Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
提交 9b3e79ac 编写于 作者: S sneaxiy
浏览文件
操作

cherry-pick mem op to release/1.3 

test=release/1.3
上级 e0bb8cce
显示空白变更内容
内联 并排
Showing with 578 addition and 74 deletion
paddle/fluid/framework/executor.cc
浏览文件 @ 9b3e79ac
......@@ -14,6 +14,10 @@ limitations under the License. */
#include "paddle/fluid/framework/executor.h"
#include <deque>
#include <memory>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
......@@ -74,11 +78,11 @@ static std::unordered_map<std::string, size_t> GetNonPersistableReferenceCounts(
ExecutorPrepareContext::ExecutorPrepareContext(
const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars)
: prog_(prog), block_id_(block_id) {
if (GetEagerDeletionThreshold() >= 0) {
global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id),
skip_ref_cnt_vars);
const std::vector<std::string>& keep_vars, bool force_disable_gc)
: prog_(prog), block_id_(block_id), force_disable_gc_(force_disable_gc) {
if (GetEagerDeletionThreshold() >= 0 && !force_disable_gc_) {
global_ref_cnts_ =
GetNonPersistableReferenceCounts(prog.Block(block_id), keep_vars);
}
}
......@@ -183,13 +187,15 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
}
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
bool create_local_scope, bool create_vars) {
bool create_local_scope, bool create_vars,
const std::vector<std::string>& skip_ref_cnt_vars,
bool force_disable_gc) {
platform::RecordBlock b(block_id);
if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
#ifdef PADDLE_WITH_NGRAPH
if (FLAGS_use_ngraph) operators::NgraphEngine::EnableNgraph(pdesc);
#endif
auto ctx = Prepare(pdesc, block_id);
auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc);
RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
}
......@@ -356,9 +362,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars) {
std::unique_ptr<ExecutorPrepareContext> ctx(
new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars));
const std::vector<std::string>& skip_ref_cnt_vars, bool force_disable_gc) {
std::unique_ptr<ExecutorPrepareContext> ctx(new ExecutorPrepareContext(
program, block_id, skip_ref_cnt_vars, force_disable_gc));
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
auto& block = program.Block(block_id);
for (auto& op_desc : block.AllOps()) {
......@@ -369,7 +375,8 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars) {
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars,
bool force_disable_gc) {
PADDLE_ENFORCE(
skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(),
"skip_ref_cnt_vars should be either empty or equals to block number %d",
......@@ -379,9 +386,11 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
for (auto& bid : block_ids) {
ExecutorPrepareContext* ctx;
if (skip_ref_cnt_vars.empty()) {
ctx = new ExecutorPrepareContext(program, bid);
ctx = new ExecutorPrepareContext(program, bid, std::vector<std::string>(),
force_disable_gc);
} else {
ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]);
ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx],
force_disable_gc);
}
PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
auto& block = program.Block(bid);
......@@ -408,8 +417,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
int64_t max_memory_size = GetEagerDeletionThreshold();
std::unique_ptr<GarbageCollector> gc;
// skip while_op and while_grad_op temporarily
if (max_memory_size >= 0 && !keep_kids) {
// FIXME(zjl): recurrent_op is rather complex, we would
// disable gc forcely in recurrent_op
if (!ctx->force_disable_gc_ && max_memory_size >= 0 && !keep_kids) {
ctx->ResetReferenceCount();
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place_)) {
......
