Merge branch 'feature/is_nan' into feature/check_nan_executor

del_some_in_makelist
Yang Yu 8 years ago
commit 5f3dd266f3

@ -252,6 +252,11 @@ first_seq
.. autoclass:: paddle.v2.layer.first_seq
:noindex:
sub_seq
---------
.. autoclass:: paddle.v2.layer.sub_seq
:noindex:
concat
------
.. autoclass:: paddle.v2.layer.concat

@ -68,12 +68,6 @@ scale
:noindex:
reshape
---------
.. autofunction:: paddle.v2.fluid.layers.reshape
:noindex:
transpose
---------
.. autofunction:: paddle.v2.fluid.layers.transpose

@ -79,7 +79,7 @@ class Optimizer(object):
def minimize(self, loss, parameter_list):
"""Add operations to minimize `loss` by updating `parameter_list`.
This method combines interface `append_backward_ops()` and
This method combines interface `append_backward()` and
`create_optimization_pass()` into one.
"""
params_grads = self.create_backward_pass(loss, parameter_list)

@ -5,7 +5,11 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
if (WITH_GPU)
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place paddle_memory device_context)
else()
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place paddle_memory device_context)
endif ()
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
cc_test(tensor_util_test SRCS tensor_util_test.cc DEPS tensor)
@ -37,7 +41,7 @@ cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
py_proto_compile(framework_py_proto SRCS framework.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/framework/data_transform.h"
#include "paddle/framework/lod_tensor.h"
namespace paddle {
namespace framework {

@ -27,7 +27,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
using DataTransformFN =
using DataTransformFn =
std::function<void(const std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out)>;
using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
@ -47,7 +47,7 @@ struct KernelTypePairHash {
};
using DataTransformMap =
std::unordered_map<KernelTypePair, DataTransformFN, KernelTypePairHash>;
std::unordered_map<KernelTypePair, DataTransformFn, KernelTypePairHash>;
class DataTransformFnMap {
public:
@ -58,25 +58,25 @@ class DataTransformFnMap {
}
void Insert(const OpKernelType& left, const OpKernelType& right,
const DataTransformFN& data_tranform_fn) {
const DataTransformFn& data_tranform_fn) {
Insert(std::make_pair(left, right), data_tranform_fn);
}
void Insert(const KernelTypePair& kernel_type_pair,
const DataTransformFN& data_tranform_fn) {
const DataTransformFn& data_tranform_fn) {
PADDLE_ENFORCE(!Has(kernel_type_pair),
"KernelTypePair %s has been registered", "");
map_.insert({kernel_type_pair, data_tranform_fn});
}
const DataTransformFN& Get(const KernelTypePair& key_pair) const {
const DataTransformFn& Get(const KernelTypePair& key_pair) const {
auto data_transformer = GetNullable(key_pair);
PADDLE_ENFORCE_NOT_NULL(data_transformer,
"DataTransformFN should not be NULL");
"DataTransformFn should not be NULL");
return *data_transformer;
}
const DataTransformFN* GetNullable(const KernelTypePair& key_pair) const {
const DataTransformFn* GetNullable(const KernelTypePair& key_pair) const {
auto it = map_.find(key_pair);
if (it == map_.end()) {
return nullptr;

