Merge pull request #2945 from reyoung/feature/grouped_ops

Skeleton Of fully connected operator

paddle/framework/attr_checker.h

@@ -4,6 +4,7 @@
 #include <functional>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 #include "paddle/framework/enforce.h"
 
@@ -41,6 +42,35 @@ class DefaultValueSetter {
   T default_value_;
 };
 
+template <typename T>
+class EnumInContainer {
+ public:
+  explicit EnumInContainer(const std::unordered_set<T>& c) : container_(c) {}
+  void operator()(T& val) const {
+    PADDLE_ENFORCE(container_.find(val) != container_.end(),
+                   "Value %s is not in enum container %s", val,
+                   ContainerDebugString());
+  }
+
+ private:
+  std::string ContainerDebugString() const {
+    std::ostringstream sout;
+    sout << "[";
+    size_t cnt = 0;
+    for (auto& v : container_) {
+      sout << v;
+      ++cnt;
+      if (cnt != container_.size()) {
+        sout << " ,";
+      }
+    }
+    sout << "]";
+    return sout.str();
+  }
+
+  std::unordered_set<T> container_;
+};
+
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
@@ -50,6 +80,11 @@ class TypedAttrChecker {
  public:
   TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
 
+  TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
+    value_checkers_.push_back(EnumInContainer<T>(range));
+    return *this;
+  }
+
   TypedAttrChecker& LargerThan(const T& lower_bound) {
     value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
     return *this;

paddle/framework/net.cc

@@ -19,7 +19,10 @@
 namespace paddle {
 namespace framework {
 
-void PlainNet::CompleteAddOp() {
+void PlainNet::CompleteAddOp(bool calc) {
+  add_op_done_ = true;
+  if (!calc) return;
+
   std::unordered_set<std::string> input_set;
   std::unordered_set<std::string> output_set;
   std::unordered_set<std::string> temp_output;
@@ -52,7 +55,6 @@ void PlainNet::CompleteAddOp() {
   }
 
   attrs_["temporary_index"] = tmp_index;
-  add_op_done_ = true;
 }
 
 std::string PlainNet::DebugString() const {

paddle/framework/net.h

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include <paddle/framework/op_desc.pb.h>
 #include <paddle/framework/operator.h>
-#include "paddle/framework/net_proto.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/scope.h"
@@ -41,7 +40,7 @@ namespace framework {
 class Net : public OperatorBase {
  public:
   virtual void AddOp(const OperatorPtr& op) = 0;
-  virtual void CompleteAddOp() = 0;
+  virtual void CompleteAddOp(bool calc) = 0;
 };
 
 using NetPtr = std::shared_ptr<Net>;
@@ -86,7 +85,7 @@ class PlainNet : public Net {
     ops_.push_back(op);
   }
 
-  void CompleteAddOp() override;
+  void CompleteAddOp(bool calculate = true) override;
 
   std::string DebugString() const override;
 

paddle/operators/CMakeLists.txt

@@ -27,7 +27,8 @@ function(op_library TARGET)
     endif()
 
     list(LENGTH cu_srcs cu_srcs_len)
-    if (${cu_srcs_len} EQUAL 0)
+    list(LENGTH op_library_DEPS dep_len)
+    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
         message(WARNING "The op library ${TARGET} not support GPU!")
     endif()
 
@@ -47,3 +48,6 @@ op_library(mul_op SRCS mul_op.cc mul_op.cu)
 op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
 op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc)
 op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
+
+op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op
+        softmax_op net)

paddle/operators/fc_op.cc

@@ -0,0 +1,76 @@
+/* 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/net.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+class FullyConnectedOp : public framework::PlainNet {
+public:
+  void Init() override {
+    AddOp(framework::OpRegistry::CreateOp("mul",
+                                          {
+                                              Input("X"), Input("W"),
+                                          },
+                                          {Output("before_act")},
+                                          {}));
+    auto b = Input("b");
+    if (b != framework::OperatorBase::EMPTY_VAR_NAME()) {
+      AddOp(framework::OpRegistry::CreateOp("rowwise_add",
+                                            {Output("before_act"), Input("b")},
+                                            {Output("before_act")},
+                                            {}));
+    }
+
+    auto activation = GetAttr<std::string>("activation");
+    AddOp(framework::OpRegistry::CreateOp(
+        activation, {Output("before_act")}, {Output("Y")}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  FullyConnectedOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input of fc operator");
+    AddInput("W", "the weight of fc operator");
+    AddInput("b", "the bias of fc operator");
+
+    AddOutput("Y", "the output of fc operator");
+    AddOutput(
+        "before_act", "the before activation output of fc operator", true);
+    AddAttr<std::string>("activation", "The activation key for fc layer")
+        .SetDefault("sigmoid")
+        .InEnum({"sigmoid", "softmax"});
+
+    //! TODO(yuyang18): Complete comment;
+    AddComment("FullyConnected Operator");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(mul);
+USE_OP(rowwise_add);
+USE_OP(sigmoid);
+USE_OP(softmax);
+
+REGISTER_OP(fc,
+            paddle::operators::FullyConnectedOp,
+            paddle::operators::FullyConnectedOpMaker);

paddle/pybind/CMakeLists.txt

@@ -1,2 +1,2 @@
 cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python
-        add_op mul_op rowwise_add_op sigmoid_op softmax_op)
+        add_op fc_op)

paddle/pybind/pybind.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include <Python.h>
 #include <paddle/framework/op_registry.h>
+#include <paddle/framework/operator.h>
 #include <paddle/framework/scope.h>
 #include <paddle/pybind/tensor_bind.h>
 #include <pybind11/numpy.h>
@@ -26,10 +27,7 @@ namespace py = pybind11;
 namespace pd = paddle::framework;
 
