run(self.program,
+                      feed=self.feed_map,
+                      fetch_list=self.fetch_list)
+        np.isclose(np.array(out[0]), x_np, rtol=1e-5)
+
+
+class RNNMemoryHelperGradOpTest(unittest.TestCase):
+    def setUp(self):
+        self.program = Program()
+        self.place = core.CPUPlace()
+
+        self.input_names = ['X', 'Out', 'Out@GRAD']
+        self.input_vars = {
+            name: self.program.global_block().create_var(
+                name=name, shape=[2, 3], dtype='float32')
+            for name in self.input_names
+        }
+
+        self.output_names = ['X@GRAD']
+        self.output_vars = {
+            name: self.program.global_block().create_var(
+                name=name, shape=[2, 3], dtype='float32')
+            for name in self.output_names
+        }
+
+        self.program.global_block().append_op(
+            type='rnn_memory_helper_grad',
+            inputs=self.input_vars,
+            outputs=self.output_vars,
+            attrs={})
+
+    def test_backward(self):
+        self.feed_map = {
+            name: create_tensor(
+                np.random.normal(size=(2, 3)).astype("float32"), self.place)
+            for name in self.input_names
+        }
+        self.fetch_list = [self.output_vars['X@GRAD']]
+
+        exe = Executor(self.place)
+        out = exe.run(self.program,
+                      feed=self.feed_map,
+                      fetch_list=self.fetch_list)
+        np.isclose(np.array(out[0]), self.feed_map['Out@GRAD'], rtol=1e-5)
+
+
+class RNNMemoryHelperGradOpWithoutInputTest(unittest.TestCase):
+    def setUp(self):
+        self.program = Program()
+        self.fake_program = Program()
+        self.place = core.CPUPlace()
+
+        self.input_names = ['X', 'Out']
+        self.input_vars = {
+            name: self.program.global_block().create_var(
+                name=name, shape=[2, 3], dtype='float32')
+            for name in self.input_names
+        }
+        self.input_vars["Out@GRAD"] = \
+            self.fake_program.global_block().create_var(
+                name="Out@GRAD", shape=[2, 3], dtype='float32')
+
+        self.output_names = ['X@GRAD']
+        self.output_vars = {
+            name: self.program.global_block().create_var(
+                name=name, shape=[2, 3], dtype='float32')
+            for name in self.output_names
+        }
+
+        self.program.global_block().append_op(
+            type='rnn_memory_helper_grad',
+            inputs=self.input_vars,
+            outputs=self.output_vars,
+            attrs={})
+
+    def test_backward(self):
+        self.feed_map = {
+            name: create_tensor(
+                np.random.normal(size=(2, 3)).astype("float32"), self.place)
+            for name in ['X', 'Out']
+        }
+        self.fetch_list = [self.output_vars['X@GRAD']]
+
+        exe = Executor(self.place)
+        out = exe.run(self.program,
+                      feed=self.feed_map,
+                      fetch_list=self.fetch_list)
+        np.isclose(
+            np.array(out[0]),
+            np.zeros(shape=(2, 3)).astype("float32"),
+            rtol=1e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()