add rnn op interfaces #2775
add rnn op interfaces #2775
Conversation
|
|
||
| protected: | ||
| std::vector<std::string> inputs_; | ||
| std::vector<std::string> outputs_; |
|
TODO 最终测试时，添加简单RNN需要的几个op，debug |
| std::vector<std::string> outputs_; | ||
| } | ||
|
|
||
| class RecurrentForwardOp { |
There was a problem hiding this comment.
- RecurrentGroupOp
- RecurrentGroupGradientOp
| std::vector<StateAttr> states_; | ||
| }; | ||
|
|
||
| class RecurrentBackwardOp; |
| /* | ||
| * Prepare inputs for each stepnet. | ||
| */ | ||
| void ApplyInLinks(Scope* scope); |
There was a problem hiding this comment.
ScatterInLinks(Scope* scope)
GatherOutLinks(Scope* scope)
| /* | ||
| * Process outputs of stepnets and merge to variables. | ||
| */ | ||
| void ApplyOutLinks(Scope* scope); |
| /* | ||
| * Build a `Net` which is shared across all steps. | ||
| */ | ||
| void BuildStepNet(Scope* scope); |
| * step scopes as the father scope. The step scopes will be stored in the | ||
| * father scope as a variable. | ||
| */ | ||
| void CreateScopes(Scope* scope); |
| */ | ||
|
|
||
| // State of a RNN (same as the role of `Momory` in PaddlePaddle) | ||
| struct StateAttr { |
| * NOTE the context's scope is not given until `Run` called, so step scopes' | ||
| * father should be set/updated in this method. | ||
| */ | ||
| virtual void Run(OpRunContext* contex) const = 0; |
| public: | ||
| virtual ~OperatorBase() {} | ||
| virtual void Run(OpRunContext* context) const = 0; | ||
| virtual void InferShape(const Scope* scope) const = 0; |
There was a problem hiding this comment.
what does InferShape do?
There was a problem hiding this comment.
I think the purpose of InferShape is to inference the size of inputs/outputs from some of them that we already know the size.
There was a problem hiding this comment.
InferShape will set the output variable dim according to the input variable dim.
There was a problem hiding this comment.
RNNOp.InferShape will just call its step net's InferShape, and will
- check input variable/tensors' shape, raise an error if wrong
- update all outputs' variable/tensors' shape according to this mini-batch of input
It is offered as a public method because we want to keep checking dynamically during user adding operators.
| * NOTE the context's scope is not given until `Run` called, so step scopes' | ||
| * father should be set/updated in this method. | ||
| */ | ||
| virtual void Run(OpRunContext* contex) const { |
There was a problem hiding this comment.
Yes, will move to .cpp later.
We are working on a simple implementation to verify the whole process and will give a version soon.
| std::vector<std::string> outputs_; | ||
| } | ||
|
|
||
| class RecurrentGroupForwardOp { |
There was a problem hiding this comment.
RecurrentGroupForwardOp => RecurrentOp
There was a problem hiding this comment.
good, short enough.
the backward op's name?
RecurrentBackwardOp- or
RecurrentGradientOp?
| /* | ||
| * Prepare inputs for each stepnet. | ||
| */ | ||
| void ScatterInLinks(Scope* scope); |
There was a problem hiding this comment.
ScatterInLinks => SegmentInputs. Let us use accurate English wording.
| /* | ||
| * Process outputs of stepnets and merge to variables. | ||
| */ | ||
| void GatherOutLinks(Scope* scope); |
There was a problem hiding this comment.
GatherOutLinks => ConcatenateOutputs
| /* | ||
| * Build a `Net` which is shared across all steps. | ||
| */ | ||
| void BuildStepNet(Scope* scope); |
There was a problem hiding this comment.
BuildStepNet => CreateStepNet
|
|
||
| class OperatorBase { | ||
| public: | ||
| virtual ~OperatorBase() {} |
There was a problem hiding this comment.
