Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 4c9699c5 编写于 作者: D dongzhihong
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into mul_op

上级 e0395a53 33d502e7
隐藏空白更改
内联 并排
Showing with 609 addition and 648 deletion
CMakeLists.txt
浏览文件 @ 4c9699c5
......@@ -36,8 +36,8 @@ include(simd)
################################ Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." OFF)
option(WITH_MKLML "Compile PaddlePaddle with mklml package." OFF)
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
......
Dockerfile
浏览文件 @ 4c9699c5
......@@ -34,9 +34,6 @@ RUN apt-get update && \
net-tools && \
apt-get clean -y
# paddle is using numpy.flip, which is introduced since 1.12.0
RUN pip --no-cache-dir install 'numpy>=1.12.0'
# Install Go and glide
RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
tar -xz -C /usr/local && \
......@@ -58,13 +55,16 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
# FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
# version util jupyter fixes this issue.
RUN pip install --upgrade pip && \
pip install -U 'protobuf==3.1.0' && \
pip install -U wheel pillow BeautifulSoup && \
pip install -U wheel && \
pip install -U docopt PyYAML sphinx && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark
RUN pip install pre-commit 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
pip install opencv-python
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
# To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
# the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
......
cmake/external/openblas.cmake
浏览文件 @ 4c9699c5
......@@ -73,10 +73,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
ADD_LIBRARY(cblas STATIC ${dummyfile})
IF(${CBLAS_PROVIDER} MATCHES MKL)
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
ENDIF()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
ELSE()
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_DEPENDENCIES(cblas mklml)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
paddle/framework/backward.cc
浏览文件 @ 4c9699c5
......@@ -22,7 +22,7 @@ namespace paddle {
namespace framework {
template <typename Map, typename T>
static void ForEachVarName(Map& names, T callback) {
static void ForEachVarName(const Map& names, T callback) {
for (auto& name : names) {
for (auto& n : name.second) {
if (callback(n)) return;
......@@ -30,6 +30,7 @@ static void ForEachVarName(Map& names, T callback) {
}
}
// return whether all the names + suffixes in the set
static bool AllInSet(
const std::map<std::string, std::vector<std::string>>& names,
const std::string& suffix, const std::unordered_set<std::string>& set) {
......@@ -43,12 +44,12 @@ static bool AllInSet(
static std::shared_ptr<OperatorBase> NOP() {
auto net_op = std::make_shared<operators::NetOp>();
net_op->type_ = "@NOP@";
net_op->SetType("@NOP@");
net_op->CompleteAddOp();
return net_op;
}
// Get backward operator from a forward operator, recursively implementation.
// Get backward operator from a forward operator, a recursive implementation.
//
// no_grad_names the gradient variable names without gradient calculating.
//
......@@ -56,28 +57,31 @@ static std::shared_ptr<OperatorBase> NOP() {
// BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
// pass `uniq_id` through recursive calling.
//
// returns The backward operator. For simple situation, it is a simple
// operator. For complex situation, it is a NetOp.
// returns The backward operator. In a simple situation, it may be a simple
// operator, in a complex situation, it maybe a NetOp.
//
// See Backward.h for details
static std::shared_ptr<OperatorBase> BackwardRecursive(
const OperatorBase& forwardOp,
std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
std::shared_ptr<OperatorBase> BackwardRecursive(
const OperatorBase& forwardOp,
std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
// If all input gradients of forwarding operator do not need to calculate,
// just return an NOP. Not return null ptr because NOP does not take
// too much time for calculation, but it is useful for simplifying logic.
if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
return NOP();
}
// All output gradients of forwarding operator do not need to calculate.
// Then all input gradients cannot be computed at all, and we put them into
// `no_grad_names` set. Return an NOP.
if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
ForEachVarName(forwardOp.inputs_,
if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
ForEachVarName(forwardOp.Inputs(),
[&no_grad_names](const std::string& name) -> bool {
no_grad_names.insert(GradVarName(name));
return false;
......@@ -93,17 +97,17 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
// Map from output gradient variable name to operator's indices in
// backward net. That operator generates that variable.
// backward net's ops_. That operator generates that variable.
std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
size_t local_op_id = 0;
// reversely travel forwardNet
// reversely travel forwardNet and collect all duplicate outputs.
for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
++it, ++local_op_id) {
auto fwd = *it;
auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
net->AddOp(bwd);
ForEachVarName(bwd->outputs_,
ForEachVarName(bwd->Outputs(),
[&dup_output_ops, local_op_id](const std::string& out) {
dup_output_ops[out].emplace_back(local_op_id);
return false;
......@@ -112,45 +116,51 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
// Get unique ID for this method.
auto uid = uniq_id++;
// TODO(dzh): more comment
// multiple operators which have the same output (y for example) may
// overwrite the same y variable when backward, special operations are token
// to handle this case. For each duplicate output, rename it to an alias
// (original name with a offset), append an `add` op for its operator,
// and finally sum all the alias variable to the final output variable y.
using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
std::list<Pos> insert_position;
for (auto& dup_output_op : dup_output_ops) {
const std::string& name = dup_output_op.first;
auto& dup_op = dup_output_op.second;
// no duplicate output
if (dup_op.size() == 1) continue;
std::vector<std::string> dup_outputs;
// process the duplicate outputs
std::vector<std::string> dup_outputs;
for (size_t i = 0; i < dup_op.size(); ++i) {
// rename each duplicate output to an alias
auto op_offset = dup_op[i];
dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
std::to_string(i));
net->ops_[op_offset]->Rename(name, dup_outputs.back());
}
// collect all the offset to append `add` op for each alias
insert_position.push_back(
{dup_op.back(),
OpRegistry::CreateOp(
"add", {{"X", {dup_outputs}}}, {{"Out", {name}}},
{{"input_format",
std::vector<int>{0, static_cast<int>(dup_outputs.size())}}})});
{dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
{{"Out", {name}}}, {})});
}
// make sure the inserted `add` ops follow the BFS order.
insert_position.sort(
[](const Pos& l, const Pos& r) { return l.first > r.first; });
for (auto& pos : insert_position) {
net->InsertOp(pos.first + 1, pos.second);
}
} else {
std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
ForEachVarName(grad_op->inputs_, [&no_grad_names,
&net](std::string& grad_input) {
ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net,
grad_op](const std::string& grad_input) {
if (no_grad_names.count(grad_input)) {
// +1 for \0
std::string prefix = grad_input.substr(
0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
grad_input = prefix + kZeroVarSuffix;
grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
......@@ -160,10 +170,10 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
return false;
});
ForEachVarName(grad_op->outputs_,
[&no_grad_names](std::string& grad_output) {
ForEachVarName(grad_op->Outputs(),
[&no_grad_names, &grad_op](const std::string& grad_output) {
if (no_grad_names.count(grad_output)) {
grad_output = kEmptyVarName;
grad_op->Rename(grad_output, kEmptyVarName);
}
return false;
});
......@@ -173,10 +183,10 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
}
net->AddOp(grad_op);
}
net->type_ = "@GENERATED_BACKWARD@";
net->SetType("@GENERATED_BACKWARD@");
net->CompleteAddOp();
return net;
}
} // namespace framework
// See header for comments
std::shared_ptr<OperatorBase> Backward(
......