 USE_OP(add_two);
-USE_OP(softmax);
-USE_OP(mul);
-USE_OP(rowwise_add);
-USE_OP(sigmoid);
+USE_OP_WITHOUT_KERNEL(fc);
 
 PYBIND11_PLUGIN(core) {
   py::module m("core", "C++ core of Paddle Paddle");
@@ -53,7 +51,9 @@ PYBIND11_PLUGIN(core) {
              self.mutable_data<int>(paddle::platform::CPUPlace());
            })
       .def("set", paddle::pybind::PyTensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyTensorSetFromArray<int>);
+      .def("set", paddle::pybind::PyTensorSetFromArray<int>)
+      .def("shape",
+           [](pd::Tensor& self) { return pd::vectorize(self.dims()); });
 
   py::class_<pd::Variable>(m, "Variable", R"DOC(Variable Class.
 
@@ -83,15 +83,16 @@ All parameter, weight, gradient are variables in Paddle.
 
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
-  m.def("get_all_op_protos", []() -> std::vector<std::string> {
+  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
     auto& protos = pd::OpRegistry::protos();
-    std::vector<std::string> ret_values;
+    std::vector<py::bytes> ret_values;
     for (auto it = protos.begin(); it != protos.end(); ++it) {
       PADDLE_ENFORCE(it->second.IsInitialized(),
                      "OpProto must all be initialized");
-      ret_values.emplace_back();
-      PADDLE_ENFORCE(it->second.SerializeToString(&ret_values.back()),
+      std::string str;
+      PADDLE_ENFORCE(it->second.SerializeToString(&str),
                      "Serialize OpProto Error. This could be a bug of Paddle.");
+      ret_values.push_back(py::bytes(str));
     }
     return ret_values;
   });
@@ -101,17 +102,26 @@ All parameter, weight, gradient are variables in Paddle.
       .def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
       .def("temp", pd::OperatorBase::TMP_VAR_NAME);
 
+  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
+      .def_static("cpu_context", []() -> paddle::platform::DeviceContext* {
+        return new paddle::platform::CPUDeviceContext();
+      });
+
   py::class_<pd::OperatorBase, pd::OperatorPtr>(m, "Operator")
       .def("__str__", &pd::OperatorBase::DebugString)
-      .def_static("create", [](const std::string& protobin) {
-        pd::OpDesc desc;
-        PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                       "Cannot parse user input to OpDesc");
-        PADDLE_ENFORCE(desc.IsInitialized(),
-                       "User OpDesc is not initialized, reason %s",
-                       desc.InitializationErrorString());
-        return pd::OpRegistry::CreateOp(desc);
-      });
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    pd::OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return pd::OpRegistry::CreateOp(desc);
+                  })
+      .def("infer_shape", &pd::OperatorBase::InferShape)
+      .def("run", &pd::OperatorBase::Run)
+      .def("outputs", [](const pd::OperatorPtr& op) { return op->outputs_; });
 
   return m.ptr();
 }

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -1,3 +1,3 @@
 add_python_test(test_framework test_protobuf.py test_scope.py
     test_default_scope_funcs.py test_op_creation_methods.py
-    test_tensor.py)
+    test_tensor.py test_fc_op.py)

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

@@ -0,0 +1,43 @@
+import paddle.v2.framework.core as core
+import unittest
+import numpy
+import paddle.v2.framework.create_op_creation_methods as creation
+
+
+class TestFc(unittest.TestCase):
+    def test_fc(self):
+        scope = core.Scope(None)
+        x = scope.create_var("X")
+        x_tensor = x.get_tensor()
+        x_tensor.set_dims([1000, 784])
+        x_tensor.alloc_float()
+
+        w = scope.create_var("W")
+        w_tensor = w.get_tensor()
+        w_tensor.set_dims([784, 100])
+        w_tensor.alloc_float()
+
+        w_tensor.set(numpy.random.random((784, 100)).astype("float32"))
+
+        # Set a real numpy array here.
+        # x_tensor.set(numpy.array([]))
+
+        op = creation.op_creations.fc(X="X", Y="Y", W="W")
+
+        for out in op.outputs():
+            if scope.get_var(out) is None:
+                scope.create_var(out).get_tensor()
+
+        tensor = scope.get_var("Y").get_tensor()
+        op.infer_shape(scope)
+        self.assertEqual([1000, 100], tensor.shape())
+
+        ctx = core.DeviceContext.cpu_context()
+
+        op.run(scope, ctx)
+
+        # After complete all ops, check Y is expect or not.
+
+
+if __name__ == '__main__':
+    unittest.main()