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
提交 07ea9ade 编写于 作者: Y Yan Chunwei 提交者： GitHub
浏览文件
操作

feature/dynamic recurrent op forward and backward (#4799)

上级 5380a547
隐藏空白更改
内联 并排
Showing with 478 addition and 283 deletion
doc/design/block.md
浏览文件 @ 07ea9ade
......@@ -189,7 +189,7 @@ OpDesc {
inputs = {0} // the index of x in vars of BlockDesc above
outputs = {5, 3} // indices of act and hidden_out in vars of BlockDesc above
attrs {
"memories" : {1} // the index of h
"states" : {1} // the index of h
"step_net" : <above step net>
}
};
......
paddle/framework/backward.cc
浏览文件 @ 07ea9ade
......@@ -21,6 +21,7 @@
#include "paddle/framework/block_desc.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/dynamic_recurrent_op.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
......@@ -220,8 +221,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// process recurrent gradient op as a special operator.
if (forwardOp.Type() == "recurrent") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
// or
// this will result in infinite loop.
// or this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::RecurrentOp*>(&forwardOp);
auto rnn_grad_op =
......@@ -231,6 +231,18 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// create stepnet's gradient op
rnn_grad_op->set_stepnet(
BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
} else if (forwardOp.Type() == "dynamic_recurrent") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
// or this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::DynamicRecurrentOp*>(&forwardOp);
auto rnn_grad_op =
static_cast<operators::DynamicRecurrentGradientOp*>(grad_op.get());
const auto& stepnet_op =
*static_cast<const OperatorBase*>(&rnnop.rnn.GetStepUnit());
// create stepnet's gradient op
rnn_grad_op->rnn.SetStepUnit(
BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
}
if (net->ops_.empty()) { // Current no aux op is added to network
......
paddle/operators/dynamic_recurrent_op.cc
浏览文件 @ 07ea9ade
......@@ -23,6 +23,7 @@ using framework::Scope;
using framework::TensorArray;
using framework::LoDTensor;
using framework::Variable;
using framework::OperatorBase;
using framework::DySeqMetaBatch;
namespace detail {
......@@ -43,10 +44,9 @@ inline void CreateVariables(Scope& scope,
* be reordered, but the RNN op should not change the `boot_state` as an input
* variable's content.
*/
template <typename T>
inline void ReorderBootState(const DySeqMetaBatch& metas,
const LoDTensor& boot_state, LoDTensor* tensor,
const platform::Place& dst_place) {
inline void ReorderInitialState(const DySeqMetaBatch& metas,
const LoDTensor& boot_state, LoDTensor* tensor,
const platform::Place& dst_place) {
for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) {
auto slice = tensor->Slice(seq_id, seq_id + 1);
auto boot_slice =
......@@ -56,58 +56,60 @@ inline void ReorderBootState(const DySeqMetaBatch& metas,
}
}
} // namespace detail
class DynamicRecurrentOpProtoAndCheckerMaker
: public framework::OpProtoAndCheckerMaker {
public:
DynamicRecurrentOpProtoAndCheckerMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
const auto& name = DynamicRecurrentOp::kArgName;
// inputs and outputs stored in proto
AddInput(name.inlinks,
"the inputs that need to be segmented for each step.")
.AsDuplicable();
AddInput(name.boot_memories, "variables to initialize memories.")
.AsDuplicable();
AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
.AsDuplicable();
AddOutput(name.step_scopes, "step scopes");
// Attributes stored in AttributeMap
AddAttr<std::vector<std::string>>(name.pre_memories,
"names of pre-memories");
AddAttr<std::vector<std::string>>(name.memories, "names of memories");
AddComment("This is a RNN operator for varience-length sequences.");
inline void RestoreInitialState(const DySeqMetaBatch& metas,
const LoDTensor& tensor, LoDTensor* boot_state,
const platform::Place& dst_place) {
for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) {
auto slice = tensor.Slice(seq_id, seq_id + 1);
auto boot_slice =
boot_state->Slice(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1);
boot_slice.CopyFrom(slice, dst_place, platform::CPUDeviceContext());
}
};
}
void DynamicRecurrentOp::Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const {
cache_.Init(kArgName, *this, scope, &arg_);
} // namespace detail
// Implementation for forward propagation.
template <>
void RNNAlgorithm::Run<RNNAlgorithm::ComputeMode::kForward>(
const framework::Scope& scope, const framework::OperatorBase& op,
const platform::DeviceContext& dev_ctx) {
SetComputeMode(ComputeMode::kForward);
cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_);
SplitInputs();
CreateScopes();
WriteStepInputs();
InitStates();
WriteStepOutputs();
RunSteps();
ConcatOutputs();
}
// call stepnet in all the time steps
for (size_t step = 0; step < cache_.num_steps; step++) {
auto& step_scope = cache_.GetScope(step);
stepnet_->Run(step_scope, dev_ctx);
// Implementation for backward propagation.
template <>
void RNNAlgorithm::Run<RNNAlgorithm::ComputeMode::kBackward>(
const framework::Scope& scope, const framework::OperatorBase& op,
const platform::DeviceContext& dev_ctx) {
SetComputeMode(ComputeMode::kBackward);
cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_);
SplitInputs();
WriteStepInputs();
InitStates();
WriteStepOutputs();
RunSteps();
// copy boot-states' gradients back.