paddle/framework/backward_test.cc
浏览文件 @ 4c9699c5
......@@ -28,13 +28,6 @@ using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
class EmptyOp : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
......@@ -155,27 +148,24 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
f::NOP);
REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
f::NOP);
TEST(Backward, simple_op_grad) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
ASSERT_NE(fwd, nullptr);
auto gop = f::OpRegistry::CreateGradOp(*fwd);
ASSERT_EQ(1UL, gop->inputs_.size());
ASSERT_EQ("rowwise_add_grad", gop->type_);
ASSERT_EQ(1UL, gop->Inputs().size());
ASSERT_EQ("rowwise_add_grad", gop->Type());
ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
}
......@@ -211,13 +201,13 @@ TEST(Backward, net_fc_backward_normal) {
ASSERT_EQ(3UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_add = *net->ops_[1];
ASSERT_EQ("rowwise_add_grad", d_add.type_);
ASSERT_EQ("rowwise_add_grad", d_add.Type());
f::OperatorBase &d_mul = *net->ops_[2];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_fc_backward_not_have_b) {
......@@ -237,10 +227,10 @@ TEST(Backward, net_fc_backward_not_have_b) {
ASSERT_EQ(2UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_mul = *net->ops_[1];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_input_of_network_not_need_grad) {
......@@ -294,7 +284,7 @@ TEST(Backward, net_shared_weight) {
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
ASSERT_EQ("add", bwd_net->ops_[2]->type_);
ASSERT_EQ("add", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_register_grad_not_for_network) {
......@@ -335,15 +325,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
ASSERT_EQ(net->ops_.size(), 2UL);
auto &fill_zero = *net->ops_[0];
ASSERT_EQ("fill_zeros_like", fill_zero.type_);
ASSERT_EQ("fill_zeros_like", fill_zero.Type());
ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
ASSERT_EQ("Z", fill_zero.Input("Src"));
ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.type_);
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG
ASSERT_EQ("many_output_op_grad", d_many_out.Type());
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size()); // I/O/OG
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
d_many_out.Input(f::GradVarName("z")));
ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
......@@ -355,9 +345,9 @@ TEST(Backward, op_part_of_input_are_not_need) {
{{"Out", {"out"}}}, {});
auto backward = f::Backward(*fwd, {"a"});
auto &grad_mul = *backward;
ASSERT_EQ(grad_mul.type_, "mul_grad");
ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
ASSERT_EQ(grad_mul.Type(), "mul_grad");
ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
......@@ -395,18 +385,18 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
auto &grad_fc = *bwd_net->ops_[0];
const char *all = paddle::operators::NetOp::kAll;
EXPECT_EQ(grad_fc.inputs_[all].size(),
EXPECT_EQ(grad_fc.Inputs(all).size(),
2UL /* external input number */
+ 1UL /* external output number*/
+ 1UL /* number of gradient of external output*/
+ 2U /* internal variable number*/);
EXPECT_EQ(grad_fc.outputs_[all].size(),
EXPECT_EQ(grad_fc.Outputs(all).size(),
2UL /* input number of mul*/
+ 2UL /* input number of rowwise_add
*/
+ 1UL /* input number of sigmod */);
EXPECT_EQ(bwd_net->ops_[1]->inputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->outputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->inputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->outputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
}
paddle/framework/grad_op_builder.cc
浏览文件 @ 4c9699c5
......@@ -13,23 +13,20 @@ express or implied. See the License for the specific language governing
permissions and limitations under the License. */
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace framework {
enum class OpArgType { IN, OUT };
static void TransOpArg(const OperatorBase* src_op,
OperatorBase::VarNameMap* vars,
const OpArgType& src_type, bool is_grad) {
static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
bool is_grad, OperatorBase::VarNameMap* vars) {
const auto& src_inout =
src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
auto& dst_inout = *vars;
const OpProto& proto = OpProtos().at(src_op->type_);
const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
const auto& src_arg_list =
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
for (const auto& arg : src_arg_list) {
if (arg.no_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
......@@ -43,22 +40,26 @@ static void TransOpArg(const OperatorBase* src_op,
}
OperatorBase* BuildGradOp(const OperatorBase* op) {
auto gop_type_it = OpRegistry::grad_ops().find(op->type_);
PADDLE_ENFORCE(gop_type_it != OpRegistry::grad_ops().end(),
"Operator %s do not register gradient type", op->type_);
auto& grad_op_type = gop_type_it->second;
auto it = OpRegistry::op_info_map().find(op->Type());
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", op->Type());
PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
op->Type());
std::string grad_op_type = it->second.grad_op_type_;
PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
op->Type());
OperatorBase::VarNameMap inputs;
OperatorBase::VarNameMap outputs;
TransOpArg(op, &inputs, OpArgType::IN, false); // I
TransOpArg(op, &inputs, OpArgType::OUT, false); // O
TransOpArg(op, &inputs, OpArgType::OUT, true); // OG
TransOpArg(op, &outputs, OpArgType::IN, true); // IG
auto gop_it = OpRegistry::op_creators().find(grad_op_type);
PADDLE_ENFORCE(gop_it != OpRegistry::op_creators().end(),
"Operator %s 's Gradient %s's creator cannot be found",
op->type_, grad_op_type);
TransOpArg(op, OpArgType::IN, false, &inputs); // I
TransOpArg(op, OpArgType::OUT, false, &inputs); // O
TransOpArg(op, OpArgType::OUT, true, &inputs); // OG
TransOpArg(op, OpArgType::IN, true, &outputs); // IG
return gop_it->second(grad_op_type, inputs, outputs, op->attrs_);
it = OpRegistry::op_info_map().find(grad_op_type);
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", grad_op_type);
return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
}
} // namespace framework
......
paddle/framework/grad_op_builder_test.cc
浏览文件 @ 4c9699c5
......@@ -8,14 +8,6 @@ USE_OP(add_two);
namespace paddle {
namespace framework {
class NOP : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope,
const platform::DeviceContext &dev_ctx) const override {}
};
class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
......@@ -52,8 +44,8 @@ TEST(GradOpBuilder, AddTwo) {
"add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_add_op =
f::OpRegistry::CreateGradOp(*add_op);
EXPECT_EQ(grad_add_op->inputs_.size(), 4UL);
EXPECT_EQ(grad_add_op->outputs_.size(), 2UL);
EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
EXPECT_EQ(grad_add_op->Input("X"), "x");
EXPECT_EQ(grad_add_op->Input("Y"), "y");
EXPECT_EQ(grad_add_op->Input("Out"), "out");
......@@ -62,10 +54,8 @@ TEST(GradOpBuilder, AddTwo) {
EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
}
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
TEST(GradOpBuilder, MutiInOut) {
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
......@@ -76,7 +66,7 @@ TEST(GradOpBuilder, MutiInOut) {
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
ASSERT_EQ(grad_test_op->inputs_.size(), 3UL + 2UL + 2UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
......@@ -90,7 +80,7 @@ TEST(GradOpBuilder, MutiInOut) {
std::vector<std::string>(
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_test_op->outputs_.size(), 3UL);
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
......@@ -109,7 +99,7 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
f::OpRegistry::CreateGradOp(*test_op);
// 'In2' and 'Out2' are ignored in gradient calculating
ASSERT_EQ(grad_test_op->inputs_.size(), 2UL + 1UL + 2UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
std::vector<std::string>({"in3_1", "in3_2"}));
......@@ -121,7 +111,7 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
f::GradVarName("out2"));
ASSERT_EQ(grad_test_op->outputs_.size(), 3UL);
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>(
......
paddle/framework/op_registry.h
浏览文件 @ 4c9699c5
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <algorithm>
#include <atomic>
#include <type_traits>
#include <typeinfo>
#include <unordered_map>
#include <unordered_set>
#include "paddle/framework/attribute.h"
......@@ -119,6 +120,12 @@ class OpProtoAndCheckerMaker {
bool validated_{false};
};
class NOPMaker : public OpProtoAndCheckerMaker {
public:
NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {}
};
class OpRegistry {
using VarNameMap = OperatorBase::VarNameMap;
using OpCreator = std::function<OperatorBase*(
......@@ -126,46 +133,56 @@ class OpRegistry {
const VarNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
public:
template <typename OpType, typename ProtoMakerType>
static void RegisterOp(const std::string& op_type) {
op_creators()[op_type] = [](
const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs) {
return new OpType(type, inputs, outputs, attrs);
};
OpAttrChecker& op_checker = op_checkers()[op_type];
OpProto& op_proto = OpProtos()[op_type];
auto maker = ProtoMakerType(&op_proto, &op_checker);
maker.Validate();
op_proto.set_type(op_type);
PADDLE_ENFORCE(
op_proto.IsInitialized(),
"Fail to initialize %s's OpProto, because %s is not initialized",
op_type, op_proto.InitializationErrorString());
}
struct OpInfo {
OpCreator creator_;
std::string grad_op_type_;
OpProto* proto_;
OpAttrChecker* checker_;
};
template <typename GradOpType>
static void RegisterGradOp(const std::string& op_type,
const std::string& grad_op_type) {
op_creators()[grad_op_type] = [](
const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs) {
return new GradOpType(type, inputs, outputs, attrs);
template <typename OpType, typename ProtoMakerType, typename GradOpType>
static void RegisterOp(const std::string& op_type,
const std::string& grad_op_type) {
PADDLE_ENFORCE(op_info_map().count(op_type) == 0,
"'%s' is registered more than once.", op_type);
OpInfo op_info;
op_info.creator_ = [](const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs,
const AttributeMap& attrs) {
return new OpType(type, inputs, outputs, attrs);
};
grad_ops()[op_type] = grad_op_type;
op_info.grad_op_type_ = grad_op_type;
if (std::type_index(typeid(ProtoMakerType)) !=
std::type_index(typeid(NOPMaker))) {
op_info.proto_ = new OpProto;
op_info.checker_ = new OpAttrChecker;
auto maker = ProtoMakerType(op_info.proto_, op_info.checker_);
maker.Validate();
op_info.proto_->set_type(op_type);
PADDLE_ENFORCE(
op_info.proto_->IsInitialized(),
"Fail to initialize %s's OpProto, because %s is not initialized",
op_type, op_info.proto_->InitializationErrorString());
} else {
op_info.proto_ = nullptr;
op_info.checker_ = nullptr;
}
op_info_map().insert(std::make_pair(op_type, op_info));
// register gradient op
if (!grad_op_type.empty()) {
RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
}
}
static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
const VarNameMap& inputs,
const VarNameMap& outputs,
AttributeMap attrs) {
auto op_create_it = op_creators().find(type);
PADDLE_ENFORCE(op_create_it != op_creators().end(),
"Operator %s cannot be found.", type);
op_checkers().at(type).Check(attrs);
auto op = op_create_it->second(type, inputs, outputs, attrs);
auto it = op_info_map().find(type);
PADDLE_ENFORCE(it != op_info_map().end(),
"Operator '%s' has not been registered.", type);
it->second.checker_->Check(attrs);
auto op = it->second.creator_(type, inputs, outputs, attrs);
return std::shared_ptr<OperatorBase>(op);
}
......@@ -200,49 +217,32 @@ class OpRegistry {
return grad_op;
}
static std::unordered_map<std::string, std::string>& grad_ops() {
static std::unordered_map<std::string, std::string> grad_ops_;
return grad_ops_;
}
static std::unordered_map<std::string, OpCreator>& op_creators() {
static std::unordered_map<std::string, OpCreator> op_creators_;
return op_creators_;
}
private:
static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
return op_checkers_;
static std::unordered_map<std::string, const OpInfo>& op_info_map() {
static std::unordered_map<std::string, const OpInfo> op_info_map_;
return op_info_map_;
}
};
class Registrar {
public:
// In our design, various kinds of classes, e.g., operators and kernels, have
// their corresponding registry and registrar. The action of registration is
// in the constructor of a global registrar variable, which, however, are not
// used in the code that calls package framework, and would be removed from
// the generated binary file by the linker. To avoid such removal, we add
// Touch to all registrar classes and make USE_OP macros to call this
// method. So, as long as the callee code calls USE_OP, the global
// In our design, various kinds of classes, e.g., operators and kernels,
// have their corresponding registry and registrar. The action of
// registration is in the constructor of a global registrar variable, which,
// however, are not used in the code that calls package framework, and would
// be removed from the generated binary file by the linker. To avoid such
// removal, we add Touch to all registrar classes and make USE_OP macros to
// call this method. So, as long as the callee code calls USE_OP, the global
// registrar variable won't be removed by the linker.