@ -11,36 +11,61 @@ 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 <array>
#include <vector>
#include "paddle/framework/data_transform.h"
#include <gtest/gtest.h>
#include "paddle/framework/data_transform.h"
namespace paddle {
namespace framework {
using namespace platform;
/**
* @brief cross validation of different kernel type transform
* We use four bit map represent different combination.
* If the field has multiple possible value, only choose two of them.
* For DataType, only test the FP32(float), FP64(double).
* e.g. 0000 -> FP32, CPUPlace, kNHWC, kPlain
* 1111 -> FP64, GPUPlace, kNCHW, kMKLDNN
*/
std::array<proto::DataType, 2> kDataType = {
{proto::DataType::FP32, proto::DataType::FP64}};
std::array<Place, 2> kPlace = {{CPUPlace(), CUDAPlace(0)}};
std::array<DataLayout, 2> kDataLayout = {
{DataLayout::kNHWC, DataLayout::kNCHW}};
std::array<LibraryType, 2> kLibraryType = {
{LibraryType::kPlain, LibraryType::kMKLDNN}};
OpKernelType GenFromBit(const std::vector<bool> bits) {
return OpKernelType(kDataType[bits[0]], kPlace[bits[1]], kDataLayout[bits[2]],
kLibraryType[bits[3]]);
}
int test_value = 0;
OpKernelType kernel_type_1(proto::DataType::FP32, CPUPlace(), DataLayout::kNCHW,
LibraryType::kCUDNN);
OpKernelType kernel_type_2(proto::DataType::FP32, CUDAPlace(0),
DataLayout::kNCHW, LibraryType::kCUDNN);
OpKernelType kernel_type_3(proto::DataType::FP16, CUDAPlace(0),
DataLayout::kNCHW, LibraryType::kCUDNN);
auto kernel0 = GenFromBit({0, 0, 0, 0});
auto kernel1 = GenFromBit({0, 0, 0, 1});
auto kernel2 = GenFromBit({0, 0, 1, 0});
auto kernel3 = GenFromBit({0, 0, 1, 1});
void type1_to_type2(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
void TransDataType_t(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
test_value++;
}
void type2_to_type3(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
void TransDataLayout_t(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
test_value--;
}
void type1_to_type3(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
void TransLibraryType_t(std::vector<platform::DeviceContext*> ctx,
const Variable& in, Variable* out) {
test_value += 2;
}
@ -49,30 +74,25 @@ void type1_to_type3(std::vector<platform::DeviceContext*> ctx,
namespace frw = paddle::framework;
REGISTER_DATA_TRANSFORM_FN(frw::kernel_type_1, frw::kernel_type_2,
frw::type1_to_type2);
REGISTER_DATA_TRANSFORM_FN(frw::kernel_type_2, frw::kernel_type_3,
frw::type2_to_type3);
REGISTER_DATA_TRANSFORM_FN(frw::kernel_type_1, frw::kernel_type_3,
frw::type1_to_type3);
REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel1, frw::TransDataType_t);
REGISTER_DATA_TRANSFORM_FN(frw::kernel1, frw::kernel2, frw::TransDataLayout_t);
REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel2, frw::TransLibraryType_t);
TEST(DataTransform, Register) {
using namespace paddle::framework;
using namespace paddle::platform;
auto& instance = DataTransformFnMap::Instance();
ASSERT_EQ(instance.Map().size(), 3UL);
std::vector<DeviceContext*> ctx;
paddle::framework::Variable in;
paddle::framework::Variable out;
instance.Get(std::make_pair(frw::kernel_type_1, frw::kernel_type_2))(ctx, in,
&out);
instance.Get(std::make_pair(frw::kernel0, frw::kernel1))(ctx, in, &out);
ASSERT_EQ(test_value, 1);
instance.Get(std::make_pair(frw::kernel_type_2, frw::kernel_type_3))(ctx, in,
&out);
instance.Get(std::make_pair(frw::kernel1, frw::kernel2))(ctx, in, &out);
ASSERT_EQ(test_value, 0);
instance.Get(std::make_pair(frw::kernel_type_1, frw::kernel_type_3))(ctx, in,
&out);
instance.Get(std::make_pair(frw::kernel0, frw::kernel2))(ctx, in, &out);
ASSERT_EQ(test_value, 2);
}

@ -123,7 +123,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
}
}
}
if (create_local_scope) {
if (create_vars && create_local_scope) {
scope->DeleteScope(local_scope);
}
}

@ -20,7 +20,11 @@ namespace framework {
// For more details about the design of LibraryType, Please refer to
// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md#library
enum class LibraryType { kPlain = 0, kMKLDNN = 1, kCUDNN = 2 };
enum class LibraryType {
kPlain = 0,
kMKLDNN = 1,
kCUDNN = 2,
};
inline std::string LibraryTypeToString(const LibraryType& library_type) {
switch (library_type) {
@ -31,7 +35,26 @@ inline std::string LibraryTypeToString(const LibraryType& library_type) {
case LibraryType::kCUDNN:
return "CUDNN";
default:
PADDLE_THROW("unknown LibraryType %d", library_type);
PADDLE_THROW("unknown LibraryType %d", static_cast<int>(library_type));
}
}
inline LibraryType StringToLibraryType(const char* ctype) {
std::string s(ctype);
if (s == std::string("PLAIN")) {
return LibraryType::kPlain;
} else if (s == std::string("MKLDNN")) {
return LibraryType::kMKLDNN;
} else if (s == std::string("CUDNN")) {
return LibraryType::kCUDNN;
// To be compatible with register macro.
// CPU, CUDA, PLAIN are same library type.
} else if (s == std::string("CPU")) {
return LibraryType::kPlain;
} else if (s == std::string("CUDA")) {
return LibraryType::kPlain;
} else {
PADDLE_THROW("Unknown LibraryType %s", s.c_str());
}
}

