Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2298
Star 20931
Fork 5422
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 b26f5050 编写于 作者: Y Yang Yang(Tony) 提交者： GitHub
浏览文件
操作

remove net op and cond_op (#9663) 

* remove net op and cond_op

* fix cpplint

* fix dependency

* delete backward_test; fix compile

* disable batch_norm backward

* rm test_net.py

* make batchnorm test independent of backward.cc

* make test_layer_norm_op independent of backward.cc

* make test_layer_norm_op independent of backward.cc

* delete unused code

* clean up
上级 8d3ce01f
显示空白变更内容
内联 并排
Showing with 205 addition and 2351 deletion
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ b26f5050
......@@ -79,8 +79,7 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
COMMENT "Copy generated python proto into directory paddle/fluid/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context fill_constant_op)
cc_library(backward SRCS backward.cc DEPS operator)
cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
......
paddle/fluid/framework/backward.cc
浏览文件 @ b26f5050
......@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/backward.h"
#include "paddle/fluid/operators/net_op.h"
#include <deque>
#include <list>
......@@ -22,7 +21,6 @@ limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/net_op.h"
namespace paddle {
namespace framework {
......@@ -60,12 +58,7 @@ static inline std::unique_ptr<OperatorBase> CreateGradOp(
if (grad_ops.size() == 1) {
return std::move(grad_ops[0]);
} else {
auto net_op = new operators::NetOp();
for (auto& grad_op : grad_ops) {
net_op->AppendOp(std::move(grad_op));
}
net_op->CompleteAddOp();
return std::unique_ptr<OperatorBase>(net_op);
PADDLE_THROW("Unexpected Branch");
}
}
......@@ -91,10 +84,7 @@ static bool AllInSet(
}
static std::unique_ptr<OperatorBase> NOP() {
auto net_op = new operators::NetOp();
net_op->SetType("@NOP@");
net_op->CompleteAddOp();
return std::unique_ptr<OperatorBase>(net_op);
PADDLE_THROW("Unexpected Branch");
}
// Get backward operator from a forward operator, a recursive implementation.
......@@ -136,110 +126,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
}
// Returned gradient network
auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
if (forwardOp.IsNetOp()) {
// Because forwardOp is a net op, it can static_cast.
auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
// Map from output gradient variable name to operator's indices in
// backward net's ops_. That operator generates that variable.
std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
size_t local_op_id = 0;
// reversely travel forwardNet and collect all duplicate outputs.
for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
++it, ++local_op_id) {
auto& fwd = *it;
auto bwd = BackwardRecursive(*fwd, no_grad_names, grad_to_var, uniq_id);
ForEachVarName(bwd->Outputs(),
[&dup_output_ops, local_op_id](const std::string& out) {
dup_output_ops[out].emplace_back(local_op_id);
return false;
});
net->AppendOp(std::move(bwd));
}
// Get unique ID for this method.
auto uid = uniq_id++;
// TODO(dzh): more comment
// multiple operators which have the same output (y for example) may
// overwrite the same y variable when backward, special operations are token
// to handle this case. For each duplicate output, rename it to an alias
// (original name with a offset), append an `add` op for its operator,
// and finally sum all the alias variable to the final output variable y.
using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
std::list<Pos> insert_position;
for (auto& dup_output_op : dup_output_ops) {
const std::string& name = dup_output_op.first;
// duplicate @Empty@ don't need to be added
if (name == kEmptyVarName) continue;
auto& dup_op = dup_output_op.second;
// no duplicate output
if (dup_op.size() == 1) continue;
// process the duplicate outputs
std::vector<std::string> dup_outputs;
for (size_t i = 0; i < dup_op.size(); ++i) {
// rename each duplicate output to an alias
auto op_offset = dup_op[i];
dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
std::to_string(i));
net->ops_[op_offset]->Rename(name, dup_outputs.back());
}
// collect all the offset for each alias,
// insert a sum operator to add all aliases to output
insert_position.push_back(
{dup_op.back(),
OpRegistry::CreateOp("sum", {{"X", dup_outputs}}, {{"Out", {name}}},
AttributeMap{})});
}
// make sure the inserted `sum` ops follow the BFS order.
insert_position.sort(
[](const Pos& l, const Pos& r) { return l.first > r.first; });
for (auto& pos : insert_position) {
net->InsertOp(pos.first + 1, std::move(pos.second));
}
} else {
std::unique_ptr<OperatorBase> grad_op(
CreateGradOp(forwardOp, no_grad_names, grad_to_var));
ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
const std::string& grad_input) {
if (no_grad_names.count(grad_input)) {
// +1 for \0
std::string prefix = grad_input.substr(
0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
{{"Out", {grad_input}}},
AttributeMap{}));
}
return false;
});
ForEachVarName(grad_op->Outputs(),
[&no_grad_names, &grad_op](const std::string& grad_output) {
if (no_grad_names.count(grad_output)) {
grad_op->Rename(grad_output, kEmptyVarName);
}
return false;
});
if (net->ops_.empty()) { // Current no aux op is added to network
return grad_op;
}
net->AppendOp(std::move(grad_op));
}
net->SetType("@GENERATED_BACKWARD@");
net->CompleteAddOp();
return std::unique_ptr<OperatorBase>(
static_cast<OperatorBase*>(net.release()));
PADDLE_THROW("Unexpected Branch");
}
// See header for comments
......
paddle/fluid/framework/backward_test.cc 已删除 100644 → 0
浏览文件 @ 8d3ce01f
// 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/framework/backward.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/operators/net_op.h"
USE_NO_KERNEL_OP(fill_constant);
namespace paddle {
namespace framework {
using DeviceContext = platform::DeviceContext;
class NoneOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext *ctx) const override {}
};
template <typename Place, typename T>
class NoneKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "Input X of Add");
AddInput("b", "Bias of Add");
AddOutput("Out", "Out of Add");
AddComment("Add Op");
}
};
class RowWiseAddGradMaker : public SingleGradOpDescMaker {
public:
using SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<OpDesc> Apply() const override {
auto grad_op = new OpDesc();
grad_op->SetInput(GradVarName("Out"), OutputGrad("Out"));
grad_op->SetOutput(GradVarName("X"), InputGrad("X"));
grad_op->SetOutput(GradVarName("b"), InputGrad("b"));
grad_op->SetType("rowwise_add_grad");
return std::unique_ptr<OpDesc>(grad_op);
}
};
class MulOpMaker : public OpProtoAndCheckerMaker {
public:
MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "A");
AddInput("Y", "B");
AddOutput("Out", "Out");
AddAttr<int>("x_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
AddAttr<int>("y_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
AddComment("Mul");
}
};
class SigmoidOpMaker : public OpProtoAndCheckerMaker {
public:
SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "X");
AddOutput("Out", "Y");
AddComment("Sigmoid");
}
};
class NoGradOpMaker : public OpProtoAndCheckerMaker {
public:
NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "X input");
AddOutput("Out", "Y output");
AddComment("NoGradOp, same input output. no Grad");
}
};
class FcOp : public operators::NetOp {
public:
FcOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(OpRegistry::CreateOp(
"mul", {{"X", {Input("X")}}, {"Y", {Input("W")}}},
{{"Out", {Output("mul_result")}}}, AttributeMap{}));
auto input_b = Inputs("b");
std::string before_act = "mul_result";
if (input_b.size() != 0) {
AppendOp(OpRegistry::CreateOp(
"rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
{{"Out", {Output("add_result")}}}, AttributeMap{}));
before_act = "add_result";
} else {
auto out_varname = Output("add_result");
if (out_varname != kEmptyVarName) {
this->Rename(out_varname, kEmptyVarName);
}
}
AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
{{"Out", {Output("Out")}}}, AttributeMap{}));
CompleteAddOp(false);
}
};
class FcOpMaker : public OpProtoAndCheckerMaker {
public:
FcOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x");
AddInput("W", "w");
AddInput("b", "b");
AddOutput("mul_result", "").AsIntermediate();
AddOutput("add_result", "").AsIntermediate();
AddOutput("Out", "");
AddComment("");
}
};
class ManyOutputOpMaker : public OpProtoAndCheckerMaker {
public:
ManyOutputOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("x", "x");
AddOutput("y", "y");
AddOutput("z", "z");
AddComment("");
}
};
class FillZeroOpMaker : public OpProtoAndCheckerMaker {
public:
FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x");
AddOutput("Out", "out");
AddComment("");
}
};
class SumOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "the input tensors of sum operator.").AsDuplicable();
AddOutput("Out", "the output tensor of sum operator.");
AddComment("");
}
};
class MultInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MultInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x");
AddInput("H", "h");
AddOutput("Y", "y");
AddOutput("Z", "z");
AddComment("");
}
};
class MinusGradOpDescMaker : public GradOpDescMakerBase {
public:
using GradOpDescMakerBase::GradOpDescMakerBase;
std::vector<std::unique_ptr<OpDesc>> operator()() const override {
std::vector<std::unique_ptr<OpDesc>> retv;
auto x_g = InputGrad("X");
if (!x_g.empty()) {
auto *op_desc = new OpDesc();
op_desc->SetType("scale");
op_desc->SetInput("X", OutputGrad("Out"));
op_desc->SetOutput("Out", x_g);
op_desc->SetAttr("scale", 1.0f);
retv.emplace_back(op_desc);
}
auto y_g = InputGrad("Y");
if (!y_g.empty()) {
auto *op_desc = new OpDesc();
op_desc->SetType("scale");
op_desc->SetInput("X", OutputGrad("Out"));
op_desc->SetOutput("Out", y_g);
op_desc->SetAttr("scale", -1.0f);
retv.emplace_back(op_desc);
}
return retv;
}
};
class MinusOpMaker : public OpProtoAndCheckerMaker {