for (const auto& state : arg_.states) {
ExportInitialStateGradient(state);
}
ConcatOutputs();
}
void DynamicRecurrentOp::SplitInputs() const {
void RNNAlgorithm::SplitInputs() {
// TODO(superjom) make level a config
// TODO(superjom) check all the inputs has the same LoD
int level = 0;
for (const auto& item : cache_.inlinks) {
for (const auto& item : cache_.inputs) {
const auto& var = item.second;
const auto& tensor = var->Get<LoDTensor>();
TensorArray& ta = step_inputs_[item.first];
......@@ -124,8 +126,8 @@ void DynamicRecurrentOp::SplitInputs() const {
}
}
void DynamicRecurrentOp::WriteStepInputs() const {
for (const auto& item : cache_.inlinks) {
void RNNAlgorithm::WriteStepInputs() {
for (const auto& item : cache_.inputs) {
auto ta_it = step_inputs_.find(item.first);
PADDLE_ENFORCE(ta_it != step_inputs_.end(),
"step_inputs_ not compatible with memory set");
......@@ -142,15 +144,15 @@ void DynamicRecurrentOp::WriteStepInputs() const {
}
}
void DynamicRecurrentOp::WriteStepOutputs() const {
void RNNAlgorithm::WriteStepOutputs() {
// initialize step outputs
for (const auto& item : cache_.outlinks) {
for (const auto& item : cache_.outputs) {
step_outputs_.emplace(item.first, TensorArray());
}
PADDLE_ENFORCE_GT(step_outputs_.size(), 0UL);
}
void DynamicRecurrentOp::CreateScopes() const {
void RNNAlgorithm::CreateScopes() {
PADDLE_ENFORCE_GT(cache_.num_steps, 0);
// resize scopes
size_t num_scopes_need_create = cache_.num_steps - cache_.scopes->size();
......@@ -159,19 +161,19 @@ void DynamicRecurrentOp::CreateScopes() const {
}
// init temporary inputs
PADDLE_ENFORCE_NOT_NULL(stepnet_, "stepnet should be set first");
std::vector<std::string> memories;
std::vector<std::string> pre_memories;
std::vector<std::string> stepnet_outputs;
std::transform(arg_.memories.begin(), arg_.memories.end(),
std::back_inserter(memories),
[](const rnn::MemoryAttr& m) { return m.var; });
std::transform(arg_.memories.begin(), arg_.memories.end(),
std::back_inserter(pre_memories),
[](const rnn::MemoryAttr& m) { return m.pre_var; });
for (const auto& item : stepnet_->Outputs()) {
PADDLE_ENFORCE_NOT_NULL(step_unit_, "stepnet should be set first");
std::vector<std::string> states;
std::vector<std::string> ex_states;
std::vector<std::string> step_unit_outputs;
std::transform(arg_.states.begin(), arg_.states.end(),
std::back_inserter(states),
[](const rnn::StateAttr& m) { return m.var; });
std::transform(arg_.states.begin(), arg_.states.end(),
std::back_inserter(ex_states),
[](const rnn::StateAttr& m) { return m.pre_var; });
for (const auto& item : step_unit_->Outputs()) {
for (const auto& var : item.second) {
stepnet_outputs.push_back(var);
step_unit_outputs.push_back(var);
}
}
......@@ -179,13 +181,13 @@ void DynamicRecurrentOp::CreateScopes() const {
auto& scope = cache_.GetScope(step);
detail::CreateVariables(scope, arg_.inlinks);
detail::CreateVariables(scope, arg_.outlinks);
detail::CreateVariables(scope, memories);
detail::CreateVariables(scope, pre_memories);
detail::CreateVariables(scope, stepnet_outputs);
detail::CreateVariables(scope, states);
detail::CreateVariables(scope, ex_states);
detail::CreateVariables(scope, step_unit_outputs);
}
}
void DynamicRecurrentOp::ConcatOutputs() const {
void RNNAlgorithm::ConcatOutputs() {
// TODO(superjom) transform this to a config
int level = 0;
for (size_t step = 0; step < cache_.num_steps; step++) {
......@@ -198,31 +200,45 @@ void DynamicRecurrentOp::ConcatOutputs() const {
item.second.WriteShared(step, *tensor);
}
}
// the inlinks' lods should be the same, so randomly get one lod.
// the inputs' lods should be the same, so randomly get one lod.
const auto& some_lod =
cache_.scope->FindVar(arg_.inlinks.front())->Get<LoDTensor>().lod();
const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
for (auto& item : step_outputs_) {
auto tensor = item.second.Pack(level, some_meta, some_lod);
auto* output = cache_.outlinks[item.first]->GetMutable<LoDTensor>();
auto* output = cache_.outputs[item.first]->GetMutable<LoDTensor>();
const_cast<LoDTensor*>(output)->ShareDataWith(tensor);
}
}
void DynamicRecurrentOp::InitStates() const {
void RNNAlgorithm::RunSteps() {
if (IsBackward()) {
// call stepnet in all the time steps reversely
for (int step = cache_.num_steps - 1; step >= 0; step--) {
auto& step_scope = cache_.GetScope(step);
step_unit_->Run(step_scope, *cache_.dev_ctx);
}
} else {
for (size_t step = 0; step < cache_.num_steps; step++) {
auto& step_scope = cache_.GetScope(step);
step_unit_->Run(step_scope, *cache_.dev_ctx);
}
}
}
void RNNAlgorithm::InitStates() {
for (size_t step = 0; step < cache_.num_steps; step++) {
for (const auto& memory : arg_.memories) {
CreateState(memory, step);
LinkState(memory, step);
for (const auto& state : arg_.states) {
CreateState(state, step);
LinkState(state, step);
}
}
}
void DynamicRecurrentOp::CreateState(const rnn::MemoryAttr& memory,
size_t step) const {
void RNNAlgorithm::CreateState(const rnn::StateAttr& state_attr, size_t step) {
auto& scope = cache_.GetScope(step);
auto& state = *cache_.GetTensor(scope, memory.var);
auto& boot_state = *cache_.GetTensor(*cache_.scope, memory.boot_var);
auto& state = *cache_.GetTensor(scope, state_attr.var);
auto& boot_state = *cache_.GetTensor(*cache_.scope, state_attr.boot_var);
size_t num_instances =
step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
......@@ -231,56 +247,79 @@ void DynamicRecurrentOp::CreateState(const rnn::MemoryAttr& memory,
state.Resize(dims);
state.mutable_data<value_type>(platform::CPUPlace());
states_[memory.var].WriteShared(step, state);
states_[state_attr.var].WriteShared(step, state);
}
void DynamicRecurrentOp::LinkState(const rnn::MemoryAttr& memory,
size_t step) const {
void RNNAlgorithm::LinkState(const rnn::StateAttr& state, size_t step) {
auto& scope = cache_.GetScope(step);
auto& state_pre = *cache_.GetTensor(scope, memory.pre_var);
auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
// process the first state's boot-state(the 0-step in forward mode or the
// last step in backward mode)
// Only forward mode need to link the boot-state to the `pre-state` in first
// time step. In backward mode, need to copy the gradient of `pre-state` in
// first time step to the gradient of `boot-state`.
if (step == 0 && IsForward()) {
LinkInitialState(state);
} else {
size_t num_instances =
step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
auto* pre_state = cache_.GetTensor(cache_.GetScope(step - 1), state.var);
// shink and share from previous state
auto shrinked_pre_state = pre_state->Slice(0, num_instances);
state_pre.ShareDataWith(shrinked_pre_state);
}
}
void RNNAlgorithm::LinkInitialState(const rnn::StateAttr& state) {
// all the step_inputs' metas should be the same, just randomly select one
// and get the dyseq meta.