paddle/fluid/framework/executor.h
浏览文件 @ 9b3e79ac
......@@ -15,7 +15,9 @@ limitations under the License. */
#pragma once
#include <map>
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/op_info.h"
......@@ -30,7 +32,8 @@ namespace framework {
struct ExecutorPrepareContext {
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
std::vector<std::string>(),
bool force_disable_gc = false);
~ExecutorPrepareContext();
......@@ -38,6 +41,7 @@ struct ExecutorPrepareContext {
const framework::ProgramDesc& prog_;
size_t block_id_;
bool force_disable_gc_;
std::vector<std::unique_ptr<OperatorBase>> ops_;
std::unordered_map<std::string, size_t> global_ref_cnts_;
......@@ -66,7 +70,10 @@ class Executor {
* Scope
*/
void Run(const ProgramDesc& prog, Scope* scope, int block_id,
bool create_local_scope = true, bool create_vars = true);
bool create_local_scope = true, bool create_vars = true,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>(),
bool force_disable_gc = false);
// This API is very slow.
void Run(const ProgramDesc& program, Scope* scope,
......@@ -79,12 +86,14 @@ class Executor {
static std::unique_ptr<ExecutorPrepareContext> Prepare(
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
std::vector<std::string>(),
bool force_disable_gc = false);
static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars =
std::vector<std::vector<std::string>>());
std::vector<std::vector<std::string>>(),
bool force_disable_gc = false);
void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
......
paddle/fluid/operators/cross_entropy_op.cc
浏览文件 @ 9b3e79ac
......@@ -13,18 +13,21 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/cross_entropy_op.h"
#include <memory>
#include <string>
#include <unordered_map>
namespace paddle {
namespace operators {
class CrossEntropyOp : public framework::OperatorWithKernel {
class CrossEntropyOpBase : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
auto x_dims = ctx->GetInputDim("X");
......@@ -32,14 +35,24 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
int rank = x_dims.size();
PADDLE_ENFORCE_EQ(rank, label_dims.size(),
"Input(X) and Input(Label) shall have the same rank.");
bool check = true;
if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
framework::product(label_dims) <= 0)) {
check = false;
}
if (check) {
PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
framework::slice_ddim(label_dims, 0, rank - 1),
"Input(X) and Input(Label) shall have the same shape "
"except the last dimension.");
if (ctx->Attrs().Get<bool>("soft_label")) {
}
if (IsSoftLabel(ctx)) {
if (check) {
PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1],
"If Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal.");
}
} else {
PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1UL,
"If Attr(softLabel) == false, the last dimension of "
......@@ -60,21 +73,24 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
ctx.device_context());
}
virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
return ctx->Attrs().Get<bool>("soft_label");
}
};
class CrossEntropyGradientOp : public framework::OperatorWithKernel {
class CrossEntropyGradientOpBase : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
void InferShape(framework::InferShapeContext* ctx) const {
PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
"Input(Y@GRAD) shoudl be not null.");
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
"Output(X@GRAD) should be not null.");
auto x_dims = ctx->GetInputDim("X");
auto x_dims = GetXDim(ctx);
auto label_dims = ctx->GetInputDim("Label");
auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
int rank = x_dims.size();
......@@ -82,6 +98,14 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
"Input(Y@Grad) and Input(X) should have the same rank.");
PADDLE_ENFORCE_EQ(label_dims.size(), rank,
"Input(Label) and Input(X) should have the same rank.");
bool check = true;
if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
framework::product(label_dims) <= 0)) {
check = false;
}
if (check) {
PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1),
framework::slice_ddim(label_dims, 0, rank - 1),
"The Input(X) and Input(Label) should have the same "
......@@ -90,19 +114,24 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
framework::slice_ddim(dy_dims, 0, rank - 1),
"The Input(X) and Input(Y@Grad) should have the same "
"shape except the last dimension.");
PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
"The last dimension of Input(Y@Grad) should be 1.");
if (ctx->Attrs().Get<bool>("soft_label")) {
PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1],
}
if (IsSoftLabel(ctx)) {
if (check) {
PADDLE_ENFORCE_EQ(
x_dims[rank - 1], label_dims[rank - 1],
"When Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal.");
}
} else {
PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1,
"When Attr(soft_label) == false, the last dimension of "
"Input(Label) should be 1.");
}
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
ctx->ShareLoD("X", framework::GradVarName("X"));
PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1,
"The last dimension of Input(Y@Grad) should be 1.");
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
ctx->ShareLoD(VarNameWithXLoD(), framework::GradVarName("X"));
}
protected:
......@@ -110,9 +139,29 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
return framework::OpKernelType(
ctx.Input<Tensor>(framework::GradVarName("Y"))->type(),
ctx.device_context());
}
virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const {
return ctx->GetInputDim("X");
}
virtual const char* VarNameWithXLoD() const { return "X"; }
virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
return ctx->Attrs().Get<bool>("soft_label");
}
};
class CrossEntropyOpInferVarType
: public framework::PassInDtypeAndVarTypeToOutput {
protected:
std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
const override {
return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
}
};
class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
......@@ -179,22 +228,132 @@ or not. But the output only shares the LoD information with input X.
