Fix/Fix memory leak in dygraph #17394
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -112,13 +112,13 @@ void AddGradBySort(BackwardSumMap* bck_map, VarBase* target) { | |
| return a.first > b.first; | ||
| }); | ||
| for (auto& var_pair : current.second) { | ||
| Variable* origin_grad = target->var_.get(); | ||
| Variable* grad_to_add = var_pair.second->var_.get(); | ||
| VLOG(2) << "add origin_grad: " << target->Name(); | ||
| VLOG(2) << "added grad: " << var_pair.second->Name() | ||
| << " trace id is: " << var_pair.first; | ||
| AddTo(grad_to_add, origin_grad, current.first); | ||
| delete var_pair.second; | ||
| var_pair.second = nullptr; | ||
| } | ||
| } | ||
|
|
@@ -132,8 +132,8 @@ class Autograd { | |
| return; | ||
| } | ||
| VLOG(3) << "start autograd"; | ||
| bck_map = BackwardSumMap(); | ||
| grad_ref = GradientRef(); | ||
| std::deque<OpBase*> ready; | ||
| ready.push_back(var->PreOp()); | ||
|
|
||
|
|
@@ -144,7 +144,7 @@ class Autograd { | |
| OpBase* ready_op = ready.front(); | ||
| ready.pop_front(); | ||
| std::map<std::string, std::vector<VarBase*>> input_grads = | ||
| ready_op->ApplyGrad(&bck_map, &grad_ref, bck_stratedy); | ||
|
|
||
| for (auto it = input_grads.rbegin(); it != input_grads.rend(); ++it) { | ||
| const std::vector<VarBase*>& ingrads = it->second; | ||
|
|
@@ -185,12 +185,12 @@ class Autograd { | |
| for (const auto& map : candidate->grad_output_vars_) { | ||
| for (const auto& it : map) { | ||
| for (const auto& vb : it.second) { | ||
| if (grad_ref.find(vb) == grad_ref.end()) { | ||
| grad_ref.insert(std::make_pair(vb, 1)); | ||
|
There was a problem hiding this comment. C++ guarantees zero-initialization of primitive types. See here. You can just write |
||
| } else { | ||
| // add ref count by 1 when we find grad_var can be generated by | ||
| // one grad_op | ||
| grad_ref.at(vb) += 1; | ||
| } | ||
| } | ||
| } | ||
|
|
@@ -213,8 +213,8 @@ class Autograd { | |
| return ret; | ||
| } | ||
|
|
||
| BackwardSumMap bck_map; | ||
| GradientRef grad_ref; | ||
|
There was a problem hiding this comment. No need to be data members. |
||
| }; | ||
|
|
||
| std::unique_ptr<VarBase> VarBase::NewVarBase(const platform::Place& dst_place, | ||
|
|
@@ -230,14 +230,16 @@ std::unique_ptr<VarBase> VarBase::NewVarBase(const platform::Place& dst_place, | |
| new_var->var_->GetMutable<framework::LoDTensor>(); | ||
| tensor->set_lod(var_->Get<framework::LoDTensor>().lod()); | ||
|
|
||
| framework::TensorCopy(var_->Get<framework::LoDTensor>(), dst_place, tensor); | ||
| if (blocking) { | ||
| platform::DeviceContext* dst_dev_ctx = | ||
| platform::DeviceContextPool::Instance().Get(dst_place); | ||
|
|
||
| platform::DeviceContext* src_dev_ctx = | ||
| platform::DeviceContextPool::Instance().Get( | ||
| var_->Get<framework::LoDTensor>().place()); | ||
| dst_dev_ctx->Wait(); | ||
| src_dev_ctx->Wait(); | ||
| } else { | ||
| framework::TensorCopy(var_->Get<framework::LoDTensor>(), dst_place, tensor); | ||
| } | ||
|
|
@@ -326,7 +328,7 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad( | |
| PADDLE_ENFORCE_NOT_NULL(grad_inp->var_, "op %s input %s nullptr", | ||