public:
MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "");
AddInput("Y", "");
AddOutput("Out", "");
AddComment("minus for unittest");
}
};
} // namespace framework
} // namespace paddle
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
// rowwise_add
REGISTER_OPERATOR(rowwise_add, f::NoneOp, f::RowWiseAddOpMaker,
f::RowWiseAddGradMaker);
REGISTER_OP_CPU_KERNEL(rowwise_add,
f::NoneKernel<paddle::platform::CPUPlace, float>);
REGISTER_OPERATOR(rowwise_add_grad, f::NoneOp);
REGISTER_OP_CPU_KERNEL(rowwise_add_grad,
f::NoneKernel<paddle::platform::CPUPlace, float>);
// mul
REGISTER_OP(mul, f::NoneOp, f::MulOpMaker, mul_grad, f::NoneOp);
REGISTER_OP_CPU_KERNEL(mul, f::NoneKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul_grad,
f::NoneKernel<paddle::platform::CPUPlace, float>);
// sigmoid
REGISTER_OP(sigmoid, f::NoneOp, f::SigmoidOpMaker, sigmoid_grad, f::NoneOp);
REGISTER_OP_CPU_KERNEL(sigmoid,
f::NoneKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NoneOp, f::NoGradOpMaker);
// fill_zeros_like
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NoneOp, f::FillZeroOpMaker);
REGISTER_OP_CPU_KERNEL(fill_zeros_like,
f::NoneKernel<paddle::platform::CPUPlace, float>);
// sum
REGISTER_OP(sum, f::NoneOp, f::SumOpMaker, sum_grad, f::NoneOp);
REGISTER_OP_CPU_KERNEL(sum, f::NoneKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(sum_grad,
f::NoneKernel<paddle::platform::CPUPlace, float>);
// fc
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
// many_output_op
REGISTER_OP(many_output_op, f::NoneOp, f::ManyOutputOpMaker,
many_output_op_grad, f::NoneOp);
// mult_in_out
REGISTER_OP(mult_in_out, f::NoneOp, f::MultInOutOpMaker, mult_in_out_grad,
f::NoneOp);
REGISTER_OP_CPU_KERNEL(mult_in_out,
f::NoneKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mult_in_out_grad,
f::NoneKernel<paddle::platform::CPUPlace, float>);
// minus
REGISTER_OPERATOR(minus, f::NoneOp, f::MinusOpMaker, f::MinusGradOpDescMaker);
REGISTER_OP_CPU_KERNEL(minus, f::NoneKernel<paddle::platform::CPUPlace, float>);
// scale
REGISTER_OPERATOR(scale, f::NoneOp);
REGISTER_OP_CPU_KERNEL(scale, f::NoneKernel<paddle::platform::CPUPlace, float>);
TEST(Backward, simple_op_not_need_grad) {
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
auto gop = f::Backward(*fwd, {"x"});
ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName);
auto no_input_gop = f::Backward(*fwd, {"x", "b"});
ASSERT_NE(no_input_gop, nullptr);
ASSERT_TRUE(no_input_gop->IsNetOp());
ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
}
TEST(Backward, net_fc_backward_normal) {
std::shared_ptr<f::OperatorBase> fwd =
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
{{"mul_result", {"mul_res"}},
{"add_result", {"add_re"}},
{"Out", {"out"}}},
f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop =
f::Backward(*fwd, std::unordered_set<std::string>{});
ASSERT_TRUE(gop->IsNetOp());
auto net = static_cast<ops::NetOp *>(gop.get());
ASSERT_NO_THROW(net->DebugString());
ASSERT_EQ(3UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_add = *net->ops_[1];
ASSERT_EQ("rowwise_add_grad", d_add.Type());
f::OperatorBase &d_mul = *net->ops_[2];
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_fc_backward_not_have_b) {
std::shared_ptr<f::OperatorBase> fwd =
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}},
{{"mul_result", {"mul_res"}},
{"add_result", {"add_res"}},
{"Out", {"tmp"}}},
f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop =
f::Backward(*fwd, std::unordered_set<std::string>{});
ASSERT_TRUE(gop->IsNetOp());
auto net = static_cast<ops::NetOp *>(gop.get());
ASSERT_NO_THROW(net->DebugString());
ASSERT_EQ(2UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_mul = *net->ops_[1];
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_input_of_network_not_need_grad) {
ops::NetOp net;
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
{{"mul_result", {"mul_tmp_0"}},
{"add_result", {"add_tmp_0"}},
{"Out", {"hidden0"}}},
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_tmp_1"}},
{"add_result", {"add_tmp_1"}},
{"Out", {"hidden1"}}},
f::AttributeMap{}));
net.CompleteAddOp();
auto bwd = Backward(net, {"x"}); // x@GRAD is not need.
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
auto output_vars = bwd_net->OutputVars(true);
std::unordered_set<std::string> all_outputs =
std::unordered_set<std::string>(output_vars.begin(), output_vars.end());
all_outputs.erase(f::kEmptyVarName);
for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end());
}
// Not Generated X
ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end());
ASSERT_EQ(2UL, bwd_net->ops_.size());
ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
ASSERT_EQ(3UL, first_fc_grad->ops_.size());
ASSERT_EQ(f::kEmptyVarName,
first_fc_grad->ops_[2]->Output(f::GradVarName("X")));
}
TEST(Backward, net_shared_weight) {
ops::NetOp net;
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
{{"Out", {"out"}}}, f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
{{"Out", {"FinalOut"}}},
f::AttributeMap{}));
net.CompleteAddOp();
auto bwd = f::Backward(net, std::unordered_set<std::string>{});
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
ASSERT_EQ("sum", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_all_input_are_not_need) {
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"x", "b"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_TRUE(net->ops_.empty());
}
TEST(Backward, op_all_output_are_not_need) {
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"out"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_TRUE(net->ops_.empty());
}
TEST(Backward, op_part_of_output_are_not_need) {
auto fwd =
f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
{{"y", {"Y"}}, {"z", {"Z"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"Z"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_EQ(net->ops_.size(), 2UL);
auto &fill_zero = *net->ops_[0];
ASSERT_EQ("fill_zeros_like", fill_zero.Type());
ASSERT_EQ(1UL, fill_zero.Inputs("X").size());
ASSERT_EQ("Z", fill_zero.Input("X"));
ASSERT_EQ(1UL, fill_zero.Outputs("Out").size());
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Out"));
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.Type());
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size()); // I/O/OG
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
d_many_out.Input(f::GradVarName("z")));
ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
}
TEST(Backward, op_part_of_input_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"a"});
auto &grad_mul = *backward;
ASSERT_EQ(grad_mul.Type(), "mul_grad");
ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
ASSERT_EQ(grad_mul.Input("X"), "a");
ASSERT_EQ(grad_mul.Input("Y"), "b");
ASSERT_EQ(grad_mul.Input("Out"), "out");
}
TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
ops::NetOp net;
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"mul_result", {"mul_out1"}},
{"add_result", {"add_out1"}},
{"Out", {"out1"}}},
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_out2"}},
{"add_result", {"tmp_out2"}},
{"Out", {"out2"}}},
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
{{"mul_result", {"mul_out3"}},
{"add_result", {"tmp_out3"}},
{"Out", {"out3"}}},
f::AttributeMap{}));
net.CompleteAddOp();
auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
ASSERT_TRUE(backward->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(backward.get());
ASSERT_EQ(bwd_net->ops_.size(), 3UL);
auto &grad_fc = *bwd_net->ops_[0];
const char *all = paddle::operators::NetOp::kAll;
EXPECT_EQ(grad_fc.Inputs(all).size(),
2UL /* external input number */
+ 1UL /* external output number*/
+ 1UL /* number of gradient of external output*/
+ 2UL /* internal variable number*/
);
EXPECT_EQ(grad_fc.Outputs(all).size(),
2UL /* input number of mul*/
+ 2UL /* input number of rowwise_add*/
+ 1UL /* input number of sigmod */
- 1UL /* out2 is not needed*/);
EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
}
TEST(Backward, simple_single_op) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op = block->AppendOp();
op->SetType("rowwise_add");
op->SetInput("X", {"x"});
op->SetInput("b", {"b"});
op->SetOutput("Out", {"out"});
auto target = f::VarDesc("out");
target.SetShape({1});
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 3UL);
f::OpDesc *fill_op = block->AllOps()[1];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op = block->AllOps()[2];
EXPECT_EQ(grad_op->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op->InputNames().size(), 1UL);
ASSERT_EQ(grad_op->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b")}));
EXPECT_EQ(var_to_grad.size(), 3UL);
EXPECT_EQ(var_to_grad.at("b"), f::GradVarInfo(f::GradVarName("b"), 0, 2));
EXPECT_EQ(var_to_grad.at("x"), f::GradVarInfo(f::GradVarName("x"), 0, 2));
EXPECT_TRUE(block->HasVar(f::GradVarName("b")));
EXPECT_TRUE(block->HasVar(f::GradVarName("x")));
}
TEST(Backward, default_attribute) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op = block->AppendOp();
op->SetType("mul");
op->SetInput("X", {"x"});
op->SetInput("Y", {"y"});
op->SetOutput("Out", {"out"});
op->CheckAttrs();
auto target = f::VarDesc("out");
target.SetShape({1});
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 3UL);
EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
EXPECT_EQ(boost::get<int>(op->GetAttr("y_num_col_dims")), 1);
f::OpDesc *fill_op = block->AllOps()[1];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op = block->AllOps()[2];
ASSERT_EQ(grad_op->Type(), "mul_grad");
EXPECT_EQ(boost::get<int>(grad_op->GetAttr("x_num_col_dims")), 1);
EXPECT_EQ(boost::get<int>(grad_op->GetAttr("y_num_col_dims")), 1);
}
TEST(Backward, simple_mult_op) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDesc *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"out1"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDesc *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out2"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