}
};
class CrossEntropyOpInferVarType
: public framework::PassInDtypeAndVarTypeToOutput {
class CrossEntropyGradientOp : public CrossEntropyGradientOpBase {
public:
using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
CrossEntropyGradientOpBase::InferShape(ctx);
}
};
class CrossEntropyOp2 : public CrossEntropyOpBase {
public:
using CrossEntropyOpBase::CrossEntropyOpBase;
void InferShape(framework::InferShapeContext* ctx) const override {
CrossEntropyOpBase::InferShape(ctx);
PADDLE_ENFORCE(ctx->HasOutput("XShape"),
"Output(XShape) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("MatchX"),
"Output(MatchX) should be not null.");
auto x_dims = ctx->GetInputDim("X");
auto x_dims_vec = framework::vectorize(x_dims);
x_dims_vec.push_back(0);
ctx->SetOutputDim("XShape", framework::make_ddim(x_dims_vec));
x_dims[x_dims.size() - 1] = 1;
ctx->SetOutputDim("MatchX", x_dims);
ctx->ShareLoD("X", /*->*/ "XShape");
}
protected:
std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
const override {
return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
bool IsSoftLabel(framework::InferShapeContext* ctx) const override {
return false;
}
};
class CrossEntropyGradientOp2 : public CrossEntropyGradientOpBase {
public:
using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("MatchX"), "Input(MatchX) must exist");
CrossEntropyGradientOpBase::InferShape(ctx);
}
protected:
virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const {
auto x_shape = ctx->GetInputDim("XShape");
return framework::DDim(x_shape.Get(), x_shape.size() - 1);
}
virtual const char* VarNameWithXLoD() const { return "XShape"; }
virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const {
return false;
}
};
class CrossEntropyOpMaker2 : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X",
"(Tensor, default Tensor<float>), a tensor whose last dimension "
"size is equal to 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 tensor which represents the ground truth. It has the "
"same shape with 'X' except the last dimension. One hot Tensor.");
AddOutput("Y",
"(Tensor, default Tensor<float>), a tensor whose shape is same "
"with 'X' except that the last dimension size is 1. It "
"represents the cross entropy loss.");
AddOutput("XShape", "Temporaily variable to save shape and LoD of X.");
AddOutput("MatchX",
"X value that matches label, used for gradient computation.");
AddAttr<int>("ignore_index",
"(int, default -100), Specifies a target value that is"
"ignored and does not contribute to the input gradient."
"Only valid if soft_label is set to False")
.SetDefault(-100);
AddComment(R"DOC(
Hard-label CrossEntropy Operator.
The input 'X' and 'Label' will first be logically flattened to 2-D matrixs.
The matrix's second dimension(row length) is as same as the original last
dimension, and the first dimension(column length) is the product of all other
original dimensions. Then the softmax computation will take palce on each raw
of flattened matrixs.
Only support hard label.
Both the input X and Label can carry the LoD (Level of Details) information,
or not. But the output only shares the LoD information with input X.