const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
size_t num_instances =
step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
auto& scope = cache_.GetScope(0);
auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
auto* pre_state = cache_.GetTensor(*cache_.scope, state.boot_var);
pre_state->mutable_data<float>(platform::CPUPlace());
// allocate state
state_pre.Resize(pre_state->dims());
state_pre.mutable_data<value_type>(platform::CPUPlace());
detail::ReorderInitialState(some_meta, *pre_state, &state_pre,
pre_state->place());
}
LoDTensor* pre_state{nullptr};
if (step == 0) {
pre_state = cache_.GetTensor(*cache_.scope, memory.boot_var);
pre_state->mutable_data<float>(platform::CPUPlace());
// allocate memory
state_pre.Resize(pre_state->dims());
state_pre.mutable_data<value_type>(platform::CPUPlace());
detail::ReorderBootState<value_type>(some_meta, *pre_state, &state_pre,
pre_state->place());
} else {
pre_state = cache_.GetTensor(cache_.GetScope(step - 1), memory.var);
}
void RNNAlgorithm::ExportInitialStateGradient(const rnn::StateAttr& state) {
// all the step_inputs' metas should be the same, just randomly select one
// and get the dyseq meta.
const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
auto& scope = cache_.GetScope(0);
// shink and share from previous state
auto shrinked_pre_state = pre_state->Slice(0, num_instances);
state_pre.ShareDataWith(shrinked_pre_state);
auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
auto& pre_state = *cache_.GetTensor(*cache_.scope, state.boot_var);
pre_state.Resize(state_pre.dims());
detail::RestoreInitialState(some_meta, state_pre, &pre_state,
pre_state.place());
}
void DynamicRecurrentOp::ArgCache::Init(
const rnn::ArgumentName& name, const paddle::framework::OperatorBase& op,
const paddle::framework::Scope& scope, rnn::Argument* arg) {
void RNNAlgorithm::ArgCache::Init(const rnn::ArgumentName& name,
const paddle::framework::OperatorBase& op,
const paddle::framework::Scope& scope,
platform::DeviceContext const* dev_ctx,
rnn::Argument* arg) {
this->scope = &scope;
InitArgument(name, op, arg);
CacheScopes(scope, *arg);
CacheInlinks(scope, arg->inlinks);
CacheOutlinks(scope, arg->outlinks);
this->dev_ctx = dev_ctx;
}
void DynamicRecurrentOp::ArgCache::InitArgument(const rnn::ArgumentName& name,
const OperatorBase& op,
rnn::Argument* arg) {
void RNNAlgorithm::ArgCache::InitArgument(const rnn::ArgumentName& name,
const OperatorBase& op,
rnn::Argument* arg) {
rnn::InitArgument(name, arg, op, false /*is_grad*/);
}
void DynamicRecurrentOp::ArgCache::CacheScopes(const Scope& scope,
const rnn::Argument& arg) {
void RNNAlgorithm::ArgCache::CacheScopes(const Scope& scope,
const rnn::Argument& arg) {
auto scopes_var = scope.FindVar(arg.step_scopes);
PADDLE_ENFORCE(scopes_var != nullptr,
"the step_scopes output argument [%s] should be created first "
......@@ -289,45 +328,85 @@ void DynamicRecurrentOp::ArgCache::CacheScopes(const Scope& scope,
this->scopes = scopes_var->GetMutable<std::vector<Scope*>>();
}
void DynamicRecurrentOp::ArgCache::CacheInlinks(
void RNNAlgorithm::ArgCache::CacheInlinks(
const Scope& scope, const std::vector<std::string>& names) {
for (auto name : names) {
auto* var = GetVariable(scope, name);
inlinks[name] = var;
inputs[name] = var;
}
}
void DynamicRecurrentOp::ArgCache::CacheOutlinks(
void RNNAlgorithm::ArgCache::CacheOutlinks(
const Scope& scope, const std::vector<std::string>& names) {
for (auto name : names) {
auto* var = GetVariable(scope, name);
outlinks[name] = var;
outputs[name] = var;
}
}
Variable* DynamicRecurrentOp::ArgCache::GetVariable(const Scope& scope,
const std::string& name) {
Variable* RNNAlgorithm::ArgCache::GetVariable(const Scope& scope,
const std::string& name) {
auto* var = scope.FindVar(name);
PADDLE_ENFORCE_NOT_NULL(var, "variable [%s] not exist in scope", name);
return var;
}
LoDTensor* DynamicRecurrentOp::ArgCache::GetTensor(
const framework::Scope& scope, const std::string& name) {
LoDTensor* RNNAlgorithm::ArgCache::GetTensor(const framework::Scope& scope,
const std::string& name) {
auto* var = GetVariable(scope, name);
return var->GetMutable<LoDTensor>();
}
const rnn::ArgumentName DynamicRecurrentOp::kArgName{
"step_net", "step_scopes", "inlinks", "outlinks",
"memories", "pre_memories", "boot_memories"};
const std::array<rnn::ArgumentName, 2> RNNAlgorithm::kArgNames{
rnn::ArgumentName{"step_unit", "step_scopes", "inputs", "outputs", "states",
"ex_states", "initial_states"},
rnn::ArgumentName{"step_unit", "step_scopes@GRAD", "outputs@GRAD",
"inputs@GRAD", "states", "ex_states",
"initial_states@GRAD"}};
void DynamicRecurrentOp::Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const {
rnn.Run<RNNAlgorithm::ComputeMode::kForward>(
scope, *dynamic_cast<const OperatorBase*>(this), dev_ctx);
}
void DynamicRecurrentGradientOp::Run(
const Scope& scope, const platform::DeviceContext& dev_ctx) const {}
const Scope& scope, const platform::DeviceContext& dev_ctx) const {
rnn.Run<RNNAlgorithm::ComputeMode::kBackward>(
scope, *dynamic_cast<const OperatorBase*>(this), dev_ctx);
}
class DynamicRecurrentOpProtoAndCheckerMaker
: public framework::OpProtoAndCheckerMaker {
public:
DynamicRecurrentOpProtoAndCheckerMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
const auto& name =
RNNAlgorithm::kArgNames[RNNAlgorithm::ComputeMode::kForward];
// inputs and outputs stored in proto
AddInput(name.inlinks,
"the inputs that need to be segmented for each step.")
.AsDuplicable();
AddInput(name.initial_states, "variables to initialize states.")
.AsDuplicable();
AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
.AsDuplicable();
AddOutput(name.step_scopes, "step scopes");
// Attributes stored in AttributeMap
AddAttr<std::vector<std::string>>(name.ex_states, "names of ex_states");
AddAttr<std::vector<std::string>>(name.states, "names of states");
AddComment("This is a RNN operator for varience-length sequences.");
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_WITHOUT_GRADIENT(
dynamic_recurrent, paddle::operators::DynamicRecurrentOp,
paddle::operators::DynamicRecurrentOpProtoAndCheckerMaker);
REGISTER_OP(dynamic_recurrent, paddle::operators::DynamicRecurrentOp,
paddle::operators::DynamicRecurrentOpProtoAndCheckerMaker,
dynamic_recurrent_grad,
paddle::operators::DynamicRecurrentGradientOp);
paddle/operators/dynamic_recurrent_op.h
浏览文件 @ 07ea9ade
......@@ -27,47 +27,39 @@
namespace paddle {
namespace operators {
class DynamicRecurrentOp : public framework::OperatorBase {
class RNNAlgorithm {
public:
static const rnn::ArgumentName kArgName;
enum ComputeMode { kForward = 0, kBackward = 1 };
static const std::array<rnn::ArgumentName, 2> kArgNames;
using value_type = float;
DynamicRecurrentOp(const std::string& type,
const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
DynamicRecurrentOp(const DynamicRecurrentOp& o)
: framework::OperatorBase(
static_cast<const framework::OperatorBase&>(o)) {
// TODO(yuyang18): Implement copy ctor well.
PADDLE_THROW("Not implemented");
}
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const override;
/*
* Different `Run` method for forward and backward, `_` is just for template
* specifialization.
*/
template <ComputeMode _>
void Run(const framework::Scope& scope, const framework::OperatorBase& op,
const platform::DeviceContext& dev_ctx);
/*
* Split the inputs(LoDTensors) to segments for each time step.
*/
void SplitInputs() const;
void SplitInputs();
/*
* Create step-scopes to store temporary outputs in each time steps.
*/
void CreateScopes() const;
void CreateScopes();
/*
* Link TensorArray steps to the corresponding variables located in
* step-scopes.
*/
void WriteStepInputs() const;
void WriteStepInputs();
/*
* Write output of each step to the corresponding TensorArray.
*/
void WriteStepOutputs() const;
void WriteStepOutputs();
/*
* Initialize the states, each state will have a corresponding pre-state,
......@@ -75,54 +67,83 @@ class DynamicRecurrentOp : public framework::OperatorBase {
* pre-state in the first time step will be initialized with an zero tensor or
* a tensor in parent scope if is provided.
*/
void InitStates() const;
void InitStates();
/*
* Create state variables for each time step.
*/
void CreateState(const rnn::MemoryAttr& memory, size_t step) const;
void CreateState(const rnn::StateAttr& state, size_t step);
/*
* Link pre-state variable in current scope to the state variable in the
* previous time step (scope).
* previous time step (scope) by reference.
*/
void LinkState(const rnn::StateAttr& state, size_t step);
/*
* Link the pre-state of the first time step to the `boot-state` in parent's
* scope.
*/
void LinkInitialState(const rnn::StateAttr& state);
/*
* Copy the gradient from `pre-state` in the first step-scope to the
* `boot-state` in parent's scope.
*/
void ExportInitialStateGradient(const rnn::StateAttr& state);
/*
* Calculate time steps.
*/
void LinkState(const rnn::MemoryAttr& memory, size_t step) const;
void RunSteps();
/*
* Concatenate outputs in each time step and generate a LoDTensor.
*/
void ConcatOutputs() const;
void ConcatOutputs();
void SetComputeMode(ComputeMode mode) { mode_ = mode; }
bool IsForward() const { return mode_ == ComputeMode::kForward; }
bool IsBackward() const { return mode_ == ComputeMode::kBackward; }
/*
* set a stepnet that is created according to a RecurrentOp's stepnet.
* set a step unit that is created according to a RecurrentOp's step unit.
*/
void SetStepNet(std::unique_ptr<OperatorBase> net) {
PADDLE_ENFORCE_NOT_NULL(net);
stepnet_ = std::move(net);
void SetStepUnit(std::unique_ptr<framework::OperatorBase> step_unit) {
PADDLE_ENFORCE_NOT_NULL(step_unit);
step_unit_ = std::move(step_unit);
}
const OperatorBase& GetStepNet() const { return *stepnet_; }
const framework::OperatorBase& GetStepUnit() const { return *step_unit_; }
const framework::TensorArray& state(const std::string& name) const {
return states_[name];
auto it = states_.find(name);
PADDLE_ENFORCE(it != states_.end());
return it->second;
}
const framework::TensorArray& step_input(const std::string& name) const {
return step_inputs_[name];
auto it = step_inputs_.find(name);
PADDLE_ENFORCE(it != step_inputs_.end());
return it->second;
}
const framework::TensorArray& step_output(const std::string& name) const {
return step_outputs_[name];
auto it = step_outputs_.find(name);
PADDLE_ENFORCE(it != step_outputs_.end());
return it->second;
}
protected:
struct ArgCache {
framework::Scope const* scope;
std::vector<framework::Scope*>* scopes;
std::map<std::string, framework::Variable*> inlinks;
std::map<std::string, framework::Variable*> outlinks;
std::map<std::string, framework::Variable*> inputs;
std::map<std::string, framework::Variable*> outputs;
platform::DeviceContext const* dev_ctx;
size_t num_steps{0};
void Init(const rnn::ArgumentName& name, const OperatorBase& op,
const framework::Scope& scope, rnn::Argument* arg);
void Init(const rnn::ArgumentName& name, const framework::OperatorBase& op,
const framework::Scope& scope,
platform::DeviceContext const* dev_ctx, rnn::Argument* arg);
framework::Scope& GetScope(size_t index) {
PADDLE_ENFORCE_LT(index, num_steps);
......@@ -133,8 +154,8 @@ class DynamicRecurrentOp : public framework::OperatorBase {
const std::string& name);
private:
void InitArgument(const rnn::ArgumentName& name, const OperatorBase& op,
rnn::Argument* arg);
void InitArgument(const rnn::ArgumentName& name,
const framework::OperatorBase& op, rnn::Argument* arg);
void CacheScopes(const framework::Scope& scope, const rnn::Argument& arg);
void CacheInlinks(const framework::Scope& scope,
const std::vector<std::string>& names);
......@@ -145,27 +166,49 @@ class DynamicRecurrentOp : public framework::OperatorBase {
};
private:
std::unique_ptr<OperatorBase> stepnet_;
mutable std::map<std::string, framework::TensorArray> states_;
mutable std::map<std::string, framework::TensorArray> step_inputs_;
mutable std::map<std::string, framework::TensorArray> step_outputs_;
mutable std::map<std::string, std::vector<framework::DySeqMeta>>
dy_seq_metas_;
mutable rnn::Argument arg_;
mutable ArgCache cache_;
std::unique_ptr<framework::OperatorBase> step_unit_;
std::map<std::string, framework::TensorArray> states_;
std::map<std::string, framework::TensorArray> step_inputs_;
std::map<std::string, framework::TensorArray> step_outputs_;
std::map<std::string, std::vector<framework::DySeqMeta>> dy_seq_metas_;
rnn::Argument arg_;
ArgCache cache_;
ComputeMode mode_{ComputeMode::kForward};
#ifdef PADDLE_WITH_TESTING
friend class DynamicRecurrentOpTestHelper;
FRIEND_TEST(DynamicRecurrentOpTestHelper, SplitInputs);
FRIEND_TEST(DynamicRecurrentOpTestHelper, CreateCache);
FRIEND_TEST(DynamicRecurrentOpTestHelper, CreateScopes);
FRIEND_TEST(DynamicRecurrentOpTestHelper, WriteStepInputs);
FRIEND_TEST(DynamicRecurrentOpTestHelper, WriteStepOutputs);
FRIEND_TEST(DynamicRecurrentOpTestHelper, InitStates);
FRIEND_TEST(DynamicRecurrentOpTestHelper, ConcatOutputs);
// test forward
friend class RNNAlgorithmTestHelper;
FRIEND_TEST(RNNAlgorithmTestHelper, SplitInputs);
FRIEND_TEST(RNNAlgorithmTestHelper, CreateCache);
FRIEND_TEST(RNNAlgorithmTestHelper, CreateScopes);
FRIEND_TEST(RNNAlgorithmTestHelper, WriteStepInputs);
FRIEND_TEST(RNNAlgorithmTestHelper, WriteStepOutputs);
FRIEND_TEST(RNNAlgorithmTestHelper, InitStates);
FRIEND_TEST(RNNAlgorithmTestHelper, ConcatOutputs);
// TODO(superjom) test backward
#endif
};
class DynamicRecurrentOp : public framework::OperatorBase {
public:
DynamicRecurrentOp(const std::string& type,
const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
DynamicRecurrentOp(const DynamicRecurrentOp& o)
: framework::OperatorBase(
static_cast<const framework::OperatorBase&>(o)) {
PADDLE_THROW("Not implemented");
}
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const override;
mutable RNNAlgorithm rnn;
};
class DynamicRecurrentGradientOp : public framework::OperatorBase {
public:
DynamicRecurrentGradientOp(const std::string& type,
......@@ -174,8 +217,16 @@ class DynamicRecurrentGradientOp : public framework::OperatorBase {
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
DynamicRecurrentGradientOp(const DynamicRecurrentGradientOp& o)
: framework::OperatorBase(
static_cast<const framework::OperatorBase&>(o)) {
PADDLE_THROW("Not implemented");
}
void Run(const framework::Scope& scope,
const platform::DeviceContext& dev_ctx) const override;
mutable RNNAlgorithm rnn;
};
} // namespace operators
......
paddle/operators/dynamic_recurrent_op_test.cc
浏览文件 @ 07ea9ade
......@@ -43,16 +43,16 @@ LoDTensor* CreateVar(Scope& scope, std::string name, framework::DDim dims,
return tensor;
}
class DynamicRecurrentOpTestHelper : public ::testing::Test {
class RNNAlgorithmTestHelper : public ::testing::Test {
protected:
const rnn::ArgumentName argname = DynamicRecurrentOp::kArgName;
const rnn::ArgumentName argname = RNNAlgorithm::kArgNames[0];
virtual void SetUp() override {
CreateGlobalVariables();
auto op_desc = CreateOpDesc();
op = paddle::framework::OpRegistry::CreateOp(op_desc, nullptr);
dop = dynamic_cast<DynamicRecurrentOp*>(op.get());
dop = &(dynamic_cast<DynamicRecurrentOp*>(op.get())->rnn);
InitCacheManually();
InitStepNet();
}
......@@ -63,20 +63,20 @@ class DynamicRecurrentOpTestHelper : public ::testing::Test {
op_desc.set_type("dynamic_recurrent");
OpDescNewVar(argname.inlinks, {"in0"}, op_desc.add_inputs());
OpDescNewVar(argname.boot_memories, {"boot_mem"}, op_desc.add_inputs());
OpDescNewVar(argname.initial_states, {"boot_mem"}, op_desc.add_inputs());
OpDescNewVar(argname.step_scopes, {"step_scopes"}, op_desc.add_outputs());
OpDescNewVar(argname.outlinks, {"out0"}, op_desc.add_outputs());
// set pre-memories
// set pre-states
auto pre_memories = op_desc.mutable_attrs()->Add();
pre_memories->set_name(argname.pre_memories);
pre_memories->set_name(argname.ex_states);
pre_memories->set_type(paddle::framework::AttrType::STRINGS);
auto pre_memories_item = pre_memories->add_strings();
*pre_memories_item = "mem@pre";
// set memories
// set states
auto memories = op_desc.mutable_attrs()->Add();
memories->set_name(argname.memories);
memories->set_name(argname.states);
memories->set_type(paddle::framework::AttrType::STRINGS);
auto memories_item = memories->add_strings();
*memories_item = "mem";
......@@ -113,32 +113,33 @@ class DynamicRecurrentOpTestHelper : public ::testing::Test {
}
void InitCacheManually() {
dop->cache_.Init(DynamicRecurrentOp::kArgName, *dop, scope, &dop->arg_);
dop->cache_.Init(RNNAlgorithm::kArgNames[0], *op, scope, &device_context,
&dop->arg_);
}
void InitStepNet() {
std::unique_ptr<framework::OperatorBase> stepnet{new NetOp};
dynamic_cast<NetOp*>(stepnet.get())
->AppendOp(std::unique_ptr<TestOp>(new TestOp(
"test", {{"inlinks", {"in0"}}, {"boot_memories", {"boot_mem"}}},
{{"outlinks", {"out0"}}, {"step_scopes", {"step_scopes"}}}, {})));
dop->SetStepNet(std::move(stepnet));
"test", {{"inputs", {"in0"}}, {"initial_states", {"boot_mem"}}},
{{"outputs", {"out0"}}, {"step_scopes", {"step_scopes"}}}, {})));
dop->SetStepUnit(std::move(stepnet));
}
protected:
DynamicRecurrentOp* dop;
RNNAlgorithm* dop;
std::unique_ptr<framework::OperatorBase> op;
paddle::platform::CPUDeviceContext device_context;
paddle::framework::Scope scope;
};
TEST_F(DynamicRecurrentOpTestHelper, CreateCache) {
TEST_F(RNNAlgorithmTestHelper, CreateCache) {
const rnn::Argument& arg = dop->arg_;
ASSERT_EQ(arg.inlinks.size(), 1UL);
ASSERT_EQ(arg.outlinks.size(), 1UL);
}
TEST_F(DynamicRecurrentOpTestHelper, SplitInputs) {
TEST_F(RNNAlgorithmTestHelper, SplitInputs) {
dop->SplitInputs();
auto& in0_ta = dop->step_inputs_["in0"];
ASSERT_EQ(in0_ta.size(), 4UL);
......@@ -153,14 +154,14 @@ TEST_F(DynamicRecurrentOpTestHelper, SplitInputs) {
EXPECT_EQ(batch3.dims()[0], 1);
}
TEST_F(DynamicRecurrentOpTestHelper, CreateScopes) {
TEST_F(RNNAlgorithmTestHelper, CreateScopes) {
dop->SplitInputs();
dop->CreateScopes();
ASSERT_EQ(dop->cache_.num_steps, 4UL);
ASSERT_EQ(dop->cache_.scopes->size(), 4UL);
}
TEST_F(DynamicRecurrentOpTestHelper, WriteStepInputs) {
TEST_F(RNNAlgorithmTestHelper, WriteStepInputs) {
dop->SplitInputs();
dop->CreateScopes();
dop->WriteStepInputs();
......@@ -173,7 +174,7 @@ TEST_F(DynamicRecurrentOpTestHelper, WriteStepInputs) {
}
}
TEST_F(DynamicRecurrentOpTestHelper, WriteStepOutputs) {
TEST_F(RNNAlgorithmTestHelper, WriteStepOutputs) {
dop->SplitInputs();
dop->CreateScopes();
dop->WriteStepInputs();
......@@ -187,11 +188,12 @@ TEST_F(DynamicRecurrentOpTestHelper, WriteStepOutputs) {
}
}
TEST_F(DynamicRecurrentOpTestHelper, ConcatOutputs) {
TEST_F(RNNAlgorithmTestHelper, ConcatOutputs) {
// Let's leave this test to python unittest.
}
TEST_F(DynamicRecurrentOpTestHelper, InitStates) {
TEST_F(RNNAlgorithmTestHelper, InitStates) {
dop->SetComputeMode(RNNAlgorithm::ComputeMode::kForward);
dop->SplitInputs();
dop->CreateScopes();
dop->WriteStepInputs();
......@@ -208,12 +210,6 @@ TEST_F(DynamicRecurrentOpTestHelper, InitStates) {
auto* boot_state = scope.FindVar("boot_mem");
ASSERT_TRUE(boot_state != nullptr);
if (step == 0) {
// check pre_state is a reference of boot_state
ASSERT_EQ(boot_state->Get<LoDTensor>().data<float>(),
pre_state->Get<LoDTensor>().data<float>());
}
}
}
......
paddle/operators/recurrent_op.cc
浏览文件 @ 07ea9ade
......@@ -42,7 +42,7 @@ void RecurrentAlgorithm::Run(const Scope& scope,
for (size_t step_id = 0; step_id < seq_len; step_id++) {
if (step_id > 0) {
rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1);
rnn::LinkMemories(step_scopes, arg_->states, step_id, -1);
}
(*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
}
......@@ -59,7 +59,8 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope,
// Now all variables in scope must be created outside of op.
PADDLE_ENFORCE_NOT_NULL(stepnet_);
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs");
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(),
"step_unit_ op has no outputs");
if (seq_len > step_scopes->size()) {
for (size_t i = step_scopes->size(); i < seq_len; ++i) {
......@@ -86,7 +87,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope,
}
void RecurrentAlgorithm::InitMemories(Scope* step_scope) const {
for (auto& attr : arg_->memories) {
for (auto& attr : arg_->states) {
auto* pre_mem = step_scope->Var(attr.pre_var)->GetMutable<LoDTensor>();
PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
"memory [%s]'s boot variable [%s] not exists", attr.var,
......@@ -100,12 +101,12 @@ void RecurrentAlgorithm::InitMemories(Scope* step_scope) const {
}
const rnn::ArgumentName RecurrentOp::kArgName{
"step_net", "step_scopes", "inlinks", "outlinks",
"memories", "pre_memories", "boot_memories"};
"step_net", "step_scopes", "inputs", "outputs",
"states", "ex_states", "initial_states"};
const rnn::ArgumentName RecurrentGradientOp::kArgName{
"step_net", "step_scopes@GRAD", "outlinks@GRAD", "inlinks@GRAD",
"memories", "pre_memories", "boot_memories@GRAD"};
"step_net", "step_scopes@GRAD", "outputs@GRAD", "inputs@GRAD",
"states", "ex_states", "initial_states@GRAD"};
RecurrentOp::RecurrentOp(const std::string& type,
const framework::VariableNameMap& inputs,
......@@ -127,7 +128,7 @@ class RecurrentAlgorithmProtoAndCheckerMaker
AddInput(name.inlinks,
"the inputs that need to be segmented for each step.")
.AsDuplicable();
AddInput(name.boot_memories, "variables to initialize memories.")
AddInput(name.initial_states, "variables to initialize states.")
.AsDuplicable();
AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
......@@ -135,9 +136,8 @@ class RecurrentAlgorithmProtoAndCheckerMaker
AddOutput(name.step_scopes, "step scopes");
// Attributes stored in AttributeMap
AddAttr<std::vector<std::string>>(name.pre_memories,
"names of pre-memories");
AddAttr<std::vector<std::string>>(name.memories, "names of memories");
AddAttr<std::vector<std::string>>(name.ex_states, "names of pre-states");
AddAttr<std::vector<std::string>>(name.states, "names of states");
AddComment("This is a recurrent group operator.");
}
......@@ -152,7 +152,7 @@ void RecurrentGradientAlgorithm::Run(
rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len);
for (int step_id = seq_len - 1; step_id >= 0; --step_id) {
if (static_cast<size_t>(step_id) != seq_len - 1) {
rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
rnn::LinkMemories(step_scopes, arg_->states, step_id, 1);
}
(*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
}
......@@ -162,7 +162,7 @@ void RecurrentGradientAlgorithm::Run(
void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
Scope* step_scope) const {
for (auto& attr : arg_->memories) {
for (auto& attr : arg_->states) {
PADDLE_ENFORCE(step_scope->FindVar(attr.var) != nullptr,
"memory variable [%s] does not exists", attr.var);
PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
......
