tangwei12
7 years ago
committed by
GitHub
commit
5a4a24cce1
File diff suppressed because it is too large
Load Diff
|
Before Width: | Height: | Size: 94 KiB |
@ -1,131 +0,0 @@
|
||||
// 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 <string>
|
||||
|
||||
#include "paddle/contrib/tape/tape.h"
|
||||
#include "paddle/contrib/tape/variable.h"
|
||||
#include "paddle/fluid/framework/type_defs.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace tape {
|
||||
|
||||
class Function {};
|
||||
|
||||
class Fill {
|
||||
public:
|
||||
Fill(const std::string &initializer, const framework::AttributeMap &attrs)
|
||||
: initializer_(initializer), attrs_(attrs) {}
|
||||
|
||||
void operator()(VariableHandle var) {
|
||||
get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_);
|
||||
}
|
||||
|
||||
private:
|
||||
const std::string initializer_;
|
||||
const framework::AttributeMap attrs_;
|
||||
};
|
||||
|
||||
class Mean {
|
||||
public:
|
||||
VariableHandle operator()(VariableHandle var) {
|
||||
VariableHandle out(new Variable("mean"));
|
||||
get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {});
|
||||
return out;
|
||||
}
|
||||
};
|
||||
|
||||
class Linear {
|
||||
public:
|
||||
Linear(int in_dim, int out_dim, const std::string &act)
|
||||
: w_(new Variable("LinearWeight")),
|
||||
b_(new Variable("LinearBias")),
|
||||
act_(act) {
|
||||
Tape init_tape;
|
||||
|
||||
std::string initializer = "fill_constant";
|
||||
framework::AttributeMap attrs;
|
||||
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
|
||||
attrs["shape"] = std::vector<int>{in_dim, out_dim};
|
||||
attrs["value"] = 1.0f;
|
||||
init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
|
||||
|
||||
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
|
||||
attrs["shape"] = std::vector<int>{out_dim};
|
||||
attrs["value"] = 1.0f;
|
||||
init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
|
||||
|
||||
init_tape.Forward();
|
||||
}
|
||||
|
||||
VariableHandle operator()(VariableHandle input) {
|
||||
VariableHandle pre_bias(new Variable("linear"));
|
||||
get_global_tape().AddOp("mul",
|
||||
{{"X", {input}}, {"Y", {w_}}},
|
||||
{{"Out", {pre_bias}}},
|
||||
{{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
|
||||
VariableHandle pre_act(new Variable("linear"));
|
||||
get_global_tape().AddOp("elementwise_add",
|
||||
{{"X", {pre_bias}}, {"Y", {b_}}},
|
||||
{{"Out", {pre_act}}},
|
||||
{{"axis", 1}});
|
||||
VariableHandle post_act(new Variable("linear"));
|
||||
get_global_tape().AddOp(
|
||||
act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {});
|
||||
return post_act;
|
||||
}
|
||||
|
||||
std::vector<VariableHandle> Params() { return {w_, b_}; }
|
||||
|
||||
private:
|
||||
VariableHandle w_;
|
||||
VariableHandle b_;
|
||||
std::string act_;
|
||||
};
|
||||
|
||||
class SGD {
|
||||
public:
|
||||
SGD(float learning_rate) : learning_rate_(new Variable("sgd")) {
|
||||
Tape init_tape;
|
||||
|
||||
std::string initializer = "fill_constant";
|
||||
framework::AttributeMap attrs;
|
||||
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
|
||||
attrs["shape"] = std::vector<int>{1};
|
||||
attrs["value"] = learning_rate;
|
||||
init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs);
|
||||
|
||||
init_tape.Forward();
|
||||
}
|
||||
|
||||
void operator()(VariableHandle input) {
|
||||
PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(),
|
||||
"optimization must happen after the backward");
|
||||
Tape temp_tape;
|
||||
temp_tape.AddOp("sgd",
|
||||
{{"Param", {input}},
|
||||
{"LearningRate", {learning_rate_}},
|
||||
