lujun
7 years ago
parent
bed0ecf3d2
commit
622fe6a56b
@ -0,0 +1,24 @@
|
|||||||
|
/* Copyright (c) 2018 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/fluid/operators/load_combine_op.h"
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
load_combine,
|
||||||
|
ops::LoadCombineOpKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::LoadCombineOpKernel<paddle::platform::CUDADeviceContext, double>,
|
||||||
|
ops::LoadCombineOpKernel<paddle::platform::CUDADeviceContext, int>,
|
||||||
|
ops::LoadCombineOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
|
||||||
@ -0,0 +1,102 @@
|
|||||||
|
/* Copyright (c) 2018 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 <fstream>
|
||||||
|
#include <string>
|
||||||
|
#include <vector>
|
||||||
|
|
||||||
|
#include "paddle/fluid/framework/data_type.h"
|
||||||
|
#include "paddle/fluid/framework/data_type_transform.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/platform/device_context.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class LoadCombineOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext &ctx) const override {
|
||||||
|
auto place = ctx.GetPlace();
|
||||||
|
auto filename = ctx.Attr<std::string>("file_path");
|
||||||
|
auto load_as_fp16 = ctx.Attr<bool>("load_as_fp16");
|
||||||
|
auto model_from_memory = ctx.Attr<bool>("model_from_memory");
|
||||||
|
auto &out_var_names = ctx.Outputs("Out");
|
||||||
|
|
||||||
|
PADDLE_ENFORCE_GT(
|
||||||
|
static_cast<int>(out_var_names.size()), 0,
|
||||||
|
"The number of output variables should be greater than 0.");
|
||||||
|
if (!model_from_memory) {
|
||||||
|
std::ifstream fin(filename, std::ios::binary);
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(fin),
|
||||||
|
"Cannot open file %s for load_combine op", filename);
|
||||||
|
LoadParamsFromBuffer(ctx, place, &fin, load_as_fp16, out_var_names);
|
||||||
|
} else {
|
||||||
|
PADDLE_ENFORCE(!filename.empty(), "Cannot load file from memory");
|
||||||
|
std::stringstream fin(filename, std::ios::in | std::ios::binary);
|
||||||
|
LoadParamsFromBuffer(ctx, place, &fin, load_as_fp16, out_var_names);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
void LoadParamsFromBuffer(
|
||||||
|
const framework::ExecutionContext &context, const platform::Place &place,
|
||||||
|
std::istream *buffer, bool load_as_fp16,
|
||||||
|
const std::vector<std::string> &out_var_names) const {
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
auto out_vars = context.MultiOutputVar("Out");
|
||||||
|
|
||||||
|
for (size_t i = 0; i < out_var_names.size(); i++) {
|
||||||
|
PADDLE_ENFORCE(out_vars[i] != nullptr,
|
||||||
|
"Output variable %s cannot be found", out_var_names[i]);
|
||||||
|
|
||||||
|
auto *tensor = out_vars[i]->GetMutable<framework::LoDTensor>();
|
||||||
|
|
||||||
|
// Error checking
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(*buffer), "Cannot read more");
|
||||||
|
|
||||||
|
// Get data from fin to tensor
|
||||||
|
DeserializeFromStream(*buffer, tensor, dev_ctx);
|
||||||
|
|
||||||
|
auto in_dtype = tensor->type();
|
||||||
|
auto out_dtype =
|
||||||
|
load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
|
||||||
|
|
||||||
|
if (in_dtype != out_dtype) {
|
||||||
|
// convert to float16 tensor
|
||||||
|
auto in_kernel_type = framework::OpKernelType(in_dtype, place);
|
||||||
|
auto out_kernel_type = framework::OpKernelType(out_dtype, place);
|
||||||
|
framework::LoDTensor fp16_tensor;
|
||||||
|
// copy LoD info to the new tensor
|
||||||
|
fp16_tensor.set_lod(tensor->lod());
|
||||||
|
framework::TransDataType(in_kernel_type, out_kernel_type, *tensor,
|
||||||
|
&fp16_tensor);
|
||||||
|
|
||||||
|
// reset output tensor
|
||||||
|
out_vars[i]->Clear();
|
||||||
|
tensor = out_vars[i]->GetMutable<framework::LoDTensor>();
|
||||||
|
tensor->set_lod(fp16_tensor.lod());
|
||||||
|
tensor->ShareDataWith(fp16_tensor);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
buffer->peek();
|
||||||
|
PADDLE_ENFORCE(buffer->eof(),
|
||||||
|
"You are not allowed to load partial data via "
|
||||||
|
"load_combine_op, use load_op instead.");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,23 @@
|
|||||||
|
/* 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/fluid/operators/load_op.h"
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
load, ops::LoadOpKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::LoadOpKernel<paddle::platform::CUDADeviceContext, double>,
|
||||||
|
ops::LoadOpKernel<paddle::platform::CUDADeviceContext, int>,
|
||||||
|
ops::LoadOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
|
||||||
@ -0,0 +1,102 @@
|
|||||||
|
/* 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. */
|
||||||
|
|
||||||
|
#pragma once
|
||||||
|
|
||||||
|
#include <fstream>
|
||||||
|
#include <string>
|
||||||
|
|
||||||
|
#include "paddle/fluid/framework/data_type_transform.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/platform/device_context.h"
|
||||||
|
#include "paddle/fluid/platform/profiler.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class LoadOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext &ctx) const override {
|
||||||
|
auto place = ctx.GetPlace();
|
||||||
|
// FIXME(yuyang18): We save variable to local file now, but we should change
|
||||||
|
// it to save an output stream.