void Touch() {}
};
template <typename OpType, typename ProtoMakerType>
template <typename OpType, typename ProtoMakerType, typename GradOpType>
class OpRegistrar : public Registrar {
public:
explicit OpRegistrar(const char* op_type) {
OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
}
};
template <typename GradOpType>
class GradOpRegistrar : public Registrar {
public:
GradOpRegistrar(const char* op_type, const char* grad_op_type) {
OpRegistry::RegisterGradOp<GradOpType>(op_type, grad_op_type);
explicit OpRegistrar(const char* op_type) { OpRegistrar(op_type, ""); }
OpRegistrar(const char* op_type, const char* grad_op_type) {
OpRegistry::RegisterOp<OpType, ProtoMakerType, GradOpType>(op_type,
grad_op_type);
}
};
......@@ -268,30 +268,20 @@ class OpKernelRegistrar : public Registrar {
/**
* Macro to register Operator.
*/
#define REGISTER_OP(op_type, op_class, op_maker_class) \
#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \
grad_op_class) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
static ::paddle::framework::OpRegistrar<op_class, op_maker_class> \
__op_registrar_##op_type##__(#op_type); \
static ::paddle::framework::OpRegistrar<op_class, op_maker_class, \
grad_op_class> \
__op_registrar_##op_type##__(#op_type, #grad_op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
return 0; \
}
/**
* Macro to register Gradient Operator.
*/
#define REGISTER_GRADIENT_OP(op_type, grad_op_type, grad_op_class) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_gradient_op__##op_type##_##grad_op_type, \
"REGISTER_GRADIENT_OP must be called in global namespace"); \
static ::paddle::framework::GradOpRegistrar<grad_op_class> \
__op_gradient_registrar_##op_type##_##grad_op_type##__(#op_type, \
#grad_op_type); \
int TouchOpGradientRegistrar_##op_type() { \
__op_gradient_registrar_##op_type##_##grad_op_type##__.Touch(); \
return 0; \
}
#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP)
/**
* Macro to register OperatorKernel.
......@@ -307,14 +297,6 @@ class OpKernelRegistrar : public Registrar {
return 0; \
}
/**
* Macro to Forbid user register Gradient Operator.
*/
#define NO_GRADIENT(op_type) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_gradient_op__##op_type##_##op_type##_grad, \
"NO_GRADIENT must be called in global namespace")
#define REGISTER_OP_GPU_KERNEL(op_type, ...) \
REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
......@@ -333,23 +315,6 @@ class OpKernelRegistrar : public Registrar {
static int use_op_itself_##op_type##_ __attribute__((unused)) = \
TouchOpRegistrar_##op_type()
// TODO(fengjiayi): Most ops' gradient op have not been compeleted. So we use
// `NO_GRAD` to disable micro USE_OP_GRADIENT(op_type). Otherwise the code can't
// be compiled. `NO_GRAD` should be removed after all gradient ops are
// compeleted.
#define NO_GRAD
#ifndef NO_GRAD
#define USE_OP_GRADIENT(op_type) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_gradient_##op_type, \
"USE_OP_GRADIENT must be called in global namespace"); \
extern int TouchOpGradientRegistrar_##op_type(); \
static int use_op_gradient_##op_type##_ __attribute__((unused)) = \
TouchOpGradientRegistrar_##op_type()
#else
#define USE_OP_GRADIENT(op_type)
#endif
#define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__use_op_kernel_##op_type##_##DEVICE_TYPE##__, \
......@@ -369,18 +334,13 @@ class OpKernelRegistrar : public Registrar {
USE_OP_DEVICE_KERNEL(op_type, GPU)
#endif
#define USE_NO_GRAD_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_KERNEL(op_type)
#define USE_CPU_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_DEVICE_KERNEL(op_type, CPU); \
USE_OP_GRADIENT(op_type)
#define USE_CPU_ONLY_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_DEVICE_KERNEL(op_type, CPU);
#define USE_OP(op_type) \
USE_NO_GRAD_OP(op_type); \
USE_OP_GRADIENT(op_type)
#define USE_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_KERNEL(op_type)
} // namespace framework
} // namespace paddle
paddle/framework/op_registry_test.cc
浏览文件 @ 4c9699c5
......@@ -59,11 +59,10 @@ static void BuildVar(const std::string& param_name,
var->add_arguments(arg_name);
}
}
REGISTER_OP(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
TEST(OpRegistry, CreateOp) {
paddle::framework::OpDesc op_desc;
......
paddle/framework/operator.cc
浏览文件 @ 4c9699c5
......@@ -33,14 +33,6 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
}
#endif
static std::unordered_map<std::string, OpProto>* g_op_protos = nullptr;
std::unordered_map<std::string, OpProto>& OpProtos() {
if (g_op_protos == nullptr) {
g_op_protos = new std::unordered_map<std::string, OpProto>();
}
return *g_op_protos;
}
const std::string& OperatorBase::Input(const std::string& name) const {
auto& ins = Inputs(name);
PADDLE_ENFORCE_EQ(ins.size(), 1UL,
......@@ -149,14 +141,18 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
}
return ret_val;
}
auto it = OpProtos().find(type_);
auto it = OpRegistry::op_info_map().find(type_);
PADDLE_ENFORCE(
it != OpProtos().end(),
it != OpRegistry::op_info_map().end(),
"Operator %s not registered, cannot figure out intermediate outputs",
type_);
PADDLE_ENFORCE(
it->second.proto_ != nullptr,
"Operator %s has no OpProto, cannot figure out intermediate outputs",
type_);
// get all OpProto::Var for outputs
for (auto& o : it->second.outputs()) {
for (auto& o : it->second.proto_->outputs()) {
// ignore all intermediate output
if (o.intermediate()) continue;
auto out = outputs_.find(o.name());
......
paddle/framework/operator.h
浏览文件 @ 4c9699c5
......@@ -50,8 +50,6 @@ inline std::string GradVarName(const std::string& var_name) {
return var_name + kGradVarSuffix;
}
extern std::unordered_map<std::string, OpProto>& OpProtos();
class OperatorBase;
class InferShapeContext;
class ExecutionContext;
......@@ -99,6 +97,8 @@ class OperatorBase {
/// rename inputs outputs name
void Rename(const std::string& old_name, const std::string& new_name);
const VarNameMap& Inputs() const { return inputs_; }
const VarNameMap& Outputs() const { return outputs_; }
//! Get a input with argument's name described in `op_proto`
const std::string& Input(const std::string& name) const;
//! Get a input which has multiple variables.