)DOC");
}
};
class CrossEntropyGradOpDescMaker2 : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
op->SetType("cross_entropy_grad2");
op->SetInput("Label", Input("Label"));
op->SetInput("MatchX", Output("MatchX"));
op->SetInput("XShape", Output("XShape"));
op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
op->SetAttrMap(Attrs());
return op;
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
using CPUCtx = paddle::platform::CPUDeviceContext;
REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
ops::CrossEntropyOpInferVarType,
REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOpBase,
ops::CrossEntropyOpMaker, ops::CrossEntropyOpInferVarType,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp);
REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,
......@@ -202,3 +361,14 @@ REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,
REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
ops::CrossEntropyGradientOpKernel<CPUCtx, float>,
ops::CrossEntropyGradientOpKernel<CPUCtx, double>);
REGISTER_OPERATOR(cross_entropy2, ops::CrossEntropyOp2,
ops::CrossEntropyOpMaker2, ops::CrossEntropyOpInferVarType,
ops::CrossEntropyGradOpDescMaker2);
REGISTER_OPERATOR(cross_entropy_grad2, ops::CrossEntropyGradientOp2);
REGISTER_OP_CPU_KERNEL(cross_entropy2,
ops::CrossEntropyOpKernel2<CPUCtx, float>,
ops::CrossEntropyOpKernel2<CPUCtx, double>);
REGISTER_OP_CPU_KERNEL(cross_entropy_grad2,
ops::CrossEntropyGradientOpKernel2<CPUCtx, float>,
ops::CrossEntropyGradientOpKernel2<CPUCtx, double>);
paddle/fluid/operators/cross_entropy_op.cu
浏览文件 @ 9b3e79ac
......@@ -27,3 +27,13 @@ REGISTER_OP_CUDA_KERNEL(
cross_entropy_grad, ops::CrossEntropyGradientOpKernel<CUDACtx, float>,
ops::CrossEntropyGradientOpKernel<CUDACtx, double>,
ops::CrossEntropyGradientOpKernel<CUDACtx, plat::float16>);
REGISTER_OP_CUDA_KERNEL(cross_entropy2,
ops::CrossEntropyOpKernel2<CUDACtx, float>,
ops::CrossEntropyOpKernel2<CUDACtx, double>,
ops::CrossEntropyOpKernel2<CUDACtx, plat::float16>);
REGISTER_OP_CUDA_KERNEL(
cross_entropy_grad2, ops::CrossEntropyGradientOpKernel2<CUDACtx, float>,
ops::CrossEntropyGradientOpKernel2<CUDACtx, double>,
ops::CrossEntropyGradientOpKernel2<CUDACtx, plat::float16>);
paddle/fluid/operators/cross_entropy_op.h
浏览文件 @ 9b3e79ac
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math.h"
#include "paddle/fluid/operators/math/cross_entropy.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/for_range.h"
......@@ -137,5 +138,124 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
}
};
template <typename T>
struct HardLabelCrossEntropyForwardFunctor {
HardLabelCrossEntropyForwardFunctor(const T* x, T* y, T* match_x,
const int64_t* label,
int64_t ignore_index,
int64_t feature_size)
: x_(x),
y_(y),
match_x_(match_x),
label_(label),
ignore_index_(ignore_index),
feature_size_(feature_size) {}
HOSTDEVICE void operator()(int64_t idx) const {
auto label = label_[idx];
if (label != ignore_index_) {
auto match_x = x_[idx * feature_size_ + label];
y_[idx] = -math::TolerableValue<T>()(real_log(match_x));
match_x_[idx] = match_x;
} else {
y_[idx] = 0;
match_x_[idx] = 0; // any value is ok
}
}
const T* x_;
T* y_;
T* match_x_;
const int64_t* label_;
int64_t ignore_index_;
int64_t feature_size_;
};
template <typename T>
struct HardLabelCrossEntropyBackwardFunctor {
HardLabelCrossEntropyBackwardFunctor(T* dx, const T* dy, const T* match_x,
const int64_t* label,
int64_t ignore_index,
int64_t feature_size)
: dx_(dx),
dy_(dy),
match_x_(match_x),
label_(label),
ignore_index_(ignore_index),
feature_size_(feature_size) {}
HOSTDEVICE void operator()(int64_t idx) const {
auto row_idx = idx / feature_size_;
auto col_idx = idx % feature_size_;
auto label = label_[row_idx];
if (label == col_idx && label != ignore_index_) {
dx_[idx] = -dy_[row_idx] / match_x_[row_idx];
} else {
dx_[idx] = 0;
}
}
T* dx_;
const T* dy_;
const T* match_x_;
const int64_t* label_;
int64_t ignore_index_;
int64_t feature_size_;
};
template <typename DeviceContext, typename T>
class CrossEntropyOpKernel2 : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<Tensor>("X");
auto* label = ctx.Input<Tensor>("Label");
auto* y = ctx.Output<Tensor>("Y");
auto* match_x = ctx.Output<Tensor>("MatchX");
auto& x_dims = x->dims();
auto feature_size = x_dims[x_dims.size() - 1];
auto batch_size = framework::product(x->dims()) / feature_size;
auto* p_x = x->data<T>();
auto* p_label = label->data<int64_t>();
auto* p_y = y->mutable_data<T>(ctx.GetPlace());
auto* p_match_x = match_x->mutable_data<T>(ctx.GetPlace());
auto ignore_index = ctx.Attr<int>("ignore_index");
platform::ForRange<DeviceContext> for_range(
ctx.template device_context<DeviceContext>(), batch_size);
for_range(HardLabelCrossEntropyForwardFunctor<T>(
p_x, p_y, p_match_x, p_label, ignore_index, feature_size));
}
};
template <typename DeviceContext, typename T>
class CrossEntropyGradientOpKernel2 : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto* match_x = ctx.Input<Tensor>("MatchX");
auto* label = ctx.Input<Tensor>("Label");
auto* p_dx = dx->mutable_data<T>(ctx.GetPlace());
auto* p_dy = dy->data<T>();
auto* p_match_x = match_x->data<T>();
auto* p_label = label->data<int64_t>();
int64_t ignore_index = ctx.Attr<int>("ignore_index");
int rank = dx->dims().size();
int64_t feature_size = dx->dims()[rank - 1];
int64_t batch_size = framework::product(dx->dims()) / feature_size;
platform::ForRange<DeviceContext> for_range(
ctx.template device_context<DeviceContext>(),
batch_size * feature_size);
for_range(HardLabelCrossEntropyBackwardFunctor<T>(
p_dx, p_dy, p_match_x, p_label, ignore_index, feature_size));
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/expand_op.cc
浏览文件 @ 9b3e79ac
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/expand_op.h"
#include <memory>
#include <vector>
namespace paddle {
......@@ -138,15 +139,35 @@ class ExpandGradOp : public framework::OperatorWithKernel {
}
};
class ExpandGradOpDescMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
op->SetType("expand_grad");
op->SetInput("X", Input("X"));
op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
op->SetAttrMap(Attrs());
return op;
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(expand, ops::ExpandOp, ops::ExpandOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
ops::ExpandGradOpDescMaker);
REGISTER_OPERATOR(expand_grad, ops::ExpandGradOp);
REGISTER_OP_CPU_KERNEL(
expand, ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>);
expand, ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>,
ops::ExpandKernel<paddle::platform::CPUDeviceContext, double>,
ops::ExpandKernel<paddle::platform::CPUDeviceContext, int>,
ops::ExpandKernel<paddle::platform::CPUDeviceContext, bool>);
REGISTER_OP_CPU_KERNEL(
expand_grad,
ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>);
ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, double>);
paddle/fluid/operators/expand_op.cu
浏览文件 @ 9b3e79ac
......@@ -15,7 +15,11 @@ limitations under the License. */
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
expand, ops::ExpandKernel<paddle::platform::CUDADeviceContext, float>);
expand, ops::ExpandKernel<paddle::platform::CUDADeviceContext, float>,
ops::ExpandKernel<paddle::platform::CUDADeviceContext, double>,
ops::ExpandKernel<paddle::platform::CUDADeviceContext, int>,
ops::ExpandKernel<paddle::platform::CUDADeviceContext, bool>);
REGISTER_OP_CUDA_KERNEL(
expand_grad,
ops::ExpandGradKernel<paddle::platform::CUDADeviceContext, float>);
ops::ExpandGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::ExpandGradKernel<paddle::platform::CUDADeviceContext, double>);
paddle/fluid/operators/math.h 0 → 100644
浏览文件 @ 9b3e79ac
// Copyright (c) 2019 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.