auto target = f::VarDesc("out3");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 6UL + 1);
f::OpDesc *fill_op = block->AllOps()[forward_len];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op1 = block->AllOps()[6];
EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
f::OpDesc *grad_op2 = block->AllOps()[5];
EXPECT_EQ(grad_op2->Type(), "mul_grad");
ASSERT_EQ(grad_op2->InputNames().size(), 4UL);
ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op2->Input("Out"), std::vector<std::string>({"out2"}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
f::OpDesc *grad_op3 = block->AllOps()[4];
EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out3")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b3")}));
EXPECT_EQ(var_to_grad.size(), 7UL);
EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 6));
EXPECT_EQ(var_to_grad.at("b1"), f::GradVarInfo(f::GradVarName("b1"), 0, 6));
EXPECT_EQ(var_to_grad.at("out1"),
f::GradVarInfo(f::GradVarName("out1"), 0, 5));
EXPECT_EQ(var_to_grad.at("y2"), f::GradVarInfo(f::GradVarName("y2"), 0, 5));
EXPECT_EQ(var_to_grad.at("out2"),
f::GradVarInfo(f::GradVarName("out2"), 0, 4));
EXPECT_EQ(var_to_grad.at("b3"), f::GradVarInfo(f::GradVarName("b3"), 0, 4));
EXPECT_TRUE(block->HasVar(f::GradVarName("x1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("b1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("out1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("y2")));
EXPECT_TRUE(block->HasVar(f::GradVarName("out2")));
EXPECT_TRUE(block->HasVar(f::GradVarName("b3")));
}
TEST(Backward, intermedia_var_no_grad) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDesc *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"x2"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDesc *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out2"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
f::OpDesc *op4 = block->AppendOp();
op4->SetType("mul");
op4->SetInput("X", {"out1"});
op4->SetInput("Y", {"out3"});
op4->SetOutput("Out", {"out4"});
auto target = f::VarDesc("out4");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad = AppendBackward(program, target, {"out3"});
ASSERT_EQ(block->AllOps().size(), 7UL);
f::OpDesc *fill_op = block->AllOps()[forward_len];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op1 = block->AllOps()[6];
EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
f::OpDesc *grad_op4 = block->AllOps()[5];
EXPECT_EQ(grad_op4->Type(), "mul_grad");
ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"out3"}));
EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out4"}));
EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out4")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), std::vector<std::string>());
EXPECT_EQ(var_to_grad.size(), 4UL);
EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 6));
EXPECT_EQ(var_to_grad.at("b1"), f::GradVarInfo(f::GradVarName("b1"), 0, 6));
EXPECT_EQ(var_to_grad.at("out1"),
f::GradVarInfo(f::GradVarName("out1"), 0, 5));
EXPECT_TRUE(block->HasVar(f::GradVarName("x1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("b1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("out1")));
}
TEST(Backward, var_no_grad) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op1 = block->AppendOp();
op1->SetType("mult_in_out");
op1->SetInput("X", {"x1"});
op1->SetInput("H", {"h1"});
op1->SetOutput("Y", {"y1"});
op1->SetOutput("Z", {"z1"});
f::OpDesc *op2 = block->AppendOp();
op2->SetType("mult_in_out");
op2->SetInput("X", {"y1"});
op2->SetInput("H", {"z1"});
op2->SetOutput("Y", {"y2"});
op2->SetOutput("Z", {"z2"});
auto target = f::VarDesc("z2");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad = AppendBackward(program, target, {"z1"});
ASSERT_EQ(block->AllOps().size(), 6UL);
f::OpDesc *fill_op = block->AllOps()[forward_len];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op2 = block->AllOps()[3];
ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad");
ASSERT_EQ(grad_op2->InputNames().size(), 6UL);
ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"y1"}));
EXPECT_EQ(grad_op2->Input("H"), std::vector<std::string>({"z1"}));
EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op2->Input("Z"), std::vector<std::string>({"z2"}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Z")),
std::vector<std::string>({f::GradVarName("z2")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("y1")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), std::vector<std::string>());
f::OpDesc *fill_zero_op = block->AllOps()[4];
ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like");
ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL);
ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL);
EXPECT_EQ(fill_zero_op->Input("X"), std::vector<std::string>({"z1"}));
EXPECT_EQ(fill_zero_op->Output("Out"),
std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
f::OpDesc *grad_op1 = block->AllOps()[5];
ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 6UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input("X"), std::vector<std::string>({"x1"}));
EXPECT_EQ(grad_op1->Input("H"), std::vector<std::string>({"h1"}));
EXPECT_EQ(grad_op1->Input("Y"), std::vector<std::string>({"y1"}));
EXPECT_EQ(grad_op1->Input("Z"), std::vector<std::string>({"z1"}));
EXPECT_EQ(grad_op1->Input(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y1")}));
EXPECT_EQ(grad_op1->Input(f::GradVarName("Z")),
std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("H")),
std::vector<std::string>({f::GradVarName("h1")}));
EXPECT_EQ(var_to_grad.size(), 4UL);
EXPECT_EQ(var_to_grad.at("y1"), f::GradVarInfo(f::GradVarName("y1"), 0, 3));
EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 5));
EXPECT_EQ(var_to_grad.at("h1"), f::GradVarInfo(f::GradVarName("h1"), 0, 5));
EXPECT_TRUE(block->HasVar(f::GradVarName("y1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("x1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("h1")));
}
TEST(Backward, shared_var) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
f::OpDesc *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDesc *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"out1"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDesc *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out1"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
auto target = f::VarDesc("out3");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 8UL);
f::OpDesc *fill_op = block->AllOps()[forward_len];
EXPECT_EQ(fill_op->Type(), "fill_constant");
f::OpDesc *grad_op3 = block->AllOps()[4];
ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out3")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0"}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b3")}));
f::OpDesc *grad_op4 = block->AllOps()[5];
ASSERT_EQ(grad_op4->Type(), "mul_grad");
ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out2"}));
EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@1"}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
f::OpDesc *sum_op = block->AllOps()[6];
ASSERT_EQ(sum_op->Type(), "sum");
ASSERT_EQ(sum_op->InputNames().size(), 1UL);
ASSERT_EQ(sum_op->OutputNames().size(), 1UL);
EXPECT_EQ(sum_op->Input("X"),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0",
f::GradVarName("out1") + "@RENAME@1"}));
EXPECT_EQ(sum_op->Output("Out"),
std::vector<std::string>({f::GradVarName("out1")}));
f::OpDesc *grad_op1 = block->AllOps()[7];
ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
EXPECT_EQ(var_to_grad.size(), 6UL);
EXPECT_EQ(var_to_grad.at("b3"), f::GradVarInfo(f::GradVarName("b3"), 0, 4));
EXPECT_EQ(var_to_grad.at("y2"), f::GradVarInfo(f::GradVarName("y2"), 0, 5));
EXPECT_EQ(var_to_grad.at("out1"),
f::GradVarInfo(f::GradVarName("out1"), 0, 6));
EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 7));
EXPECT_EQ(var_to_grad.at("b1"), f::GradVarInfo(f::GradVarName("b1"), 0, 7));
EXPECT_TRUE(block->HasVar(f::GradVarName("b3")));
EXPECT_TRUE(block->HasVar(f::GradVarName("y2")));
EXPECT_TRUE(block->HasVar(f::GradVarName("out1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("x1")));
EXPECT_TRUE(block->HasVar(f::GradVarName("b1")));
}
TEST(Backward, half_backward) {
f::ProgramDesc program;
f::BlockDesc *block = program.MutableBlock(0);
auto *op1 = block->AppendOp();
op1->SetType("minus");
op1->SetInput("X", {"a"});
op1->SetInput("Y", {"b"});
op1->SetOutput("Out", {"out"});
auto target = f::VarDesc("out");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad = AppendBackward(program, target, {"b"});
f::OpDesc *fill_op = block->AllOps()[forward_len];
EXPECT_EQ(fill_op->Type(), "fill_constant");
auto ops = block->AllOps();
ASSERT_EQ(3UL, ops.size());
EXPECT_EQ(var_to_grad.size(), 2UL);
EXPECT_EQ(var_to_grad.at("a"),
f::GradVarInfo(f::GradVarName("a"), 0, forward_len + 1));
}
paddle/fluid/framework/prune_test.cc
浏览文件 @ b26f5050
......@@ -14,18 +14,17 @@ limitations under the License. */
#include "paddle/fluid/framework/prune.h"
#include <gtest/gtest.h>
#include <string>
#include "paddle/fluid/framework/attribute.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/net_op.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/program_desc.h"
#include <gtest/gtest.h>
namespace f = paddle::framework;
namespace ops = paddle::operators;
void AddOp(const std::string &type, const f::VariableNameMap &inputs,
const f::VariableNameMap &outputs, f::AttributeMap attrs,
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ b26f5050
......@@ -100,7 +100,7 @@ function(op_library TARGET)
endif()
# Define operators that don't need pybind here.
foreach(manual_pybind_op "net_op" "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()
......@@ -199,7 +199,6 @@ else()
set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op)
endif()
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)
......@@ -259,7 +258,6 @@ 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(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op)
......