@ -88,6 +88,14 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
need_update_ = true;
}
void OpDesc::CopyFrom(const OpDesc &op_desc) {
desc_.set_type(op_desc.Type());
inputs_ = op_desc.inputs_;
outputs_ = op_desc.outputs_;
attrs_ = op_desc.attrs_;
need_update_ = true;
}
OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog)
: desc_(desc), need_update_(false) {
// restore inputs_

@ -35,6 +35,8 @@ class OpDesc {
OpDesc(const proto::OpDesc &desc, ProgramDesc *prog);
void CopyFrom(const OpDesc &op_desc);
proto::OpDesc *Proto();
std::string Type() const { return desc_.type(); }

@ -68,6 +68,8 @@ struct OpKernelType {
data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
library_type_ == o.library_type_;
}
bool operator!=(const OpKernelType& o) const { return !(*this == o); }
};
inline std::ostream& operator<<(std::ostream& os,
@ -78,5 +80,11 @@ inline std::ostream& operator<<(std::ostream& os,
return os;
}
inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
std::ostringstream stream;
stream << kernel_key;
return stream.str();
}
} // namespace framework
} // namespace paddle

@ -26,10 +26,8 @@ TEST(OpKernelType, ToString) {
OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
LibraryType::kCUDNN);
std::ostringstream stream;
stream << op_kernel_type;
ASSERT_EQ(
stream.str(),
paddle::framework::KernelTypeToString(op_kernel_type),
"data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
}
@ -48,4 +46,4 @@ TEST(OpKernelType, Hash) {
OpKernelType::Hash hasher;
ASSERT_NE(hasher(op_kernel_type_1), hasher(op_kernel_type_2));
}
}

@ -79,30 +79,31 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
using KERNEL_TYPE =
typename std::tuple_element<I, std::tuple<KernelTypes...>>::type;
void operator()(const char* op_type) const {
void operator()(const char* op_type, const char* library_type) const {
using T = typename KERNEL_TYPE::ELEMENT_TYPE;
OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType());
OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
DataLayout::kAnyLayout, StringToLibraryType(library_type));
OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>
func;
func(op_type);
func(op_type, library_type);
}
};
template <typename PlaceType, size_t I, typename... KernelType>
struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
void operator()(const char* op_type) const {}
void operator()(const char* op_type, const char* library_type) const {}
};
// User can register many kernel in one place. The data type could be different.
template <typename PlaceType, typename... KernelType>
class OpKernelRegistrar : public Registrar {
public:
explicit OpKernelRegistrar(const char* op_type) {
explicit OpKernelRegistrar(const char* op_type, const char* library_type) {
OpKernelRegistrarFunctor<PlaceType, false, 0, KernelType...> func;
func(op_type);
func(op_type, library_type);
}
};
@ -181,7 +182,8 @@ class OpKernelRegistrar : public Registrar {
__reg_op_kernel_##op_type##_##DEVICE_TYPE##__, \
"REGISTER_OP_KERNEL must be called in global namespace"); \
static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__> \
__op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type); \
__op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type, \
#DEVICE_TYPE); \
int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() { \
__op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch(); \
return 0; \

@ -1,3 +1,17 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/framework/op_registry.h"
#include <gtest/gtest.h>
@ -182,3 +196,71 @@ TEST(OperatorRegistrar, Test) {
using namespace paddle::framework;
OperatorRegistrar<CosineOpComplete, CosineOpProtoAndCheckerMaker> reg("cos");
}
namespace paddle {
namespace framework {
class OpKernelTestMaker : public OpProtoAndCheckerMaker {
public:
OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddComment("NoGradOp, same input output. no Grad");
}
};
class OpWithKernelTest : public OperatorWithKernel {
public:
using OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(InferShapeContext* ctx) const override {}
framework::OpKernelType GetActualKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(proto::DataType::FP32, ctx.device_context());
}
};
template <typename DeviceContext, typename T>
class OpKernelTest : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const {}
};
} // namespace framework
} // namespace paddle
REGISTER_OP_WITHOUT_GRADIENT(op_with_kernel,
paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMaker);
REGISTER_OP_CPU_KERNEL(
op_with_kernel,
paddle::framework::OpKernelTest<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(op_with_kernel,
paddle::framework::OpKernelTest<
paddle::platform::CUDADeviceContext, float>);
TEST(OperatorRegistrar, CPU) {
paddle::framework::proto::OpDesc op_desc;
paddle::platform::CPUPlace cpu_place;
paddle::framework::Scope scope;
op_desc.set_type("op_with_kernel");
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_place);
}
#ifdef PADDLE_WITH_CUDA
TEST(OperatorRegistrar, CUDA) {
paddle::framework::proto::OpDesc op_desc;
paddle::platform::CUDAPlace cuda_place(0);
paddle::framework::Scope scope;
op_desc.set_type("op_with_kernel");
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cuda_place);
}
#endif

@ -413,37 +413,51 @@ void OperatorWithKernel::Run(const Scope& scope,
}
if (actual_kernel_key == expected_kernel_key) {
kernel_iter->second->Compute(ctx);
PADDLE_ENFORCE_EQ(actual_kernel_key.place_, expected_kernel_key.place_,
"Currently, model parallelism is only supported between "
"CPU and other devices. For example, multi-GPU model "
"parallelism will failed.");
} else {
Scope& op_scope = scope.NewScope();
auto input_vars = this->InputVars();
for (auto var_name : input_vars) {
op_scope.Var(var_name);
}
// TODO(qijun) get appropriate DeviceContext from DeviceContext pool
platform::DeviceContext* trans_dev_ctx = nullptr;
std::vector<platform::DeviceContext*> trans_dev_ctx_vec{trans_dev_ctx};
const DataTransformFn* trans_fun =
DataTransformFnMap::Instance().GetNullable(
std::make_pair(actual_kernel_key, expected_kernel_key));
if (trans_fun) {
auto input_vars = this->InputVars();
// TODO(qijun) filter the input vars that do not need to be transformed
// filter vars that has been transformed
std::vector<std::string> need_trans;
for (auto var_name : input_vars) {
auto var_name_trans =
var_name + framework::KernelTypeToString(expected_kernel_key);
if (!scope.FindVar(var_name_trans)) {
const_cast<Scope&>(scope).Var(var_name_trans);
need_trans.push_back(var_name);
}
}
// TODO(qijun) get appropriate DataTransformFN from global map
framework::DataTransformFN trans_fun = nullptr;
if (!need_trans.empty()) {
// TODO(qijun) get appropriate DeviceContext from DeviceContext pool
platform::DeviceContext* trans_dev_ctx = nullptr;
std::vector<platform::DeviceContext*> trans_dev_ctx_vec{trans_dev_ctx};
// Wait for transform starting
dev_ctx->Wait();
// Wait for transform starting
dev_ctx->Wait();
for (auto var_name : input_vars) {
trans_fun(trans_dev_ctx_vec, *(scope.FindVar(var_name)),
op_scope.FindVar(var_name));
}
// Wait for data transform finishing
for (auto ctx : trans_dev_ctx_vec) {
ctx->Wait();
for (auto var_name : need_trans) {
(*trans_fun)(trans_dev_ctx_vec, *(scope.FindVar(var_name)),
scope.FindVar(var_name + framework::KernelTypeToString(
expected_kernel_key)));
}
// Wait for data transform finishing
for (auto ctx : trans_dev_ctx_vec) {
ctx->Wait();
}
}
}
// Create a new ExecutionContext
ExecutionContext op_ctx(*this, op_scope, *dev_ctx);
kernel_iter->second->Compute(op_ctx);
}
kernel_iter->second->Compute(ctx);
}
OpKernelType OperatorWithKernel::GetActualKernelType(

@ -178,7 +178,7 @@ class Tensor {
DDim dims_;
/**
* @brief the layout of memory block, default is NCHW.
* @brief the layout of memory block, default is NHWC.
*
* @note the memory allocation order, describe how weight/data is stored
* For example, in 4-D Tensor(rank=4), there are three commonly

