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

revert-3824-remove_grad_op_type
dongzhihong 8 years ago
commit 4c9699c5f8

@ -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)

@ -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

@ -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})

@ -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(

@ -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);
}

@ -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

@ -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>(

File diff suppressed because it is too large Load Diff

@ -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;

@ -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());

@ -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>();
});

@ -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);

@ -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);

@ -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)

@ -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>);

@ -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>);

@ -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>);

@ -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)) {

@ -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);
}

@ -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>);

@ -1,13 +1,8 @@
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)

@ -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>);

@ -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>);

@ -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);
}

@ -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());

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save