paddle/fluid/operators/cond_op.cc 已删除 100644 → 0
浏览文件 @ 8d3ce01f
/* 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/cond_op.h"
#include "paddle/fluid/operators/gather.h"
#include "paddle/fluid/operators/scatter.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace operators {
using Scope = framework::Scope;
using Variable = framework::Variable;
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
using DDim = framework::DDim;
framework::Scope& CondOp::AddSubScope(const Scope& scope) const {
auto sub_scopes_var = scope.FindVar("SubScopes");
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
"Output(SubScopes) of CondOp should not be null.");
auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
auto& sub_scope = scope.NewScope();
sub_scopes->push_back(&sub_scope);
return sub_scope;
}
std::vector<framework::Scope*>& CondOp::GetSubScopes(
const framework::Scope& scope) const {
auto sub_scopes_var = scope.FindVar("SubScopes");
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
"Output(SubScopes) of CondOp should not be null.");
return *sub_scopes_var->GetMutable<std::vector<framework::Scope*>>();
}
LoDTensor& CondOp::AddIndexTensor(const Scope& scope) const {
auto index_tensors_var = scope.FindVar("IndexTensors");
PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
"Output(IndexTensors) of CondOp should not be null.");
auto& index_tensors =
*index_tensors_var->GetMutable<std::vector<LoDTensor>>();
index_tensors.push_back(LoDTensor());
return index_tensors.back();
}
std::vector<framework::LoDTensor>& CondOp::GetIndexTensors(
const framework::Scope& scope) const {
auto* index_tensors_var = scope.FindVar("IndexTensors");
PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
"Output(IndexTensors) of CondOp should not be null.");
return *index_tensors_var->GetMutable<std::vector<framework::LoDTensor>>();
}
void CondOp::PrepareDataForSubnet(
const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const {
PADDLE_ENFORCE(!Inputs("Xs").empty(), "Inputs(Xs) of CondOp can't be empty.");
for (int i = 0; i < BRANCH_NUM; ++i) {
// Create two sub scopes for true and false branches
// sub_scopes[0] for the true branch
// sub_scopes[1] for the false branch
AddSubScope(scope);
// Create two tensors for true and false indices:
// index_tensors[0] for the true branch
// index_tensors[1] for the false branch
AddIndexTensor(scope);
}
Variable* cond_var = scope.FindVar(Input("Cond"));
PADDLE_ENFORCE_NOT_NULL(cond_var,
"Input(Cond) of CondOp should not be null.");
const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
// get the true/false index at runtime according to cond tensor
// index_vectors[0]: vector<int>, contains all index for cond[i] == true
// index_vectors[1]: vector<int>, contains all index for cond[i] == false
std::vector<std::vector<int>> index_vectors;
index_vectors.resize(BRANCH_NUM);
const int* cond_data = cond->data<int>();
for (int i = 0; i < cond->dims()[0]; ++i) {
if (cond_data[i])
index_vectors[TRUE_BRANCH].push_back(i);
else
index_vectors[FALSE_BRANCH].push_back(i);
}
// put index_vectors[0] and index_vectors[1] into two tensors:
// index_tensors[0] and index_tensors[1]
std::vector<framework::LoDTensor>& index_tensors = GetIndexTensors(scope);
std::vector<framework::Scope*>& sub_scopes = GetSubScopes(scope);
for (int i = 0; i < BRANCH_NUM; ++i) {
DDim dim = {static_cast<int64_t>(index_vectors[i].size())};
int* index_tensor_data_ptr =
index_tensors[i].mutable_data<int>(dim, platform::CPUPlace());
memcpy(index_tensor_data_ptr, index_vectors[i].data(),
dim[0] * sizeof(int));
}
// create input in subscopes according to index_vectors
for (auto& input : Inputs("Xs")) {
Variable* var_parent = scope.FindVar(input);
PADDLE_ENFORCE_NOT_NULL(var_parent);
const auto* tensor_parent = &var_parent->Get<LoDTensor>();
for (int i = 0; i < BRANCH_NUM; ++i) {
Variable* var_child = sub_scopes[i]->FindVar(input);
PADDLE_ENFORCE_NOT_NULL(var_child);
auto* tensor_child = var_child->GetMutable<LoDTensor>();
// Resize child
DDim dim = tensor_parent->dims();
dim[0] = index_tensors[i].dims()[0];
tensor_child->mutable_data<float>(dim, platform::CPUPlace());
CPUGather<float>(dev_ctx, *tensor_parent, index_tensors[i], tensor_child);
}
}
// create output_tensors in subscope for sub_net
for (int i = 0; i < BRANCH_NUM; ++i) {
for (auto& output : (*sub_net_op_[i]).Outputs()) {
for (auto& var_name : output.second) {
sub_scopes[i]->Var(var_name);
}
}
}
}
void CondOp::MergeDataFromSubnet(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const {
std::vector<framework::Scope*>& sub_scopes = GetSubScopes(scope);
const std::vector<framework::LoDTensor>& index_tensors =
GetIndexTensors(scope);
// Infer the output dim, out_dim[0] = true_dim[0] + false_dim[0]
PADDLE_ENFORCE(!Outputs("Outs").empty(),
"Outputs(Outs) of CondOp can't be empty.");
for (auto& output : Outputs("Outs")) {
const LoDTensor* tensor_t_out =
&sub_scopes[TRUE_BRANCH]->FindVar(output)->Get<LoDTensor>();
PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL");
const LoDTensor* tensor_f_out =
&sub_scopes[FALSE_BRANCH]->FindVar(output)->Get<LoDTensor>();
PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL");
auto* var_out = scope.FindVar(output);
PADDLE_ENFORCE_NOT_NULL(var_out, "Output not found");
LoDTensor* tensor_out = var_out->GetMutable<LoDTensor>();
PADDLE_ENFORCE_NOT_NULL(tensor_t_out,
"True output tensor should not be NULL");
DDim true_dim = tensor_t_out->dims();
DDim false_dim = tensor_f_out->dims();
true_dim[0] = 0;
false_dim[0] = 0;
PADDLE_ENFORCE_EQ(true_dim, false_dim,
"Outputs not of the same shape except the first dim");
DDim out_dim = tensor_t_out->dims();
out_dim[0] = tensor_t_out->dims()[0] + tensor_f_out->dims()[0];
tensor_out->Resize(out_dim);
tensor_out->mutable_data<float>(platform::CPUPlace());
}
// merge output results:
// output_tensor = true_output_tensor + false_output_tensor
for (auto& output : Outputs("Outs")) {
Variable* var_parent = scope.FindVar(output);
PADDLE_ENFORCE_NOT_NULL(var_parent);
auto* tensor_parent = var_parent->GetMutable<LoDTensor>();
for (int i = 0; i < BRANCH_NUM; ++i) {
Variable* var_child = sub_scopes[i]->FindVar(output);
PADDLE_ENFORCE_NOT_NULL(var_child);
auto* tensor_child = &var_child->Get<LoDTensor>();
ScatterAssign<float>(dev_ctx, *tensor_child, index_tensors[i],
tensor_parent);
}
}
}
void CondOp::RunImpl(const Scope& scope, const platform::Place& place) const {
// get device context from pool
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto& dev_ctx = *pool.Get(place);
PrepareDataForSubnet(scope, dev_ctx);
std::vector<framework::Scope*>& sub_scopes = GetSubScopes(scope);
for (int i = 0; i < BRANCH_NUM; ++i) {
sub_net_op_[i]->Run(*sub_scopes[i], place);
}
MergeDataFromSubnet(scope, dev_ctx);
}
class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
public:
CondOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Cond", "The condition, which is a bool vector");
AddInput("Xs", "Inputs of Subnets").AsDuplicable();
AddOutput("Outs", "Outputs of Cond_Op after merge").AsDuplicable();
AddOutput("SubScopes", "sub scopes for true and false branches");
AddOutput("IndexTensors", "Index Tensors contains indices for true/false");
AddComment(R"DOC(
Sample Dependent Conditional Operator.
Given Cond[i] as a 1/0 vector to indicate true/false:
Out[i] = subnet_true[i], if Cond[i] == true
Out[i] = subnet_false[i], if Cond[i] == false
)DOC");
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_WITHOUT_GRADIENT(cond, paddle::operators::CondOp,
paddle::operators::CondOpProtoAndCheckerMaker);
paddle/fluid/operators/cond_op.h 已删除 100644 → 0
浏览文件 @ 8d3ce01f
/* 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 <string>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/net_op.h"
namespace paddle {
namespace operators {
/*
* @brief CondOp is a dynamic if-else Operator
*
* It has a input tensor named cond indicating which netop each instance will
* run.
*
* if cond == 1, it will run true_net, which is a NetOp.
*
* if cond == 0, it will run false_net, which is another NetOp.
*/
class CondOp : public framework::OperatorBase {
public:
CondOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {
sub_net_op_.resize(BRANCH_NUM);
}
CondOp(const CondOp& o)
: framework::OperatorBase(
static_cast<const framework::OperatorBase&>(o)) {
// TODO(yuyang18): Implement copy ctor well.
PADDLE_THROW("Not implemented");
}
framework::Scope& AddSubScope(const framework::Scope& scope) const;
std::vector<framework::Scope*>& GetSubScopes(
const framework::Scope& scope) const;
framework::LoDTensor& AddIndexTensor(const framework::Scope& scope) const;
std::vector<framework::LoDTensor>& GetIndexTensors(
const framework::Scope& scope) const;
void PrepareDataForSubnet(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const;
void MergeDataFromSubnet(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const;
/*
* Set True Block
*/
void set_truenet(std::unique_ptr<OperatorBase>&& net) {
sub_net_op_[TRUE_BRANCH] = std::move(net);
}
/*
* Set False Block
*/
void set_falsenet(std::unique_ptr<OperatorBase>&& net) {
sub_net_op_[FALSE_BRANCH] = std::move(net);
}
private:
void RunImpl(const framework::Scope& scope,
const platform::Place& place) const override;
private:
const int TRUE_BRANCH = 0;
const int FALSE_BRANCH = 1;
const int BRANCH_NUM = 2;
// sub_net_op_[0]: subnet_t
// sub_net_op_[1]: subnet_f
std::vector<std::unique_ptr<framework::OperatorBase>> sub_net_op_;
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/minus_op.cc
浏览文件 @ b26f5050
......@@ -13,9 +13,9 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/minus_op.h"
#include <string>
#include <vector>
#include "paddle/fluid/operators/net_op.h"
namespace paddle {
namespace operators {
......
paddle/fluid/operators/net_op.cc 已删除 100644 → 0
浏览文件 @ 8d3ce01f
// 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/net_op.h"
#include <set>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
const char NetOp::kAll[] = "all";
void NetOp::CompleteAddOp(bool calc) {
add_op_done_ = true;
if (!calc) return;
std::set<std::string> input_set;
std::set<std::string> output_set;
for (auto& op : ops_) {
for (auto& ipt : op->Inputs()) {
for (auto& var_name : ipt.second) {
// If input variable has been in output set, then it will be
// added into intermediate_outputs_. Otherwise, it will be
// added into input set.
if (Contains(output_set, var_name)) {
intermediate_outputs_.insert(var_name);
} else {
input_set.insert(var_name);
}
}
}
for (auto& opt : op->Outputs()) {
for (auto& var_name : opt.second) {
output_set.insert(var_name);
}
}
}
auto& inputs = inputs_[kAll];
inputs.reserve(input_set.size());
std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs));
auto& outputs = outputs_[kAll];
outputs.reserve(output_set.size());
std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs));
}
std::string NetOp::DebugStringEx(const framework::Scope* scope) const {
std::ostringstream os;
os << OperatorBase::DebugStringEx(scope) << std::endl;
for (auto& op : ops_) {
std::istringstream is(op->DebugStringEx(scope));
for (std::string line; std::getline(is, line);) {
os << " " << line << std::endl;
}
}
return os.str();
}
bool NetOp::IsNetOp() const { return true; }
std::vector<std::string> NetOp::OutputVars(bool has_intermediate) const {
std::vector<std::string> all;
for (auto& pair : this->outputs_) {
for (auto& var_name : pair.second) {
all.push_back(var_name);
}
}
if (has_intermediate) {
return all;
}
std::vector<std::string> ret_val;
for (auto& each : all) {
if (!Contains(intermediate_outputs_, each)) {
ret_val.push_back(each);
}
}
return ret_val;
}
NetOp::NetOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: framework::OperatorBase(type, inputs, outputs, attrs) {}
std::unique_ptr<framework::OperatorBase> NetOp::Clone() const {
PADDLE_ENFORCE(
add_op_done_,
"Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone");
return std::unique_ptr<OperatorBase>(new NetOp(*this));
}
} // namespace operators
} // namespace paddle
paddle/fluid/operators/net_op.h 已删除 100644 → 0
浏览文件 @ 8d3ce01f
/* 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 <set>
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
/**
* @brief Network is also a type of Operator
*
* It will manage the operators it has.