paddle/operators/rnn/recurrent_op_utils.cc
浏览文件 @ 07ea9ade
......@@ -36,7 +36,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
LoDTensor* input = input_var->GetMutable<LoDTensor>();
f::DDim dims = input->dims();
PADDLE_ENFORCE_EQ(static_cast<size_t>(dims[0]), seq_len,
"all the inlinks be the same length");
"all the inputs be the same length");
f::DDim step_dims = slice_ddim(dims, 1, dims.size());
for (size_t j = 0; j < seq_len; j++) {
Tensor* step_input =
......@@ -78,7 +78,7 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
}
void LinkMemories(const std::vector<Scope*>& scopes,
const std::vector<rnn::MemoryAttr>& memories,
const std::vector<rnn::StateAttr>& memories,
const size_t step_id, const int offset) {
PADDLE_ENFORCE_LT(step_id, scopes.size(),
"step [%d] is out of range of step scopes' size [%d]",
......@@ -106,26 +106,26 @@ void InitArgument(const ArgumentName& name, Argument* arg,
arg->inlinks = op.Inputs(name.inlinks);
arg->outlinks = op.Outputs(name.outlinks);
auto& boot_memories =
is_grad ? op.Outputs(name.boot_memories) : op.Inputs(name.boot_memories);
auto& boot_memories = is_grad ? op.Outputs(name.initial_states)
: op.Inputs(name.initial_states);
// attributes
auto& memories = op.Attr<std::vector<std::string>>(name.memories);
auto& pre_memories = op.Attr<std::vector<std::string>>(name.pre_memories);
auto& memories = op.Attr<std::vector<std::string>>(name.states);
auto& pre_memories = op.Attr<std::vector<std::string>>(name.ex_states);
PADDLE_ENFORCE(memories.size() == boot_memories.size(),
"the size of memories, boot_memories don't match:%d,%d",
"the size of states, initial_states don't match:%d,%d",
memories.size(), boot_memories.size());
PADDLE_ENFORCE(pre_memories.size() == boot_memories.size(),
"the size of pre_memories, boot_memories don't match:%d,%d",
"the size of ex_states, initial_states don't match:%d,%d",
pre_memories.size(), boot_memories.size());
PADDLE_ENFORCE(memories.size() > 0, "more than 1 memories should be set");
PADDLE_ENFORCE(memories.size() > 0, "more than 1 states should be set");
for (size_t i = 0; i < memories.size(); ++i) {
rnn::MemoryAttr mem_attr;
rnn::StateAttr mem_attr;
mem_attr.var = memories[i];
mem_attr.pre_var = pre_memories[i];
mem_attr.boot_var = boot_memories[i];
(arg->memories).push_back(mem_attr);
(arg->states).push_back(mem_attr);
}
}
......
paddle/operators/rnn/recurrent_op_utils.h
浏览文件 @ 07ea9ade
......@@ -31,7 +31,7 @@ using Scope = framework::Scope;
* boot memories in father scope. Other attributes are copied from Op's proto
* attributes.
*/
struct MemoryAttr {
struct StateAttr {
// name of current state variable
std::string var;
// name of previous step's state variable
......@@ -46,7 +46,7 @@ struct Argument {
std::string step_scopes;
std::vector<std::string> inlinks;
std::vector<std::string> outlinks;
std::vector<rnn::MemoryAttr> memories;
std::vector<rnn::StateAttr> states;
};
struct ArgumentName {
......@@ -54,9 +54,9 @@ struct ArgumentName {
std::string step_scopes;
std::string inlinks;
std::string outlinks;
std::string memories; // the memory name
std::string pre_memories; // the previous memory name
std::string boot_memories; // the boot memory name
std::string states; // the memory name
std::string ex_states; // the previous memory name
std::string initial_states; // the boot memory name
};
/**
......@@ -74,7 +74,7 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
const size_t seq_len, const platform::DeviceContext& ctx);
void LinkMemories(const std::vector<Scope*>& step_scopes,
const std::vector<MemoryAttr>& memories, const size_t step_id,
const std::vector<StateAttr>& memories, const size_t step_id,
const int offset);
void InitArgument(const ArgumentName& name, Argument* arg,
......
paddle/pybind/pybind.cc
浏览文件 @ 07ea9ade
......@@ -413,18 +413,18 @@ All parameter, weight, gradient are variables in Paddle.
return static_cast<operators::DynamicRecurrentOp *>(
rnn_op.release());
})
.def("set_stepnet",
.def("set_step_unit",
[](operators::DynamicRecurrentOp &self, const operators::NetOp &net)
-> void { self.SetStepNet(net.Clone()); })
-> void { self.rnn.SetStepUnit(net.Clone()); })
.def("get_state",
[](operators::DynamicRecurrentOp &self, const std::string &name)
-> const TensorArray & { return self.state(name); })
-> const TensorArray & { return self.rnn.state(name); })
.def("get_step_input",
[](operators::DynamicRecurrentOp &self, const std::string &name)
-> const TensorArray & { return self.step_input(name); })
-> const TensorArray & { return self.rnn.step_input(name); })
.def("get_step_output",
[](operators::DynamicRecurrentOp &self, const std::string &name)
-> const TensorArray & { return self.step_output(name); });
-> const TensorArray & { return self.rnn.step_output(name); });
// cond_op
py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
......