|
||||||
|
auto filename = ctx.Attr<std::string>("file_path");
|
||||||
|
std::ifstream fin(filename, std::ios::binary);
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s for load op",
|
||||||
|
filename);
|
||||||
|
|
||||||
|
auto out_var_name = ctx.Outputs("Out").data();
|
||||||
|
auto *out_var = ctx.OutputVar("Out");
|
||||||
|
|
||||||
|
PADDLE_ENFORCE(out_var != nullptr, "Output variable %s cannot be found ",
|
||||||
|
out_var_name);
|
||||||
|
|
||||||
|
PADDLE_ENFORCE(out_var != nullptr, "Output variable cannot be found ");
|
||||||
|
|
||||||
|
if (out_var->IsType<framework::LoDTensor>()) {
|
||||||
|
LoadLodTensor(fin, place, out_var, ctx);
|
||||||
|
} else if (out_var->IsType<framework::SelectedRows>()) {
|
||||||
|
LoadSelectedRows(fin, place, out_var);
|
||||||
|
} else {
|
||||||
|
PADDLE_ENFORCE(
|
||||||
|
false,
|
||||||
|
"Load only support LoDTensor and SelectedRows, %s has wrong type",
|
||||||
|
out_var_name);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
void LoadLodTensor(std::istream &fin, const platform::Place &place,
|
||||||
|
framework::Variable *var,
|
||||||
|
const framework::ExecutionContext &ctx) const {
|
||||||
|
// get device context from pool
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
auto *tensor = var->GetMutable<framework::LoDTensor>();
|
||||||
|
DeserializeFromStream(fin, tensor, dev_ctx);
|
||||||
|
|
||||||
|
auto load_as_fp16 = ctx.Attr<bool>("load_as_fp16");
|
||||||
|
auto in_dtype = tensor->type();
|
||||||
|
auto out_dtype = load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
|
||||||
|
|
||||||
|
if (in_dtype != out_dtype) {
|
||||||
|
// convert to float16 tensor
|
||||||
|
auto in_kernel_type = framework::OpKernelType(in_dtype, place);
|
||||||
|
auto out_kernel_type = framework::OpKernelType(out_dtype, place);
|
||||||
|
framework::LoDTensor fp16_tensor;
|
||||||
|
// copy LoD info to the new tensor
|
||||||
|
fp16_tensor.set_lod(tensor->lod());
|
||||||
|
framework::TransDataType(in_kernel_type, out_kernel_type, *tensor,
|
||||||
|
&fp16_tensor);
|
||||||
|
|
||||||
|
// reset output tensor
|
||||||
|
var->Clear();
|
||||||
|
tensor = var->GetMutable<framework::LoDTensor>();
|
||||||
|
tensor->set_lod(fp16_tensor.lod());
|
||||||
|
tensor->ShareDataWith(fp16_tensor);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
void LoadSelectedRows(std::istream &fin, const platform::Place &place,
|
||||||
|
framework::Variable *var) const {
|
||||||
|
auto *selectedRows = var->GetMutable<framework::SelectedRows>();
|
||||||
|
// get device context from pool
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
framework::DeserializeFromStream(fin, selectedRows, dev_ctx);
|
||||||
|
selectedRows->SyncIndex();
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,24 @@
|
|||||||
|
/* Copyright (c) 2018 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/fluid/operators/save_combine_op.h"
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
save_combine,
|
||||||
|
ops::SaveCombineOpKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::SaveCombineOpKernel<paddle::platform::CUDADeviceContext, double>,
|
||||||
|
ops::SaveCombineOpKernel<paddle::platform::CUDADeviceContext, int>,
|
||||||
|
ops::SaveCombineOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
|
||||||
@ -0,0 +1,95 @@
|
|||||||
|
/* Copyright (c) 2018 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 <stdint.h>
|
||||||
|
#include <fstream>
|
||||||
|
#include <numeric>
|
||||||
|
#include <sstream>
|
||||||
|
#include <string>
|
||||||
|
|
||||||
|
#include "paddle/fluid/framework/data_type.h"
|
||||||
|
#include "paddle/fluid/framework/data_type_transform.h"
|
||||||
|
#include "paddle/fluid/framework/framework.pb.h"
|
||||||
|
#include "paddle/fluid/framework/lod_tensor.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/platform/device_context.h"
|
||||||
|
#include "paddle/fluid/platform/port.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class SaveCombineOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext &ctx) const override {