@ -0,0 +1,115 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/framework/tensor_util.h"
namespace paddle {
namespace framework {
template <typename Predicate, typename DevCtx>
struct AnyDTypeVisitor {
Predicate predicate_;
const Tensor& tensor_;
const DevCtx& ctx_;
Tensor* out_;
AnyDTypeVisitor(Predicate predicate, const Tensor& tensor, const DevCtx& ctx,
Tensor* out)
: predicate_(predicate), tensor_(tensor), ctx_(ctx), out_(out) {}
template <typename T>
void operator()() const {
auto t = EigenVector<T>::Flatten(tensor_);
auto o = EigenScalar<bool>::From(*out_);
o.device(*ctx_.eigen_device()) = predicate_(t).any();
}
};
template <typename Predicate, typename DevCtx>
inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor,
const DevCtx& ctx, framework::Tensor* out) {
VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor<Predicate, DevCtx>(
predicate, tensor, ctx, out));
}
template <typename Predicate>
struct AnyVisitor : public boost::static_visitor<bool> {
const framework::Tensor& tensor_;
Predicate predicate_;
AnyVisitor(const framework::Tensor& tensor, Predicate predicate)
: tensor_(tensor), predicate_(std::move(predicate)) {}
template <typename Place>
bool operator()(const Place& place) const {
framework::Tensor out;
out.Resize({1});
out.mutable_data<bool>(place);
auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place);
AnyImpl(predicate_, tensor_, *ctx, &out);
return this->GetResult(out, place);
}
bool GetResult(const framework::Tensor& out,
const platform::CUDAPlace& gpu) const {
platform::CPUPlace cpu;
framework::Tensor tmp;
tmp.Resize({1});
tmp.mutable_data<bool>(cpu);
platform::DeviceContextPool::Instance().Get(gpu)->Wait();
CopyFrom(out, cpu, &tmp);
platform::DeviceContextPool::Instance().Get(gpu)->Wait();
return GetResult(tmp, cpu);
}
bool GetResult(const framework::Tensor& out,
const platform::CPUPlace& cpu) const {
return *out.data<bool>();
}
};
template <typename Predicate>
inline bool Any(const framework::Tensor& tensor, Predicate predicate) {
AnyVisitor<Predicate> visitor(tensor, predicate);
auto place = tensor.place();
return platform::VisitPlace(place, visitor);
}
struct HasNANPredicate {
template <typename T>
auto operator()(const T& eigen_vec) const
-> decltype(std::declval<T>().isnan()) {
return eigen_vec.isnan();
}
};
bool HasNAN(const framework::Tensor& tensor) {
HasNANPredicate predicate;
return Any(tensor, predicate);
}
struct HasInfPredicate {
template <typename T>
auto operator()(const T& eigen_vec) const
-> decltype(std::declval<T>().isinf()) {
return eigen_vec.isinf();
}
};
bool HasInf(const framework::Tensor& tensor) {
HasInfPredicate predicate;
return Any(tensor, predicate);
}
} // namespace framework
} // namespace paddle

@ -208,100 +208,8 @@ inline void CopyToVector(const Tensor& src, std::vector<T>* dst) {
src_ptr, size);
}
template <typename Predicate, typename DevCtx>
struct AnyDTypeVisitor {
Predicate predicate_;
const Tensor& tensor_;
const DevCtx& ctx_;
Tensor* out_;
AnyDTypeVisitor(Predicate predicate, const Tensor& tensor, const DevCtx& ctx,
Tensor* out)
: predicate_(predicate), tensor_(tensor), ctx_(ctx), out_(out) {}
template <typename T>
void operator()() const {
auto t = EigenVector<T>::Flatten(tensor_);
auto o = EigenScalar<bool>::From(*out_);
o.device(*ctx_.eigen_device()) = predicate_(t).any();
}
};
template <typename Predicate, typename DevCtx>
inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor,
const DevCtx& ctx, framework::Tensor* out) {
VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor<Predicate, DevCtx>(
predicate, tensor, ctx, out));
}
template <typename Predicate>
struct AnyVisitor : public boost::static_visitor<bool> {
const framework::Tensor& tensor_;
Predicate predicate_;
AnyVisitor(const framework::Tensor& tensor, Predicate predicate)
: tensor_(tensor), predicate_(std::move(predicate)) {}
template <typename Place>
bool operator()(const Place& place) const {
framework::Tensor out;
out.Resize({1});
out.mutable_data<bool>(place);
auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place);
AnyImpl(predicate_, tensor_, *ctx, &out);
return this->GetResult(out, place);
}
bool GetResult(const framework::Tensor& out,
const platform::CUDAPlace& gpu) const {
platform::CPUPlace cpu;
framework::Tensor tmp;
tmp.Resize({1});
tmp.mutable_data<bool>(cpu);
platform::DeviceContextPool::Instance().Get(gpu)->Wait();
CopyFrom(out, cpu, &tmp);
platform::DeviceContextPool::Instance().Get(gpu)->Wait();
return GetResult(tmp, cpu);
}
bool GetResult(const framework::Tensor& out,
const platform::CPUPlace& cpu) const {
return *out.data<bool>();
}
};
template <typename Predicate>
inline bool Any(const framework::Tensor& tensor, Predicate predicate) {
AnyVisitor<Predicate> visitor(tensor, predicate);
auto place = tensor.place();
return platform::VisitPlace(place, visitor);
}
struct HasNANPredicate {
template <typename T>
auto operator()(const T& eigen_vec) const
-> decltype(std::declval<T>().isnan()) {
return eigen_vec.isnan();
}
};
inline bool HasNAN(const framework::Tensor& tensor) {
HasNANPredicate predicate;
return Any(tensor, predicate);
}
struct HasInfPredicate {
template <typename T>
auto operator()(const T& eigen_vec) const
-> decltype(std::declval<T>().isinf()) {
return eigen_vec.isinf();
}
};
inline bool HasInf(const framework::Tensor& tensor) {
HasInfPredicate predicate;
return Any(tensor, predicate);
}
extern bool HasNAN(const framework::Tensor& tensor);
extern bool HasInf(const framework::Tensor& tensor);
} // namespace framework
} // namespace paddle

@ -74,7 +74,7 @@ const proto::TensorDesc &VarDesc::tensor_desc() const {
case proto::VarDesc::LOD_TENSOR_ARRAY:
return desc_.tensor_array().tensor();
default:
PADDLE_THROW("Unexpected branch.");
PADDLE_THROW("The type of var '", this->Name(), "' is unsupported.");
}
}

@ -1,5 +1,6 @@
file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
set(DEPS_OPS "")
set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt. DO NOT EDIT!\n\n")
function(op_library TARGET)
@ -48,6 +49,11 @@ function(op_library TARGET)
message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
endif()
list(LENGTH op_library_DEPS op_library_DEPS_len)
if (${op_library_DEPS_len} GREATER 0)
set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
@ -181,55 +187,26 @@ endfunction()
add_subdirectory(math)
add_subdirectory(nccl)
set(DEPS_OPS
cond_op
cross_entropy_op
recurrent_op
softmax_with_cross_entropy_op
softmax_op
sequence_softmax_op
sum_op
pool_op
maxout_op
unpool_op
pool_with_index_op
conv_op
conv_transpose_op
nccl_op
sequence_conv_op
sequence_pool_op
lod_rank_table_op
lod_tensor_to_array_op
array_to_lod_tensor_op
max_sequence_len_op
lstm_op
tensor_array_read_write_op
gru_op
adagrad_op
sgd_op
save_op
load_op
send_op
recv_op)
if(WITH_GPU)
op_library(nccl_op DEPS nccl_common)
else()
set(DEPS_OPS ${DEPS_OPS} nccl_op)
endif()
if(WITH_DISTRIBUTE)
add_subdirectory(detail)
op_library(send_op SRCS send_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
set_source_files_properties(
send_op.cc
PROPERTIES
COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
op_library(recv_op SRCS recv_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
set_source_files_properties(
recv_op.cc
PROPERTIES
COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
add_subdirectory(detail)
set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
op_library(send_op DEPS ${DISTRIBUTE_DEPS})
set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
op_library(recv_op DEPS ${DISTRIBUTE_DEPS})
set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
else()
set(DEPS_OPS ${DEPS_OPS} send_op recv_op)
endif()
op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
op_library(cond_op DEPS framework_proto tensor net_op)
op_library(cross_entropy_op DEPS cross_entropy)
op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
op_library(softmax_op DEPS softmax)
@ -242,21 +219,16 @@ op_library(pool_op DEPS pooling)
op_library(maxout_op DEPS maxouting)
op_library(unpool_op DEPS unpooling)
op_library(pool_with_index_op DEPS pooling)
op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op)
op_library(array_to_lod_tensor_op SRCS array_to_lod_tensor_op.cc DEPS lod_rank_table_op)
op_library(max_sequence_len_op SRCS max_sequence_len_op.cc DEPS lod_rank_table)
op_library(tensor_array_read_write_op SRCS tensor_array_read_write_op.cc)
if(WITH_GPU)
op_library(nccl_op DEPS nccl_common)
endif()
op_library(lod_rank_table_op DEPS lod_rank_table)
op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
op_library(array_to_lod_tensor_op DEPS lod_rank_table_op)
op_library(max_sequence_len_op DEPS lod_rank_table)
op_library(sequence_conv_op DEPS context_project)
op_library(sequence_pool_op DEPS sequence_pooling)
op_library(lstm_op DEPS sequence2batch lstm_compute)
op_library(conv_transpose_op DEPS vol2col)
op_library(gru_op DEPS sequence2batch gru_compute)
op_library(recurrent_op SRCS recurrent_op.cc DEPS executor)
op_library(recurrent_op DEPS executor)
# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
@ -269,13 +241,12 @@ endforeach()
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
if(WITH_GPU)
cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
endif()
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)