*
* Network is the container and controller of a set of operators.
* A network object knows all Operators belonging to this network. Variables,
* which are inputs and outputs of these operators, are created and managed by a
* hierarchy of Scope objects.
*
* This is the base class of network, all the networks should implement the APIs
* it defines.
*/
class NetOp : public framework::OperatorBase {
public:
static const char kAll[];
NetOp()
: framework::OperatorBase("plain_net", framework::VariableNameMap{},
framework::VariableNameMap{},
framework::AttributeMap{}) {}
NetOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs);
NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
this->ops_.reserve(o.ops_.size());
std::transform(
o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
[](const std::unique_ptr<framework::OperatorBase>& op) {
return std::unique_ptr<framework::OperatorBase>(op->Clone());
});
this->CompleteAddOp();
}
bool SupportGPU() const override {
for (auto& op : ops_) {
if (!op->SupportGPU()) {
return false;
}
}
return true;
}
void AppendOp(const framework::OperatorBase& op) { AppendOp(op.Clone()); }
/**
* @brief Add an operator by ptr
*/
void AppendOp(std::unique_ptr<framework::OperatorBase> op) {
PADDLE_ENFORCE(!add_op_done_,
"Cannot AppendOp when this network is sealed");
PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
ops_.push_back(std::move(op));
}
void InsertOp(size_t pos, std::unique_ptr<framework::OperatorBase> op) {
PADDLE_ENFORCE(!add_op_done_,
"Cannot InsertOp when this network is sealed");
PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range");
ops_.insert(ops_.begin() + pos, std::move(op));
}
void InsertOp(size_t pos, const framework::OperatorBase& op) {
InsertOp(pos, op.Clone());
}
void CompleteAddOp(bool calculate = true);
std::string DebugStringEx(
const framework::Scope* scope = nullptr) const override;
bool IsNetOp() const override;
std::vector<std::string> OutputVars(bool has_intermediate) const override;
std::unique_ptr<framework::OperatorBase> Clone() const override;
std::vector<std::unique_ptr<framework::OperatorBase>> ops_;
private:
/**
* @brief Run the network.
*
* Run all the operators with the `scope`, if no scope is provided, default
* scope will be used instead. If no OpContext is provicded, default context
* will be used.
*/
void RunImpl(const framework::Scope& scope,
const platform::Place& place) const override {
for (auto& op : ops_) {
op->Run(scope, place);
}
}
bool add_op_done_{false};
std::set<std::string> intermediate_outputs_;
template <typename T, typename KeyType>
static bool Contains(T container, KeyType key) {
return container.find(key) != container.end();
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/net_op_test.cc 已删除 100644 → 0
浏览文件 @ 8d3ce01f
// 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/net_op.h"
#include <gtest/gtest.h>
namespace paddle {
namespace operators {
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
static int run_cnt = 0;
class TestOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
DEFINE_OP_CLONE_METHOD(TestOp);
private:
void RunImpl(const Scope& scope,
const platform::Place& place) const override {
++run_cnt;
}
};
template <typename T>
void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
const std::vector<T>& actual) {
ASSERT_EQ(expected.size(), actual.size());
std::unordered_set<T> expected_set;
for (auto& tmp : expected) {
expected_set.insert(tmp);
}
for (auto& act : actual) {
ASSERT_NE(expected_set.end(), expected_set.find(act));
}
}
TEST(OpKernel, all) {
auto net = std::make_shared<NetOp>();
ASSERT_NE(net, nullptr);
net->AppendOp(std::unique_ptr<TestOp>(
new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"Out", {"y"}}}, framework::AttributeMap{})));
net->AppendOp(std::unique_ptr<TestOp>(
new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
{{"Out", {"z"}}}, framework::AttributeMap{})));
net->CompleteAddOp();
AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
net->Inputs(NetOp::kAll));
AssertSameVectorWithoutOrder({"y", "z"}, net->Outputs(NetOp::kAll));
auto final_outs = net->OutputVars(false);
ASSERT_EQ(final_outs.size(), 1UL);
ASSERT_EQ(final_outs[0], "z");
}
TEST(NetOp, insert_op) {
NetOp net;
auto op1 = std::unique_ptr<framework::NOP>(
new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"Out", {"y"}}}, framework::AttributeMap{}));
net.AppendOp(*op1);
net.InsertOp(0, *op1);
ASSERT_EQ(2UL, net.ops_.size());
net.InsertOp(2, std::move(op1));
ASSERT_EQ(3UL, net.ops_.size());
}
TEST(NetOp, Clone) {
NetOp net;
net.AppendOp(std::unique_ptr<framework::NOP>(new framework::NOP{
"empty", framework::VariableNameMap{}, framework::VariableNameMap{},
framework::AttributeMap{}}));
net.AppendOp(std::unique_ptr<framework::NOP>(new framework::NOP{
"empty2", framework::VariableNameMap{}, framework::VariableNameMap{},
framework::AttributeMap{}}));
net.CompleteAddOp(true);
auto new_net_op = net.Clone();
ASSERT_NE(new_net_op, nullptr);
ASSERT_TRUE(new_net_op->IsNetOp());
auto* new_net = static_cast<NetOp*>(new_net_op.get());
ASSERT_EQ(2UL, new_net->ops_.size());
ASSERT_EQ(new_net->ops_[0]->Type(), "empty");
ASSERT_EQ(new_net->ops_[1]->Type(), "empty2");
}
} // namespace operators
} // namespace paddle
paddle/fluid/operators/prelu_op.cc
浏览文件 @ b26f5050
......@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/prelu_op.h"
#include "paddle/fluid/operators/net_op.h"
#include <string>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/scale_op.cc
浏览文件 @ b26f5050
......@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/scale_op.h"
#include "paddle/fluid/operators/net_op.h"
#include <string>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/split_op.cc
浏览文件 @ b26f5050
......@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/split_op.h"
#include "paddle/fluid/operators/net_op.h"
namespace paddle {
namespace operators {
......
paddle/fluid/pybind/pybind.cc
浏览文件 @ b26f5050
......@@ -20,8 +20,6 @@ limitations under the License. */
#include <utility>
#include <vector>
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/framework/backward.h"
#include "paddle/fluid/framework/channel.h"
#include "paddle/fluid/framework/executor.h"
......@@ -31,18 +29,18 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/parallel_executor.h"
#include "paddle/fluid/framework/prune.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/operators/cond_op.h"
#include "paddle/fluid/operators/net_op.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/fluid/pybind/const_value.h"
#include "paddle/fluid/pybind/exception.h"
#include "paddle/fluid/pybind/pybind.h"
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h" // NOLINT
#include "paddle/fluid/pybind/recordio.h"
#include "paddle/fluid/pybind/tensor_py.h"
......@@ -239,11 +237,6 @@ All parameter, weight, gradient are variables in Paddle.
},
py::return_value_policy::reference)
#endif
.def("get_net",
[](Variable &self) -> operators::NetOp * {
return self.GetMutable<operators::NetOp>();
},
py::return_value_policy::reference)
.def("get_reader",
[](Variable &self) -> framework::ReaderHolder * {
PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
......@@ -420,42 +413,6 @@ All parameter, weight, gradient are variables in Paddle.
[](const OperatorBase &op) { return op.OutputVars(false); })
.def("support_gpu", &OperatorBase::SupportGPU);
py::class_<operators::NetOp, OperatorBase>(m, "Net")
.def_static("create",
[]() -> operators::NetOp * {
auto *retv = new operators::NetOp;
retv->SetType("plain_net");
return retv;
})
.def("append_op", [](operators::NetOp &self,
const OperatorBase &op) { self.AppendOp(op); })
.def("complete_add_op", &operators::NetOp::CompleteAddOp)
.def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
self->CompleteAddOp();
});
// cond_op
py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
.def_static("create",
[](py::bytes protobin) -> operators::CondOp * {
proto::OpDesc desc;
PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
"Cannot parse user input to OpDesc");
PADDLE_ENFORCE(desc.IsInitialized(),
"User OpDesc is not initialized, reason %s",
desc.InitializationErrorString());
auto cond_op = OpRegistry::CreateOp(desc);
return static_cast<operators::CondOp *>(cond_op.release());
})
.def("set_truenet",
[](operators::CondOp &self, const operators::NetOp &net) -> void {
self.set_truenet(net.Clone());
})
.def("set_falsenet",
[](operators::CondOp &self, const operators::NetOp &net) -> void {
self.set_falsenet(net.Clone());
});
py::class_<framework::Executor>(m, "Executor")
.def(py::init<const platform::Place &>())
.def("run",
......
paddle/fluid/recordio/chunk.cc
浏览文件 @ b26f5050
......@@ -14,13 +14,13 @@
#include "paddle/fluid/recordio/chunk.h"
#include <zlib.h>
#include <algorithm>
#include <memory>
#include <sstream>
#include "paddle/fluid/platform/enforce.h"
#include "snappy_stream/include/snappystream.hpp"
#include "zlib/include/zlib.h"
#include "snappystream.hpp"
namespace paddle {
namespace recordio {
......
paddle/fluid/recordio/header.cc
浏览文件 @ b26f5050
......@@ -13,6 +13,9 @@
// limitations under the License.