// Copyright (c) 2019 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.
#pragma once
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/platform/hostdevice.h"
#include "math.h" // NOLINT
namespace paddle {
namespace operators {
inline HOSTDEVICE platform::float16 real_exp(platform::float16 x) {
return static_cast<platform::float16>(::expf(static_cast<float>(x)));
}
inline HOSTDEVICE float real_exp(float x) { return ::expf(x); }
inline HOSTDEVICE double real_exp(double x) { return ::exp(x); }
inline HOSTDEVICE platform::float16 real_log(platform::float16 x) {
return static_cast<platform::float16>(::logf(static_cast<float>(x)));
}
inline HOSTDEVICE float real_log(float x) { return ::logf(x); }
inline HOSTDEVICE double real_log(double x) { return ::log(x); }
} // namespace operators
} // namespace paddle
paddle/fluid/operators/recurrent_op.cc
浏览文件 @ 9b3e79ac
......@@ -270,7 +270,9 @@ class RecurrentOp : public RecurrentBase {
// Every inputs are linked now, execute!
executor.Run(*program, &cur_scope, block->ID(),
false /*create_local_scope*/);
false /*create_local_scope*/, true /*create_vars*/,
std::vector<std::string>() /*skip_ref_cnt_vars*/,
true /*force_disable_gc*/);
// get device context from pool
platform::DeviceContextPool &pool =
......@@ -385,7 +387,9 @@ class RecurrentGradOp : public RecurrentBase {
VLOG(5) << "Recurrent memory linking finished ";
// Run step block with cur_scope
executor.Run(*program, &cur_scope, block->ID(),
false /*create_local_scope*/);
false /*create_local_scope*/, true /*create_vars*/,
std::vector<std::string>() /*skip_ref_cnt_vars*/,
true /*force_disable_gc*/);
VLOG(5) << "executor.Run finished ";
......
python/paddle/fluid/executor.py
浏览文件 @ 9b3e79ac
......@@ -588,7 +588,7 @@ class Executor(object):
fetch_var_name=fetch_var_name)
self._feed_data(program, feed, feed_var_name, scope)
exe.run(program.desc, scope, 0, True, True)
exe.run(program.desc, scope, 0, True, True, fetch_var_name)
outs = self._fetch_data(fetch_list, fetch_var_name, scope)
if return_numpy:
outs = as_numpy(outs)
......
python/paddle/fluid/layers/nn.py
浏览文件 @ 9b3e79ac
......@@ -1398,6 +1398,8 @@ def cross_entropy(input, label, soft_label=False, ignore_index=kIgnoreIndex):
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
"""
if not soft_label:
return cross_entropy2(input, label, ignore_index)
helper = LayerHelper('cross_entropy', **locals())
out = helper.create_variable_for_type_inference(dtype=input.dtype)
helper.append_op(
......@@ -1410,6 +1412,22 @@ def cross_entropy(input, label, soft_label=False, ignore_index=kIgnoreIndex):
return out
def cross_entropy2(input, label, ignore_index=kIgnoreIndex):
helper = LayerHelper('cross_entropy2', **locals())
out = helper.create_variable_for_type_inference(dtype=input.dtype)
xshape = helper.create_variable_for_type_inference(dtype=input.dtype)
match_x = helper.create_variable_for_type_inference(dtype=input.dtype)
helper.append_op(
type='cross_entropy2',
inputs={'X': [input],
'Label': [label]},
outputs={'Y': [out],
'MatchX': [match_x],
'XShape': [xshape]},
attrs={'ignore_index': ignore_index})
return out
def bpr_loss(input, label, name=None):
"""
Bayesian Personalized Ranking Loss Operator.