@ -50,10 +50,6 @@ class BatchNormOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE(ctx->HasOutput("SavedMean"), "");
PADDLE_ENFORCE(ctx->HasOutput("SavedVariance"), "");
const float epsilon = ctx->Attrs().Get<float>("epsilon");
PADDLE_ENFORCE_GE(epsilon, 0.0, "epsilon should be larger than 0");
PADDLE_ENFORCE_LE(epsilon, 0.001, "epsilon should not be too large");
// make sure Mean/MeanOut and Variance/VarianceOut share memory in Python
PADDLE_ENFORCE_EQ(ctx->Inputs("Mean")[0], ctx->Outputs("MeanOut")[0],
"Mean and MeanOut should share the same memory");
@ -91,7 +87,12 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
: OpProtoAndCheckerMaker(proto, op_checker) {
AddAttr<bool>("is_test", "").SetDefault(false);
AddAttr<float>("momentum", "").SetDefault(0.9);
AddAttr<float>("epsilon", "").SetDefault(1e-5);
AddAttr<float>("epsilon", "")
.SetDefault(1e-5)
.AddCustomChecker([](const float &epsilon) {
PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f,
"'epsilon' should be between 0.0 and 0.001.");
});
AddAttr<std::string>("data_layout", "").SetDefault("NCHW");
AddInput("X", "The input tensor");
AddInput("Scale",

@ -315,6 +315,10 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
} // namespace operators
} // namespace paddle
REGISTER_OP_KERNEL(conv2d, CUDNN, paddle::platform::CUDAPlace,
paddle::operators::CudnnConvOpKernel<float>,
paddle::operators::CudnnConvOpKernel<double>);
REGISTER_OP_CUDA_KERNEL(conv2d_cudnn,
paddle::operators::CudnnConvOpKernel<float>,
paddle::operators::CudnnConvOpKernel<double>);

@ -114,15 +114,15 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
CrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"(Tensor, default Tensor<float>), a 2-D tensor with shape N x D, "
"where N is the batch size and D is the number of classes. "
"(Tensor, default Tensor<float>), a 2-D tensor with shape [N x D],"
" where N is the batch size and D is the number of classes. "
"This input is a probability computed by the previous operator, "
"which is almost always the result of a softmax operator.");
AddInput("Label",
"(Tensor), the ground truth which is a 2-D tensor. When "
"soft_label is set to false, Label is a Tensor<int64> with shape "
"[N x 1]. When soft_label is set to true, Label is a "
"Tensor<float/double> with shape [N x K].");
"Tensor<float/double> with shape [N x D].");
AddOutput("Y",
"(Tensor, default Tensor<float>), a 2-D tensor with shape "
"[N x 1]. The cross entropy loss.");

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

Loading…
Cancel
Save