|
||||||
|
auto place = ctx.GetPlace();
|
||||||
|
auto filename = ctx.Attr<std::string>("file_path");
|
||||||
|
auto overwrite = ctx.Attr<bool>("overwrite");
|
||||||
|
auto save_as_fp16 = ctx.Attr<bool>("save_as_fp16");
|
||||||
|
|
||||||
|
bool is_present = FileExists(filename);
|
||||||
|
if (is_present && !overwrite) {
|
||||||
|
PADDLE_THROW("%s exists!, cannot save_combine to it when overwrite=false",
|
||||||
|
filename, overwrite);
|
||||||
|
}
|
||||||
|
|
||||||
|
MkDirRecursively(DirName(filename).c_str());
|
||||||
|
std::ofstream fout(filename, std::ios::binary);
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(fout), "Cannot open %s to write",
|
||||||
|
filename);
|
||||||
|
|
||||||
|
auto &inp_var_names = ctx.Inputs("X");
|
||||||
|
auto &inp_vars = ctx.MultiInputVar("X");
|
||||||
|
PADDLE_ENFORCE_GT(static_cast<int>(inp_var_names.size()), 0,
|
||||||
|
"The number of input variables should be greater than 0");
|
||||||
|
|
||||||
|
// get device context from pool
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
|
||||||
|
for (size_t i = 0; i < inp_var_names.size(); i++) {
|
||||||
|
PADDLE_ENFORCE(inp_vars[i] != nullptr,
|
||||||
|
"Cannot find variable %s for save_combine_op",
|
||||||
|
inp_var_names[i]);
|
||||||
|
PADDLE_ENFORCE(inp_vars[i]->IsType<framework::LoDTensor>(),
|
||||||
|
"SaveCombineOp only supports LoDTensor, %s has wrong type",
|
||||||
|
inp_var_names[i]);
|
||||||
|
|
||||||
|
auto &tensor = inp_vars[i]->Get<framework::LoDTensor>();
|
||||||
|
// Serialize tensors one by one
|
||||||
|
|
||||||
|
// Check types to see if a fp16 transformation is required
|
||||||
|
auto in_dtype = tensor.type();
|
||||||
|
auto out_dtype =
|
||||||
|
save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
|
||||||
|
|
||||||
|
if (in_dtype != out_dtype) {
|
||||||
|
auto in_kernel_type = framework::OpKernelType(in_dtype, place);
|
||||||
|
auto out_kernel_type = framework::OpKernelType(out_dtype, place);
|
||||||
|
framework::LoDTensor out;
|
||||||
|
// copy LoD info to the new tensor
|
||||||
|
out.set_lod(tensor.lod());
|
||||||
|
framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out);
|
||||||
|
framework::SerializeToStream(fout, out, dev_ctx);
|
||||||
|
} else {
|
||||||
|
framework::SerializeToStream(fout, tensor, dev_ctx);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
fout.close();
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,26 @@
|
|||||||
|
/* 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/fluid/operators/save_op.h"
|
||||||
|
#include "paddle/fluid/platform/float16.h"
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
save, ops::SaveOpKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::SaveOpKernel<paddle::platform::CUDADeviceContext, double>,
|
||||||
|
ops::SaveOpKernel<paddle::platform::CUDADeviceContext, int>,
|
||||||
|
ops::SaveOpKernel<paddle::platform::CUDADeviceContext, int64_t>,
|
||||||
|
ops::SaveOpKernel<paddle::platform::CUDADeviceContext,
|
||||||
|
paddle::platform::float16>);
|
||||||
@ -0,0 +1,133 @@
|
|||||||
|
/* 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. */
|
||||||
|
|
||||||
|
#pragma once
|
||||||
|
|
||||||
|
#include <stdint.h>
|
||||||
|
#include <fstream>
|
||||||
|
#include <numeric>
|
||||||
|
#include <string>
|
||||||
|
#include <vector>
|
||||||
|
|
||||||
|
#include "paddle/fluid/framework/data_type.h"
|
||||||
|
#include "paddle/fluid/framework/data_type_transform.h"
|
||||||
|
#include "paddle/fluid/framework/framework.pb.h"
|
||||||
|
#include "paddle/fluid/framework/lod_tensor.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/framework/selected_rows.h"
|
||||||
|
#include "paddle/fluid/framework/variable.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
// define LOOKUP_TABLE_PATH for checkpoint notify to save lookup table variables
|
||||||
|
// to directory specified.