| grad_op_desc->Type(), grad_inp->Name()); | ||
|
|
||
| grad_invars.emplace_back(grad_inp->var_.get()); | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -337,7 +339,7 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad( | |
| PADDLE_ENFORCE_NOT_NULL(grad_out->var_, "op %s output %s nullptr", | ||
| grad_op_desc->Type(), grad_out->Name()); | ||
|
|
||
| grad_outvars.emplace_back(grad_out->var_.get()); | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -396,13 +398,13 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad( | |
| grad_ref->at(origin_outputs[i])--; | ||
| } | ||
| } else { | ||
| framework::Variable* grad = outputs[i]->var_.get(); | ||
| framework::Variable* orig_grad = origin_outputs[i]->var_.get(); | ||
| VLOG(2) << "AddTo Called with orig_grad is: " | ||
| << origin_outputs[i]->name_ << " Grad to be added is " | ||
| << outputs[i]->name_; | ||
| AddTo(grad, orig_grad, place_); | ||
| delete outputs[i]; | ||
| } | ||
| } | ||
| } | ||
|
|
@@ -453,7 +455,7 @@ void PyLayer::RegisterFunc(int func_id, const py::object& py_func) { | |
|
|
||
| int PyLayer::NumFuncs() { return py_funcs_.size(); } | ||
|
|
||
| std::vector<std::unique_ptr<framework::Variable>> PyLayer::Apply( | ||
| int func_id, const std::vector<VarBase*>& inputs) { | ||
| PADDLE_ENFORCE(py_funcs_.find(func_id) != py_funcs_.end()); | ||
| return CallPythonFunc(py_funcs_[func_id], inputs); | ||
|
|
@@ -470,13 +472,13 @@ std::vector<VarBase*> PyLayer::ApplyGrad(int func_id, | |
| outs.emplace_back(new VarBase( | ||
| string::Sprintf("%s_out_%d", framework::GradVarName(PyLayer::kFwdOut), | ||
| i), | ||
| std::move(rets[i]), nullptr, true)); | ||
| } | ||
|
|
||
| return outs; | ||
| } | ||
|
|
||
| std::vector<std::unique_ptr<framework::Variable>> PyLayer::CallPythonFunc( | ||
| const py::object& callable, const std::vector<VarBase*>& ins) { | ||
| py::gil_scoped_acquire guard; | ||
| py::tuple in_args(ins.size()); | ||
|
|
@@ -490,19 +492,20 @@ std::vector<framework::Variable*> PyLayer::CallPythonFunc( | |
| auto ret = callable(in_args); | ||
| auto ret_tuple = py::cast<py::tuple>(ret); | ||
| size_t ret_num = py::len(ret_tuple); | ||
| std::vector<std::unique_ptr<framework::Variable>> outs; | ||
| outs.reserve(ret_num); | ||
| VLOG(3) << "pyfunc out " << ret_num; | ||
| for (size_t i = 0; i < ret_num; ++i) { | ||
| try { | ||
| auto* py_out_tensor = py::cast<framework::LoDTensor*>(ret_tuple[i]); | ||
| PADDLE_ENFORCE_NOT_NULL(py_out_tensor, | ||
| "Output tensor %d should not be nullptr", i); | ||
| auto var = | ||
| std::unique_ptr<framework::Variable>(new framework::Variable()); | ||
| auto* tensor = var->GetMutable<framework::LoDTensor>(); | ||
| tensor->ShareDataWith(*py_out_tensor); | ||
| tensor->set_lod(py_out_tensor->lod()); | ||
| outs.emplace_back(std::move(var)); | ||
| } catch (py::cast_error&) { | ||
| PADDLE_THROW("The %d-th output must be LoDTensor", i); | ||
| } | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
These two lines are unnecessary. BTW, it seems that it is not necessary that
bck_mapandgrad_refare data member ofAutograd. Just make them temporary variable inside the function.