#include "paddle/fluid/recordio/header.h"
#include <string>
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
......
python/paddle/fluid/tests/unittests/test_batch_norm_op.py
浏览文件 @ b26f5050
......@@ -14,23 +14,13 @@
import unittest
import numpy as np
from op_test import OpTest
import paddle.fluid.core as core
from paddle.fluid.op import Operator
import paddle.fluid as fluid
from op_test import OpTest
from paddle.fluid.framework import grad_var_name
def get_backward_op(scope, op, no_grad_set):
backward_op = core.Operator.backward(op, no_grad_set)
for input in backward_op.input_vars():
var = scope.var(input)
var.get_tensor()
for output in backward_op.output_vars():
var = scope.var(output)
var.get_tensor()
return backward_op
def _reference_testing(x, scale, offset, mean, var, epsilon, data_format):
x_shape = x.shape
if len(x_shape) == 2:
......@@ -64,11 +54,6 @@ def _reference_testing(x, scale, offset, mean, var, epsilon, data_format):
def _reference_training(x, scale, offset, epsilon, data_format):
x_shape = x.shape
if len(x_shape) == 2:
if data_format == "NCHW":
x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1))
else:
x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1]))
if data_format == "NCHW":
n, c, h, w = x.shape
......@@ -88,8 +73,6 @@ def _reference_training(x, scale, offset, epsilon, data_format):
offset_tile = np.reshape(offset, (1, c, 1, 1))
offset_tile = np.reshape(offset_tile, (1, c, 1, 1))
y = normalized * scale_tile + offset_tile
if len(x_shape) == 2:
y = np.reshape(y, (y.shape[0], y.shape[1]))
return y, mean, var
elif data_format == "NHWC":
x_square = x * x
......@@ -100,59 +83,42 @@ def _reference_training(x, scale, offset, epsilon, data_format):
var = x_square_sum / element_count - mean * mean
normalized = (x - mean) / np.sqrt(var + epsilon)
y = normalized * scale + offset
if len(x_shape) == 2:
y = np.reshape(y, x_shape)
return y, mean, var
else:
raise ValueError("Unknown data order.")
def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format):
def _reference_grad(x, y_grad, scale, mean, var, epsilon, data_format):
# Use the following formulas to calculate gradients:
# grad_scale =
# sum(grad_y * (x - mean)) * rsqrt(var + epsilon)
#
# grad_offset = sum(output_y)
#
# grad_x =
# x_grad =
# 1/N * scale * rsqrt(var + epsilon) * (N * grad_y - sum(grad_y) -
# (x - mean) * sum(grad_y * (x - mean)) / (var + epsilon))
# transfer from (N, C, H, W) to (N, H, W, C) to simplify computation
x_shape = x.shape
if len(x_shape) == 2:
if data_format == "NCHW":
x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1))
grad_y = np.reshape(grad_y,
(grad_y.shape[0], grad_y.shape[1], 1, 1))
else:
x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1]))
grad_y = np.reshape(grad_y,
(grad_y.shape[0], 1, 1, grad_y.shape[1]))
if data_format == "NCHW":
x = np.transpose(x, (0, 2, 3, 1))
grad_y = np.transpose(grad_y, (0, 2, 3, 1))
y_grad = np.transpose(y_grad, (0, 2, 3, 1))
# raise ValueError("data_format must be NHWC, got %s." % data_format)
grad_x = scale * (grad_y - np.mean(
grad_y, axis=(0, 1, 2)) - (x - mean) * np.mean(
grad_y * (x - mean), axis=(0, 1, 2)) /
x_grad = scale * (y_grad - np.mean(
y_grad, axis=(0, 1, 2)) - (x - mean) * np.mean(
y_grad * (x - mean), axis=(0, 1, 2)) /
(var + epsilon)) / np.sqrt(var + epsilon)
grad_scale = np.sum(grad_y * (x - mean) / np.sqrt(var + epsilon),
grad_scale = np.sum(y_grad * (x - mean) / np.sqrt(var + epsilon),
axis=(0, 1, 2))
grad_offset = np.sum(grad_y, axis=(0, 1, 2))
grad_offset = np.sum(y_grad, axis=(0, 1, 2))
# transfer back to N, C, H, W
if data_format == "NCHW":
grad_x = np.transpose(grad_x, (0, 3, 1, 2))
x_grad = np.transpose(x_grad, (0, 3, 1, 2))
x = np.transpose(x, (0, 3, 1, 2))
grad_y = np.transpose(grad_y, (0, 3, 1, 2))
y_grad = np.transpose(y_grad, (0, 3, 1, 2))
if len(x_shape) == 2:
grad_x = np.reshape(grad_x, x_shape)
return grad_x, grad_scale, grad_offset
return x_grad, grad_scale, grad_offset
def create_or_get_tensor(scope, var_name, var, place):
......@@ -186,7 +152,7 @@ def set_output_grad(scope, outputs, place, feed_dict=None):
__set_tensor__(output, data)
class TestBatchNormOpInference(OpTest):
class TestBatchNormOpInference(unittest.TestCase):
def setUp(self):
self.dtype = np.float32
......@@ -304,231 +270,121 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference):
self.check_with_place(place, data_format, self.dtype, [2, 3])
class TestBatchNormOpTraining(OpTest):
class TestBatchNormOpTraining(unittest.TestCase):
def __assert_close(self, tensor, np_array, msg, atol=1e-4):
if not np.allclose(np.array(tensor), np_array, atol=atol):
import pdb
pdb.set_trace()
self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg)
def test_python_testing(self):
data_format = "NHWC"
epsilon = 0.00001
n, h, w, c = 2, 3, 4, 5
x_shape = [n, h, w, c]
scale_shape = [c]
x_val = np.random.random_sample(x_shape).astype(np.float32)
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
bias_val = np.random.random_sample(scale_shape).astype(np.float32)
mean = np.zeros(scale_shape).astype(np.float32)
variance = np.ones(scale_shape).astype(np.float32)
y_out = _reference_testing(x_val, scale_val, bias_val, mean, variance,
epsilon, "NHWC")
# running N, C, H, W case
# should produce the same results
x_shape2 = [n, c, h, w]
x_val2 = np.transpose(x_val, (0, 3, 1, 2))
y_out2 = _reference_testing(x_val2, scale_val, bias_val, mean, variance,
epsilon, "NCHW")
# transfer (N, C, H, W) back to (N, H, W, C)
y_out2_trans = np.transpose(y_out2, (0, 2, 3, 1))
self.__assert_close(y_out, y_out2_trans, "inference output")
print 'python: NHWC, NCHW, inference checking passed'
def test_python_training(self):
data_format = "NHWC"
epsilon = 0.00001
momentum = 0.9
# N, H, W, C: 2, 3, 4, 2
n, h, w, c = 2, 3, 4, 5
x_shape = [n, h, w, c]
scale_shape = [c]
x_val = np.random.random_sample(x_shape).astype(np.float32)
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
bias_val = np.random.random_sample(scale_shape).astype(np.float32)
mean = np.zeros(scale_shape).astype(np.float32)
variance = np.ones(scale_shape).astype(np.float32)
# run forward
y_out, saved_mean, var_ref = _reference_training(
x_val, scale_val, bias_val, epsilon, "NHWC")
#
mean_out = saved_mean * (1. - momentum) + momentum * mean
variance_out = var_ref * (1. - momentum) + momentum * variance
saved_variance = 1. / np.sqrt(var_ref + epsilon)
# running N, C, H, W case
# should produce the same results
x_shape2 = [n, c, h, w]
x_val2 = np.transpose(x_val, (0, 3, 1, 2))
y_out2, saved_mean2, var_ref2 = _reference_training(
x_val2, scale_val, bias_val, epsilon, "NCHW")
self.__assert_close(saved_mean, saved_mean2, "batch mean")
self.__assert_close(var_ref, var_ref2, "batch variance")
# transfer (N, C, H, W) back to (N, H, W, C)
y_out2_trans = np.transpose(y_out2, (0, 2, 3, 1))
self.__assert_close(y_out, y_out2_trans, "batch output")
print 'python: NHWC, NCHW, forward checking passed'
# test backward now
# NHWC
self.y_grad = np.random.random_sample(x_shape).astype(np.float32)
y_grad = self.y_grad
# y_grad = np.ones(x_shape).astype(np.float32)
x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_grad(
x_val, y_grad, scale_val, saved_mean, var_ref, epsilon, "NHWC")
# NCHW
y_grad2 = np.transpose(y_grad, (0, 3, 1, 2))
# y_grad2 = np.ones(x_shape2).astype(np.float32)
x_grad_ref2, scale_grad_ref2, bias_grad_ref2 = _reference_grad(
x_val2, y_grad2, scale_val, saved_mean2, var_ref2, epsilon, "NCHW")
self.__assert_close(scale_grad_ref, scale_grad_ref2, "scale gradient")
self.__assert_close(bias_grad_ref, bias_grad_ref2, "bias gradient")
x_grad_transpose = np.transpose(x_grad_ref2, (0, 2, 3, 1))
self.__assert_close(x_grad_ref, x_grad_transpose, "x gradient")
print 'python: NHWC, NCHW, backward checking passed'
def test_forward_backward(self):
def test_with_place(place, data_layout, shape):
# attr
epsilon = 0.00001
momentum = 0.9
if len(shape) == 2:
x_shape = shape
c = shape[1]
if data_layout == "NCHW":
n, c, h, w = shape[0], shape[1], shape[2], shape[3]
else:
# n, h, w, c = 2, 3, 4, 2
n, h, w, c = shape[0], shape[1], shape[2], shape[3]
if data_format == "NHWC":
x_shape = [n, h, w, c]
elif data_format == "NCHW":
x_shape = [n, c, h, w]
else:
raise ValueError("Unknown data type.")
scale_shape = [c]
x_val = np.random.random_sample(x_shape).astype(np.float32)
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
bias_val = np.random.random_sample(scale_shape).astype(np.float32)
np.random.seed(123)
x = np.random.random_sample(shape).astype(np.float32)
scale = np.random.random_sample(scale_shape).astype(np.float32)
bias = np.random.random_sample(scale_shape).astype(np.float32)
mean = np.zeros(scale_shape).astype(np.float32)
variance = np.ones(scale_shape).astype(np.float32)
# run forward
y_out, saved_mean, var_ref = _reference_training(
x_val, scale_val, bias_val, epsilon, data_format)
# update moving mean and variance
y, saved_mean, var_ref = _reference_training(x, scale, bias,
epsilon, data_layout)
mean_out = saved_mean * (1. - momentum) + momentum * mean
variance_out = var_ref * (1. - momentum) + momentum * variance
saved_variance = 1. / np.sqrt(var_ref + epsilon)
# for gradient test
# y_grad = np.ones(x_shape).astype(np.float32)
y_grad = np.zeros(x_shape).astype(np.float32)
if len(y_grad.shape) == 2:
y_grad[0, 0] = 1.
else:
y_grad[0, 0, 0, 0] = 1.