......@@ -112,10 +112,11 @@ class OperatorBase {
virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
std::string Type() const { return type_; }
const std::string& Type() const { return type_; }
void SetType(const std::string& type) { type_ = type; }
const AttributeMap& Attrs() const { return attrs_; }
public:
protected:
std::string type_;
// NOTE: in case of OpGrad, inputs_ contains:
// I (Inputs)
......@@ -129,6 +130,14 @@ class OperatorBase {
AttributeMap attrs_;
};
class NOP : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {}
};
class InferShapeContext {
public:
InferShapeContext(const OperatorBase& op, const Scope& scope)
......@@ -210,7 +219,7 @@ class InferShapeContext {
[&](const std::string& sub_name) {
auto var = scope_.FindVar(sub_name);
PADDLE_ENFORCE_NOT_NULL(
var, "MultiOutput(%s:%s) should not be nullptr", name,
var, "MultiOutput(%s:%s) should not be nullptr.", name,
sub_name);
return var->GetMutable<T>();
});
......
paddle/framework/operator_test.cc
浏览文件 @ 4c9699c5
......@@ -65,8 +65,9 @@ static void BuildVar(const std::string& param_name,
}
}
REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
TEST(OperatorBase, all) {
paddle::framework::OpDesc op_desc;
......@@ -184,8 +185,9 @@ class CPUKernalMultiInputsTest : public OpKernel {
} // namespace framework
} // namespace paddle
REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_with_kernel,
paddle::framework::CPUKernelTest<float, float>);
......@@ -210,8 +212,9 @@ TEST(OpKernel, all) {
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
}
REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
paddle::framework::CPUKernalMultiInputsTest);
......
paddle/framework/pybind.cc
浏览文件 @ 4c9699c5
......@@ -20,6 +20,7 @@ limitations under the License. */
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor_py.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "paddle/string/to_string.h"
......@@ -30,8 +31,8 @@ limitations under the License. */
namespace py = pybind11;
USE_OP(add_two);
USE_CPU_OP(onehot_cross_entropy);
USE_NO_GRAD_OP(sgd);
USE_CPU_ONLY_OP(onehot_cross_entropy);
USE_OP(sgd);
USE_OP(mul);
USE_OP(mean);
USE_OP(sigmoid);
......@@ -53,15 +54,15 @@ void ExposeOperator(ClassType &m) {
.def("run", &ClassType::type::Run)
.def("type",
[](const typename ClassType::type &op) -> std::string {
return op.type_;
return op.Type();
})
.def("outputs",
[](const typename ClassType::type &op)
-> std::map<std::string, std::vector<std::string>> {
return op.outputs_;
return op.Outputs();
})
.def("inputs",
[](const typename ClassType::type &op) { return op.inputs_; })
[](const typename ClassType::type &op) { return op.Inputs(); })
.def("__str__", &ClassType::type::DebugString)
.def("no_intermediate_outputs",
[](const typename ClassType::type &op) {
......@@ -160,13 +161,16 @@ 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<py::bytes> {
auto &protos = OpProtos();
auto &op_info_map = OpRegistry::op_info_map();
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");
for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) {
const OpProto *proto = it->second.proto_;
if (proto == nullptr) {
continue;
}
PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized");
std::string str;
PADDLE_ENFORCE(it->second.SerializeToString(&str),
PADDLE_ENFORCE(proto->SerializeToString(&str),
"Serialize OpProto Error. This could be a bug of Paddle.");
ret_values.push_back(py::bytes(str));
}
......@@ -229,7 +233,7 @@ All parameter, weight, gradient are variables in Paddle.
net.def_static("create",
[]() -> std::shared_ptr<operators::NetOp> {
auto retv = std::make_shared<operators::NetOp>();
retv->type_ = "plain_net";
retv->SetType("plain_net");
return retv;
})
.def("add_op", &operators::NetOp::AddOp)
......@@ -238,6 +242,11 @@ All parameter, weight, gradient are variables in Paddle.
const std::shared_ptr<operators::NetOp> &net) -> void {
self.AddOp(std::static_pointer_cast<OperatorBase>(net));
})
.def("add_op",
[](operators::NetOp &self,
const std::shared_ptr<operators::RecurrentOp> &rnn) -> void {
self.AddOp(std::static_pointer_cast<OperatorBase>(rnn));
})
.def("complete_add_op", &operators::NetOp::CompleteAddOp)
.def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
self->CompleteAddOp();
......@@ -245,6 +254,29 @@ All parameter, weight, gradient are variables in Paddle.
ExposeOperator(net);
// recurrent_op
py::class_<operators::RecurrentOp, std::shared_ptr<operators::RecurrentOp>>
rnn(m, "RecurrentOp");
rnn.def_static(
"create",
[](py::bytes protobin) -> std::shared_ptr<operators::RecurrentOp> {
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());
auto rnn_op = OpRegistry::CreateOp(desc);
return std::dynamic_pointer_cast<operators::RecurrentOp>(rnn_op);
})
.def("set_stepnet",
[](operators::RecurrentOp &self,
const std::shared_ptr<operators::NetOp> &net) -> void {
self.set_stepnet(net);
});
ExposeOperator(rnn);
m.def("unique_integer", UniqueIntegerGenerator);
m.def("is_compile_gpu", IsCompileGPU);
......
paddle/operators/CMakeLists.txt
浏览文件 @ 4c9699c5
......@@ -44,6 +44,8 @@ endfunction()
add_subdirectory(math)
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
cc_library(net_op SRCS net_op.cc DEPS op_registry)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
......@@ -64,6 +66,5 @@ op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS framework_proto tensor op_registry operator net_op)
cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)
op_library(uniform_random_op
SRCS uniform_random_op.cc uniform_random_op.cu)
paddle/operators/add_op.cc
浏览文件 @ 4c9699c5
......@@ -57,8 +57,7 @@ class AddOpGrad : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker);
REGISTER_GRADIENT_OP(add_two, add_two_grad, ops::AddOpGrad);
REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad);
REGISTER_OP_CPU_KERNEL(add_two,
ops::AddKernel<paddle::platform::CPUPlace, float>);
paddle/operators/cross_entropy_op.cc
浏览文件 @ 4c9699c5
......@@ -68,12 +68,11 @@ OnehotCrossEntropy Operator.
namespace ops = paddle::operators;
REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp,
ops::OnehotCrossEntropyOpMaker);
ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOp);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(onehot_cross_entropy, onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOp);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOpKernel<paddle::platform::CPUPlace, float>);
paddle/operators/fill_zeros_like_op.cc
浏览文件 @ 4c9699c5
......@@ -46,7 +46,8 @@ The output will have the same size with input.
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp,
ops::FillZerosLikeOpMaker);
REGISTER_OP_CPU_KERNEL(
fill_zeros_like,
ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);
paddle/operators/gather.h
浏览文件 @ 4c9699c5
......@@ -29,7 +29,7 @@ void CPUGather(const T* params, const int* indices, const int slice_size,
const int index_size, T* output) {
const size_t slice_bytes = slice_size * sizeof(T);
for (size_t i = 0; i < index_size; ++i) {
for (int i = 0; i < index_size; ++i) {
int index_ = indices[i];
memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes);
}
......@@ -60,7 +60,7 @@ void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
// slice size
int slice_size = 1;
for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
// Gathering
if (platform::is_cpu_place(place)) {
......
paddle/operators/gather_test.cc
浏览文件 @ 4c9699c5
......@@ -35,7 +35,7 @@ TEST(Gather, GatherData) {
p_src = src->mutable_data<int>(make_ddim({3, 4}), CPUPlace());
p_index = index->mutable_data<int>(make_ddim({2}), CPUPlace());
for (size_t i = 0; i < 12; ++i) p_src[i] = i;
for (int i = 0; i < 12; ++i) p_src[i] = i;
p_index[0] = 1;
p_index[1] = 0;
......@@ -43,6 +43,6 @@ TEST(Gather, GatherData) {
Gather<int>(CPUPlace(), src, index, output);
for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
}
paddle/operators/gaussian_random_op.cc
浏览文件 @ 4c9699c5
......@@ -81,5 +81,6 @@ Use to initialize tensor with gaussian random generator.
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(gaussian_random, ops::GaussianRandomOp, ops::GaussianRandomOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(gaussian_random, ops::GaussianRandomOp,
ops::GaussianRandomOpMaker);
REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);
paddle/operators/math/CMakeLists.txt
浏览文件 @ 4c9699c5
if(WITH_MKLML)
set(BLAS_LIB mklml)
else()
set(BLAS_LIB cblas)
endif()
if(WITH_GPU)
nv_library(math_function SRCS math_function.cc math_function.cu DEPS ${BLAS_LIB} device_context)
nv_library(math_function SRCS math_function.cc math_function.cu DEPS cblas device_context)
else()
cc_library(math_function SRCS math_function.cc DEPS ${BLAS_LIB} device_context)
cc_library(math_function SRCS math_function.cc DEPS cblas device_context)
endif()
nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)
paddle/operators/mean_op.cc
浏览文件 @ 4c9699c5
......@@ -54,9 +54,8 @@ class MeanGradOp : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker);
REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp);
REGISTER_OP_CPU_KERNEL(mean,
ops::MeanKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp);
REGISTER_OP_CPU_KERNEL(mean_grad,
ops::MeanGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/mul_op.cc
浏览文件 @ 4c9699c5
......@@ -85,9 +85,7 @@ class MulOpGrad : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad);
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/net_op.cc
浏览文件 @ 4c9699c5
......@@ -29,7 +29,7 @@ void NetOp::CompleteAddOp(bool calc) {
std::set<std::string> input_set;
std::set<std::string> output_set;
for (auto& op : ops_) {
for (auto& ipt : op->inputs_) {
for (auto& ipt : op->Inputs()) {
for (auto& var_name : ipt.second) {
if (!Contains(output_set, var_name)) { // Not other op's output
input_set.insert(var_name);
......@@ -39,7 +39,7 @@ void NetOp::CompleteAddOp(bool calc) {
}
}
for (auto& opt : op->outputs_) {
for (auto& opt : op->Outputs()) {
for (auto& var_name : opt.second) {
output_set.insert(var_name);
}
......
paddle/operators/net_op_test.cc
浏览文件 @ 4c9699c5
......@@ -20,13 +20,6 @@ class TestOp : public framework::OperatorBase {
}
};
class EmptyOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {}
};
template <typename T>
void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
const std::vector<T>& actual) {
......@@ -56,8 +49,8 @@ TEST(OpKernel, all) {
net->CompleteAddOp();
AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
net->inputs_.at(NetOp::kAll));
AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_.at(NetOp::kAll));
net->Inputs(NetOp::kAll));
AssertSameVectorWithoutOrder({"y", "z"}, net->Outputs(NetOp::kAll));
auto final_outs = net->OutputVars(false);
......@@ -67,9 +60,9 @@ TEST(OpKernel, all) {
TEST(NetOp, insert_op) {
NetOp net;
auto op1 = std::shared_ptr<EmptyOp>(
new EmptyOp("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"Out", {"y"}}}, {}));
auto op1 = std::shared_ptr<framework::NOP>(
new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"Out", {"y"}}}, {}));
net.AddOp(op1);
net.InsertOp(0, op1);
ASSERT_EQ(2UL, net.ops_.size());
......
paddle/operators/recurrent_op.cc
浏览文件 @ 4c9699c5
......@@ -36,15 +36,13 @@ void RecurrentAlgorithm::InferShape(const Scope& scope) const {
rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
true /*infer_shape_mode*/);
InitMemories(step_scopes[0], true /*infer_shape_mode*/);
Variable* net = scope.FindVar(arg_->step_net);
PADDLE_ENFORCE(net != nullptr, "failed to get step net");
for (size_t i = 0; i < seq_len_; i++) {
if (i > 0) {
rnn::LinkMemories(step_scopes, arg_->memories, i, -1,
true /*infer_shape_mode*/);
}
net->GetMutable<NetOp>()->InferShape(*step_scopes[i]);
(*stepnet_)->InferShape(*step_scopes[i]);
}
rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_,
true /*infer_shape_mode*/);
......@@ -56,7 +54,6 @@ void RecurrentAlgorithm::Run(const Scope& scope,
rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
false /*infer_shape_mode*/);
InitMemories(step_scopes[0], false /*infer_shape_mode*/);
Variable* net = scope.FindVar(arg_->step_net);
for (size_t step_id = 0; step_id < seq_len_; step_id++) {
// create output alias variables
......@@ -64,7 +61,7 @@ void RecurrentAlgorithm::Run(const Scope& scope,
rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1,
false /*infer_shape_mode*/);
}
net->GetMutable<NetOp>()->Run(*step_scopes[step_id], dev_ctx);
(*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
}
rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_,
false /*infer_shape_mode*/);
......@@ -78,18 +75,16 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
auto step_scopes = step_scopes_var->GetMutable<std::vector<Scope*>>();
// Now all variables in scope must be created outside of op.
auto net_var = scope.FindVar(arg_->step_net);
PADDLE_ENFORCE(net_var != nullptr, "no stepnet called %s in scope",
arg_->step_net);
auto net_op = net_var->GetMutable<NetOp>();
PADDLE_ENFORCE(!net_op->outputs_.empty(), "net_op has no outputs");
PADDLE_ENFORCE_NOT_NULL(stepnet_);
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs");
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs");
if (seq_len_ > step_scopes->size()) {
for (size_t i = step_scopes->size(); i < seq_len_; ++i) {
auto& step_scope = scope.NewScope();
// create step net's temp inputs
for (auto& input : net_op->inputs_) {
for (auto& input : (*stepnet_)->Inputs()) {
// the weight are located in parent scope
for (auto& var_name : input.second) {
if (!step_scope.FindVar(var_name)) {
......@@ -98,7 +93,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
}
}
// create stepnet's outputs
for (const auto& output : net_op->outputs_) {
for (const auto& output : (*stepnet_)->Outputs()) {
for (auto& var_name : output.second) {
step_scope.NewVar(var_name);
}
......@@ -140,9 +135,8 @@ RecurrentOp::RecurrentOp(const std::string& type,
const framework::OperatorBase::VarNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {
std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
rnn::InitArgument(kArgName, arg.get(), *this);
alg_.Init(std::move(arg));
rnn::InitArgument(kArgName, &arg_, *this);
alg_.Init(&arg_, &stepnet_);
}
class RecurrentAlgorithmProtoAndCheckerMaker
......@@ -158,7 +152,6 @@ class RecurrentAlgorithmProtoAndCheckerMaker
.AsDuplicable();
AddInput(name.boot_memories, "variables to initialize memories.")
.AsDuplicable();
AddInput(name.step_net, "network shared by all steps.");
AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
.AsDuplicable();
......@@ -180,14 +173,12 @@ void RecurrentGradientAlgorithm::Run(
auto step_scopes = GetStepScopes(scope);
rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
false /*infer_shape_mode*/);
Variable* net = scope.FindVar(arg_->step_net);
PADDLE_ENFORCE(net != nullptr, "failed to get step net");
for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
if (static_cast<size_t>(step_id) != seq_len_ - 1) {
rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1,
false /*infer_shape_mode*/);
}
net->GetMutable<NetOp>()->Run(*step_scopes[step_id], dev_ctx);
(*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
}
LinkBootMemoryGradients(step_scopes[0], false);
rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_,
......@@ -219,14 +210,12 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
auto step_scopes = GetStepScopes(scope);
rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
true /*infer_shape_mode*/);
Variable* net = scope.FindVar(arg_->step_net);
PADDLE_ENFORCE(net != nullptr, "failed to get step net");
for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
if (static_cast<size_t>(step_id) != seq_len_ - 1) {
rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1,
true /*infer_shape_mode*/);
}
net->GetMutable<NetOp>()->InferShape(*step_scopes[step_id]);
(*stepnet_)->InferShape(*step_scopes[step_id]);
}
rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_,
true /*infer_shape_mode*/);
......@@ -238,13 +227,13 @@ RecurrentGradientOp::RecurrentGradientOp(
const framework::OperatorBase::VarNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {
std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
rnn::InitArgument(kArgName, arg.get(), *this);
alg_.Init(std::move(arg));
rnn::InitArgument(kArgName, &arg_, *this);
alg_.Init(&arg_, &stepnet_);
}
} // namespace operators
} // namespace paddle
REGISTER_OP(recurrent_op, paddle::operators::RecurrentOp,
paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
recurrent_op, paddle::operators::RecurrentOp,
paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker);
paddle/operators/recurrent_op.h
浏览文件 @ 4c9699c5
......@@ -15,6 +15,7 @@
#pragma once
#include "paddle/framework/operator.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/rnn/recurrent_op_utils.h"
namespace paddle {
......@@ -33,7 +34,11 @@ class RecurrentAlgorithm {
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const;
void Init(std::unique_ptr<rnn::Argument> arg) { arg_ = std::move(arg); }
void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
arg_ = arg;
stepnet_ = stepnet;
}
/**
* InferShape must be called before Run.
......@@ -58,7 +63,8 @@ class RecurrentAlgorithm {
void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const;
private:
std::unique_ptr<rnn::Argument> arg_;
std::shared_ptr<NetOp>* stepnet_;
rnn::Argument* arg_;
mutable size_t seq_len_;
};
......@@ -74,7 +80,11 @@ class RecurrentGradientAlgorithm {
* operator.