python/paddle/v2/framework/tests/test_dynamic_recurrent_op.py
浏览文件 @ 07ea9ade
......@@ -4,6 +4,12 @@ import unittest
from paddle.v2.framework.op import Operator, DynamicRecurrentOp
import numpy as np
# for siplicity, just one level LoD
lod_py = [[0, 4, 7, 9, 10]]
input_dim = 30
num_sents = len(lod_py[0]) - 1
weight_dim = 15
def create_tensor(scope, name, shape, np_data):
tensor = scope.var(name).get_tensor()
......@@ -12,6 +18,17 @@ def create_tensor(scope, name, shape, np_data):
return tensor
class PyRNNStep(object):
def __init__(self):
self.x = np.random.normal(size=(lod_py[0][-1],
input_dim)).astype("float32")
self.W = np.random.normal(size=(input_dim, input_dim)).astype("float32")
self.U = np.random.normal(size=(input_dim, input_dim)).astype("float32")
self.h_boot = np.random.normal(size=(num_sents,
input_dim)).astype("float32")
class DynamicRecurrentOpTest(unittest.TestCase):
'''
Test RNNOp
......@@ -23,17 +40,13 @@ class DynamicRecurrentOpTest(unittest.TestCase):
- U
vars:
- x
memories:
states:
- h
outputs:
- h
'''
# for siplicity, just one level LoD
lod_py = [[0, 4, 7, 9, 10]]
input_dim = 30
num_sents = len(lod_py[0]) - 1
weight_dim = 15
py = PyRNNStep()
def forward(self):
self.scope = core.Scope()
......@@ -42,64 +55,55 @@ class DynamicRecurrentOpTest(unittest.TestCase):
self.create_step_net()
ctx = core.DeviceContext.create(core.CPUPlace())
self.rnnop.run(self.scope, ctx)
state = self.rnnop.get_state("h@mem")
state = self.rnnop.get_state("h@state")
print 'state size: ', state.size()
step_inputs = self.rnnop.get_step_input("x")
print "x size ", step_inputs.size()
for i in range(step_inputs.size()):
print "x %d" % i, np.array(step_inputs.read(i).get_dims())
step_outputs = self.rnnop.get_step_output('h@mem')
step_outputs = self.rnnop.get_step_output('h@state')
print 'step_outputs.size ', step_outputs.size()
output = self.scope.find_var("h@mem").get_tensor()
output = self.scope.find_var("h@state").get_tensor()
print 'output', np.array(output).shape
def create_global_variables(self):
x = np.random.normal(size=(self.lod_py[0][-1],
self.input_dim)).astype("float32")
W = np.random.normal(size=(self.input_dim,
self.input_dim)).astype("float32")
U = np.random.normal(size=(self.input_dim,
self.input_dim)).astype("float32")
h_boot = np.random.normal(size=(self.num_sents,
self.input_dim)).astype("float32")
# create inlink
x_tensor = create_tensor(self.scope, "x",
[self.num_sents, self.input_dim], x)
x_tensor.set_lod(self.lod_py)
create_tensor(self.scope, "W", [self.input_dim, self.input_dim], W)
create_tensor(self.scope, "U", [self.input_dim, self.input_dim], U)
create_tensor(self.scope, "h_boot", [self.num_sents, self.input_dim],
h_boot)
x_tensor = create_tensor(self.scope, "x", [num_sents, input_dim],
self.py.x)
x_tensor.set_lod(lod_py)
create_tensor(self.scope, "W", [input_dim, input_dim], self.py.W)
create_tensor(self.scope, "U", [input_dim, input_dim], self.py.U)
create_tensor(self.scope, "h_boot", [num_sents, input_dim],
self.py.h_boot)
self.scope.var("step_scopes")
self.scope.var("h@mem")
self.scope.var("h@state")
def create_rnn_op(self):
# create RNNOp
self.rnnop = DynamicRecurrentOp(
# inputs
inlinks=["x"],
boot_memories=["h_boot"],
step_net="stepnet",
inputs=["x"],
initial_states=["h_boot"],
step_net="step_unit",
# outputs
outlinks=["h@mem"],
outputs=["h@state"],
step_scopes="step_scopes",
# attributes
pre_memories=["h@pre"],
memories=["h@mem"])
ex_states=["h@pre"],
states=["h@state"])
def create_step_net(self):
stepnet = core.Net.create()
step_unit = core.Net.create()
x_fc_op = Operator("mul", X="x", Y="W", Out="Wx")
h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
sum_op = Operator("sum", X=["Wx", "Uh"], Out="sum")
sig_op = Operator("sigmoid", X="sum", Y="h@mem")
sig_op = Operator("sigmoid", X="sum", Y="h@state")
for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
stepnet.append_op(op)
stepnet.complete_add_op(True)
self.rnnop.set_stepnet(stepnet)
step_unit.append_op(op)
step_unit.complete_add_op(True)
self.rnnop.set_step_unit(step_unit)
def test_forward(self):
print 'test recurrent op forward'
......@@ -107,5 +111,58 @@ class DynamicRecurrentOpTest(unittest.TestCase):
print 'pd_output', pd_output
class RecurrentGradientOpTest(unittest.TestCase):
py = PyRNNStep()
def create_forward_op(self):
# create RNNOp
self.forward_op = DynamicRecurrentOp(
# inputs
inputs=["x"],
initial_states=["h_boot"],
step_net="step_unit",
# outputs
outputs=["h@state"],
step_scopes="step_scopes",
# attributes
ex_states=["h@pre"],
states=["h@state"])
def create_gradient_op(self):
a = set()
backward_op = core.DynamicRecurrentOp.backward(self.forward_op, a)
def create_step_net(self):
step_unit = core.Net.create()
x_fc_op = Operator("mul", X="x", Y="W", Out="Wx")
h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
sum_op = Operator("sum", X=["Wx", "Uh"], Out="sum")
sig_op = Operator("sigmoid", X="sum", Y="h@state")
for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
step_unit.append_op(op)
step_unit.complete_add_op(True)
self.forward_op.set_step_unit(step_unit)
def create_global_variables(self):
# create inlink
x_tensor = create_tensor(self.scope, "x", [num_sents, input_dim],
self.py.x)
x_tensor.set_lod(lod_py)
create_tensor(self.scope, "W", [input_dim, input_dim], self.py.W)
create_tensor(self.scope, "U", [input_dim, input_dim], self.py.U)
create_tensor(self.scope, "h_boot", [num_sents, input_dim],
self.py.h_boot)
self.scope.var("step_scopes")
self.scope.var("h@state")
def test_grad(self):
self.scope = core.Scope()
self.create_forward_op()
self.create_global_variables()
self.create_step_net()
self.create_gradient_op()
if __name__ == '__main__':
unittest.main()
python/paddle/v2/framework/tests/test_recurrent_op.py
浏览文件 @ 07ea9ade
......@@ -132,15 +132,15 @@ class RecurrentOpTest(unittest.TestCase):
# create RNNOp
self.rnnop = RecurrentOp(
# inputs
inlinks=["x"],
boot_memories=["h_boot"],
inputs=["x"],
initial_states=["h_boot"],
step_net="stepnet",
# outputs
outlinks=["h@mem"],
outputs=["h@mem"],
step_scopes="step_scopes",
# attributes
pre_memories=["h@pre"],
memories=["h@mem"])
ex_states=["h@pre"],
states=["h@mem"])
def create_step_net(self):
stepnet = core.Net.create()
......@@ -169,15 +169,15 @@ class RecurrentGradientOpTest(unittest.TestCase):
def create_forward_op(self):
self.forward_op = RecurrentOp(
# inputs
inlinks=["x"],
boot_memories=["h_boot"],
inputs=["x"],
initial_states=["h_boot"],
step_net="stepnet",
# outputs
outlinks=["h"],
outputs=["h"],
step_scopes="step_scopes",
# attributes
pre_memories=["h@pre"],
memories=["h@alias"])
ex_states=["h@pre"],
states=["h@alias"])
# create a stepnet for RNN
stepnet = core.Net.create()
......
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