# y_grad = np.random.random_sample(x_shape).astype(np.float32)
x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_grad(
x_val, y_grad, scale_val, saved_mean, var_ref, epsilon,
data_format)
scope = core.Scope()
# create input
x_tensor = create_or_get_tensor(scope, "x_val", x_val, place)
scale_tensor = create_or_get_tensor(scope, "scale_val", scale_val,
place)
bias_tensor = create_or_get_tensor(scope, "bias_val", bias_val,
place)
mean_tensor = create_or_get_tensor(scope, "mean", mean, place)
variance_tensor = create_or_get_tensor(scope, "variance", variance,
place)
# create output
y_tensor = create_or_get_tensor(scope, "y_out", None, place)
saved_mean_tensor = create_or_get_tensor(scope, "saved_mean", None,
place)
saved_variance_tensor = create_or_get_tensor(
scope, "saved_variance", None, place)
mean_out_tensor = mean_tensor
variance_out_tensor = variance_tensor
batch_norm_op = Operator(
"batch_norm",
# inputs
X="x_val",
Scale="scale_val",
Bias="bias_val",
Mean="mean",
Variance="variance",
# outputs
Y="y_out",
MeanOut="mean",
VarianceOut="variance",
SavedMean="saved_mean",
SavedVariance="saved_variance",
# attrs
is_test=False,
data_layout=data_layout,
momentum=momentum,
epsilon=epsilon)
batch_norm_op.run(scope, place)
# check forward result
self.__assert_close(y_tensor, y_out, "y_out")
self.__assert_close(saved_mean_tensor, saved_mean, "saved_mean")
self.__assert_close(saved_variance_tensor, saved_variance,
"saved_variance")
self.__assert_close(mean_out_tensor, mean_out, "mean_out")
if isinstance(place, core.CUDAPlace):
atol = 5e-2
else:
atol = 1e-4
self.__assert_close(variance_out_tensor, variance_out,
"variance_out", atol)
print "op test forward passed: ", str(place), data_layout
# run backward
batch_norm_op_grad = get_backward_op(scope, batch_norm_op, set())
set_output_grad(
scope,
["y_out", "mean", "variance", "saved_mean", "saved_variance"],
place,
feed_dict={"y_out": y_grad})
batch_norm_op_grad.run(scope, place)
x_grad_tensor = create_or_get_tensor(scope,
grad_var_name("x_val"), None,
place)
scale_grad_tensor = create_or_get_tensor(scope,
grad_var_name("scale_val"),
None, place)
bias_grad_tensor = create_or_get_tensor(scope,
grad_var_name("bias_val"),
None, place)
y_grad = np.random.random_sample(shape).astype(np.float32)
x_grad, scale_grad, bias_grad = _reference_grad(
x, y_grad, scale, saved_mean, var_ref, epsilon, data_format)
var_dict = locals()
var_dict['y@GRAD'] = y_grad
var_names = [
'x', 'scale', 'bias', 'mean', 'variance', 'y', 'saved_mean',
'saved_variance'
]
ground_truth = {name: var_dict[name] for name in var_names}
program = fluid.Program()
with fluid.program_guard(program):
block = program.global_block()
for name in ground_truth:
block.create_var(
name=name,
dtype='float32',
shape=ground_truth[name].shape)
bn_op = block.append_op(
type="batch_norm",
inputs={
"X": block.var('x'),
"Scale": block.var('scale'),
"Bias": block.var('bias'),
"Mean": block.var('mean'),
"Variance": block.var('variance')
},
outputs={
"Y": block.var('y'),
"MeanOut": block.var('mean'), # share the same memory
"VarianceOut":
block.var('variance'), # share the same memory
"SavedMean": block.var('saved_mean'),
"SavedVariance": block.var('saved_variance')
},
attrs={
"momentum": momentum,
"epsilon": epsilon,
"is_test": False,
"data_layout": data_layout
})
block.create_var(name='y@GRAD', dtype='float32', shape=y.shape)
# generate backward op_desc
grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(
bn_op.desc, set(), [])
grad_op_desc = grad_op_desc_list[0]
new_op_desc = block.desc.append_op()
new_op_desc.copy_from(grad_op_desc)
for var_name in grad_op_desc.output_arg_names():
block.desc.var(var_name.encode("ascii"))
grad_op_desc.infer_var_type(block.desc)
grad_op_desc.infer_shape(block.desc)
for arg in grad_op_desc.output_arg_names():
grad_var = block.desc.find_var(arg.encode("ascii"))
grad_var.set_dtype(core.VarDesc.VarType.FP32)
exe = fluid.Executor(place)
out = exe.run(
program,
feed={
name: var_dict[name]
for name in
['x', 'scale', 'bias', 'mean', 'variance', 'y@GRAD']
},
fetch_list=[
'y', 'mean', 'variance', 'saved_mean', 'saved_variance',
'x@GRAD', 'scale@GRAD', 'bias@GRAD'
])
self.__assert_close(y, out[0], "y")
self.__assert_close(mean_out, out[1], "mean")
self.__assert_close(variance_out, out[2], "variance", 1e-3)
self.__assert_close(saved_mean, out[3], "saved_mean")
self.__assert_close(saved_variance, out[4], "saved_variance", 1e-3)
self.__assert_close(x_grad, out[5], "x_grad")
self.__assert_close(scale_grad, out[6], "scale_grad")
self.__assert_close(bias_grad, out[7], "bias_grad")
# check gradient output
self.__assert_close(x_grad_tensor, x_grad_ref, "x_grad")
self.__assert_close(scale_grad_tensor, scale_grad_ref, "scale_grad")
self.__assert_close(bias_grad_tensor, bias_grad_ref, "bias_grad")
print "op test backward passed: ", str(place), data_layout
print "op test forward passed: ", str(place), data_layout
places = [core.CPUPlace()]
if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"):
......@@ -537,7 +393,6 @@ class TestBatchNormOpTraining(OpTest):
for place in places:
for data_format in ["NCHW", "NHWC"]:
test_with_place(place, data_format, [2, 3, 4, 5])
test_with_place(place, data_format, [2, 3])
if __name__ == '__main__':
......
python/paddle/fluid/tests/unittests/test_cond_op.py 已删除 100644 → 0
浏览文件 @ 8d3ce01f
# 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.
import logging
import paddle.fluid.core as core
import unittest
import numpy as np
from paddle.fluid.op import Operator, CondOp
class PySimpleCond(object):
'''
A simple implementation of dynamic if-else based on numpy
'''
def __init__(self):
array = [1] * 10
for i in range(1, 10, 2):
array[i] = 0
self.cond = np.array(array)
self.x = np.ones(shape=(10, 1)).astype("float32")
def forward(self):
self.index_t = np.where(self.cond == 1)
self.index_f = np.where(self.cond == 0)
y_t = self.x[self.index_t]
y_f = self.x[self.index_f]
y_t = y_t * 2.
y_f = y_f * (-2.)
output = np.zeros(shape=(10, 1))
output[self.index_t] = y_t
output[self.index_f] = y_f
return output
class PySimpleCondTest(unittest.TestCase):
def setUp(self):
self.condnn = PySimpleCond()
def test_forward(self):
output = self.condnn.forward()
def create_tensor(scope, name, shape, np_data):
tensor = scope.var(name).get_tensor()
tensor.set_dims(shape)
tensor.set(np_data, core.CPUPlace())
return tensor
class TestCondOp(unittest.TestCase):
'''
Test CondOp
equation:
cond = [True, False, True, False, ...]
y[index_t] = x[index_t] * 2.
y[index_f] = x[index_f] * -2.
outputs:
y
'''
def setUp(self):
self.py_cond = PySimpleCond()
def forward(self):
self.scope = core.Scope()
self.create_global_variables()
self.create_cond_op()
self.create_sub_net()
self.condop.run(self.scope, core.CPUPlace())
return np.array(self.scope.find_var("Out").get_tensor())
def create_global_variables(self):
x_np_data = self.py_cond.x
create_tensor(self.scope, "X", [10, 1], x_np_data)
cond_np_data = self.py_cond.cond.astype("int32")
create_tensor(self.scope, "cond", [10, 1], cond_np_data)
self.scope.var("SubScopes")
self.scope.var("IndexTensors")
self.scope.var("Out")
def create_cond_op(self):
self.condop = CondOp(
Cond="cond",
Xs=["X"],
Outs=["Out"],
SubScopes="SubScopes",
IndexTensors="IndexTensors")
def create_sub_net(self):
truenet = core.Net.create()
scale_op_t = Operator("scale", X='X', Out='Out', scale=2.)
truenet.append_op(scale_op_t)
truenet.complete_add_op(True)
self.condop.set_truenet(truenet)
falsenet = core.Net.create()
scale_op_t = Operator("scale", X='X', Out='Out', scale=-2.)