*/
public:
void Init(std::unique_ptr<rnn::Argument> arg) { arg_ = std::move(arg); }
void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
arg_ = std::move(arg);
stepnet_ = stepnet;
}
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const;
......@@ -95,8 +105,9 @@ class RecurrentGradientAlgorithm {
}
private:
std::unique_ptr<rnn::Argument> arg_;
rnn::Argument* arg_;
mutable size_t seq_len_;
std::shared_ptr<NetOp>* stepnet_;
};
class RecurrentOp final : public framework::OperatorBase {
......@@ -115,10 +126,15 @@ class RecurrentOp final : public framework::OperatorBase {
alg_.Run(scope, dev_ctx);
}
void set_stepnet(std::shared_ptr<NetOp> net) { stepnet_ = net; }
const NetOp* stepnet() const { return stepnet_.get(); }
static const rnn::ArgumentName kArgName;
private:
RecurrentAlgorithm alg_;
rnn::Argument arg_;
std::shared_ptr<NetOp> stepnet_;
};
class RecurrentGradientOp final : public framework::OperatorBase {
......@@ -141,8 +157,13 @@ class RecurrentGradientOp final : public framework::OperatorBase {
static const rnn::ArgumentName kArgName;
void set_stepnet(const std::shared_ptr<NetOp>& net) { stepnet_ = net; }
const NetOp* stepnet() const { return stepnet_.get(); }
private:
RecurrentGradientAlgorithm alg_;
std::shared_ptr<NetOp> stepnet_;
rnn::Argument arg_;
};
} // namespace operators
......
paddle/operators/recurrent_op_test.cc 已删除 100644 → 0
浏览文件 @ e0395a53
/*
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/operators/recurrent_op.h"
#include <glog/logging.h>
#include <gtest/gtest.h>
#include "paddle/framework/ddim.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/tensor.h"
#include "paddle/operators/net_op.h"
namespace paddle {
namespace operators {
using namespace paddle::framework;
class RecurrentGradientAlgorithmTest : public ::testing::Test {
protected:
virtual void SetUp() override {
CreateGlobalVariables();
CreateStepScopes();
CreateStepNet();
CreateRNNGradientAlgorithm();
// segment inputs
SegmentInputs();
// link forward memories
LinkeMemories();
}
virtual void TearDown() override {}
void CreateGlobalVariables() {
// inputs: x
LOG(INFO) << "create global variable x";
Variable* x = scope_.NewVar("x");
DDim dims =
make_ddim({10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/});
x->GetMutable<Tensor>()->mutable_data<float>(dims, platform::CPUPlace());
// inputs: h_boot
LOG(INFO) << "create global variable h_boot";
Variable* h_boot = scope_.NewVar("h_boot");
h_boot->GetMutable<Tensor>()->mutable_data<float>(
make_ddim({20 /*batch size*/, 30 /*input dim*/}), platform::CPUPlace());
// inputs: w
LOG(INFO) << "create global variable w";
Variable* w = scope_.NewVar("rnn/w");
w->GetMutable<Tensor>()->mutable_data<float>(make_ddim({30, 30}),
platform::CPUPlace());
// inputs: h_grad
LOG(INFO) << "create variable h_grad";
Variable* dh = scope_.NewVar("h_grad");
dh->GetMutable<Tensor>()->mutable_data<float>(make_ddim({10, 20, 30}),
platform::CPUPlace());
// inputs: step_scopes
LOG(INFO) << "create variable step_scopes";
scope_.NewVar("step_scopes");
// inputs: step_net
LOG(INFO) << "create variable step_net";
scope_.NewVar("step_net");
// outputs: w_grad
LOG(INFO) << "create global variable w_grad";
scope_.NewVar("rnn/w_grad");
// outputs: x_grad
LOG(INFO) << "create global variable x_grad";
scope_.NewVar("x_grad");
// outputs: h_boot_grad
LOG(INFO) << "create global variable h_boot_grad";
scope_.NewVar("h_boot_grad");
}
void CreateStepScopes() {
auto step_scopes =
scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
for (int i = 0; i < 10; ++i) {
auto& scope = scope_.NewScope();
auto pre_t = scope.NewVar("rnn/pre_h")->GetMutable<Tensor>();
pre_t->mutable_data<float>({20, 30}, platform::CPUPlace());
auto tensor = scope.NewVar("rnn/h")->GetMutable<Tensor>();
tensor->mutable_data<float>({20, 30}, platform::CPUPlace());
// for unit test of ConcatOutputs
auto xg = scope.NewVar("rnn/x_grad")->GetMutable<Tensor>();
xg->mutable_data<float>({20, 30}, platform::CPUPlace());
step_scopes->emplace_back(&scope);
}
// last time step
auto g = (*step_scopes)[9]->NewVar("rnn/h_pre_grad")->GetMutable<Tensor>();
g->mutable_data<float>({20, 30}, platform::CPUPlace());
}
void CreateRNNGradientAlgorithm() {
std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
arg->step_net = "step_net";
arg->step_scopes = "step_scopes";
rnn::Link inlink;
inlink.external = "h_grad";
inlink.internal = "rnn/h_grad";
arg->inlinks = std::vector<rnn::Link>{inlink};
rnn::Link outlink;
outlink.external = "x_grad";
outlink.internal = "rnn/x_grad";
arg->outlinks = std::vector<rnn::Link>{outlink};
rnn::MemoryAttr mem_attr;
mem_attr.pre_var = "rnn/h_pre_grad";
mem_attr.var = "rnn/h_grad";
mem_attr.boot_var = "h_boot_grad";
arg->memories = std::vector<rnn::MemoryAttr>{mem_attr};
rnn_grad_algo_.Init(std::move(arg));
}
void CreateStepNet() {
LOG(INFO) << "create variable step_net";
Variable* var = scope_.NewVar("step_net");
auto net = var->GetMutable<NetOp>();
// TODO(qingqing) modify backward op create for RNNOp unit test
// and the unit test will be removed to Python.
// net->AddOp(OpRegistry::CreateOp("mul", {"X", {"rnn/h_pre", "rnn/w",
// "rnn/s_grad"}}, {"Y", {"rnn/h_pre_grad", "rnn/w_grad"}}, {}));
// net->AddOp(OpRegistry::CreateOp("add_two", {"X", {"rnn/h_grad"}},
// {"Y", {"rnn/x_grad"}}, {"Out", "rnn/s_grad"}}, {}));
net->CompleteAddOp();
}
void SegmentInputs() {
LOG(INFO) << "segment inputs";
std::vector<std::string> inlinks = {"x"};
std::vector<std::string> inlinks_alias = {"rnn/x"};
rnn::Link inlink;
inlink.external = "x";
inlink.internal = "rnn/x";
auto step_scopes =
scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
rnn::SegmentInputs(*step_scopes, std::vector<rnn::Link>{inlink}, 10,
true /*infer_shape_mode*/);
}
void LinkeMemories() {
LOG(INFO) << "link memories";
rnn::MemoryAttr mem_attr;
mem_attr.pre_var = "rnn/h_pre";
mem_attr.var = "rnn/h";
mem_attr.boot_var = "boot_h";
std::vector<rnn::MemoryAttr> memories;
memories.push_back(mem_attr);
auto step_scopes =
scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
for (int i = 1; i < 10; ++i) {
rnn::LinkMemories(*step_scopes, memories, i, -1,
true /*infer_shape_mode*/);
}
}
Scope scope_;
RecurrentGradientAlgorithm rnn_grad_algo_;
};
// TEST_F(RecurrentGradientAlgorithmTest, Run) {
// platform::CPUDeviceContext ctx;
// rnn_grad_algo_.Run(scope_, ctx);
// }
} // namespace operators
} // namespace paddle
TEST(RecurrentOp, LinkMemories) {
using namespace paddle::framework;
using namespace paddle::platform;
using namespace paddle::operators;
// create and init step scopes
size_t len = 10;
std::vector<Scope*> step_scopes;
for (size_t i = 0; i < len; ++i) {
auto scope = new Scope();
scope->NewVar("pre_h");
auto tensor = scope->NewVar("h")->GetMutable<Tensor>();
float* data = tensor->mutable_data<float>({15, 20}, CPUPlace());
for (size_t j = 0; j < 15 * 20; ++j) {
data[j] = rand() * (1. / (double)RAND_MAX);
}
step_scopes.push_back(scope);
}
// create MemoryAttr
rnn::MemoryAttr mem_attr;
mem_attr.pre_var = "pre_h";
mem_attr.var = "h";
mem_attr.boot_var = "boot_h";
std::vector<rnn::MemoryAttr> memories;
memories.push_back(mem_attr);
for (size_t i = 1; i < len; ++i) {
rnn::LinkMemories(step_scopes, memories, i, -1, false
/*infer_shape_mode*/);
}
// check
for (size_t i = 0; i < len - 1; ++i) {
const float* a =
step_scopes[i]->FindVar("h")->GetMutable<Tensor>()->data<float>();
const float* b = step_scopes[i + 1]
->FindVar("pre_h")
->GetMutable<Tensor>()
->data<float>();
for (size_t j = 0; j < 15 * 20; ++j) {
ASSERT_FLOAT_EQ(a[j], b[j]);
}
}
for (int i = len - 2; i >= 0; --i) {
rnn::LinkMemories(step_scopes, memories, i, 1, false
/*infer_shape_mode*/);
}
// check
for (int i = len - 2; i >= 0; --i) {
const float* a =
step_scopes[i]->FindVar("pre_h")->GetMutable<Tensor>()->data<float>();
const float* b =
step_scopes[i + 1]->FindVar("h")->GetMutable<Tensor>()->data<float>();
for (size_t j = 0; j < 15 * 20; ++j) {
ASSERT_FLOAT_EQ(a[j], b[j]);
}
}
for (auto s : step_scopes) {
delete s;
}
}
USE_OP(add_two);
USE_OP(mul);
USE_OP_ITSELF(recurrent_op);
paddle/operators/rnn/recurrent_op_utils.cc
浏览文件 @ 4c9699c5
......@@ -106,7 +106,6 @@ void LinkMemories(const std::vector<Scope*>& scopes,
void InitArgument(const ArgumentName& name, Argument* arg,
const framework::OperatorBase& op) {
arg->step_net = op.Input(name.step_net);
arg->step_scopes = op.Output(name.step_scopes);
auto inlinks = op.Inputs(name.inlinks);
......
paddle/operators/rowwise_add_op.cc
浏览文件 @ 4c9699c5
......@@ -54,6 +54,7 @@ for i in xrange(X.shape[0]):
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(rowwise_add, ops::RowWiseAddOp, ops::RowWiseAddOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(rowwise_add, ops::RowWiseAddOp,
ops::RowWiseAddOpMaker);
REGISTER_OP_CPU_KERNEL(
rowwise_add, ops::RowWiseAddKernel<paddle::platform::CPUPlace, float>);
paddle/operators/scatter.h 0 → 100644
浏览文件 @ 4c9699c5
/* 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. */
#pragma once
#include <cstring>
#include "paddle/framework/ddim.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/place.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
// Implementation of CPU copy
template <typename T>
void CPUScatterUpdate(const paddle::framework::Tensor* src, const int* index,
const size_t index_size,
paddle::framework::Tensor* output) {
paddle::framework::DDim output_dims = output->dims();
for (size_t i = 0; i < index_size; ++i) {
int index_ = index[i];
paddle::framework::Tensor src_ = *src;
paddle::framework::Tensor output_ = *output;
if (index_size > 1) src_ = src->Slice<T>(i, i + 1);
if (output_dims[0] > 1) output_ = output->Slice<T>(index_, index_ + 1);
auto X = EigenVector<T>::Flatten(src_);
auto Y = EigenVector<T>::Flatten(output_);
Y = X + Y;
}
}
// Implementation of GPU scatter:
template <typename T>
void GPUScatterUpdate(const T* src, const int* index, const int slice_size,
const int index_size, T* output);
/**
* Return a updated tensor from source tensor, scattered according to index:
* dst[i] += src[index[i]]
* input[src]: type-T source Tensor
* input[index]: type-int index Tensor (1-D)
* return: output tensor
*/
template <typename T>
void ScatterUpdate(const platform::Place& place,
const paddle::framework::Tensor* src,
const paddle::framework::Tensor* index,
paddle::framework::Tensor* output) {
// check index of shape 1-D
PADDLE_ENFORCE(index->dims().size() == 1);
int index_size = index->dims()[0];
auto src_dims = src->dims();
auto dst_dims = output->dims();
// check src shape and dst shape should match
for (int i = 1; i < src_dims.size(); i++)
PADDLE_ENFORCE(src_dims[i] == dst_dims[i]);
// slice size
size_t slice_size = 1;
for (int i = 0; i < src_dims.size(); ++i) slice_size *= src_dims[i];
if (platform::is_cpu_place(place)) {
CPUScatterUpdate<T>(src, index->data<int>(), index_size, output);
} else {
}
}
} // namespace operators
} // namespace paddle
paddle/operators/scatter_test.cc 0 → 100644
浏览文件 @ 4c9699c5
/* 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/operators/scatter.h"
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/place.h"
#include <gtest/gtest.h>
#include <iostream>
#include <string>
TEST(scatter, ScatterUpdate) {
using namespace paddle::framework;
using namespace paddle::platform;
using namespace paddle::operators;
Tensor* src = new Tensor();
Tensor* index = new Tensor();
Tensor* output = new Tensor();
float* p_src = nullptr;
int* p_index = nullptr;
p_src = src->mutable_data<float>(make_ddim({1, 4}), CPUPlace());
p_index = index->mutable_data<int>(make_ddim({1}), CPUPlace());
for (size_t i = 0; i < 4; ++i) p_src[i] = float(i);
p_index[0] = 1;
float* p_output = output->mutable_data<float>(make_ddim({4, 4}), CPUPlace());
ScatterUpdate<float>(CPUPlace(), src, index, output);
for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], float(0));
for (size_t i = 0; i < 4; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], float(i - 4));
for (size_t i = 4; i < 8; ++i)
EXPECT_EQ(output->data<float>()[i], float(i - 4));
for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0));
for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
}
paddle/operators/sgd_op.cc
浏览文件 @ 4c9699c5
......@@ -51,6 +51,6 @@ param_out = param - learning_rate * grad;
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker);
REGISTER_OP_CPU_KERNEL(sgd,
ops::SGDOpKernel<paddle::platform::CPUPlace, float>);
paddle/operators/sigmoid_op.cc
浏览文件 @ 4c9699c5
......@@ -52,9 +52,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad);
REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad,
ops::SigmoidOpGrad);
REGISTER_OP_CPU_KERNEL(sigmoid,
ops::SigmoidKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
......
paddle/operators/softmax_op.cc
浏览文件 @ 4c9699c5
......@@ -62,9 +62,9 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
namespace ops = paddle::operators;
REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad,
ops::SoftmaxOpGrad);
REGISTER_OP_CPU_KERNEL(softmax,
ops::SoftmaxKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
REGISTER_OP_CPU_KERNEL(
softmax_grad, ops::SoftmaxGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/uniform_random_op.cc
浏览文件 @ 4c9699c5
......@@ -81,7 +81,7 @@ Used to initialize tensor with uniform random generator.
} // namespace operators
} // namespace paddle
REGISTER_OP(uniform_random, paddle::operators::UniformRandomOp,
paddle::operators::UniformRandomOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(uniform_random, paddle::operators::UniformRandomOp,
paddle::operators::UniformRandomOpMaker);
REGISTER_OP_CPU_KERNEL(uniform_random,
paddle::operators::CPUUniformRandomKernel<float>);
paddle/platform/enforce.h
浏览文件 @ 4c9699c5
......@@ -14,14 +14,21 @@ limitations under the License. */
#pragma once
#include <execinfo.h>
#include <dlfcn.h> // for dladdr
#include <execinfo.h> // for backtrace
#include <iomanip>
#include <memory>
#include <sstream>
#include <stdexcept>
#include <string>
#include "paddle/string/printf.h"
#include "paddle/string/to_string.h"
#ifdef __GNUC__
#include <cxxabi.h> // for __cxa_demangle
#endif
#ifndef PADDLE_ONLY_CPU
#include "paddle/platform/dynload/cublas.h"
......@@ -39,6 +46,19 @@ limitations under the License. */
namespace paddle {
namespace platform {
namespace {
#ifdef __GNUC__
inline std::string demangle(std::string name) {
int status = -4; // some arbitrary value to eliminate the compiler warning
std::unique_ptr<char, void (*)(void*)> res{
abi::__cxa_demangle(name.c_str(), NULL, NULL, &status), std::free};
return (status == 0) ? res.get() : name;
}
#else
inline std::string demangle(std::string name) { return name; }
#endif
}
struct EnforceNotMet : public std::exception {
std::exception_ptr exp_;
std::string err_str_;
......@@ -48,15 +68,29 @@ struct EnforceNotMet : public std::exception {
std::rethrow_exception(exp_);
} catch (const std::exception& exp) {
std::ostringstream sout;
sout << string::Sprintf("%s at [%s:%d]", exp.what(), f, l) << std::endl;
sout << "Call Stacks: " << std::endl;
sout << "PaddlePaddle Call Stacks: " << std::endl;
void* call_stack[TRACE_STACK_LIMIT];
int sz = backtrace(call_stack, TRACE_STACK_LIMIT);
auto line = backtrace_symbols(call_stack, sz);
for (int i = 0; i < sz; ++i) {
sout << line[i] << std::endl;
auto size = backtrace(call_stack, TRACE_STACK_LIMIT);
auto symbols = backtrace_symbols(call_stack, size);
Dl_info info;
for (int i = 0; i < size; ++i) {
if (dladdr(call_stack[i], &info)) {
auto demangled = demangle(info.dli_sname);
auto addr_offset = static_cast<char*>(call_stack[i]) -
static_cast<char*>(info.dli_saddr);
sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,
2 + sizeof(void*) * 2, call_stack[i],
demangled, addr_offset);
} else {
sout << string::Sprintf("%-3d %*0p %s\n", i, 2 + sizeof(void*) * 2,
call_stack[i]);
}
}
free(line);
free(symbols);
err_str_ = sout.str();
}
}
......@@ -170,7 +204,7 @@ inline void throw_on_error(T e) {
* PADDLE_ENFORCE_EQ(a, b);
*
* will raise an expression described as follows:
* "enforce a == b failed, 1 != 2" with detailed stack infomation.
* "enforce a == b failed, 1 != 2" with detailed stack information.