......
python/paddle/fluid/tests/unittests/test_cross_entropy2_op.py 0 → 100644
浏览文件 @ 9b3e79ac
# Copyright (c) 2019 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.
from op_test import OpTest
import unittest
import numpy as np
import six
class CrossEntropy2OpTestBase(OpTest):
def initParameters(self):
return [32, 64], 'float32', -100
def calc_output(self, logits, label, ignore_index):
ret = np.zeros(shape=label.shape, dtype=logits.dtype)
match_x = np.zeros(shape=label.shape, dtype=logits.dtype)
for idx in six.moves.range(label.shape[0]):
if label[idx] == ignore_index:
continue
match_x[idx] = logits[idx][label[idx]]
ret[idx] = -np.log(match_x[idx])
return ret, match_x
def setUp(self):
self.shape, self.dtype, self.ignore_index = self.initParameters()
self.op_type = 'cross_entropy2'
feature_size = int(self.shape[-1])
batch_size = int(np.prod(self.shape) / feature_size)
logits = (np.random.random(size=self.shape) + 1).astype(self.dtype)
label = np.random.random_integers(
low=0, high=feature_size - 1,
size=self.shape[0:-1] + [1]).astype('int64')
outputs, match_x = self.calc_output(
np.reshape(logits, [batch_size, feature_size]),
np.reshape(label, [batch_size, 1]), self.ignore_index)
self.inputs = {'X': logits, 'Label': label}
self.outputs = {
'Y': np.reshape(outputs, label.shape),
'MatchX': np.reshape(match_x, label.shape),
'XShape': np.zeros(
shape=logits.shape, dtype=logits.dtype)
}
self.attrs = {'ignore_index': self.ignore_index}
def test_check_output(self):
self.check_output(no_check_set=['XShape'])
def test_check_grad(self):
self.check_grad(
inputs_to_check=['X'],
output_names=['Y'],
no_grad_set=['XShape', 'MatchX', 'Label'])
class CrossEntropy2OpTest2(CrossEntropy2OpTestBase):
def initParameters(self):
return [32, 64], 'float64', 3
class CrossEntropy2OpTest3(CrossEntropy2OpTestBase):
def initParameters(self):
return [4, 8, 16, 32], 'float32', -100
class CrossEntropy2OpTest4(CrossEntropy2OpTestBase):
def initParameters(self):
return [4, 8, 16, 32], 'float32', 3
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
浏览文件 @ 9b3e79ac
......@@ -515,8 +515,8 @@ class TestLocalLookupTable(TestDistLookupTableBase):
ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'cross_entropy2', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad2', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'split_selected_rows', 'send', 'sequence_pool_grad',
'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
......@@ -555,8 +555,8 @@ class TestDistLookupTable(TestDistLookupTableBase):
ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'elementwise_add', 'cross_entropy2', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad2', 'elementwise_add_grad', 'send',
'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'lookup_table_grad', 'split_selected_rows', 'send',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
......@@ -603,8 +603,8 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'cross_entropy2', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad2', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'split_selected_rows', 'send', 'sequence_pool_grad',
'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
......@@ -643,8 +643,8 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'elementwise_add', 'cross_entropy2', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad2', 'elementwise_add_grad', 'send',
'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'lookup_table_grad', 'split_selected_rows', 'send',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
......@@ -831,8 +831,8 @@ class TestRemoteLookupTable(TestDistLookupTableBase):
ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'cross_entropy2', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad2', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'split_selected_rows', 'send', 'sequence_pool_grad',
'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
......
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