falsenet.append_op(scale_op_t)
falsenet.complete_add_op(True)
self.condop.set_falsenet(falsenet)
def test_forward(self):
print 'test cond op forward'
pd_output = self.forward()
py_output = self.py_cond.forward()
print 'pd_output', pd_output
print
print 'py_output', py_output
self.assertEqual(pd_output.shape, py_output.shape)
print 'test passed'
return 0
if __name__ == "__main__":
unittest.main()
python/paddle/fluid/tests/unittests/test_layer_norm_op.py
浏览文件 @ b26f5050
......@@ -15,10 +15,8 @@ import unittest
import numpy as np
from operator import mul
from op_test import OpTest
import paddle.fluid.core as core
from paddle.fluid.op import Operator
from paddle.fluid.framework import grad_var_name
import paddle.fluid as fluid
np.random.random(123)
......@@ -70,161 +68,93 @@ def _reference_layer_norm_grad(x, grad_y, scale, mean, var, begin_norm_axis=1):
return grad_x, d_scale, d_bias
def get_backward_op(scope, op, no_grad_set):
backward_op = core.Operator.backward(op, no_grad_set)
for input in backward_op.input_vars():
var = scope.var(input)
var.get_tensor()
for output in backward_op.output_vars():
var = scope.var(output)
var.get_tensor()
return backward_op
def create_or_get_tensor(scope, var_name, var, place):
tensor = scope.var(var_name).get_tensor()
if var is not None:
assert isinstance(var, np.ndarray)
tensor.set_lod([[]])
tensor.set_dims(var.shape)
tensor.set(var, place)
return tensor
def set_output_grad(scope, outputs, place, feed_dict=None):
def __set_tensor__(name, data=None):
out_tensor = scope.find_var(name).get_tensor()
grad_tensor = scope.var(grad_var_name(name)).get_tensor()
out_dtype = out_tensor.dtype()
if data is None:
if out_dtype == core.VarDesc.VarType.FP64:
data = np.ones(out_tensor.shape(), dtype=np.float64)
elif out_dtype == core.VarDesc.VarType.FP32:
data = np.ones(out_tensor.shape(), dtype=np.float32)
else:
raise ValueError("Not supported data type " + str(out_dtype))
grad_tensor.set(data, place)
for output in outputs:
data = None
if output in feed_dict:
data = feed_dict[output]
__set_tensor__(output, data)
class TestLayerNormdOp(OpTest):
class TestLayerNormdOp(unittest.TestCase):
def __assert_close(self, tensor, np_array, msg, atol=1e-4):
self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg)
def __assert_grad_close(self,
tensor,
np_array,
name,
place,
max_relative_error=0.02):
a = np.array(tensor)
b = np_array
abs_a = np.abs(a)
abs_a[abs_a < 1e-5] = 1
diff_mat = np.abs(a - b) / abs_a
max_diff = np.max(diff_mat)
def err_msg():
offset = np.argmax(diff_mat > max_relative_error)
return ("%s Variable %s max gradient diff %f over limit %f, "
"the first error element is %d, %f, %f") % (
"Gradient Check On %s" % str(place), name, max_diff,
max_relative_error, offset, a.flatten()[offset],
b.flatten()[offset])
self.assertLessEqual(max_diff, max_relative_error, err_msg())
def check_forward_backward(self, shape, begin_norm_axis):
def test_with_place(place, shape, begin_norm_axis=1):
# setUp
assert begin_norm_axis > 0 and begin_norm_axis < len(
shape), 'begin_norm_axis must be between 0 and len(shape)-1.'
def test_with_place(place, shape, begin_norm_axis):
# attr
epsilon = 0.00001
x_shape = shape
D = reduce(mul, x_shape[begin_norm_axis:len(x_shape)], 1)
scale_shape = [D]
x_val = np.random.random_sample(x_shape).astype(np.float32)
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
bias_val = np.random.random_sample(scale_shape).astype(np.float32)
np.random.seed(123)
x = np.random.random_sample(x_shape).astype(np.float32)
scale = np.random.random_sample(scale_shape).astype(np.float32)
bias = np.random.random_sample(scale_shape).astype(np.float32)
y_grad = np.random.random_sample(x_shape).astype(np.float32)
# run forward
y_out, saved_mean, var_ref = _reference_layer_norm_naive(
x_val, scale_val, bias_val, epsilon, begin_norm_axis)
naive_fw = {"Y": y_out, "Mean": saved_mean, "Variance": var_ref}
# get gradient
x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_layer_norm_grad(
x_val, y_grad, scale_val, saved_mean, var_ref, begin_norm_axis)
naive_grad = {
"X": x_grad_ref,
"Scale": scale_grad_ref,
"Bias": bias_grad_ref
}
scope = core.Scope()
# create input
input_map = {"X": x_val, "Scale": scale_val, "Bias": bias_val}
for i_name in input_map:
create_or_get_tensor(scope, i_name, input_map[i_name], place)
# create output
output_map = {"Y": None, "Mean": None, "Variance": None}
output_tensor = {}
for o_name in output_map:
output_tensor[o_name] = create_or_get_tensor(
scope, o_name, output_map[o_name], place)
layer_norm_op = Operator(
"layer_norm",
# inputs
X="X",
Scale="Scale",
Bias="Bias",
# outputs
Y="Y",
Mean="Mean",
Variance="Variance",
# attrs
epsilon=epsilon,
begin_norm_axis=begin_norm_axis)
layer_norm_op.run(scope, place)
# check forward result
atol = 5e-2 if isinstance(place, core.CUDAPlace) else 1e-4
for o_tensor in output_tensor:
self.__assert_close(output_tensor[o_tensor], naive_fw[o_tensor],
o_tensor, atol)
# run backward
layer_norm_op_grad = get_backward_op(scope, layer_norm_op, set())
set_output_grad(
scope, ["Y", "Mean", "Variance"],
place,
feed_dict={"Y": y_grad})
layer_norm_op_grad.run(scope, place)
# get output
grad_tensor = {}
for o_name in naive_grad:
grad_tensor[o_name] = x_ = create_or_get_tensor(
scope, grad_var_name(o_name), None, place)
# check gradient output
for o_grad in naive_grad:
self.__assert_grad_close(grad_tensor[o_grad],
naive_grad[o_grad], o_grad + "@GRAD",
place)
# reference forward & backward
y, mean, variance = _reference_layer_norm_naive(
x, scale, bias, epsilon, begin_norm_axis)
x_grad, scale_grad, bias_grad = _reference_layer_norm_grad(
x, y_grad, scale, mean, variance, begin_norm_axis)
var_dict = locals()
var_dict['y@GRAD'] = y_grad
var_names = [
'x', 'scale', 'bias', 'mean', 'variance', 'y', 'y@GRAD'
]
ground_truth = {name: var_dict[name] for name in var_names}
program = fluid.Program()
with fluid.program_guard(program):
block = program.global_block()
for name in ground_truth:
block.create_var(
name=name,
dtype='float32',
shape=ground_truth[name].shape)
layer_norm_op = block.append_op(
type="layer_norm",
inputs={
"X": block.var('x'),
"Scale": block.var('scale'),
"Bias": block.var('bias'),
},
outputs={
"Y": block.var('y'),
"Mean": block.var('mean'), # share the same memory
"Variance":
block.var('variance'), # share the same memory
},
attrs={
"epsilon": epsilon,
"begin_norm_axis": begin_norm_axis
})
# generate backward op_desc
grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(
layer_norm_op.desc, set(), [])
grad_op_desc = grad_op_desc_list[0]
new_op_desc = block.desc.append_op()
new_op_desc.copy_from(grad_op_desc)
for var_name in grad_op_desc.output_arg_names():
block.desc.var(var_name.encode("ascii"))
grad_op_desc.infer_var_type(block.desc)
grad_op_desc.infer_shape(block.desc)
for arg in grad_op_desc.output_arg_names():
grad_var = block.desc.find_var(arg.encode("ascii"))
grad_var.set_dtype(core.VarDesc.VarType.FP32)
exe = fluid.Executor(place)
out = exe.run(program,
feed={
name: var_dict[name]
for name in ['x', 'scale', 'bias', 'y@GRAD']
},
fetch_list=[
'y', 'mean', 'variance', 'x@GRAD',
'scale@GRAD', 'bias@GRAD'
])
self.__assert_close(y, out[0], "y")
self.__assert_close(mean, out[1], "mean")
self.__assert_close(variance, out[2], "variance", 1e-3)
self.__assert_close(x_grad, out[3], "x_grad")
self.__assert_close(scale_grad, out[4], "scale_grad", 1e-3)
self.__assert_close(bias_grad, out[5], "bias_grad")
places = [core.CPUPlace()]
if core.is_compiled_with_cuda() and core.op_support_gpu("layer_norm"):
......@@ -237,15 +167,6 @@ class TestLayerNormdOp(OpTest):
self.check_forward_backward(shape=[2, 3, 4, 5], begin_norm_axis=1)
self.check_forward_backward(shape=[2, 3, 4, 5], begin_norm_axis=3)
def test_check_forward_backward_with_scale(self):
pass # TODO(zcd)
def test_check_forward_backward_with_bias(self):
pass # TODO(zcd)
def test_check_forward_backward(self):
pass # TODO(zcd)
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_net.py 已删除 100644 → 0
浏览文件 @ 8d3ce01f
# 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.
import paddle.fluid.core as core
from paddle.fluid.op import Operator
import unittest
def fc(X, W, Y):
ret_v = core.Net.create()
ret_v.append_op(Operator("mul", X="X", Y="W", Out="pre_activation"))
ret_v.append_op(Operator("sigmoid", X="pre_activation", Out=Y))
ret_v.complete_add_op(True)
return ret_v
class TestNet(unittest.TestCase):
def test_net_all(self):
net = core.Net.create()
op1 = Operator("sum", X=["X", "Y"], Out="Out")
net.append_op(op1)
net2 = core.Net.create()
net2.append_op(fc(X="X", W="w", Y="fc.out"))
net2.complete_add_op(True)
net.append_op(net2)
net.complete_add_op(True)
expected = '''
Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}.
Op(sum), inputs:{X[X, Y]}, outputs:{Out[Out]}.
Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}.
Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Out[fc.out]}.
'''
self.assertEqual(expected, "\n" + str(net))
if __name__ == "__main__":
unittest.main()
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