|
||||||
|
constexpr char LOOKUP_TABLE_PATH[] = "kLookupTablePath";
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class SaveOpKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext &ctx) const override {
|
||||||
|
auto place = ctx.GetPlace();
|
||||||
|
|
||||||
|
auto *input_var = ctx.InputVar("X");
|
||||||
|
auto iname = ctx.Inputs("X").data();
|
||||||
|
PADDLE_ENFORCE(input_var != nullptr, "Cannot find variable %s for save_op",
|
||||||
|
iname);
|
||||||
|
|
||||||
|
if (input_var->IsType<framework::LoDTensor>()) {
|
||||||
|
SaveLodTensor(ctx, place, input_var);
|
||||||
|
} else if (input_var->IsType<framework::SelectedRows>()) {
|
||||||
|
SaveSelectedRows(ctx, place, input_var);
|
||||||
|
} else {
|
||||||
|
PADDLE_ENFORCE(
|
||||||
|
false,
|
||||||
|
"SaveOp only support LoDTensor and SelectedRows, %s has wrong type",
|
||||||
|
iname);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
void SaveLodTensor(const framework::ExecutionContext &ctx,
|
||||||
|
const platform::Place &place,
|
||||||
|
const framework::Variable *var) const {
|
||||||
|
auto filename = ctx.Attr<std::string>("file_path");
|
||||||
|
auto overwrite = ctx.Attr<bool>("overwrite");
|
||||||
|
|
||||||
|
if (FileExists(filename) && !overwrite) {
|
||||||
|
PADDLE_THROW("%s is existed, cannot save to it when overwrite=false",
|
||||||
|
filename, overwrite);
|
||||||
|
}
|
||||||
|
|
||||||
|
MkDirRecursively(DirName(filename).c_str());
|
||||||
|
|
||||||
|
auto &tensor = var->Get<framework::LoDTensor>();
|
||||||
|
|
||||||
|
// get device context from pool
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
|
||||||
|
// FIXME(yuyang18): We save variable to local file now, but we should change
|
||||||
|
// it to save an output stream.
|
||||||
|
std::ofstream fout(filename, std::ios::binary);
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(fout), "Cannot open %s to write",
|
||||||
|
filename);
|
||||||
|
|
||||||
|
auto save_as_fp16 = ctx.Attr<bool>("save_as_fp16");
|
||||||
|
auto in_dtype = tensor.type();
|
||||||
|
auto out_dtype = save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
|
||||||
|
|
||||||
|
if (in_dtype != out_dtype) {
|
||||||
|
auto in_kernel_type = framework::OpKernelType(in_dtype, place);
|
||||||
|
auto out_kernel_type = framework::OpKernelType(out_dtype, place);
|
||||||
|
framework::LoDTensor out;
|
||||||
|
framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out);
|
||||||
|
// copy LoD info to the new tensor
|
||||||
|
out.set_lod(tensor.lod());
|
||||||
|
framework::SerializeToStream(fout, out, dev_ctx);
|
||||||
|
} else {
|
||||||
|
framework::SerializeToStream(fout, tensor, dev_ctx);
|
||||||
|
}
|
||||||
|
fout.close();
|
||||||
|
}
|
||||||
|
|
||||||
|
void SaveSelectedRows(const framework::ExecutionContext &ctx,
|
||||||
|
const platform::Place &place,
|
||||||
|
const framework::Variable *var) const {
|
||||||
|
framework::Variable *out_put_var = ctx.OutputVar(LOOKUP_TABLE_PATH);
|
||||||
|
PADDLE_ENFORCE(
|
||||||
|
out_put_var != nullptr,
|
||||||
|
"Can not find variable kLookupTablePath for SaveSelectedRows");
|
||||||
|
auto *lt_var = out_put_var->GetMutable<std::string>();
|
||||||
|
|
||||||
|
std::string filename = lt_var->data();
|
||||||
|
VLOG(4) << "SaveSelectedRows get File name: " << filename;
|
||||||
|
|
||||||
|
MkDirRecursively(DirName(filename).c_str());
|
||||||
|
|
||||||
|
auto &selectedRows = var->Get<framework::SelectedRows>();
|
||||||
|
|
||||||
|
// get device context from pool
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(place);
|
||||||
|
|
||||||
|
// FIXME(yuyang18): We save variable to local file now, but we should change
|
||||||
|
// it to save an output stream.
|
||||||
|
std::ofstream fout(filename, std::ios::binary);
|
||||||
|
PADDLE_ENFORCE(static_cast<bool>(fout), "Cannot open %s to write",
|
||||||
|
filename);
|
||||||
|
framework::SerializeToStream(fout, selectedRows, dev_ctx);
|
||||||
|
fout.close();
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
Loading…
Reference in new issue