I looks to me that the constructor needs a parameter paddle::framework::proto::OperatorDesc so could it possble to call InferShape, which saves sizes of inputs/outputs into the desc. Only if so, we could have all necessary information for calling OperatorBase::Run:
class OperatorBase {
public:
OperatorBase(const proto::OperatorDesc& desc) : desc_(desc) {}
virtual void Run(OpRunContext* context) const = 0;
protected:
virtual void InferShape(const Scope* scope) const = 0; // needs to read from and write to desc_
proto::OperatorDesc desc_;
};So the information in proto::OperatorDesc propagates along the path:
Operator's constructor
① ↓
OperatorBase::desc_ → Operator's Run
②↓ ↑③ ④
Operator's InferShape
@Superjom @reyoung @jacquesqiao
There was a problem hiding this comment.
In the new design of Operator, OpDesc will store in Op, and InferShape can get the information from scope, but it seems that it need not store the shape into the desc
There was a problem hiding this comment.
@jacquesqiao You are right.
The first clue about input/output sizes is in training data instances, and we get the instance when we do training, i.e,. call operator's Run.
Should we just remove InferShape and let each operator defines its own shape inference methods, i.e., one method for an output, so to shorten code in its Run method like this:
template <typename Context> class MyOperator;
template <>
class MyOperator<GPUContext> : public OperatorBase {
public:
MyOperator(const proto::OperatorDesc& desc) : OperatorBase(desc) {}
virtual void Run(OpRunContext* ctx) const {
cudnnGemm(
ctx->cudnnHandle,
Output(0, ctx)->GetMutable<Tensor>(Output0Size(ctx))->mutable_data(),
Input(0, ctx)->Get<Tensor>()->data(),
Input(1, ctx)->Get<Tensor>()->data(),
);
}
private:
DDim Output0Size(OpRunContext* ctx) const { ...}
DDim Output1Size(OpRunContext* ctx) const { ...}
};
… feature/rnn_op
| for (size_t i = step_scopes->size(); i < seq_len; ++i) { | ||
| // step_scopes->push_back(std::make_shared<Scope>( | ||
| // std::shared_ptr<Scope>(scope))); | ||
| step_scopes->push_back(new Scope(std::shared_ptr<Scope>(scope))); |
There was a problem hiding this comment.
There was a problem hiding this comment.
The shared_ptr is used for the scope constructor. I created an issue.
… feature/rnn_op
| } | ||
| } | ||
|
|
||
| // void RecurrentOp::CreateStepNet(ScopePtr scope) const { |
There was a problem hiding this comment.
delete these codes
| auto dims = Input(scope, inlinks_[0])->GetMutable<Tensor>()->dims(); | ||
| size_t seq_len = dims[0]; | ||
| Variable* scopes_var = scope->GetVariable(step_scopes_name_); | ||
| auto step_scopes = scopes_var->GetMutable<std::vector<ScopePtr>>(); |
There was a problem hiding this comment.
- 116行和117行，在SegmentInput和ConcatOutput中都重复出现，可以把step_scopes作为一个变量存在.h文件中以保持代码整洁么
- seq_len也可以作为一个变量存在.h文件，名字可以叫maxSequenceLength_。因为很多地方都会用到，就不用重复取了。
| // NetDesc desc; | ||
| // desc.name_ = "rnn_step_net"; | ||
| // var->Reset<PlainNet>(new PlainNet(desc)); | ||
| // } |
| LOG(INFO) << "create scopes"; | ||
| CreateScopes(scope); | ||
| LOG(INFO) << "segment input"; | ||
| SegmentInputs(scope); |
| LOG(INFO) << "run step " << step_id; | ||
| ScopePtr step_scope = scopes[step_id]; | ||
| // TODO replace memorys' copy with reference | ||
| LinkMemories(scope, scopes, step_id); |
There was a problem hiding this comment.
LinkMemories(scope, scopes, step_id); -> LinkMemories(scopes, step_id);
scopes - > step_scopes
|
|
||
| void RecurrentOp::LinkMemories(ScopePtr scope, | ||
| std::vector<ScopePtr>& step_scopes, | ||
| size_t step) const { |
There was a problem hiding this comment.
有一个疑问，传了scope后，为什么还要传step_scopes，后者可以从前者获得吧
| #include "paddle/framework/scope.h" | ||
| #include "paddle/framework/variable.h" | ||
|
|
||
| // Remove when including operator.h |
| // TODO: | ||
| // 1. No-padding computing for sequences with indifinite length in one batch. | ||
| // 2. Hierarchical RNN for sequence with sub-sequence. | ||
| // 3. Multi-inputs with indifinate length for RecurrentOp. |
| /* | ||
| * Create a `Net` which is shared across all steps. | ||
| */ | ||
| // void CreateStepNet(ScopePtr scope) const; |
| /* | ||
| * Create memories in each step scope. | ||
| */ | ||
| // void CreateMemories(ScopePtr scope) const; |
| * Link memory in previous step scope to current scope. | ||
| */ | ||
| void LinkMemories(ScopePtr scope, std::vector<ScopePtr>& step_scopes, | ||
| size_t step) const; |
There was a problem hiding this comment.
最后一个变量step改成step_id，是否更容易理解
| outputs_.push_back(output); | ||
| } | ||
|
|
||
| name_ = op_desc.name(); |
There was a problem hiding this comment.
This name will be removed.
| for (int j = 0; j < seq_len; j++) { | ||
| Variable* input_var = step_scopes[j]->CreateVariable(input_alias[i]); | ||
| Tensor* step_input_tensor = input_var->GetMutable<Tensor>(); | ||
| *step_input_tensor = scope_input_tensor->Slice(j, j + 1); |
There was a problem hiding this comment.
The Slice operation will not change the rank of the tensor. If the input tensor shape is [10, 20, 30], the step_input_tensor shape is [1, 20, 30]. See the code: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L82
There was a problem hiding this comment.
这里不需要改变tensor的顺序，每个step的tensor取得对应的那部分就行了。
| auto& step_scopes = *scopes_var->GetMutable<std::vector<ScopePtr>>(); | ||
| auto dims = Input(scope, inlinks_[0])->GetMutable<Tensor>()->dims(); | ||
| int seq_len = dims[0]; | ||
| int batch_size = dims[1]; |
There was a problem hiding this comment.
Need to support more than 3 dimensions.
There was a problem hiding this comment.
等tensor.h中copyFrom函数merge了，这里就进行更改。
| PADDLE_ENFORCE(boot_tensor, "boot_tensor should be retrieved before"); | ||
| // copy from boot memory | ||
| std::memcpy(memory_tensor_val, boot_tensor->data<float>(), | ||
| product(attr.dims)); |
| step_scopes[step - 1]->GetVariable(attr.var)->GetMutable<Tensor>(); | ||
|
|
||
| std::memcpy(memory_tensor_val, pre_step_memory->data<float>(), | ||
| product(attr.dims)); |
|
|
||
| /* | ||
| * RecurrentOp inputs stored in proto: | ||
| * - real inputs that need to be segmented to steps. |
|
|
||
| // this op's name, used as a unique key in father scope. | ||
| // TODO repace it with OpBase's interface if supported. | ||
| std::string name_; |
| op_desc.add_inputs("rnn/h_pre"); | ||
| op_desc.add_inputs("rnn/w"); | ||
| op_desc.add_outputs("rnn/s"); | ||
| // s = h_pre * check |
There was a problem hiding this comment.
名字改成rnn.h_pre，rnn.w这样的比较自然些吧。不然s = rnn/h_pre * rnn/w会认为是除号。
| op_desc.add_inputs("rnn/x"); | ||
| op_desc.add_inputs("rnn/s"); | ||
| op_desc.add_outputs("rnn/h"); | ||
| // h = x + s |
| std::string step_scopes_name_; | ||
| // real inputs that need to be segmented. | ||
| std::vector<int> inlinks_; | ||
| std::vector<std::string> outlinks_; |
There was a problem hiding this comment.
inlinks和outlinks是对应的，为什么一个存int，一个存string？
|
|
||
| // Remove when including operator.h | ||
| #include <glog/logging.h> | ||
| #include "paddle/framework/attr_checker.h" |
| // std::ostringstream stream; | ||
| // op_desc.SerializeToOstream(&stream); | ||
| // std::string text = stream.str(); | ||
| // LOG(INFO) << text; |
| CreateRNNOp(); | ||
| } | ||
|
|
||
| virtual void TearDown() override {} |
| // PaddlePaddle), which is store in father's scope. | ||
| std::string boot_var; | ||
| // this dim will infered from boot memories's tensor in the first step. | ||
| DDim dims; |
There was a problem hiding this comment.
Remove DDim dims after using SharedDataFrom.
paddle/framework/op_desc.proto
Outdated
| @@ -51,6 +51,9 @@ message OpDesc { | |||
| // type of this Operator, such as "add", "sub", "fc". | |||
| required string type = 3; | |||
|
|
|||
| // the name of this Operator. | |||
| required string name = 4; | |||
| LOG(INFO) << "run step " << step_id; | ||
| ScopePtr step_scope = scopes[step_id]; | ||
| // TODO replace memorys' copy with reference | ||
| LinkMemories(scope, scopes, step_id); |
There was a problem hiding this comment.
LinkMemories(scope, scopes, step_id); -> LinkMemories(scopes, step_id);
scopes - > step_scopes
| LOG(INFO) << "segment input"; | ||
| SegmentInputs(scope); | ||
|
|
||
| Variable* step_scopes = scope->GetVariable(step_scopes_name_); |
| LOG(INFO) << "create scopes"; | ||
| CreateScopes(scope); | ||
| LOG(INFO) << "segment input"; | ||
| SegmentInputs(scope); |
| LOG(INFO) << "run step " << step_id; | ||
| ScopePtr step_scope = scopes[step_id]; | ||
| // TODO replace memorys' copy with reference | ||
| LinkMemories(scope, scopes, step_id); |
| auto dims = Input(scope, inlinks_[0])->GetMutable<Tensor>()->dims(); | ||
| size_t seq_len = dims[0]; | ||
| Variable* scopes_var = scope->GetVariable(step_scopes_name_); | ||
| auto step_scopes = scopes_var->GetMutable<std::vector<ScopePtr>>(); |
| for (int j = 0; j < seq_len; j++) { | ||
| Variable* input_var = step_scopes[j]->CreateVariable(input_alias[i]); | ||
| Tensor* step_input_tensor = input_var->GetMutable<Tensor>(); | ||
| *step_input_tensor = scope_input_tensor->Slice(j, j + 1); |
| #include "paddle/framework/scope.h" | ||
| #include "paddle/framework/variable.h" | ||
|
|
||
| // Remove when including operator.h |
| /* | ||
| * Create memories in each step scope. | ||
| */ | ||
| // void CreateMemories(ScopePtr scope) const; |
|
|
||
| // this op's name, used as a unique key in father scope. | ||
| // TODO repace it with OpBase's interface if supported. | ||
| std::string name_; |
| // std::ostringstream stream; | ||
| // op_desc.SerializeToOstream(&stream); | ||
| // std::string text = stream.str(); | ||
| // LOG(INFO) << text; |
| op_desc.add_inputs("rnn/h_pre"); | ||
| op_desc.add_inputs("rnn/w"); | ||
| op_desc.add_outputs("rnn/s"); | ||
| // s = h_pre * check |
| op_desc.add_inputs("rnn/x"); | ||
| op_desc.add_inputs("rnn/s"); | ||
| op_desc.add_outputs("rnn/h"); | ||
| // h = x + s |
| ConcatOutputs(scope); | ||
| } | ||
|
|
||
| void RecurrentOp::Init(const OpDesc& op_desc, AttributeMap& attrs) { |
There was a problem hiding this comment.
Init是不带参数的，op_desc和attr都可以从Op成员变量中拿出来~小问题
There was a problem hiding this comment.
这里会连到 OpBase.Run 里，
if (!is_inited) Init(...);只是还没引入 OpBase
resolve #2801