*
* extra messages is also supported, for example:
* PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2)
......
paddle/scripts/submit_local.sh.in
浏览文件 @ 4c9699c5
......@@ -18,6 +18,8 @@ function version(){
echo "PaddlePaddle @PADDLE_VERSION@, compiled with"
echo " with_avx: @WITH_AVX@"
echo " with_gpu: @WITH_GPU@"
echo " with_mkldnn: @WITH_MKLDNN"
echo " with_mklml: @WITH_MKLML@"
echo " with_double: @WITH_DOUBLE@"
echo " with_python: @WITH_PYTHON@"
echo " with_rdma: @WITH_RDMA@"
......
python/CMakeLists.txt
浏览文件 @ 4c9699c5
......@@ -21,6 +21,18 @@ if(WITH_GOLANG)
add_dependencies(copy_paddle_master paddle_master)
endif(WITH_GOLANG)
set(MKL_SHARED_LIBS "")
set(MKL_DEPENDS "")
if(WITH_MKLML)
list(APPEND MKL_SHARED_LIBS ${MKLML_LIB} ${MKLML_IOMP_LIB})
list(APPEND MKL_DEPENDS mklml)
endif()
if(WITH_MKLDNN)
list(APPEND MKL_SHARED_LIBS "${MKLDNN_LIB}" "${MKLDNN_LIB}.0")
list(APPEND MKL_DEPENDS mkldnn)
endif()
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
${CMAKE_CURRENT_BINARY_DIR}/setup.py)
......@@ -39,7 +51,7 @@ add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
DEPENDS gen_proto_py copy_paddle_pybind framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
add_custom_target(paddle_python ALL DEPENDS
${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel)
${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel ${MKL_DEPENDS})
set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/)
......
python/paddle/trainer_config_helpers/evaluators.py
浏览文件 @ 4c9699c5
......@@ -298,8 +298,8 @@ def pnpair_evaluator(
input,
label,
info,
name=None,
weight=None, ):
weight=None,
name=None, ):
"""
Positive-negative pair rate Evaluator which adapts to rank task like
learning to rank. This evaluator must contain at least three layers.
......@@ -308,27 +308,31 @@ def pnpair_evaluator(
.. code-block:: python
eval = pnpair_evaluator(input, info, label)
eval = pnpair_evaluator(input, label, info)
:param name: Evaluator name.
:type name: None|basestring
:param input: Input Layer name. The output prediction of network.
:type input: LayerOutput
:param label: Label layer name.
:type label: LayerOutput
:param info: Label layer name. (TODO, explaination)
:param info: Info layer name. (TODO, explaination)
:type info: LayerOutput
:param weight: Weight Layer name. It should be a matrix with size
[sample_num, 1]. (TODO, explaination)
:type weight: LayerOutput
:param name: Evaluator name.
:type name: None|basestring
"""
if not isinstance(input, list):
input = [input]
if label:
input.append(label)
if info:
input.append(info)
evaluator_base(
name=name,
type="pnpair",
input=input,
label=label,
info=info,
weight=weight)
type="pnpair",
weight=weight,
name=name, )
@evaluator(EvaluatorAttribute.FOR_CLASSIFICATION)
......@@ -429,12 +433,12 @@ def chunk_evaluator(
.. code-block:: text
Scheme Description
Scheme Description
plain Use the same label for the whole chunk.
IOB Two labels for chunk type X, B-X for chunk begining and I-X for chunk inside.
IOB Two labels for chunk type X, B-X for chunk begining and I-X for chunk inside.
IOE Two labels for chunk type X, E-X for chunk ending and I-X for chunk inside.
IOBES Four labels for chunk type X, B-X for chunk begining, I-X for chunk inside, E-X for chunk end and S-X for single word chunk.
IOBES Four labels for chunk type X, B-X for chunk begining, I-X for chunk inside, E-X for chunk end and S-X for single word chunk.
To make it clear, let's illustrate by an NER example.
Assuming that there are three named entity types including ORG, PER and LOC which are called 'chunk type' here,
if 'IOB' scheme were used, the label set will be extended to a set including B-ORG, I-ORG, B-PER, I-PER, B-LOC, I-LOC and O,
......@@ -451,7 +455,7 @@ def chunk_evaluator(
tagType = label % numTagType
chunkType = label / numTagType
otherChunkType = numChunkTypes
The following table shows the mapping rule between tagType and tag type in each scheme.
.. code-block:: text
......@@ -475,7 +479,7 @@ def chunk_evaluator(
O 6
In this example, chunkType has three values: 0 for ORG, 1 for PER, 2 for LOC, because the scheme is
"IOB" so tagType has two values: 0 for B and 1 for I.
"IOB" so tagType has two values: 0 for B and 1 for I.
Here we will use I-LOC to explain the above mapping rules in detail.
For I-LOC, the label id is 5, so we can get tagType=1 and chunkType=2, which means I-LOC is a part of NER chunk LOC
and the tag is I.
......@@ -486,7 +490,7 @@ def chunk_evaluator(
eval = chunk_evaluator(input, label, chunk_scheme, num_chunk_types)
:param input: The input layers.
:type input: LayerOutput
:param label: An input layer containing the ground truth label.
......
python/paddle/v2/framework/op.py
浏览文件 @ 4c9699c5
......@@ -23,7 +23,7 @@ class OpDescCreationMethod(object):
"""
A Functor object to convert user input(use key word args) to OpDesc based on
OpProto.
:param op_proto: The OpProto object.
:type op_proto: op_proto_pb2.OpProto
"""
......@@ -177,4 +177,26 @@ class OperatorFactory(object):
return self.get_op_info(type).attrs
class __RecurrentOp__(object):
__proto__ = None
type = 'recurrent_op'
def __init__(self):
# cache recurrent_op's proto
if self.__proto__ is None:
for op_proto in get_all_op_protos():
if op_proto.type == self.type:
self.__proto__ = op_proto
def __call__(self, *args, **kwargs):
if self.type not in args and 'type' not in kwargs:
kwargs['type'] = self.type
# create proto
create_method = OpDescCreationMethod(self.__proto__)
proto = create_method(*args, **kwargs)
# create rnnop
return core.RecurrentOp.create(proto.SerializeToString())
Operator = OperatorFactory() # Default global factory
RecurrentOp = __RecurrentOp__()
python/paddle/v2/framework/tests/test_recurrent_op.py
浏览文件 @ 4c9699c5
......@@ -2,7 +2,7 @@ import logging
import paddle.v2.framework.core as core
import unittest
import numpy as np
from paddle.v2.framework.op import Operator
from paddle.v2.framework.op import Operator, RecurrentOp
def py_sigmoid(x):
......@@ -98,11 +98,11 @@ class TestRecurrentOp(unittest.TestCase):
def forward(self):
self.scope = core.Scope()
self.create_global_variables()
self.create_rnn_op()
self.create_step_net()
rnn_op = self.create_rnn_op()
ctx = core.DeviceContext.create(core.CPUPlace())
rnn_op.infer_shape(self.scope)
rnn_op.run(self.scope, ctx)
self.rnnop.infer_shape(self.scope)
self.rnnop.run(self.scope, ctx)
return np.array(self.scope.find_var("h").get_tensor())
def create_global_variables(self):
......@@ -128,8 +128,7 @@ class TestRecurrentOp(unittest.TestCase):
def create_rnn_op(self):
# create RNNOp
rnnop = Operator(
"recurrent_op",
self.rnnop = RecurrentOp(
# inputs
inlinks=["x"],
boot_memories=["h_boot"],
......@@ -142,14 +141,9 @@ class TestRecurrentOp(unittest.TestCase):
outlink_alias=["h@alias"],
pre_memories=["h@pre"],
memories=["h@alias"])
return rnnop
def create_step_net(self):
var = self.scope.new_var("stepnet")
stepnet = var.get_net()
# x_fc_op = Operator("fc", X="x@alias", W="W", Y="Wx")
# h_fc_op = Operator("fc", X="h@pre", W="U", Y="Uh")
stepnet = core.Net.create()
x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
......@@ -158,6 +152,7 @@ class TestRecurrentOp(unittest.TestCase):
for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
stepnet.add_op(op)
stepnet.complete_add_op(True)
self.rnnop.set_stepnet(stepnet)
def test_forward(self):
print 'test recurrent op forward'
......
python/setup.py.in
浏览文件 @ 4c9699c5
......@@ -23,6 +23,16 @@ with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f:
if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
setup_requires+=["opencv-python"]
# the prefix is sys.prefix which should always be usr
paddle_bin_dir = 'local/opt/paddle/bin'
paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
'${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
'${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
'${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main']
paddle_rt_lib_dir = 'local/lib'
paddle_rt_libs = [] if '${MKL_SHARED_LIBS}'== '' else '${MKL_SHARED_LIBS}'.split(';')
setup(name='paddlepaddle',
version='${PADDLE_VERSION}',
description='Parallel Distributed Deep Learning',
......@@ -42,9 +52,6 @@ setup(name='paddlepaddle',
},
scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'],
distclass=BinaryDistribution,
data_files=[('/usr/local/opt/paddle/bin',
['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
'${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
'${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
'${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main'])]
data_files=[(paddle_bin_dir, paddle_bins),
(paddle_rt_lib_dir, paddle_rt_libs)]
)
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或