{"Grad", {input->Grad()}}},
|
||||
{{"ParamOut", {input}}},
|
||||
{});
|
||||
temp_tape.Forward();
|
||||
}
|
||||
|
||||
private:
|
||||
VariableHandle learning_rate_;
|
||||
};
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@ -1,64 +0,0 @@
|
||||
// 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 <map>
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "paddle/contrib/tape/variable.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace tape {
|
||||
|
||||
using VariableHandleMap = std::map<std::string, std::vector<VariableHandle>>;
|
||||
|
||||
struct OpHandle {
|
||||
OpHandle(const std::string &type,
|
||||
const VariableHandleMap &in_vars,
|
||||
const VariableHandleMap &out_vars,
|
||||
const framework::AttributeMap &attrs)
|
||||
: type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {}
|
||||
|
||||
std::string type_;
|
||||
VariableHandleMap inputs_;
|
||||
VariableHandleMap outputs_;
|
||||
framework::AttributeMap attrs_;
|
||||
};
|
||||
|
||||
class Tape {
|
||||
public:
|
||||
void AddOp(const std::string &type,
|
||||
const VariableHandleMap &in_vars,
|
||||
VariableHandleMap out_vars,
|
||||
const framework::AttributeMap &attrs);
|
||||
void Forward();
|
||||
void Backward(VariableHandle target);
|
||||
|
||||
bool HasBeenBackwarded() { return has_been_backwarded_; }
|
||||
|
||||
private:
|
||||
bool has_been_backwarded_ = false;
|
||||
size_t current_position_ = 0;
|
||||
|
||||
std::vector<OpHandle> tape_;
|
||||
std::shared_ptr<Tape> backward_tape_;
|
||||
};
|
||||
|
||||
Tape &get_global_tape();
|
||||
|
||||
void reset_global_tape();
|
||||
}
|
||||
}
|
||||
@ -1,61 +0,0 @@
|
||||
// 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 "gtest/gtest.h"
|
||||
#include "paddle/contrib/tape/function.h"
|
||||
|
||||
using namespace paddle::tape;
|
||||
|
||||
TEST(Tape, TestMLP) {
|
||||
LOG(INFO) << "TestMLP";
|
||||
Linear linear1(3, 3, "relu");
|
||||
Linear linear2(3, 3, "relu");
|
||||
Mean mean;
|
||||
|
||||
SGD sgd(0.001);
|
||||
|
||||
std::string initializer = "fill_constant";
|
||||
paddle::framework::AttributeMap attrs;
|
||||
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
|
||||
attrs["shape"] = std::vector<int>{3, 3};
|
||||
attrs["value"] = 1.0f;
|
||||
Fill filler(initializer, attrs);
|
||||
|
||||
for (int i = 0; i < 2; ++i) {
|
||||
reset_global_tape();
|
||||
|
||||
VariableHandle input(new Variable("input"));
|
||||
filler(input);
|
||||
|
||||
auto loss = mean(linear2(linear1(input)));
|
||||
|
||||
get_global_tape().Backward(loss);
|
||||
|
||||
for (auto w : linear1.Params()) {
|
||||
sgd(w);
|
||||
}
|
||||
for (auto w : linear2.Params()) {
|
||||
sgd(w);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
std::vector<paddle::platform::Place> places;
|
||||
places.emplace_back(paddle::platform::CPUPlace());
|
||||
paddle::platform::DeviceContextPool::Init(places);
|
||||
|
||||
testing::InitGoogleTest(&argc, argv);
|
||||
return RUN_ALL_TESTS();
|
||||
}
|
||||
@ -1,33 +0,0 @@
|
||||
// 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/contrib/tape/variable.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace tape {
|
||||
|
||||
void Variable::InitializeVariable() {
|
||||
LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType();
|
||||
framework::proto::VarType::Type var_type = desc_.GetType();
|
||||
if (var_type == framework::proto::VarType::LOD_TENSOR) {
|
||||
var_.GetMutable<framework::LoDTensor>();
|
||||
} else if (var_type == framework::proto::VarType::SELECTED_ROWS) {
|
||||
var_.GetMutable<framework::SelectedRows>();
|
||||
} else {
|
||||
PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]",
|
||||
var_type);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -1,85 +0,0 @@
|
||||
// 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 <memory>
|
||||
|
||||
#include "paddle/fluid/framework/operator.h" // framework::kGradVarSuffix
|
||||
#include "paddle/fluid/framework/program_desc.h"
|
||||
#include "paddle/fluid/framework/variable.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace tape {
|
||||
|
||||
class Variable;
|
||||
using VariableHandle = std::shared_ptr<Variable>;
|
||||
|
||||
/*
|
||||
* Combination of
|
||||
* framework::VarDesc desc_;
|
||||
* framework::Variable var_;
|
||||
*/
|
||||
class Variable {
|
||||
public:
|
||||
Variable(const std::string pre_fix)
|
||||
: desc_(pre_fix + std::to_string(count())) {}
|
||||
|
||||
Variable(const std::string pre_fix, bool is_grad)
|
||||
: desc_(pre_fix + (is_grad ? framework::kGradVarSuffix
|
||||
: std::to_string(count()))) {}
|
||||
|
||||
~Variable() { LOG(INFO) << "Deleting " << Name(); }
|
||||
|
||||
// Instantiate LoDTensor/SelectedRow
|
||||
void InitializeVariable();
|
||||
|
||||
VariableHandle Grad() {
|
||||
if (grad_.expired()) {
|
||||
VariableHandle new_grad(new Variable(desc_.Name(), true));
|
||||
grad_ = new_grad;
|
||||
return new_grad;
|
||||
} else {
|
||||
return VariableHandle(grad_);
|
||||
}
|
||||
}
|
||||
|
||||
// Stochastic Gradient Descent with Momentum
|
||||
// VariableHandle Momentum ();
|
||||
|
||||
// void init(const std::string& initializer,
|
||||
// const framework::AttributeMap& attrs);
|
||||
|
||||
// void value() {};
|
||||
|
||||
const framework::VarDesc& Desc() const { return desc_; }
|
||||
framework::VarDesc* MutableDesc() { return &desc_; }
|
||||
|
||||
// TODO(tonyyang-svail): No need to expose name
|
||||
std::string Name() const { return desc_.Name(); }
|
||||
|
||||
framework::Variable* Var() { return &var_; }
|
||||
|
||||
private:
|
||||
int count() {
|
||||
static int counter = 0;
|
||||
return counter++;
|
||||
}
|
||||
|
||||
framework::VarDesc desc_;
|
||||
framework::Variable var_;
|
||||
|
||||
std::weak_ptr<Variable> grad_;
|
||||
};
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,55 @@
|
||||
/* 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 <string>
|
||||
#include "paddle/fluid/operators/mean_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using framework::DataLayout;
|
||||
template <typename T>
|
||||
class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& context) const override {
|
||||
float mean = context.Attr<float>("mean");
|
||||
float std = context.Attr<float>("std");
|
||||
auto* tensor = context.Output<framework::Tensor>("Out");
|
||||
T* data = tensor->mutable_data<T>(context.GetPlace());
|
||||
|
||||
unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
|
||||
std::minstd_rand engine;
|
||||
if (seed == 0) {
|
||||
seed = std::random_device()();
|
||||
}
|
||||
engine.seed(seed);
|
||||
std::normal_distribution<T> dist(mean, std);
|
||||
int64_t size = tensor->numel();
|
||||
for (int64_t i = 0; i < size; ++i) {
|
||||
data[i] = dist(engine);
|
||||
}
|
||||
|
||||
// The format of output is set as the mkldnn's format
|
||||
// TODO(@mozga-intel) The format of matrix sets inside the another layers.
|
||||
tensor->set_layout(DataLayout::kMKLDNN);
|
||||
tensor->set_format(mkldnn::memory::format::oihw);
|
||||
}
|
||||
};
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
|
||||
REGISTER_OP_KERNEL(gaussian_random, MKLDNN, ::paddle::platform::CPUPlace,
|
||||
ops::GaussianMKLDNNKernel<float>);
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue