diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index e88084424baf7eb5cd1d67ea19966866d71ec3eb..dc308fd2592bb158f46f6eac9dd0df25787559fe 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -50,7 +50,12 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper) +if(WITH_GPU) +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info) +else() +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info) +endif() + cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass) cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info) cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index f8030c53f72bc8a6f007c1eb6a3072abd8037de2..0c823b9ca2a135101f7776bf0a8f94f4a93ffb03 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -240,7 +240,9 @@ std::unique_ptr BuildStrategy::Apply( continue; } } + VLOG(3) << "Start Apply Pass " << pass->Type(); graph = pass->Apply(std::move(graph)); + VLOG(3) << "Finish Apply Pass " << pass->Type(); } return graph; } diff --git a/paddle/fluid/framework/details/inplace_op_pass.cc b/paddle/fluid/framework/details/inplace_op_pass.cc index b0c5968499be3a959dd6103424f25056a6dc2282..c91fc81b2defc9fe6b5720ce652a9aa94b27735e 100644 --- a/paddle/fluid/framework/details/inplace_op_pass.cc +++ b/paddle/fluid/framework/details/inplace_op_pass.cc @@ -49,7 +49,7 @@ DEFINE_bool( "If this option turns on, only these op in whitelist can be inplaced." "If it turns off, all of the running op can be candidate of inplaced op." "Such as scale, elementwise_add" - "By default, it's turned on"); + "By default, it's turned off"); DECLARE_string(memory_optimize_debug); diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc index 6345ba335997ec42ebc63f90e9bf6a3ed2648edc..6126c168ccf4b52ae75648d2dfe48db770639214 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper.cc @@ -13,13 +13,19 @@ // limitations under the License. #include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include #include #include -#include +#include #include #include #include #include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/cpu_info.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/gpu_info.h" +#endif // PADDLE_WITH_CUDA namespace paddle { namespace framework { @@ -166,6 +172,11 @@ struct NodeComparator { bool operator()(ir::Node* lhs, ir::Node* rhs) const { auto* lhs_desc = FindVarDescInBlock(lhs); auto* rhs_desc = FindVarDescInBlock(rhs); + // match data type + if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) { + return false; + } + // match shape auto lhs_shape = lhs_desc->GetShape(); auto rhs_shape = rhs_desc->GetShape(); if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) || @@ -230,6 +241,27 @@ ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const { return found_node; } +ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const { + ir::Node* found_node = nullptr; + NodeComparator functor; + auto it = + std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) { + if (v.front() == prev) + return true; + else + return false; + }); + PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!"); + for (it = std::next(it); it != nodes_.end(); ++it) { + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; + break; + } + } + return found_node; +} + bool OrderedSet::Has(ir::Node* var) const { if (mark_table_.count(var->Name())) { auto& node_in_samename = mark_table_.at(var->Name()); @@ -241,10 +273,15 @@ bool OrderedSet::Has(ir::Node* var) const { return false; } +void OrderedSet::Erase(const std::string& var) { + PADDLE_ENFORCE(mark_table_.count(var)); + nodes_.erase(mark_table_[var]); + mark_table_.erase(var); +} + void OrderedSet::Erase(ir::Node* var) { - PADDLE_ENFORCE(mark_table_.count(var->Name())); - nodes_.erase(mark_table_[var->Name()]); - mark_table_.erase(var->Name()); + PADDLE_ENFORCE(var != nullptr); + Erase(var->Name()); } std::string OrderedSet::ToString() const { @@ -274,14 +311,35 @@ bool NodeCanReused(ir::Node* node) { return flag; } +int MinChunkSize() { + int size{0}; +#ifdef PADDLE_WITH_CUDA + size = platform::GpuMinChunkSize(); +#else + size = platform::CpuMinChunkSize(); +#endif // PADDLE_WITH_CUDA + return size; +} + bool NodeCanReused(const VarDesc& node) { auto type = node.GetType(); + // only these types holds bulk of gpu memory if (!(type == proto::VarType::LOD_TENSOR || type == proto::VarType::SELECTED_ROWS || type == proto::VarType::LOD_TENSOR_ARRAY)) { return false; } - if (node.Persistable() || node.GetShape().empty()) { + // persistable variable is parameter + if (node.Persistable()) { + return false; + } + // shape < min_chunk_size is meaningless. + // further more, fetched loss always has size = 1 + // which should not be reused. + auto shape = node.GetShape(); + int size = std::abs( + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies())); + if (shape.empty() || size < MinChunkSize()) { return false; } // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad @@ -461,7 +519,9 @@ ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name, for (auto* node : ops_) { if (node == op) break; for (auto& output : node->outputs) { - if (output->Name() == name) { + PADDLE_ENFORCE((output != nullptr && output->IsVar()), + "Output is empty!"); + if (output->Var() && output->Name() == name) { found_node = output; } } diff --git a/paddle/fluid/framework/details/memory_optimize_helper.h b/paddle/fluid/framework/details/memory_optimize_helper.h index 0bfaf827fea84030de48a9984197f5b39f5c9261..dba96309fdf37e6ca73780cb194843659e1bd2d1 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.h +++ b/paddle/fluid/framework/details/memory_optimize_helper.h @@ -55,6 +55,7 @@ class OrderedSet { void Insert(ir::Node* var); void Erase(ir::Node* var); + void Erase(const std::string& var); bool Has(ir::Node* var) const; void Clear() { mark_table_.clear(); @@ -62,6 +63,7 @@ class OrderedSet { } // find the bestfit shape node block with var. ir::Node* FindBestFitNode(ir::Node* var) const; + ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const; // map store non-const iterator, can not promise const int GetNodeIndexInPool(ir::Node* var); // pool all node to string diff --git a/paddle/fluid/framework/details/memory_optimize_helper_test.cc b/paddle/fluid/framework/details/memory_optimize_helper_test.cc index 5c13dda9e5491044d2bcbed4b24d438cc8b8a413..3cfe297a73cf4128b7191cbd432cdceadc6240ec 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper_test.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper_test.cc @@ -107,6 +107,52 @@ TEST(OrderedSet, Normal) { ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2] } } + +TEST(OrderedSet, FindBestFitNode) { + OrderedSet pool; + std::vector> nodes; + ProgramDesc prog; + BlockDesc* block_desc = prog.MutableBlock(0); + auto* op_desc = block_desc->AppendOp(); + op_desc->SetType("dummy"); + std::unique_ptr op = ir::CreateNodeForTest(op_desc); + + { + auto desc = block_desc->Var("a"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("b"); + desc->SetShape({128, 129}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("c"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + + for (auto& node : nodes) { + pool.Insert(node.get()); + } + + // FindNextBestFitNode + auto* n = nodes[0].get(); + auto* cache = pool.FindBestFitNode(n); + PADDLE_ENFORCE(cache->Name() == "a"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "c"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "b"); +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc index 41e4a834df0abab069ab1f6cd21c8479b911250c..b35b967c72dcec7c427f6921c094f7a87d536f25 100644 --- a/paddle/fluid/framework/details/memory_optimize_pass.cc +++ b/paddle/fluid/framework/details/memory_optimize_pass.cc @@ -69,55 +69,59 @@ std::unique_ptr MemoryOptimizePass::ApplyImpl( } for (auto& var : op->outputs) { - if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 || - skip_set_.count(var->Name())) + if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) { + VLOG(3) << "Skip set contains variable of " << var->Name() + << "disable reuse on it. skipped"; continue; - ir::Node* cache = pool_.FindBestFitNode(var); - - if (var->Name() == FLAGS_memory_optimize_debug) { - VLOG(3) << "start match var " << DebugString(var) << " of op " - << op->Name(); - VLOG(3) << pool_.ToString(); - VLOG(3) << "matched in pool : " - << ((cache == nullptr) ? "False" : "True"); } + if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) { + ir::Node* cache = pool_.FindBestFitNode(var); + while (cache != nullptr && var->Name() == cache->Name()) { + VLOG(3) << "The same cache variable is cascade reused. " + << cache->Name() << " is re-filled to the pool after " + << "the reused op is finished. Current op can not " + << "replace it again. Skip this candidate."; + cache = pool_.FindNextBestFitNode(var, cache); + } + if (var->Name() == FLAGS_memory_optimize_debug) { + VLOG(3) << "start match var " << DebugString(var) << " of op " + << op->Name(); + VLOG(3) << pool_.ToString(); + VLOG(3) << "matched in pool : " + << ((cache == nullptr) ? "False" : "True"); + } - if (cache == nullptr) continue; - if (var->Name() == cache->Name()) { - VLOG(3) << "The same cache variable is cascade reused." << var->Name() - << " is re-filled to the pool after" - << "the reused op is finished. Current op can not " - << "replace it again. Skip this candidate."; - continue; - - int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); - VLOG(3) << string::Sprintf( - "!!! %s, %s => %s, cache idx %d, pool size %d", - std::to_string(reuse_id++), DebugString(var), DebugString(cache), - node_idx_in_pool, static_cast(pool_.size())); - - // update CFG Graph on the fly. - // reused var maybe re-fill into the pool - cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx); - // NOTE(dzhwinter): we need to both update the ProgramDesc - // and IR Graph. because op_desc/var_desc is used in CreateOp, - // CreateVar when running happens. But IR Graph - // define the dependence relationship between nodes. - RenameVarInGraphDesc(var->Name(), cache->Name(), idx); - RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get()); - - pool_.Erase(cache); - } + if (cache != nullptr) { + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); + VLOG(3) << string::Sprintf( + "!!! %s, %s => %s, cache idx %d, pool size %d", + std::to_string(reuse_id++), DebugString(var), DebugString(cache), + node_idx_in_pool, static_cast(pool_.size())); + // NOTE(dzhwinter): update the ProgramDesc/IR Graph + // and the CFG Graph on the fly. + // + // IR Graph define the dependence relationship between nodes. + // + // ProgramDesc defines the input/output vars. Its used in + // CreateOp, CreateVar when running happens. + // + // CFG Graph store the liveness information, when reuse happens + // we also need to update the variable liveness. + const std::string var_name = var->Name(); + const std::string cache_name = cache->Name(); - // fill the pool - std::unordered_set unlived_vars; - for (auto var : cfg_->LiveIn(op)) { - if (cfg_->LiveOut(op).count(var) == 0) { - unlived_vars.emplace(var); + cfg_->RenameVarInCFGGraph(var_name, cache_name, idx); + RenameVarInGraphDesc(var_name, cache_name, idx); + RenameVarInGraphNode(var_name, cache_name, idx, graph.get()); + pool_.Erase(cache_name); } } - for (auto var : unlived_vars) { + } + // fill the pool + for (auto var : cfg_->LiveIn(op)) { + if (cfg_->LiveOut(op).count(var) == 0) { ir::Node* var_node = cfg_->GetNodeByName(var, op); + if (var_node == nullptr || var_node->IsCtrlVar()) continue; if (NodeCanReused(var_node) && !pool_.Has(var_node)) { pool_.Insert(var_node); } @@ -273,8 +277,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, // redirect the input to the latest version of cache_var for (auto* node : op->inputs) { if (node->Name() == var) { - ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); - var_nodes_[cache_var].emplace_back(cache_node); + ir::Node* cache_node = var_nodes_[cache_var].back(); // swap node to cache_node cache_node->outputs.insert(cache_node->outputs.end(), @@ -283,11 +286,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, auto* prev_op = node->inputs[0]; std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, cache_node); - cache_node->inputs.emplace_back(prev_op); for (auto* next_op : node->outputs) { std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } @@ -307,15 +314,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } } - - // release node of unused var in graph - for (auto* node : var_nodes_[var]) { - graph->RemoveNode(node); - } - var_nodes_.at(var).clear(); } } // namespace details diff --git a/paddle/fluid/framework/inplace_op_inference_test.cc b/paddle/fluid/framework/inplace_op_inference_test.cc index 3e4d715c6f089496d1b1f7906e3f10147a073622..bf9d1dcd380cdff886301faf13b0015fd5a2ed5c 100644 --- a/paddle/fluid/framework/inplace_op_inference_test.cc +++ b/paddle/fluid/framework/inplace_op_inference_test.cc @@ -179,11 +179,11 @@ TEST(InferInplace, SingleOpInplaceInToOut) { op->SetOutput("Out", {"test2_out"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -201,11 +201,11 @@ TEST(InferInplace, SingleGradOpInplaceInToOut) { op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -233,12 +233,12 @@ TEST(InferInplace, MultiOutInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -267,12 +267,12 @@ TEST(InferInplace, MultiGradInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); diff --git a/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc index 3b738aa159ebfd77f00c9e532fbd94542e2097db..5bdc0c5faed7131b873edf9b43c847c010b6e3f3 100644 --- a/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc +++ b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc @@ -38,9 +38,13 @@ std::unique_ptr IdentityScaleOpCleanPass::ApplyImpl( ->assert_is_op("scale") ->assert_op_attr("scale", 1.) ->assert_op_attr("bias", 0.); - auto scale_out = detector.mutable_pattern() - ->NewNode("scale_out") - ->assert_is_op_output("scale"); + auto scale_out = + detector.mutable_pattern() + ->NewNode("scale_out") + ->assert_is_op_output("scale") + // scale's output var should has only one consumer, or it can't be + // removed. + ->assert_more([](Node* x) { return x->outputs.size() == 1UL; }); pre_op->LinksTo({scale_in}); scale_op->LinksFrom({scale_in}).LinksTo({scale_out}); diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc index 47488d4dea79f285769f29c93f7888a7f783f070..8f20f0c06e043ddc629e47c6e49280c5467b0e20 100644 --- a/paddle/fluid/imperative/layer.cc +++ b/paddle/fluid/imperative/layer.cc @@ -207,7 +207,7 @@ framework::LoDTensor& VarBase::GradValue() { std::map> OpBase::ApplyGrad() { if (grad_op_descs_.empty() && backward_id_ <= 0) { - LOG(WARNING) << "op with no grad: " << op_desc_->Type(); + VLOG(3) << "op with no grad: " << op_desc_->Type(); return {}; } diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc index a64f85ee9ac1a7bb8f0ed7bb8678166bbbcd5746..96befe7f8a5d16402338ac337daa96d714b4d310 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc @@ -460,77 +460,6 @@ inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) { return node.inputs.size() == n; } -NodesTSIterator::NodesTSIterator(const std::vector &source) { - PADDLE_ENFORCE(!source.empty(), - "Start points of topological sorting should not be empty!"); - // CHECK all the inputs' in-degree is 0 - for (auto *node : source) { - PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0)); - } - - std::unordered_set visited; - std::unordered_set to_visit{source.begin(), source.end()}; - - std::vector inlink_visited; - while (!to_visit.empty()) { - std::vector queue(to_visit.begin(), to_visit.end()); - for (auto *p : queue) { - if (Agent(p).deleted()) { - visited.insert(p); - to_visit.erase(p); - } - - inlink_visited.clear(); - - std::copy_if(p->inputs.begin(), p->inputs.end(), - std::back_inserter(inlink_visited), - [&](Node *x) -> bool { return visited.count(x) != 0; }); - - if (inlink_visited.size() == p->inputs.size()) { - sorted_.push_back(p); - for (auto *_ : p->outputs) { - if (!visited.count(_)) { - to_visit.insert(_); - } - } - - to_visit.erase(p); - visited.insert(p); - } - } - } -} - -NodesTSIterator::NodesTSIterator(const NodesTSIterator &other) - : sorted_(other.sorted_), cursor_(other.cursor_) {} - -Node &NodesTSIterator::operator*() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return *sorted_[cursor_]; -} - -NodesTSIterator &NodesTSIterator::operator++() { - if (++cursor_ >= sorted_.size()) { - sorted_.clear(); - cursor_ = 0; - } - return *this; -} -NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) { - cursor_ = other.cursor_; - sorted_ = other.sorted_; - return *this; -} - -bool NodesTSIterator::operator==(const NodesTSIterator &other) { - return sorted_ == other.sorted_ && cursor_ == other.cursor_; -} - -Node *NodesTSIterator::operator->() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return sorted_[cursor_]; -} - } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h index ea88edd042aa9d46f66af1aa92f2cb273696c118..5d11c217b69f11d45c6fb6d552dc404fa8313daf 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h @@ -30,6 +30,7 @@ namespace inference { namespace analysis { using framework::ir::Graph; +using framework::ir::NodesTSIterator; const char kIsFunctionNode[] = "__is_function_node__"; const char kFunctionNodeSubGraph[] = "__function_node_sub_graph__"; @@ -132,32 +133,6 @@ struct Agent { framework::ir::Node *x_; }; -// Topological sorting iterator on nodes. -struct NodesTSIterator - : public std::iterator { - NodesTSIterator() = default; - explicit NodesTSIterator(const std::vector &source); - NodesTSIterator(NodesTSIterator &&other) - : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) { - other.cursor_ = 0; - } - NodesTSIterator(const NodesTSIterator &other); - - framework::ir::Node &operator*(); - NodesTSIterator &operator++(); - // TODO(Superjomn) current implementation just compare the first - // element, need to compare the graph and all the elements in the queue and - // set. - NodesTSIterator &operator=(const NodesTSIterator &other); - bool operator==(const NodesTSIterator &other); - bool operator!=(const NodesTSIterator &other) { return !(*this == other); } - framework::ir::Node *operator->(); - - private: - std::vector sorted_; - size_t cursor_{0}; -}; - // The nodes those have no input will be treated as start points. static std::vector ExtractStartPoints(const Graph &g) { std::vector result; diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h index 3591681fc3f6951dfc8d73e8edce38180b771eaf..42137215e21af1a529563ecc995a54d610120beb 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.h +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -72,7 +72,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif - for (int i = 0; i < fixed_ratios.size(); i++) { + for (size_t i = 0; i < fixed_ratios.size(); i++) { sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i])); } @@ -115,11 +115,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { } } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; var_t.mutable_data( @@ -141,7 +140,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #pragma omp parallel for collapse(2) #endif for (int i = 0; i < box_num; ++i) { - for (int j = 0; j < variances.size(); ++j) { + for (size_t j = 0; j < variances.size(); ++j) { e_vars(i, j) = variances[j]; } } diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h index 4e226abbb51c271502f0ca5419d488643b5a1a82..f84405664596ebe25983e5acbbb82bfc18c38124 100644 --- a/paddle/fluid/operators/detection/prior_box_op.h +++ b/paddle/fluid/operators/detection/prior_box_op.h @@ -46,13 +46,6 @@ inline void ExpandAspectRatios(const std::vector& input_aspect_ratior, } } -template -struct ClipFunctor { - HOSTDEVICE inline T operator()(T in) const { - return std::min(std::max(in, 0.), 1.); - } -}; - template class PriorBoxOpKernel : public framework::OpKernel { public: @@ -101,31 +94,30 @@ class PriorBoxOpKernel : public framework::OpKernel { boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - auto e_boxes = framework::EigenTensor::From(*boxes); + T* b_t = boxes->data(); for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { T center_x = (w + offset) * step_width; T center_y = (h + offset) * step_height; T box_width, box_height; - int idx = 0; for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; if (min_max_aspect_ratios_order) { box_width = box_height = min_size / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } // priors with different aspect ratios for (size_t r = 0; r < aspect_ratios.size(); ++r) { @@ -135,11 +127,11 @@ class PriorBoxOpKernel : public framework::OpKernel { } box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } else { // priors with different aspect ratios @@ -147,21 +139,21 @@ class PriorBoxOpKernel : public framework::OpKernel { float ar = aspect_ratios[r]; box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } } @@ -169,11 +161,10 @@ class PriorBoxOpKernel : public framework::OpKernel { } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; diff --git a/paddle/fluid/operators/group_norm_op.cc b/paddle/fluid/operators/group_norm_op.cc index e18d9841bb87c6a684d53e1bceb6c20a37dcfcfa..cbdffa0db8277dbf7257c3b3c1d03c1b459d5b2b 100644 --- a/paddle/fluid/operators/group_norm_op.cc +++ b/paddle/fluid/operators/group_norm_op.cc @@ -170,13 +170,48 @@ class GroupNormGradMaker : public framework::SingleGradOpDescMaker { } }; +class GroupNormInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + return {{"X", "Y"}}; + } +}; + +class GroupNormGradInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + return {{framework::GradVarName("Y"), framework::GradVarName("X")}}; + } +}; + +class GroupNormOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return {{"X", /*->*/ "Y"}}; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(group_norm, ops::GroupNormOp, ops::GroupNormOpMaker, - ops::GroupNormGradMaker); -REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp); + ops::GroupNormOpInferVarType, ops::GroupNormGradMaker, + ops::GroupNormInplaceInToOut); +REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp, + ops::GroupNormGradInplaceInToOut); REGISTER_OP_CPU_KERNEL( group_norm, ops::GroupNormKernel, ops::GroupNormKernel); diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc index 97ddf223aefcdfaf8a488f93a152336c1ed458f4..77a2d04ebf176ca0807fa155f87261fcd838da04 100644 --- a/paddle/fluid/operators/jit/benchmark.cc +++ b/paddle/fluid/operators/jit/benchmark.cc @@ -339,6 +339,71 @@ void BenchSoftmaxKernel() { } } +template +void BenchLayerNormKernel() { + const T epsilon = 9.99999975e-06; + for (int n : {1, 2, 10}) { + for (int x_dim_0 : {1, 9, 17, 50}) { + int left = n * x_dim_0; + for (int x_dim_1 : TestSizes()) { + int right = x_dim_1; + int sz = left * right; + Tensor x, mean, var, scale, bias, out; + x.Resize({n, x_dim_0, x_dim_1}); + out.Resize({n, x_dim_0, x_dim_1}); + mean.Resize({n, x_dim_0}); + var.Resize({n, x_dim_0}); + scale.Resize({x_dim_1}); + bias.Resize({x_dim_1}); + + RandomVec(sz, x.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(left, mean.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(left, var.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(right, scale.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(right, bias.mutable_data(PlaceType()), -2.f, 2.f); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* out_data = out.mutable_data(PlaceType()); + + BenchAllImpls, PlaceType>( + right, x_data, out_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + } + } + } +} + +template +void BenchCRFDecodingKernel() { + constexpr int state_trans_base_idx = 2; + for (int seq_len : {1, 11, 17, 50}) { + for (int tag_num : TestSizes()) { + int x_sz = seq_len * tag_num; + int w_sz = (tag_num + state_trans_base_idx) * tag_num; + Tensor x, w, alpha, track; + x.Resize({seq_len, tag_num}); + w.Resize({tag_num + state_trans_base_idx, tag_num}); + alpha.Resize({seq_len, tag_num}); + track.Resize({seq_len, tag_num}); + + RandomVec(x_sz, x.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(w_sz, w.mutable_data(PlaceType()), -2.f, 2.f); + + const T* x_data = x.data(); + const T* w_data = w.data(); + T* alpha_data = alpha.mutable_data(PlaceType()); + int* track_data = track.mutable_data(PlaceType()); + + BenchAllImpls, PlaceType>( + tag_num, seq_len, x_data, w_data, alpha_data, track_data, tag_num); + } + } +} + using T = float; using CPUPlace = paddle::platform::CPUPlace; @@ -382,6 +447,16 @@ BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel(); } // softmax BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel(); } +// layernorm +BENCH_FP32_CPU(kLayerNorm) { + BenchLayerNormKernel(); +} + +// crfdecoding +BENCH_FP32_CPU(kCRFDecoding) { + BenchCRFDecodingKernel(); +} + // Benchmark all jit kernels including jitcode, mkl and refer. // To use this tool, run command: ./benchmark [options...] // Options: diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc index 237e588d35cc3b33658a830db34676967818aab6..85b50b79d95070a56da384d88c356294b7ed9f9f 100644 --- a/paddle/fluid/operators/jit/test.cc +++ b/paddle/fluid/operators/jit/test.cc @@ -292,6 +292,63 @@ struct TestFuncWithRefer, std::vector, std::vector, } }; +template +struct TestFuncWithRefer, std::vector, + std::vector, std::vector, std::vector, + std::vector, std::vector, int, float, int> { + void operator()(const typename jit::LayerNormTuples::func_type tgt, + std::vector& x, std::vector& outref, // NOLINT + std::vector& mean, std::vector& var, // NOLINT + const std::vector& scale, const std::vector& bias, + int left, const float epsilon, int right) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(left * right)); + EXPECT_EQ(outref.size(), static_cast(left * right)); + EXPECT_EQ(mean.size(), static_cast(left)); + EXPECT_EQ(var.size(), static_cast(left)); + EXPECT_EQ(scale.size(), static_cast(right)); + EXPECT_EQ(bias.size(), static_cast(right)); + std::vector outtgt(outref.size()); + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + T* outtgt_data = outtgt.data(); + + tgt(x_data, outtgt_data, mean_data, var_data, scale_data, bias_data, left, + epsilon, right); + ExpectEQ(outtgt_data, outref_data, left * right); + } +}; + +template +struct TestFuncWithRefer, int, std::vector, + std::vector, std::vector, std::vector, + int> { + void operator()(const typename jit::CRFDecodingTuples::func_type tgt, + const int seq_len, const std::vector& x, + const std::vector& w, std::vector& alpharef, // NOLINT + std::vector& trackref, int tag_num) { // NOLINT + constexpr int state_trans_base_idx = 2; + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(w.size(), + static_cast((tag_num + state_trans_base_idx) * tag_num)); + EXPECT_EQ(alpharef.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(trackref.size(), static_cast(seq_len * tag_num)); + std::vector alphatgt(alpharef.size()); + std::vector tracktgt(trackref.size()); + + memcpy(trackref.data(), tracktgt.data(), tag_num * sizeof(int)); + tgt(seq_len, (const T*)x.data(), (const T*)w.data(), alphatgt.data(), + tracktgt.data(), tag_num); + ExpectEQ(alpharef.data(), alphatgt.data(), seq_len * tag_num); + ExpectEQ(trackref.data(), tracktgt.data(), seq_len * tag_num); + } +}; + template void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { @@ -640,6 +697,71 @@ void TestNCHW16CMulNCKernel() { } } +template +void TestLayerNormKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + const T epsilon = 9.99999975e-06; + for (int n : {1, 2, 10}) { + for (int x_dim_0 : {1, 9, 17, 50}) { + int left = n * x_dim_0; + for (int x_dim_1 : TestSizes()) { + int right = x_dim_1; + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + int sz = left * right; + std::vector x(sz), mean(left), var(left), scale(right), bias(right), + outref(sz); + RandomVec(sz, x.data(), -2.f, 2.f); + RandomVec(left, mean.data(), -2.f, 2.f); + RandomVec(left, var.data(), -2.f, 2.f); + RandomVec(right, scale.data(), -2.f, 2.f); + RandomVec(right, bias.data(), -2.f, 2.f); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + + ref(x_data, outref_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + + TestAllImpls, PlaceType, std::vector, + std::vector, std::vector, std::vector, + std::vector, std::vector, int, float>( + right, x, outref, mean, var, scale, bias, left, epsilon, right); + } + } + } +} + +template +void TestCRFDecodingKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + constexpr int state_trans_base_idx = 2; + for (int seq_len : {1, 11, 17, 50}) { + for (int tag_num : TestSizes()) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + int x_sz = seq_len * tag_num; + int w_sz = (tag_num + state_trans_base_idx) * tag_num; + std::vector x(x_sz), w(w_sz), alpharef(x_sz); + std::vector trackref(x_sz); + RandomVec(x_sz, x.data(), -2.f, 2.f); + RandomVec(w_sz, w.data(), -2.f, 2.f); + + ref(seq_len, (const T*)x.data(), (const T*)w.data(), alpharef.data(), + trackref.data(), tag_num); + + TestAllImpls, PlaceType, int, + std::vector, std::vector, std::vector, + std::vector, int>(tag_num, seq_len, x, w, alpharef, + trackref, tag_num); + } + } +} + // XYZNTuple TEST(JITKernel, kVMul) { TestXYZNKernel(); @@ -761,7 +883,16 @@ TEST(JITKernel, kNCHW16CMulNC) { TestNCHW16CMulNCKernel(); } -// TODO(yihua/TJ): add crf decoding and layer norm unit tests +TEST(JITKernel, kLayerNorm) { + TestLayerNormKernel(); + TestLayerNormKernel(); +} + +TEST(JITKernel, kCRFDecoding) { + TestCRFDecodingKernel(); + TestCRFDecodingKernel(); +} TEST(JITKernel, pool) { // TODO(TJ): add some test diff --git a/paddle/fluid/operators/load_combine_op.cc b/paddle/fluid/operators/load_combine_op.cc index c4a2282e16483dbe78a32a4148c5bc4349dde3dc..f5c802986e0573e81b3ab6187b57657b52b37215 100644 --- a/paddle/fluid/operators/load_combine_op.cc +++ b/paddle/fluid/operators/load_combine_op.cc @@ -64,7 +64,7 @@ class LoadCombineOp : public framework::OperatorBase { auto *tensor = out_var->GetMutable(); // Error checking - PADDLE_ENFORCE(static_cast(buffer), "Cannot read more"); + PADDLE_ENFORCE(static_cast(*buffer), "Cannot read more"); // Get data from fin to tensor DeserializeFromStream(*buffer, tensor, dev_ctx); @@ -90,6 +90,10 @@ class LoadCombineOp : public framework::OperatorBase { tensor->ShareDataWith(fp16_tensor); } } + buffer->peek(); + PADDLE_ENFORCE(buffer->eof(), + "You are not allowed to load partial data via " + "load_combine_op, use load_op instead."); } }; diff --git a/paddle/fluid/operators/lstm_op.h b/paddle/fluid/operators/lstm_op.h index 7d62d2d020ec2e3a29ad8720a8f04fead3a90a63..3f110024b285d41ccfe305e35c8efca5ed5ee0fe 100644 --- a/paddle/fluid/operators/lstm_op.h +++ b/paddle/fluid/operators/lstm_op.h @@ -311,6 +311,10 @@ class LSTMGradKernel : public framework::OpKernel { lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data() : nullptr; } + // lstm_value.output_value not used in bp, set to nullptr + // lstm_grad.state_active_grad not used in bp, set to nullptr + lstm_value.output_value = nullptr; + lstm_grad.state_active_grad = nullptr; int cur_batch_size = bend - bstart; math::LstmUnitGradFunctor::compute( device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size, diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h index 370dd04d1449a8e211febf9a4f9e90e6f5008e20..1f11e57dcb721012c7b8e50d7e138355685053da 100644 --- a/paddle/fluid/operators/lstmp_op.h +++ b/paddle/fluid/operators/lstmp_op.h @@ -405,6 +405,11 @@ class LSTMPGradKernel : public framework::OpKernel { } int cur_batch_size = bend - bstart; + // lstmp_value.output_value not used in bp, set to null + // lstmp_grad.state_active_grad not used in bp, set to null + lstmp_value.output_value = nullptr; + lstmp_grad.state_active_grad = nullptr; + math::LstmUnitGradFunctor::compute( device_ctx, lstmp_value, lstmp_grad, frame_size, cur_batch_size, gate_act, cell_act, cand_act); diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 10b1b0c899d833d70fa6afe51998fe210899e3c3..d283bddbe9f974ac6835ee91d5a7851453687b80 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -109,23 +109,23 @@ from future subsequences in a computationally efficient manner to improve unidirectional recurrent neural networks. The row convolution operator is different from the 1D sequence convolution, and is computed as follows: -Given an input sequence $in$ of length $t$ and input dimension $d$, -and a filter ($W$) of size $context \times d$, +Given an input sequence $X$ of length $t$ and input dimension $D$, +and a filter ($W$) of size $context \times D$, the output sequence is convolved as: $$ -out_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\cdot W_{i-j, :} +out_{i} = \\sum_{j=i}^{i + context - 1} X_{j} \\cdot W_{j-i} $$ In the above equation: * $Out_{i}$: The i-th row of output variable with shape [1, D]. -* $\\tau$: Future context size. +* $context$: Future context size. * $X_{j}$: The j-th row of input variable with shape [1, D]. -* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D]. +* $W_{j-i}$: The (j-i)-th row of parameters with shape [1, D]. More details about row_conv please refer to the design document diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index 142d38f0609d963ce3ff45c595b8432b0e5edd21..d32f9c8667d34246e55b039165bec05b7bec0376 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -233,9 +233,11 @@ inline void throw_on_error(ncclResult_t stat, const std::string& msg) { #endif // __APPLE__ and windows #endif // PADDLE_WITH_CUDA -#define PADDLE_THROW(...) \ - throw ::paddle::platform::EnforceNotMet( \ - ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__) +#define PADDLE_THROW(...) \ + do { \ + throw ::paddle::platform::EnforceNotMet( \ + ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \ + } while (0) #define PADDLE_ENFORCE(COND, ...) \ do { \ @@ -270,23 +272,25 @@ inline void throw_on_error(ncclResult_t stat, const std::string& msg) { * extra messages is also supported, for example: * PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2) */ -#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...) \ - do { \ - if (UNLIKELY(nullptr == (__VAL))) { \ - PADDLE_THROW(#__VAL " should not be null\n%s", \ - paddle::string::Sprintf("" __VA_ARGS__)); \ - } \ +#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...) \ + do { \ + if (UNLIKELY(nullptr == (__VAL))) { \ + PADDLE_THROW(#__VAL " should not be null\n%s", \ + ::paddle::string::Sprintf(__VA_ARGS__)); \ + } \ } while (0) #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ do { \ - if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) { \ + auto __cond1__ = (__VAL0); \ + auto __cond2__ = (__VAL1); \ + if (UNLIKELY(!((__cond1__)__CMP(__cond2__)))) { \ PADDLE_THROW("Enforce failed. Expected %s " #__CMP \ " %s, but received %s:%s " #__INV_CMP " %s:%s.\n%s", \ #__VAL0, #__VAL1, #__VAL0, \ - paddle::string::to_string(__VAL0), #__VAL1, \ - paddle::string::to_string(__VAL1), \ - paddle::string::Sprintf("" __VA_ARGS__)); \ + ::paddle::string::to_string(__cond1__), #__VAL1, \ + ::paddle::string::to_string(__cond2__), \ + ::paddle::string::Sprintf(__VA_ARGS__)); \ } \ } while (0) diff --git a/paddle/fluid/pybind/ir.cc b/paddle/fluid/pybind/ir.cc index 24059140ab20e24917b93a5f60936b1087797ff9..1cd1be8e8d9da8c6a82ceefc3284084bfeda0252 100644 --- a/paddle/fluid/pybind/ir.cc +++ b/paddle/fluid/pybind/ir.cc @@ -13,10 +13,12 @@ // limitations under the License. #include "paddle/fluid/pybind/ir.h" +#include #include #include #include #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" #include "paddle/fluid/framework/ir/node.h" #include "paddle/fluid/framework/op_desc.h" @@ -27,6 +29,10 @@ namespace py = pybind11; using paddle::framework::ir::Graph; using paddle::framework::ir::Node; using paddle::framework::ir::GraphSafeRemoveNodes; +using paddle::framework::ir::HasCircle; +using paddle::framework::ir::GraphNum; +using paddle::framework::ir::TopologySortOperations; +using paddle::framework::ir::BuildOperationAdjList; using paddle::framework::OpDesc; using paddle::framework::ProgramDesc; using paddle::framework::VarDesc; @@ -36,6 +42,12 @@ namespace paddle { namespace pybind { void BindGraph(py::module *m) { m->def("graph_safe_remove_nodes", GraphSafeRemoveNodes); + m->def("has_circle", HasCircle); + m->def("graph_num", GraphNum); + m->def("topology_sort", TopologySortOperations, + return_value_policy::reference); + m->def("build_adjacency_list", BuildOperationAdjList, + return_value_policy::reference); py::class_>( *m, "Graph", "The graph is a Directed Acyclic Single Static Assignment Graph, see " @@ -46,7 +58,6 @@ void BindGraph(py::module *m) { .def("get_float", &Graph::Get) .def("get_double", &Graph::Get) .def("get_string", &Graph::Get) - .def("get_program", &Graph::Get) .def("get_marked_nodes", &Graph::Get>) .def("set", [](Graph &self, const std::string &attr_name, int attr) { return self.Set(attr_name, new int(attr)); }) @@ -63,11 +74,6 @@ void BindGraph(py::module *m) { [](Graph &self, const std::string &attr_name, double attr) { return self.Set(attr_name, new double(attr)); }) - .def("set", - [](Graph &self, const std::string &attr_name, - const ProgramDesc &attr) { - return self.Set(attr_name, new ProgramDesc(attr)); - }) .def("set", [](Graph &self, const std::string &attr_name, const std::unordered_set &attr) { @@ -108,42 +114,42 @@ void BindNode(py::module *m) { .def("is_op", &Node::IsOp) .def("is_var", &Node::IsVar) .def("is_ctrl_var", &Node::IsCtrlVar) + .def("clear_inputs", [](Node &self) { self.inputs.clear(); }) .def("inputs_remove", [](Node &self, int node_id) { - for (auto it = self.inputs.begin(); it != self.inputs.end(); - it++) { - if ((*it)->id() == node_id) { - self.inputs.erase(it); - } + auto pos = std::find_if( + self.inputs.begin(), self.inputs.end(), + [&node_id](const Node *n) { return n->id() == node_id; }); + if (pos != self.inputs.end()) { + self.inputs.erase(pos); } }) .def("inputs_remove", [](Node &self, Node &node) { - for (auto it = self.inputs.begin(); it != self.inputs.end(); - it++) { - if (*it == &node) { - self.inputs.erase(it); - } + auto pos = + std::find(self.inputs.begin(), self.inputs.end(), &node); + if (pos != self.inputs.end()) { + self.inputs.erase(pos); } }) .def("inputs_append", [](Node &self, Node &node) { self.inputs.push_back(&node); }) + .def("clear_outputs", [](Node &self) { self.outputs.clear(); }) .def("outputs_remove", [](Node &self, int node_id) { - for (auto it = self.outputs.begin(); it != self.outputs.end(); - it++) { - if ((*it)->id() == node_id) { - self.outputs.erase(it); - } + auto pos = std::find_if( + self.outputs.begin(), self.outputs.end(), + [&node_id](const Node *n) { return n->id() == node_id; }); + if (pos != self.outputs.end()) { + self.outputs.erase(pos); } }) .def("outputs_remove", [](Node &self, Node &node) { - for (auto it = self.outputs.begin(); it != self.outputs.end(); - it++) { - if (*it == &node) { - self.outputs.erase(it); - } + auto pos = + std::find(self.outputs.begin(), self.outputs.end(), &node); + if (pos != self.outputs.end()) { + self.outputs.erase(pos); } }) .def("outputs_append", diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 351513712cc4297bf7fbe67878aeba162ef66e4d..a4a01ad647b038bd2bfea00fefa30abb19f58b66 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -829,8 +829,7 @@ All parameter, weight, gradient are variables in Paddle. m.def("disable_profiler", platform::DisableProfiler); m.def("is_profiler_enabled", platform::IsProfileEnabled); m.def("reset_profiler", platform::ResetProfiler); - m.def("get_pass", [](const py::bytes &binary_str) { - std::string pass_type(binary_str); + m.def("get_pass", [](const std::string &pass_type) { auto pass = framework::ir::PassRegistry::Instance().Get(pass_type); return std::shared_ptr(std::move(pass)); }); @@ -838,10 +837,9 @@ All parameter, weight, gradient are variables in Paddle. py::class_> pass(m, "Pass"); pass.def(py::init()) .def("has", &ir::Pass::Has) - .def("set", - [](ir::Pass &self, const std::string &attr_name, - const ProgramDesc &attr) { - return self.Set(attr_name, new ProgramDesc(attr)); + .def("set_not_owned", + [](ir::Pass &self, const std::string &attr_name, ProgramDesc &attr) { + self.SetNotOwned(attr_name, &attr); }) .def( "set", @@ -850,7 +848,6 @@ All parameter, weight, gradient are variables in Paddle. }) .def("set", [](ir::Pass &self, const std::string &name, int val) { self.Set(name, new int(val)); }) - .def("get_program", &ir::Pass::Get) .def("type", &ir::Pass::Type) .def("apply", [](ir::Pass &self, std::shared_ptr graph) { std::unique_ptr origin_graph(graph.get()); diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index 90b8fd1a0aab159eb1a829d67485c845182d295b..bcc997ff4511db45d2a775092c0798d7c1e9be06 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -64,6 +64,7 @@ if (WITH_TESTING) add_subdirectory(paddle/dataset/tests) add_subdirectory(paddle/fluid/tests) add_subdirectory(paddle/fluid/contrib/tests) + add_subdirectory(paddle/fluid/contrib/slim/tests) endif() install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR} DESTINATION opt/paddle/share/wheels diff --git a/python/paddle/fluid/compiler.py b/python/paddle/fluid/compiler.py index ef0242942838fcca737a10fafbafa61bf520b532..b24cec044f1918ea9962dc10fc3939444bb340e8 100644 --- a/python/paddle/fluid/compiler.py +++ b/python/paddle/fluid/compiler.py @@ -177,7 +177,10 @@ class CompiledProgram(object): # FIXME(dzhwinter): enable_inplace should be after memory_optimize # if turn on python memory optimize, turn off the inplace_pass. - self._build_strategy.enable_inplace = False if self._program._is_mem_optimized else True + if self._build_strategy.memory_optimize is None: + self._build_strategy.memory_optimize = False if main._is_mem_optimized else True + if self._build_strategy.enable_inplace is None: + self._build_strategy.enable_inplace = False if main._is_mem_optimized else True if self._build_strategy.num_trainers > 1 and trainers_endpoints: assert self._build_strategy.num_trainers == len( diff --git a/python/paddle/fluid/contrib/int8_inference/README.md b/python/paddle/fluid/contrib/int8_inference/README.md index a9691dad4494f5eacf427b2806b2393baa57dc1e..460ae393f158ae320c93601365a68b8cfe2ba50e 100644 --- a/python/paddle/fluid/contrib/int8_inference/README.md +++ b/python/paddle/fluid/contrib/int8_inference/README.md @@ -63,10 +63,10 @@ Notes: ## 4. How to reproduce the results * Small dataset ```bash -python python/paddle/fluid/contrib/tests/test_calibration.py +FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py ``` * Full dataset ```bash -DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py +FLAGS_use_mkldnn=true DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py ``` diff --git a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py index 266a106bc507104c0a8db1c882b55ac59e88195e..18b58e6f388bbe9495333b12f32d63b74fddcb3a 100644 --- a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py +++ b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py @@ -13,14 +13,19 @@ # limitations under the License. import collections +import numpy as np +import six +from ..... import compat as cpt from .... import core from ....framework import IrGraph from ....framework import Program -from ....framework import Variable from ....initializer import Constant from .... import unique_name -__all__ = ['QuantizationTransformPass'] +__all__ = [ + 'QuantizationTransformPass', 'QuantizationFreezePass', 'ConvertToInt8Pass', + 'TransformForMobilePass' +] class QuantizationTransformPass(object): @@ -35,7 +40,13 @@ class QuantizationTransformPass(object): """ Convert and rewrite the IrGraph according to weight and activation quantization type. + Args: + scope(fluid.Scope): When activation use 'range_abs_max' as the quantize + type, this pass will create some new parameters. The scope is used to + initialize these new parameters. + program_exe(fluid.Executor): program_exe is used to initialize new + parameters described above. weight_bits (int): quantization bit number for weights, the bias is not quantized. activation_bits (int): quantization bit number for activation. @@ -49,6 +60,7 @@ class QuantizationTransformPass(object): support 'abs_max'. The 'range_abs_max' usually is not used for weight, since weights are fixed once the model is well trained. window_size (int): the window size for 'range_abs_max' quantization. + Examples: .. code-block:: python # The original graph will be rewrite. @@ -88,31 +100,35 @@ class QuantizationTransformPass(object): self._quantizable_grad_ops = [ '%s_grad' % (op) for op in self._quantizable_ops ] - self._fake_quant_op_types = [ - 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' - ] - self._fake_dequant_op_types = ['fake_dequantize_max_abs'] self._is_test = None self._global_step = None def apply(self, graph): + """ + Quantize the graph for training process. According to weight and + activation quantization type, the graph will be added some fake + quantize operators and fake dequantize operators. + + Args: + graph(IrGraph): the applied graph. + """ assert isinstance(graph, IrGraph), 'graph must be the instance of IrGraph.' self._need_initialized.clear() self._is_test = graph.is_test() # marked the variable which has been dequantized. dequantized_vars = collections.OrderedDict() - params = [p.name() for p in graph.all_parameters()] + persistable_vars = [p.name() for p in graph.all_persistable_vars()] def _transform_forward(graph, op): for var_node in op.inputs: if var_node.name() in dequantized_vars: dequant_var_node = dequantized_vars[var_node.name()] else: - quant_bits = self._weight_bits if var_node.name() in params \ + quant_bits = self._weight_bits if var_node.name() in persistable_vars \ else self._activation_bits quant_type = self._weight_quantize_type if var_node.name() \ - in params else self._activation_quantize_type + in persistable_vars else self._activation_quantize_type quant_var_node, scale_var_node = self._insert_quant_op( graph, var_node, quant_bits, quant_type) dequant_var_node = self._insert_dequant_op( @@ -150,9 +166,14 @@ class QuantizationTransformPass(object): assert self._program_exe is not None, \ 'The program_exe cannot be set None when activation_quantize_type equals to range_abs_max.' init_program = Program() - for var_desc, initializer in self._need_initialized.iteritems(): - var = Variable(init_program.global_block()) - var._set_desc(var_desc) + for var_desc, initializer in six.iteritems(self._need_initialized): + var = init_program.global_block().create_var( + name=var_desc.name(), + shape=var_desc.shape(), + dtype=var_desc.dtype(), + type=var_desc.type(), + lod_level=var_desc.lod_level(), + persistable=var_desc.persistable()) initializer(var, init_program.global_block()) self._program_exe.run(program=init_program, scope=self._scope) @@ -161,7 +182,7 @@ class QuantizationTransformPass(object): def _create_global_step(self, graph): if self._weight_quantize_type == 'range_abs_max' or \ self._activation_quantize_type == 'range_abs_max': - counter_name = '@STEP_COUNTER@' + counter_name = cpt.to_text('@STEP_COUNTER@') for node in graph.all_vars(): if node.name() == counter_name: self._global_step = node @@ -175,9 +196,14 @@ class QuantizationTransformPass(object): Constant(value=0, force_cpu=True) global_step_out = graph.create_var_node_from_desc( global_step_in.var()) + # The attribute of `op_role` is needed by ParallelExecutor. increment_op = graph.create_op_node( op_type='increment', - attrs={'step': 1.0}, + attrs={ + 'step': 1.0, + 'op_role': + core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': global_step_in}, outputs={'Out': global_step_out}) graph.link_to(global_step_in, increment_op) @@ -212,7 +238,10 @@ class QuantizationTransformPass(object): var_dtype=var_node.var().dtype()) quant_op_node = graph.create_op_node( op_type='fake_quantize_abs_max', - attrs={'bit_length': quant_bits}, + attrs={ + 'bit_length': quant_bits, + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': var_node}, outputs={'Out': quant_var_node, 'OutScale': scale_var_node}) @@ -257,7 +286,8 @@ class QuantizationTransformPass(object): attrs = { 'window_size': self._window_size, 'bit_length': quant_bits, - 'is_test': self._is_test + 'is_test': self._is_test, + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward } quant_op_node = graph.create_op_node( op_type='fake_quantize_range_abs_max', @@ -290,7 +320,10 @@ class QuantizationTransformPass(object): max_range = (1 << (quant_bits - 1)) - 1 dequant_op_node = graph.create_op_node( op_type='fake_dequantize_max_abs', - attrs={'max_range': float(max_range)}, + attrs={ + 'max_range': float(max_range), + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': var_node, 'Scale': scale_var_node}, outputs={'Out': dequant_var_node}) @@ -316,3 +349,330 @@ class QuantizationTransformPass(object): Return the scale name of quantized variable for the input `var_name`. """ return "%s.scale" % (var_name) + + +class QuantizationFreezePass(object): + """ + The freeze pass is used to adjust the quantize operator order, for example: + 1) `activation -> quant -> dequant -> conv2d` will be freezed into + `activation -> quant -> conv2d -> dequant` + 2) `weight -> quant -> dequant -> conv2d` will be freezed into `weight -> conv2d`, + and weight will be sacled offline. + + Args: + scope(fluid.Scope): scope is used to get the weight tensor values. + place(fluid.CPUPlace|fluid.CUDAPlace): place is used to restore the weight tensors. + weight_bits (int): quantization bit number for weights. + activation_bits (int): quantization bit number for activation. + weight_quantize_type (str): quantization type for weights, support 'abs_max'. + The 'range_abs_max' usually is not used for weight, since weights are fixed once the + model is well trained. + """ + + def __init__(self, + scope, + place, + weight_bits=8, + activation_bits=8, + weight_quantize_type='abs_max'): + assert scope is not None, \ + 'The scope cannot be set None.' + assert place is not None, \ + 'The place cannot be set None.' + self._scope = scope + self._place = place + self._weight_bits = weight_bits + self._activation_bits = activation_bits + self._weight_quantize_type = weight_quantize_type + self._quantizable_ops = ['conv2d', 'depthwise_conv2d', 'mul'] + self._fake_quant_op_names = [ + 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' + ] + self._fake_dequant_op_names = ['fake_dequantize_max_abs'] + self._op_input_rename_map = collections.OrderedDict() + self._op_output_rename_map = collections.OrderedDict() + self._var_scale_map = collections.OrderedDict() + + def apply(self, graph): + """ + Adjust quantize/dequantize operators order for the inference process. + + Args: + graph(IrGraph): the applied graph. + """ + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._fake_quant_op_names: + input_arg_name = op_node.op().input('X')[0] + if input_arg_name in persistable_vars: + if self._weight_quantize_type == 'abs_max': + param = self._load_var(input_arg_name) + scale_v = np.max(np.abs(param)) + else: + scale_v = self._load_var(op_node.op().output('OutScale') + [0])[0] + self._var_scale_map[input_arg_name] = scale_v + else: + scale_v = graph.var_node(op_node.op().output('OutScale')[0]) + self._var_scale_map[input_arg_name] = scale_v + if input_arg_name in persistable_vars: + self._remove_fake_quant_and_dequant_op(graph, op_node) + # quantize weight and restore + param_v = self._load_var(input_arg_name) + quantized_param_v = self._quant(param_v, scale_v, + self._weight_bits) + self._restore_var(input_arg_name, quantized_param_v) + + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._fake_dequant_op_names: + self._remove_fake_quant_and_dequant_op(graph, op_node) + + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._quantizable_ops: + self._insert_post_dequant_op(graph, op_node) + + for op_node in ops: + # insert dequant_op after fc/conv, need to rename inputs of the followed ops + for var_node in op_node.inputs: + name = var_node.name() + if name in self._op_output_rename_map: + old_in = graph.var_node(name) + new_in = self._op_output_rename_map[name] + graph.update_input_link(old_in, new_in, op_node) + + # remove the unused var node in the graph + self._remove_unused_var_nodes(graph) + return graph + + def _remove_fake_quant_and_dequant_op(self, graph, op_node): + k = op_node.op().output('Out')[0] + v = op_node.op().input('X')[0] + if v not in self._op_input_rename_map: + self._op_input_rename_map[k] = v + else: + self._op_input_rename_map[k] = self._op_input_rename_map[v] + graph.safe_remove_nodes(op_node) + + def _insert_post_dequant_op(self, graph, op_node): + max_range = None + scale_var_node = None + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + for var_node in op_node.inputs: + name = var_node.name() + if name in self._op_input_rename_map: + old_in = graph.var_node(name) + new_in = graph.var_node(self._op_input_rename_map[name]) + new_in.clear_outputs() + graph.update_input_link(old_in, new_in, op_node) + original_var_name = self._original_var_name(name) + scale_v = self._var_scale_map[original_var_name] + if original_var_name in persistable_vars: + param_range = (1 << (self._weight_bits - 1)) - 1 + act_range = (1 << (self._activation_bits - 1)) - 1 + assert self._is_float( + scale_v), 'The scale of parameter %s is not a float.' % ( + original_var_name) + max_range = param_range * act_range / scale_v + else: + assert isinstance(scale_v, core.Node) + scale_var_node = self._var_scale_map[original_var_name] + + if len(op_node.outputs) != 1: + raise ValueError("Only support one output, but op %s has" + " more than one output." % (op_node.name())) + + output_var_node = op_node.outputs[0] + dequant_var_node = graph.create_var_node( + name=self._dequantized_var_name(output_var_node.name()), + var_type=output_var_node.var().type(), + shape=output_var_node.var().shape(), + var_dtype=output_var_node.var().dtype()) + dequant_op_node = graph.create_op_node( + op_type='fake_dequantize_max_abs', + attrs={ + 'max_range': float(max_range), + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, + inputs={'X': output_var_node, + 'Scale': scale_var_node}, + outputs={'Out': dequant_var_node}) + graph.link_to(output_var_node, dequant_op_node) + graph.link_to(scale_var_node, dequant_op_node) + graph.link_to(dequant_op_node, dequant_var_node) + self._op_output_rename_map[output_var_node.name()] = dequant_var_node + return dequant_var_node + + def _load_var(self, name): + return np.array(self._scope.find_var(name).get_tensor()) + + def _restore_var(self, name, array): + tensor = self._scope.find_var(name).get_tensor() + tensor.set(array, self._place) + + def _remove_unused_var_nodes(self, graph): + all_used_vars = set() + ops = graph.all_ops() + for op_node in ops: + for input_node in op_node.inputs: + all_used_vars.add(input_node) + for output_node in op_node.outputs: + all_used_vars.add(output_node) + + all_unused_vars = graph.all_vars() - all_used_vars + graph.safe_remove_nodes(all_unused_vars) + + def _original_var_name(self, var_name): + """ + Return the original variable name. + """ + if var_name.endswith('.quantized.dequantized'): + return var_name[:-len('.quantized.dequantized')] + if var_name.endswith('.quantized'): + return var_name[:-len('.quantized')] + if var_name.endswith('.dequantized'): + return var_name[:-len('.dequantized')] + if var_name.endswith('.scale'): + return var_name[:-len('.scale')] + else: + return var_name + + def _dequantized_var_name(self, var_name): + """ + Return dequantized variable name for the input `var_name`. + """ + return "%s.dequantized" % (var_name) + + def _is_float(self, v): + return isinstance(v, float) or isinstance(v, np.float32) \ + or isinstance(v, np.float64) + + def _quant(self, x, scale, num_bits): + return np.round(x / scale * ((1 << (num_bits - 1)) - 1)) + + +class ConvertToInt8Pass(object): + """ + Convert the weights into int8_t type. + + Args: + scope(fluid.Scope): scope is used to get the weight tensor values. + place(fluid.CPUPlace|fluid.CUDAPlace): place is used to restore the + 8bits weight tensors. + """ + + def __init__(self, scope, place): + assert scope is not None, \ + 'The scope cannot be set None.' + assert place is not None, \ + 'The place cannot be set None.' + self._scope = scope + self._place = place + self._quantizable_ops = ['conv2d', 'depthwise_conv2d', 'mul'] + + def apply(self, graph): + """ + Convert weights' tpye of the graph. After that, the data type of the + graph weigths is int8_t. + + Args: + graph(IrGraph): the applied graph. + """ + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + ops = graph.all_ops() + input_map = {} + for op_node in ops: + op_name = op_node.name() + if op_name in self._quantizable_ops: + for var_node in op_node.inputs: + name = var_node.name() + if name in persistable_vars: + if name not in input_map: + int8_var_node = self._convert_to_int8(graph, + var_node) + input_map[name] = int8_var_node + graph.update_input_link(var_node, input_map[name], + op_node) + + # remove the unused var node in the graph + self._remove_unused_var_nodes(graph) + return graph + + def _convert_to_int8(self, graph, var_node): + int8_var_node_name = var_node.name() + ".int8" + int8_var_node = graph.create_param_node( + name=cpt.to_text(int8_var_node_name), + var_type=var_node.var().type(), + shape=var_node.var().shape(), + var_dtype=core.VarDesc.VarType.INT8) + array = self._load_var(var_node.name()) + self._scope.var(int8_var_node_name) + self._store_var(int8_var_node_name, array, np.int8) + return int8_var_node + + def _load_var(self, name): + return np.array(self._scope.find_var(name).get_tensor()) + + def _store_var(self, name, array, dtype): + tensor = self._scope.find_var(name).get_tensor() + tensor.set(array.astype(dtype), self._place) + + def _remove_unused_var_nodes(self, graph): + all_used_vars = set() + ops = graph.all_ops() + for op_node in ops: + for input_node in op_node.inputs: + all_used_vars.add(input_node) + for output_node in op_node.outputs: + all_used_vars.add(output_node) + + all_unused_vars = graph.all_vars() - all_used_vars + graph.safe_remove_nodes(all_unused_vars) + + +class TransformForMobilePass(object): + """ + This pass is used to convert the freezed graph for paddle-mobile execution. + """ + + def __init__(self): + self._fake_quant_op_names = [ + 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' + ] + self._fake_dequant_op_names = ['fake_dequantize_max_abs'] + + def apply(self, graph): + """ + Because paddle-mobile use `quantize` an `dequantize` as the names of + quantize operator and dequantize operator, the `apply` function just + realize this logic. + + Args: + graph(IrGraph): the graph will be transformed. + """ + ops = graph.all_ops() + for op_node in ops: + name = op_node.name() + if name in self._fake_quant_op_names: + op_node.op().set_type('quantize') + quant_node = graph.create_op_node_from_desc(op_node.op()) + for input_node in op_node.inputs: + graph.link_to(input_node, quant_node) + for output_node in op_node.outputs: + graph.link_to(quant_node, output_node) + graph.safe_remove_nodes(op_node) + if name in self._fake_dequant_op_names: + op_node.op().set_type('dequantize') + dequant_node = graph.create_op_node_from_desc(op_node.op()) + for input_node in op_node.inputs: + graph.link_to(input_node, dequant_node) + for output_node in op_node.outputs: + graph.link_to(dequant_node, output_node) + graph.safe_remove_nodes(op_node) + + return graph diff --git a/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..79bec8c4ad34d682895250bc29b1fddb3a569bd4 --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt @@ -0,0 +1,6 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/contrib/slim/unitest/__init__.py b/python/paddle/fluid/contrib/slim/tests/__init__.py similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/__init__.py rename to python/paddle/fluid/contrib/slim/tests/__init__.py diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/config.yaml b/python/paddle/fluid/contrib/slim/tests/configs/config.yaml similarity index 88% rename from python/paddle/fluid/contrib/slim/unitest/configs/config.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/config.yaml index db488b96330210df15b02b19d90abd5c9101f844..d9b49029d3e34d487ad65fe0f7e54e2cee1d5838 100644 --- a/python/paddle/fluid/contrib/slim/unitest/configs/config.yaml +++ b/python/paddle/fluid/contrib/slim/tests/configs/config.yaml @@ -1,5 +1,5 @@ version: 1.0 -include: ["./unitest/configs/pruners.yaml", "./unitest/configs/pruners_0.yaml"] +include: ["./configs/pruners.yaml", "./configs/pruners_0.yaml"] pruners: pruner_1: class: 'RatioPruner' diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/pruners.yaml b/python/paddle/fluid/contrib/slim/tests/configs/pruners.yaml similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/configs/pruners.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/pruners.yaml diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/pruners_0.yaml b/python/paddle/fluid/contrib/slim/tests/configs/pruners_0.yaml similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/configs/pruners_0.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/pruners_0.yaml diff --git a/python/paddle/fluid/contrib/slim/unitest/test_factory.py b/python/paddle/fluid/contrib/slim/tests/test_factory.py similarity index 95% rename from python/paddle/fluid/contrib/slim/unitest/test_factory.py rename to python/paddle/fluid/contrib/slim/tests/test_factory.py index 07f28aac905d1a2813dbde6143235c7916fd9278..2fc72b6475e6bdd977dafb57696046a1100d0087 100644 --- a/python/paddle/fluid/contrib/slim/unitest/test_factory.py +++ b/python/paddle/fluid/contrib/slim/tests/test_factory.py @@ -18,7 +18,7 @@ import unittest class TestFactory(unittest.TestCase): def test_parse(self): - factory = ConfigFactory('./unitest/configs/config.yaml') + factory = ConfigFactory('./configs/config.yaml') pruner = factory.instance('pruner_1') self.assertEquals(pruner.ratios['conv1_1.w'], 0.3) diff --git a/python/paddle/fluid/contrib/slim/tests/test_graph.py b/python/paddle/fluid/contrib/slim/tests/test_graph.py new file mode 100644 index 0000000000000000000000000000000000000000..75e0c95b5c3cc06d66eab9de0b85e5d7ed110837 --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/test_graph.py @@ -0,0 +1,80 @@ +# 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. + +from __future__ import print_function +import unittest +import paddle.fluid as fluid +import six +from paddle.fluid.framework import IrGraph +from paddle.fluid import core + + +def residual_block(num): + def conv_bn_layer(input, + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) + + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) + short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) + hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') + fc = fluid.layers.fc(input=hidden, size=10) + loss = fluid.layers.cross_entropy(input=fc, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestGraph(unittest.TestCase): + def test_graph_functions(self): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = residual_block(2) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + graph = IrGraph(core.Graph(main.desc), for_test=False) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('conv2d') > -1: + marked_nodes.add(op) + graph.draw('.', 'residual', marked_nodes) + self.assertFalse(graph.has_circle()) + self.assertEqual(graph.graph_num(), 1) + nodes = graph.topology_sort() + self.assertEqual(len(nodes), len(graph.all_ops())) + nodes_map = graph.build_adjacency_list() + self.assertEqual(len(nodes_map), len(graph.all_ops())) + nodes_num = len(graph.all_nodes()) + graph.safe_remove_nodes(marked_nodes) + self.assertEqual(len(graph.all_nodes()), nodes_num - len(marked_nodes)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py b/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..2f291132f3049af21420f863972792c1a862b9ad --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py @@ -0,0 +1,372 @@ +# 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 unittest +import random +import numpy as np +import paddle.fluid as fluid +import six +import paddle +from paddle.fluid.framework import IrGraph +from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass +from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass +from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass +from paddle.fluid.contrib.slim.quantization import TransformForMobilePass +from paddle.fluid import core + + +def linear_fc(num): + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + hidden = fluid.layers.fc(hidden, size=128, act='relu') + loss = fluid.layers.cross_entropy(input=hidden, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def residual_block(num): + def conv_bn_layer(input, + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) + + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) + short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) + hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') + fc = fluid.layers.fc(input=hidden, size=10) + loss = fluid.layers.cross_entropy(input=fc, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def conv_net(img, label): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + return avg_loss + + +class TestQuantizationTransformPass(unittest.TestCase): + def setUp(self): + self.quantizable_op_and_inputs = { + 'conv2d': ['Input', 'Filter'], + 'depthwise_conv2d': ['Input', 'Filter'], + 'mul': ['X', 'Y'] + } + self.quantizable_grad_op_inputs = { + 'conv2d_grad': ['Input', 'Filter'], + 'depthwise_conv2d_grad': ['Input', 'Filter'], + 'mul_grad': ['X', 'Y'] + } + + def check_program(self, transform_pass, program): + quantized_ops = set() + for block in program.blocks: + for op in block.ops: + # check forward + if op.type in self.quantizable_op_and_inputs: + for arg_name in op.input_arg_names: + self.assertTrue( + arg_name.endswith('.quantized.dequantized')) + quantized_ops.add(arg_name) + + for op in block.ops: + # check backward + if op.type in self.quantizable_grad_op_inputs: + for pname in self.quantizable_grad_op_inputs[op.type]: + arg_name = op.input(pname)[0] + self.assertTrue( + arg_name.endswith('.quantized.dequantized')) + self.assertTrue(arg_name in quantized_ops) + + def linear_fc_quant(self, quant_type): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = linear_fc(3) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + exe = fluid.Executor(fluid.CPUPlace()) + graph = IrGraph(core.Graph(main.desc), for_test=False) + transform_pass = QuantizationTransformPass( + scope=fluid.global_scope(), + program_exe=exe, + activation_quantize_type=quant_type) + transform_pass.apply(graph) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes) + program = graph.to_program() + self.check_program(transform_pass, program) + val_graph = IrGraph(core.Graph(program.desc), for_test=False) + val_marked_nodes = set() + for op in val_graph.all_ops(): + if op.name().find('quantize') > -1: + val_marked_nodes.add(op) + val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes) + + def test_linear_fc_quant_abs_max(self): + self.act_quant_op_type = 'fake_quantize_abs_max' + self.linear_fc_quant('abs_max') + + def test_linear_fc_quant_range_abs_max(self): + self.act_quant_op_type = 'fake_quantize_range_abs_max' + self.linear_fc_quant('range_abs_max') + + def residual_block_quant(self, quant_type): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = residual_block(2) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + exe = fluid.Executor(fluid.CPUPlace()) + graph = IrGraph(core.Graph(main.desc), for_test=False) + transform_pass = QuantizationTransformPass( + scope=fluid.global_scope(), + program_exe=exe, + activation_quantize_type=quant_type) + transform_pass.apply(graph) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes) + program = graph.to_program() + self.check_program(transform_pass, program) + val_graph = IrGraph(core.Graph(program.desc), for_test=False) + val_marked_nodes = set() + for op in val_graph.all_ops(): + if op.name().find('quantize') > -1: + val_marked_nodes.add(op) + val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes) + + def test_residual_block_abs_max(self): + self.act_quant_op_type = 'fake_quantize_abs_max' + self.residual_block_quant('abs_max') + + def test_residual_block_range_abs_max(self): + self.act_quant_op_type = 'fake_quantize_range_abs_max' + self.residual_block_quant('range_abs_max') + + +class TestQuantizationFreezePass(unittest.TestCase): + def freeze_graph(self, use_cuda, seed, quant_type): + def build_program(main, startup, is_test): + main.random_seed = seed + startup.random_seed = seed + with fluid.unique_name.guard(): + with fluid.program_guard(main, startup): + img = fluid.layers.data( + name='image', shape=[1, 28, 28], dtype='float32') + label = fluid.layers.data( + name='label', shape=[1], dtype='int64') + loss = conv_net(img, label) + if not is_test: + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + return [img, label], loss + + random.seed(0) + np.random.seed(0) + + main = fluid.Program() + startup = fluid.Program() + test_program = fluid.Program() + feeds, loss = build_program(main, startup, False) + build_program(test_program, startup, True) + test_program = test_program.clone(for_test=True) + main_graph = IrGraph(core.Graph(main.desc), for_test=False) + test_graph = IrGraph(core.Graph(test_program.desc), for_test=True) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + scope = fluid.Scope() + with fluid.scope_guard(scope): + exe.run(startup) + transform_pass = QuantizationTransformPass( + scope=scope, program_exe=exe, activation_quantize_type=quant_type) + transform_pass.apply(main_graph) + transform_pass.apply(test_graph) + dev_name = '_gpu_' if use_cuda else '_cpu_' + marked_nodes = set() + for op in main_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + main_graph.draw('.', 'main' + dev_name + quant_type, marked_nodes) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test' + dev_name + quant_type, marked_nodes) + + quantized_main_program = main_graph.to_program() + quantized_test_program = test_graph.to_program() + iters = 5 + batch_size = 8 + + #train_exe = fluid.ParallelExecutor( + # main_program=quantized_main_program, + # use_cuda=bool(use_cuda), + # loss_name=loss.name, + # scope=scope) + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + feeder = fluid.DataFeeder(feed_list=feeds, place=place) + with fluid.scope_guard(scope): + for _ in range(iters): + data = next(train_reader()) + loss_v = exe.run(program=quantized_main_program, + feed=feeder.feed(data), + fetch_list=[loss]) + #loss_v = train_exe.run(feed=feeder.feed(data), + # fetch_list=[loss.name]) + #print('{}: {}'.format('loss' + dev_name + quant_type, loss_v)) + + test_data = next(test_reader()) + with fluid.program_guard(quantized_test_program): + w_var = fluid.framework._get_var('conv2d_1.w_0.quantized', + quantized_test_program) + # Testing + with fluid.scope_guard(scope): + test_loss1, w_quant = exe.run(program=quantized_test_program, + feed=feeder.feed(test_data), + fetch_list=[loss, w_var]) + + # Freeze graph for inference, but the weight of fc/conv is still float type. + freeze_pass = QuantizationFreezePass(scope=scope, place=place) + freeze_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_freeze' + dev_name + quant_type, + marked_nodes) + + server_program = test_graph.to_program() + with fluid.scope_guard(scope): + test_loss2, = exe.run(program=server_program, + feed=feeder.feed(test_data), + fetch_list=[loss]) + self.assertAlmostEqual(test_loss1, test_loss2, delta=5e-3) + #print('{}: {}'.format('test_loss1' + dev_name + quant_type, test_loss1)) + #print('{}: {}'.format('test_loss2' + dev_name + quant_type, test_loss2)) + w_freeze = np.array(scope.find_var('conv2d_1.w_0').get_tensor()) + # Maybe failed, this is due to the calculation precision + # self.assertAlmostEqual(np.sum(w_freeze), np.sum(w_quant)) + #print('{}: {}'.format('w_freeze' + dev_name + quant_type, + # np.sum(w_freeze))) + #print('{}: {}'.format('w_quant' + dev_name + quant_type, + # np.sum(w_quant))) + + # Convert parameter to 8-bit. + convert_int8_pass = ConvertToInt8Pass(scope=scope, place=place) + convert_int8_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_int8' + dev_name + quant_type, marked_nodes) + server_program_int8 = test_graph.to_program() + # Save the 8-bit parameter and model file. + with fluid.scope_guard(scope): + fluid.io.save_inference_model('server_int8' + dev_name + quant_type, + ['image', 'label'], [loss], exe, + server_program_int8) + # Test whether the 8-bit parameter and model file can be loaded successfully. + [infer, feed, fetch] = fluid.io.load_inference_model( + 'server_int8' + dev_name + quant_type, exe) + # Check the loaded 8-bit weight. + w_8bit = np.array(scope.find_var('conv2d_1.w_0.int8').get_tensor()) + self.assertEqual(w_8bit.dtype, np.int8) + self.assertEqual(np.sum(w_8bit), np.sum(w_freeze)) + #print('{}: {}'.format('w_8bit' + dev_name + quant_type, np.sum(w_8bit))) + #print('{}: {}'.format('w_freeze' + dev_name + quant_type, + # np.sum(w_freeze))) + + mobile_pass = TransformForMobilePass() + mobile_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_mobile' + dev_name + quant_type, + marked_nodes) + + mobile_program = test_graph.to_program() + with fluid.scope_guard(scope): + fluid.io.save_inference_model('mobile_int8' + dev_name + quant_type, + ['image', 'label'], [loss], exe, + mobile_program) + + def test_freeze_graph_cuda_dynamic(self): + if fluid.core.is_compiled_with_cuda(): + with fluid.unique_name.guard(): + self.freeze_graph(True, seed=1, quant_type='abs_max') + + def test_freeze_graph_cpu_dynamic(self): + with fluid.unique_name.guard(): + self.freeze_graph(False, seed=2, quant_type='abs_max') + + def test_freeze_graph_cuda_static(self): + if fluid.core.is_compiled_with_cuda(): + with fluid.unique_name.guard(): + self.freeze_graph(True, seed=1, quant_type='range_abs_max') + + def test_freeze_graph_cpu_static(self): + with fluid.unique_name.guard(): + self.freeze_graph(False, seed=2, quant_type='range_abs_max') + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py b/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py deleted file mode 100644 index 1bd4b95d6b90b7f16d507061190f0b463f6c4cc5..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py +++ /dev/null @@ -1,175 +0,0 @@ -# copyright (c) 2018 paddlepaddle authors. all rights reserved. -# -# licensed under the apache license, version 2.0 (the "license"); -# you may not use this file except in compliance with the license. -# you may obtain a copy of the license at -# -# http://www.apache.org/licenses/license-2.0 -# -# unless required by applicable law or agreed to in writing, software -# distributed under the license is distributed on an "as is" basis, -# without warranties or conditions of any kind, either express or implied. -# see the license for the specific language governing permissions and -# limitations under the license. - -import unittest -import random -import numpy as np -import paddle.fluid as fluid -import six -from paddle.fluid.framework import Program -from paddle.fluid.framework import IrGraph -from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass -from paddle.fluid import core - - -def linear_fc(num): - data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - hidden = data - for _ in six.moves.xrange(num): - hidden = fluid.layers.fc(hidden, size=128, act='relu') - loss = fluid.layers.cross_entropy(input=hidden, label=label) - loss = fluid.layers.mean(loss) - return loss - - -def residual_block(num): - def conv_bn_layer(input, - ch_out, - filter_size, - stride, - padding, - act='relu', - bias_attr=False): - tmp = fluid.layers.conv2d( - input=input, - filter_size=filter_size, - num_filters=ch_out, - stride=stride, - padding=padding, - act=None, - bias_attr=bias_attr) - return fluid.layers.batch_norm(input=tmp, act=act) - - data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - hidden = data - for _ in six.moves.xrange(num): - conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) - short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) - hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') - fc = fluid.layers.fc(input=hidden, size=10) - loss = fluid.layers.cross_entropy(input=fc, label=label) - loss = fluid.layers.mean(loss) - return loss - - -class TestQuantizationTransformPass(unittest.TestCase): - def setUp(self): - self.quantizable_op_and_inputs = { - 'conv2d': ['Input', 'Filter'], - 'depthwise_conv2d': ['Input', 'Filter'], - 'mul': ['X', 'Y'] - } - self.quantizable_grad_op_inputs = { - 'conv2d_grad': ['Input', 'Filter'], - 'depthwise_conv2d_grad': ['Input', 'Filter'], - 'mul_grad': ['X', 'Y'] - } - - def check_program(self, transform_pass, program): - quantized_ops = set() - for block in program.blocks: - for op in block.ops: - # check forward - if op.type in self.quantizable_op_and_inputs: - for arg_name in op.input_arg_names: - self.assertTrue( - arg_name.endswith('.quantized.dequantized')) - quantized_ops.add(arg_name) - - for op in block.ops: - # check backward - if op.type in self.quantizable_grad_op_inputs: - for pname in self.quantizable_grad_op_inputs[op.type]: - arg_name = op.input(pname)[0] - self.assertTrue( - arg_name.endswith('.quantized.dequantized')) - self.assertTrue(arg_name in quantized_ops) - - def linear_fc_quant(self, quant_type): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - loss = linear_fc(3) - opt = fluid.optimizer.Adam(learning_rate=0.001) - opt.minimize(loss) - exe = fluid.Executor(fluid.CPUPlace()) - graph = IrGraph(core.Graph(main.desc), for_test=False) - transform_pass = QuantizationTransformPass( - scope=fluid.global_scope(), - program_exe=exe, - activation_quantize_type=quant_type) - transform_pass.apply(graph) - marked_nodes = set() - for op in graph.all_ops(): - if op.name().find('quantize') > -1: - marked_nodes.add(op) - graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes) - program = graph.to_program() - self.check_program(transform_pass, program) - val_graph = IrGraph(core.Graph(program.desc), for_test=False) - val_marked_nodes = set() - for op in val_graph.all_ops(): - if op.name().find('quantize') > -1: - val_marked_nodes.add(op) - val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes) - - def test_linear_fc_quant_abs_max(self): - self.act_quant_op_type = 'fake_quantize_abs_max' - self.linear_fc_quant('abs_max') - - def test_linear_fc_quant_range_abs_max(self): - self.act_quant_op_type = 'fake_quantize_range_abs_max' - self.linear_fc_quant('range_abs_max') - - def residual_block_quant(self, quant_type): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - loss = residual_block(2) - opt = fluid.optimizer.Adam(learning_rate=0.001) - opt.minimize(loss) - exe = fluid.Executor(fluid.CPUPlace()) - graph = IrGraph(core.Graph(main.desc), for_test=False) - transform_pass = QuantizationTransformPass( - scope=fluid.global_scope(), - program_exe=exe, - activation_quantize_type=quant_type) - transform_pass.apply(graph) - marked_nodes = set() - for op in graph.all_ops(): - if op.name().find('quantize') > -1: - marked_nodes.add(op) - graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes) - program = graph.to_program() - self.check_program(transform_pass, program) - val_graph = IrGraph(core.Graph(program.desc), for_test=False) - val_marked_nodes = set() - for op in val_graph.all_ops(): - if op.name().find('quantize') > -1: - val_marked_nodes.add(op) - val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes) - - def test_residual_block_abs_max(self): - self.act_quant_op_type = 'fake_quantize_abs_max' - self.residual_block_quant('abs_max') - - def test_residual_block_range_abs_max(self): - self.act_quant_op_type = 'fake_quantize_range_abs_max' - self.residual_block_quant('range_abs_max') - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/contrib/tests/CMakeLists.txt b/python/paddle/fluid/contrib/tests/CMakeLists.txt index 81aee1233d1db756686d1a934b94672dc5c770fe..a2c59416467e5dbe66f058666633807eb0e45047 100644 --- a/python/paddle/fluid/contrib/tests/CMakeLists.txt +++ b/python/paddle/fluid/contrib/tests/CMakeLists.txt @@ -6,5 +6,9 @@ if(APPLE OR WIN32 OR NOT WITH_MKL) endif() foreach(src ${TEST_OPS}) - py_test(${src} SRCS ${src}.py) + if(src MATCHES "test_calibration") + py_test(${src} SRCS ${src}.py ENVS FLAGS_use_mkldnn=true) + else() + py_test(${src} SRCS ${src}.py) + endif() endforeach() diff --git a/python/paddle/fluid/contrib/tests/test_calibration.py b/python/paddle/fluid/contrib/tests/test_calibration.py index 424ea245a0f2dff0d437ace386f2e4e0fa6b517d..b9f938bebed71dc9611df8d743a066858ea38bca 100644 --- a/python/paddle/fluid/contrib/tests/test_calibration.py +++ b/python/paddle/fluid/contrib/tests/test_calibration.py @@ -199,7 +199,6 @@ class TestCalibrationForResnet50(unittest.TestCase): def run_program(self, model_path, generate_int8=False, algo='direct'): image_shape = [3, 224, 224] - os.environ['FLAGS_use_mkldnn'] = 'True' fluid.memory_optimize(fluid.default_main_program()) @@ -241,9 +240,6 @@ class TestCalibrationForResnet50(unittest.TestCase): label = label.reshape([-1, 1]) running_program = calibrator.sampling_program.clone( ) if generate_int8 else infer_program.clone() - for op in running_program.current_block().ops: - if op.has_attr("use_mkldnn"): - op._set_attr("use_mkldnn", True) t1 = time.time() _, acc1, _ = exe.run( diff --git a/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py b/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py index 86fa84ad4bd7a55fb27f4e43128f0bfda6dfe6db..77fdf0087b93c3ad44a2492de68f8f57ce243ef3 100644 --- a/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py +++ b/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py @@ -204,9 +204,11 @@ class TestQuantizeTranspiler(unittest.TestCase): build_program(test_program, startup, True) test_program = test_program.clone(for_test=True) - quant_transpiler = QuantizeTranspiler() - quant_transpiler.training_transpile(main) - quant_transpiler.training_transpile(test_program) + quant_type = 'range_abs_max' # 'range_abs_max' or 'abs_max' + quant_transpiler = QuantizeTranspiler( + activation_quantize_type=quant_type) + quant_transpiler.training_transpile(main, startup) + quant_transpiler.training_transpile(test_program, startup) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 832c97c7deb49b4e118e15989ab7a34da6ce57a0..ef304b11106628f8541b348fb263274a0c4b31e9 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -16,6 +16,8 @@ from __future__ import print_function import collections from collections import defaultdict +from collections import Iterable +import contextlib from .wrapped_decorator import signature_safe_contextmanager import os import re @@ -1529,12 +1531,16 @@ class Block(object): class IrGraph(object): """ - IrGraph uses core.Graph as the delegation to accomplish the manipulation. + Python IrGraph. Beneath it is a core.Graph, which is used for + create a c++ Ir Pass Graph. An IrGraph is just a graph view of + a Program. In an IrGraph, both Variables and Operators are graph + nodes. """ def __init__(self, graph, for_test=False): """ - Construct the IrGraph using core.Graph. + Construct an IrGraph using core.Graph. + Args: graph(core.Graph): C++ Graph. for_test(bool): True for the test graph and false for the train graph. @@ -1545,23 +1551,81 @@ class IrGraph(object): self._for_test = for_test def is_test(self): + """ + If the graph is used for testing, the function returns true. Otherwise, returns false. + """ return self._for_test - def all_parameters(self): - param_nodes = set() - for node in self.graph.nodes(): - if node.is_var() and node.var() is not None and node.var( - ).persistable(): - param_nodes.add(node) - return param_nodes + def all_nodes(self): + """ + Return all nodes included in the graph as a set. + """ + return {node for node in self.graph.nodes()} def all_vars(self): + """ + Return all variable nodes included in the graph as a set. + """ return {node for node in self.graph.nodes() if node.is_var()} + def all_persistable_vars(self): + """ + Return all persistable variable nodes included in the graph as a set. + """ + persistable_nodes = set() + for node in self.graph.nodes(): + if node.is_var() and node.var() is not None and node.var( + ).persistable(): + persistable_nodes.add(node) + return persistable_nodes + def all_ops(self): + """ + Return all operator nodes included in the graph as a set. + """ return {node for node in self.graph.nodes() if node.is_op()} + def var_node(self, name): + """ + Get a variable node by name from the graph. + + Args: + name(str): the name of the variable node. + + Raises: + ValueError: The If input's type is not str, or this graph + doesn't have a variable with the giving name. + + Returns: + core.Node: the variable node with the giving name. + """ + if not isinstance(name, six.string_types): + raise TypeError( + "var require string as parameter, but get %s instead." % + (type(name))) + target_var_node = None + var_nodes = self.all_vars() + for var_node in var_nodes: + if var_node.name() == name: + target_var_node = var_node + if target_var_node is None: + raise ValueError("var_node %s not in this graph" % name) + return target_var_node + def create_param_node(self, name, var_type, shape, var_dtype): + """ + Create a persistable variable node in the graph. In IrGraph, + it can not distinguish between persistable variables and parameters. + + Args: + name(str): the name of the persistable variable node. + vart_type(core.VarDesc.VarType): the type of the persistable variable node. + shape(list): the shape of the persistable variable node. + var_dtype(core.VarDesc.VarType): the data type of the persistable variable node. + + Returns: + core.Node: the created persistable variable node. + """ var_desc = core.VarDesc(name) var_desc.set_type(var_type) var_desc.set_shape(shape) @@ -1570,6 +1634,20 @@ class IrGraph(object): return self.graph.create_var_node(var_desc) def create_var_node(self, name, var_type, shape, var_dtype): + """ + Create a variable node in the graph. The created variable node is + not persistable. + + Args: + name(str): the name of the variable node. + vart_type(core.VarDesc.VarType): the type of the variable node. + shape(list): the shape of the variable node. + var_dtype(core.VarDesc.VarType): the data type of the variable node. + + Returns: + core.Node: the created variable node. + """ + var_desc = core.VarDesc(name) var_desc.set_type(var_type) var_desc.set_shape(shape) @@ -1577,19 +1655,41 @@ class IrGraph(object): return self.graph.create_var_node(var_desc) def create_var_node_from_desc(self, var_desc): + """ + Create a variable node by using an existing VarDesc in the graph. + Depend on the giving VarDesc, the created variable node may be persistable. + + Args: + var_desc(core.VarDesc): the giving variable description. + + Returns: + core.Node: the created variable node. + """ return self.graph.create_var_node(var_desc) def create_op_node(self, op_type, attrs, inputs, outputs): + """ + Create a operator node in the graph. + + Args: + op_type(str): the type of the operator node. + attrs(dict): the attributes of the operator node. + inputs(dict): the inputs of the operator node. + outputs(dict): the outpus of the operator node. + + Returns: + core.Node: the created operator node. + """ op_desc = core.OpDesc() op_desc.set_type(op_type) - for attr, value in attrs.iteritems(): + for attr, value in six.iteritems(attrs): self._update_desc_attr(op_desc, attr, value) - for input_name, var_nodes in inputs.iteritems(): + for input_name, var_nodes in six.iteritems(inputs): if not isinstance(var_nodes, list): var_nodes = [var_nodes] op_desc.set_input(input_name, [var_node.name() for var_node in var_nodes]) - for output_name, var_nodes in outputs.iteritems(): + for output_name, var_nodes in six.iteritems(outputs): if not isinstance(var_nodes, list): var_nodes = [var_nodes] op_desc.set_output(output_name, @@ -1597,11 +1697,29 @@ class IrGraph(object): return self.graph.create_op_node(op_desc) def create_op_node_from_desc(self, op_desc): + """ + Create a operator node by using an existing OpDesc in the graph. + + Args: + op_desc(core.VarDesc): the giving operator description. + + Returns: + core.Node: the created operator node. + """ return self.graph.create_op_node(op_desc) def update_input_link(self, old_input_node, new_input_node, op_node): - assert old_input_node in self.graph.nodes() and new_input_node in self.graph.nodes() and \ - op_node in self.graph.nodes(), 'Th three arguments must be in the graph nodes.' + """ + Update the input's link of a operator node. + + Args: + old_input_node(core.Node): the old input node of the giving op_node. + new_input_node(core.Node): the new input node of the giving op_node. + op_node(core.Node): the operator node that is needed to update input's link. + """ + assert old_input_node in self.graph.nodes() and new_input_node in \ + self.graph.nodes() and op_node in self.graph.nodes(), \ + 'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.' old_input_node.outputs_remove(op_node) op_node.inputs_remove(old_input_node) new_input_node.outputs_append(op_node) @@ -1609,17 +1727,85 @@ class IrGraph(object): op_node.op()._rename_input(old_input_node.name(), new_input_node.name()) def link_to(self, node_in, node_out): + """ + Connect two nodes. + + Args: + node_in(core.Node): the input node. + node_out(core.Node): the output node. + """ assert node_in in self.graph.nodes() and node_out in self.graph.nodes(), \ - 'Th two arguments must be in the graph nodes.' + 'The two arguments(node_in&node_out) must be in the graph nodes.' node_in.outputs_append(node_out) node_out.inputs_append(node_in) def safe_remove_nodes(self, remove_nodes): + """ + Remove nodes safely since links connected to these removed nodes are + also removed. + + Args: + remove_nodes(set): the nodes prepared to be removed. + """ if not isinstance(remove_nodes, set): - remove_nodes = set(remove_nodes) + if isinstance(remove_nodes, Iterable): + remove_nodes = set(remove_nodes) + else: + remove_nodes = {remove_nodes} core.graph_safe_remove_nodes(self.graph, remove_nodes) - def draw(self, save_path, name, marked_nodes=None): + def has_circle(self): + """ + Check if the graph has a circle. + + Returns: + bool: True if the graph has a circle else False. + """ + return core.has_circle(self.graph) + + def graph_num(self): + """ + Count the number of unconnected graphs in this graph. + + Returns: + int: the number of unconnected graphs. + """ + return core.graph_num(self.graph) + + def topology_sort(self): + """ + Perform the topology sort operation on the graph. + + Notes: the `graph` cannot contain a circle. + + Returns: + set(core.Node): nodes in topology order. + """ + return core.topology_sort(self.graph) + + def build_adjacency_list(self): + """ + Build an adjacency list of operations for the `graph`. + + Returns: + dict{core.Node: set(core.Node)}: the adjacency list. + """ + return core.build_adjacency_list(self.graph) + + def draw(self, save_path, name, marked_nodes=None, remove_ctr_var=True): + """ + Draw the graph. If `dot` command is installed, the drawn graph + will be saved as pdf file type, otherwise dot file type is used. + + Args: + save_path(str): the save path of drawn graph. + name(str): the name of drawn graph. + marked_nodes(set(core.Node)): nodes that are needed to be marked. + Default value is None. + remove_ctr_var(bool): If it is set True, all control variable nodes + in the graph will be removed. Default value is True. + """ + def _convert_to_pdf(dot_file_path): pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf' exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \ @@ -1629,15 +1815,17 @@ class IrGraph(object): print('The {} is saved as the dot filetype.'.format( dot_file_path)) - remove_ctr_vars = set() + if remove_ctr_var: + remove_ctr_vars = set() + for node in self.graph.nodes(): + if node.is_ctrl_var(): + remove_ctr_vars.add(node) + self.safe_remove_nodes(remove_ctr_vars) ops_num = 0 for node in self.graph.nodes(): - if node.is_ctrl_var(): - remove_ctr_vars.add(node) - elif node.is_op(): + if node.is_op(): ops_num += 1 print('Total ops num = {}.'.format(ops_num)) - self.safe_remove_nodes(remove_ctr_vars) if marked_nodes is not None: if not isinstance(marked_nodes, set): marked_nodes = set(marked_nodes) @@ -1652,10 +1840,20 @@ class IrGraph(object): _convert_to_pdf(viz_dot_path) def to_program(self): + """ + Convert the graph into a Program. + + Notes: When the graph includes backward operator nodes, the + conversion process may be failed. Usually, this function is + only used to convert a test graph. + + Returns: + Program: a program converted from the graph. + """ convert_pass = core.get_pass('graph_to_program_pass') - convert_pass.set('program', Program().desc) + desc = core.ProgramDesc() + convert_pass.set_not_owned('program', desc) convert_pass.apply(self.graph) - desc = convert_pass.get_program('program') program = Program._construct_from_desc(desc) return program diff --git a/python/paddle/fluid/imperative/layers.py b/python/paddle/fluid/imperative/layers.py index 71ff95bdea36967c1fa6b5c94cc7ca305e7a544a..59fe6bbf74b80c2260c5b4881fee8807482c9c68 100644 --- a/python/paddle/fluid/imperative/layers.py +++ b/python/paddle/fluid/imperative/layers.py @@ -12,6 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. +import collections import contextlib import sys import numpy as np @@ -30,31 +31,45 @@ class Layer(core.Layer): def __init__(self, dtype=core.VarDesc.VarType.FP32, name=None): self._built = False self._dtype = dtype + self._parameters = collections.OrderedDict() + self._sub_layers = collections.OrderedDict() + + def parameters(self, include_sublayers=True): + """Returns a list of Parameters from current and sub-layers. + + Args: + include_sublayers: If true, also include the parameters from + sublayers. + + Returns a list of Parameters. + """ + ret = [p for p in self._parameters.values()] + if include_sublayers: + for l in self._sub_layers.values(): + for p in l.parameters(include_sublayers): + ret.append(p) + return ret + + def sublayers(self, include_sublayers=True): + """Returns a list of sub layers. - def parameters(self): - params = [] - for key in self.__dict__.keys(): - value = self.__dict__[key] - if isinstance(value, framework.Parameter): - params.append(value) - elif isinstance(value, core.Layer): - params.extend(value.parameters()) - elif isinstance(value, collections.Container): - if len(value) == 0: - continue - if isinstance(value[0], framework.Parameter): - params.extend(value) - elif isinstance(value[0], core.Layer): - for v in value: - params.extend(v.parameters()) - - return params + Args: + include_sublayers: If true, also include the layers from sublayers. + + Returns a list of sub layers. + """ + ret = [l for l in self._sub_layers.values()] + if include_sublayers: + for l in self._sub_layers.values(): + for sub_l in l.sublayers(include_sublayers): + ret.append(sub_l) + return ret def clear_gradients(self): for p in self.parameters(): p._clear_gradient() - def _build_once(self, inputs): + def _build_once(self, *args): pass def __call__(self, *inputs): @@ -71,6 +86,66 @@ class Layer(core.Layer): def backward(self, *inputs): raise ValueError("Layer shouldn't implement backward") + def add_sublayer(self, name, sublayer): + """Adds a sub Layer instance. + + Added sublayer can be access like self.name. + + Args: + name: name of this sublayer. + sublayer: an instance of Layer. + Returns: + the sublayer passed in. + """ + assert isinstance(sublayer, core.Layer) + self._sub_layers[name] = sublayer + return sublayer + + def add_parameter(self, name, parameter): + """Adds a Parameter instance. + + Added parameter can be access like self.name. + + Args: + name: name of this sublayer. + parameter: an instance of Parameter. + Returns: + the parameter passed in. + """ + assert isinstance(parameter, framework.Parameter) + self._parameters[name] = parameter + return parameter + + def __getattr__(self, name): + if name in self._parameters: + return self._parameters[name] + elif name in self._sub_layers: + return self._sub_layers[name] + + def __setattr__(self, name, value): + if isinstance(value, framework.Parameter): + params = self.__dict__.get('_parameters', None) + if params is None: + raise ValueError( + "super(YourLayer, self).__init__() should be called first") + params[name] = value + elif isinstance(value, core.Layer): + layers = self.__dict__.get('_sub_layers', None) + if layers is None: + raise ValueError( + "super(YourLayer, self).__init__() should be called first") + layers[name] = value + else: + object.__setattr__(self, name, value) + + def __delattr__(self, name): + if name in self._parameters: + del self._parameters[name] + elif name in self._sub_layers: + del self._sub_layers[name] + else: + object.__delattr__(self, name) + class PyLayer(core.PyLayer): """Layers composed of user-defined python codes.""" diff --git a/python/paddle/fluid/imperative/nn.py b/python/paddle/fluid/imperative/nn.py index 6c5961cc63d1c140e0a6f33aac054acdbbe8e8e0..c86a373ae4a92053538c93386003f9014c32841f 100644 --- a/python/paddle/fluid/imperative/nn.py +++ b/python/paddle/fluid/imperative/nn.py @@ -225,9 +225,6 @@ class FC(layers.Layer): act=act, name=name) - def parameters(self): - return [self._w, self._b] - def _build_once(self, input): input_shape = input.shape param_shape = [ @@ -478,9 +475,6 @@ class Embedding(layers.Layer): dtype=self._dtype, is_bias=False) - def parameters(self): - return [self._w] - def forward(self, input): out = self._helper.create_variable_for_type_inference(self._dtype) self._helper.append_op( diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 69885fd17af9e568adfcc50779691f20d80e7bdf..d72921dc00232e35cacf453e8266471a943f5062 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -3236,7 +3236,7 @@ def group_norm(input, # create output mean_out = helper.create_variable(dtype=dtype, stop_gradient=True) variance_out = helper.create_variable(dtype=dtype, stop_gradient=True) - group_norm_out = helper.create_variable(dtype) + group_norm_out = helper.create_variable(dtype=dtype) helper.append_op( type="group_norm", @@ -5936,13 +5936,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None): than :attr:`shape`. act (str): The non-linear activation to be applied to the reshaped tensor variable. - inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple - operators. If this flag is set :attr:`True`, reuse input - :attr:`x` to reshape, which will change the shape of - tensor variable :attr:`x` and might cause errors when - :attr:`x` is used in multiple operators. If :attr:`False`, - preserve the shape :attr:`x` and create a new output tensor - variable whose data is copied from input x but reshaped. + inplace(bool): If ``inplace`` is `True`, the input and output of ``layers.reshape`` + are the same variable, otherwise, the input and output of + ``layers.reshape`` are different variables. Note that if :attr:`x` + is more than one layer's input, ``inplace`` must be :attr:`False`. name (str): The name of this layer. It is optional. Returns: @@ -8335,6 +8332,46 @@ def stack(x, axis=0): If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`. If :code:`axis` is None, it would be replaced with 0. + For Example: + + .. code-block:: text + + Case 1: + Input: + x[0].data = [ [1.0 , 2.0 ] ] + x[0].dims = [1, 2] + x[1].data = [ [3.0 , 4.0 ] ] + x[1].dims = [1, 2] + x[2].data = [ [5.0 , 6.0 ] ] + x[2].dims = [1, 2] + + Attrs: + axis = 0 + + Output: + Out.data =[ [ [1.0, 2.0] ], + [ [3.0, 4.0] ], + [ [5.0, 6.0] ] ] + Out.dims = [3, 1, 2] + + Case 2: + Given + x[0].data = [ [1.0 , 2.0 ] ] + x[0].dims = [1, 2] + x[1].data = [ [3.0 , 4.0 ] ] + x[1].dims = [1, 2] + x[2].data = [ [5.0 , 6.0 ] ] + x[2].dims = [1, 2] + + Attrs: + axis = 1 or axis = -2 + + Output: + Out.data =[ [ [1.0, 2.0] + [3.0, 4.0] + [5.0, 6.0] ] ] + Out.dims = [1, 3, 2] + Args: x (Variable|list(Variable)|tuple(Variable)): Input variables. axis (int|None): The axis along which all inputs are stacked. diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 2153ca254f0e286a77160a2d53473e1bc76109d5..af747c3cecac66492bb2e2642a88f66a5cfae3db 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -567,7 +567,7 @@ def ones(shape, dtype, force_cpu=False): It also sets *stop_gradient* to True. Args: - shape(tuple|list|None): Shape of output tensor + shape(tuple|list): Shape of output tensor dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor Returns: @@ -578,6 +578,10 @@ def ones(shape, dtype, force_cpu=False): data = fluid.layers.ones(shape=[1], dtype='int64') """ + assert isinstance(shape, list) or isinstance( + shape, tuple), "The shape's type should be list or tuple." + assert reduce(lambda x, y: x * y, + shape) > 0, "The shape is invalid: %s." % (str(shape)) return fill_constant(value=1.0, **locals()) diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 22212ae9a216acaab3f295f1f8d091829a0aa471..8586670c2481a0f997e84f33860b6df28b3223ae 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -148,6 +148,8 @@ class ParallelExecutor(object): else framework.default_main_program() # FIXME(dzhwinter): enable_inplace should be after memory_optimize # if turn on python memory optimize, turn off the inplace_pass. + if build_strategy.memory_optimize is None: + build_strategy.memory_optimize = False if main._is_mem_optimized else True if build_strategy.enable_inplace is None: build_strategy.enable_inplace = False if main._is_mem_optimized else True scope = scope if scope is not None else executor.global_scope() diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 534411219b500723f3799a08fdf1b7796534376b..289a48aac9cf7ff52295feb722a8116939442425 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -77,6 +77,7 @@ list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op) list(REMOVE_ITEM TEST_OPS test_nearest_interp_op) list(REMOVE_ITEM TEST_OPS test_imperative_resnet) list(REMOVE_ITEM TEST_OPS test_imperative_optimizer) +list(REMOVE_ITEM TEST_OPS test_ir_memory_optimize_transformer) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) @@ -107,6 +108,9 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450) py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) +if(NOT WIN32) +py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL) +endif() if(NOT APPLE) py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) if(CMAKE_BUILD_TYPE STREQUAL "Debug") diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index c429c8af7d37cb4e209edc41f704868afe054829..a94487e67dc90d4df935867f841bc567c37c8aa2 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -79,7 +79,7 @@ class TestParallelExecutorBase(unittest.TestCase): if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv - build_strategy.memory_optimize = use_ir_memory_optimize + build_strategy.memory_optimize = False if memory_opt else use_ir_memory_optimize # python memory optimization is conflict with inplace pass. # Use ir graph memory optimization after inplace pass is the correct way. build_strategy.enable_inplace = False if memory_opt else enable_inplace diff --git a/python/paddle/fluid/tests/unittests/test_base_layer.py b/python/paddle/fluid/tests/unittests/test_base_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..bf00698d63624d4e20a0853641219a2735d89d25 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_base_layer.py @@ -0,0 +1,82 @@ +# 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 unittest +import numpy as np + +import paddle.fluid as fluid +from paddle.fluid.layer_helper import LayerHelper + + +class L1(fluid.imperative.Layer): + def __init__(self): + super(L1, self).__init__() + self._helper = LayerHelper( + 'MyLayer', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.1))) + + self.w1 = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=[2, 2], + dtype='float32', + is_bias=False) + self.w2 = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=[2, 2], + dtype='float32', + is_bias=False) + + def forward(self): + return self.w1 + self.w2 + + +class L2(fluid.imperative.Layer): + def __init__(self): + super(L2, self).__init__() + self.layer1 = L1() + self.layer2 = L1() + + def forward(self): + return self.layer1() + self.layer2() + + +class L3(fluid.imperative.Layer): + def __init__(self): + super(L3, self).__init__() + self.layer1 = L2() + self.layer2 = L2() + + def forward(self): + return self.layer1() + self.layer2() + + +class TestBaseLayer(unittest.TestCase): + def test_one_level(self): + with fluid.imperative.guard(): + l = L1() + ret = l() + self.assertEqual(l.w1.name, "MyLayer_0.w_0") + self.assertEqual(l.w2.name, "MyLayer_0.w_1") + self.assertTrue(np.allclose(ret._numpy(), 0.2 * np.ones([2, 2]))) + + def test_three_level(self): + with fluid.imperative.guard(): + l = L3() + ret = l() + self.assertTrue(np.allclose(ret._numpy(), 0.8 * np.ones([2, 2]))) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index 3566fed215229223f4d2ecd1bbb66cb297dd7716..12132477d28c74c7da718321140a3ddef784fc30 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -22,6 +22,9 @@ import six import unittest import numpy as np +import gc +gc.set_debug(gc.DEBUG_COLLECTABLE) + import paddle.fluid as fluid @@ -99,6 +102,12 @@ class TranspilerTest(unittest.TestCase): with fluid.unique_name.guard(): with fluid.program_guard(main, startup): self.transpiler_test_impl() + # NOTE: run gc.collect to eliminate pybind side objects to + # prevent random double-deallocate when inherited in python. + del self.transpiler + del main + del startup + gc.collect() class TestBasicModel(TranspilerTest): @@ -797,6 +806,7 @@ class TestNCCL2Transpile(TranspilerTest): print([op.type for op in startup.global_block().ops]) self.assertEqual(startup.global_block().ops[-1].type, "gen_nccl_id") self.assertIsNotNone(startup.global_block().vars.get("NCCLID")) + gc.collect() else: pass diff --git a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py index 03471a4432f2b6bf6220e79e99aa506628b1535b..c1fb53ecf52d953fa470998c120930b2bec6325b 100644 --- a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py +++ b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py @@ -121,6 +121,8 @@ class TestMNIST(TestParallelExecutorBase): regularization=fluid.regularizer.L2Decay(1e-6)) return optimizer + # NOTE(dzh): + # need to make it compatible with elewise fuse act not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence( model, feed_dict={"image": img, @@ -128,6 +130,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=False, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence( model, @@ -136,6 +139,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=True, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss): diff --git a/python/paddle/fluid/tests/unittests/test_imperative.py b/python/paddle/fluid/tests/unittests/test_imperative.py index baaddf9f2e5b123300f1d083b33ea644665348fd..c54e998ea875e1bd27f9816f88db0e38bc488459 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative.py +++ b/python/paddle/fluid/tests/unittests/test_imperative.py @@ -333,6 +333,18 @@ class TestImperative(unittest.TestCase): self.assertTrue(np.allclose(dy_out, static_out)) self.assertTrue(np.allclose(dy_grad, static_grad)) + params = mlp.parameters(True) + self.assertEqual("FC_0.w_0", params[0].name) + self.assertEqual("FC_0.b_0", params[1].name) + self.assertEqual("FC_1.w_0", params[2].name) + self.assertEqual("FC_1.b_0", params[3].name) + self.assertEqual(len(params), 4) + + sublayers = mlp.sublayers(True) + self.assertEqual(mlp._fc1, sublayers[0]) + self.assertEqual(mlp._fc2, sublayers[1]) + self.assertEqual(len(sublayers), 2) + def test_rnn(self): np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0], [10.0, 11.0, 12.0]]) diff --git a/python/paddle/fluid/tests/unittests/test_imperative_gan.py b/python/paddle/fluid/tests/unittests/test_imperative_gan.py index 681661bfc63db95653be371688a047efe96f3866..33c196d1ab52b393491561e75054e6c323fce18d 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_gan.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_gan.py @@ -33,9 +33,6 @@ class Discriminator(fluid.imperative.Layer): self._fc1 = FC(size=32, act='elu', name="d_fc1") self._fc2 = FC(size=1, name="d_fc2") - def parameters(self): - return self._fc1.parameters() + self._fc2.parameters() - def forward(self, inputs): x = self._fc1(inputs) return self._fc2(x) @@ -48,10 +45,6 @@ class Generator(fluid.imperative.Layer): self._fc2 = FC(size=64, act='elu', name="g_fc2") self._fc3 = FC(size=1, name="g_fc3") - def parameters(self): - return self._fc1.parameters() + self._fc2.parameters( - ) + self._fc3.parameters() - def forward(self, inputs): x = self._fc1(inputs) x = self._fc2(x) diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py index afe990e74ff96dfbca4f335b561f9bbe7d295246..82aff18b728f45388cd747e3bce19c1c9d6f91cc 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py @@ -75,16 +75,6 @@ class SimpleLSTMRNN(fluid.imperative.Layer): self.hidden_array.append(pre_hidden) self.cell_array.append(pre_cell) - def parameters(self): - parameters = list() - for param in self.weight_1_arr: - parameters.append(param) - for param in self.weight_2_arr: - parameters.append(param) - for bias in self.bias_arr: - parameters.append(bias) - return parameters - def forward(self, input_embedding, init_hidden=None, init_cell=None): res = [] for index in range(self._num_steps): @@ -177,12 +167,6 @@ class PtbModel(fluid.imperative.Layer): def _build_once(self, input, label, init_hidden, init_cell): pass - def parameters(self): - parameters = self.simple_lstm_rnn.parameters() + [ - self.softmax_weight, self.softmax_bias - ] + self.embedding.parameters() - return parameters - def forward(self, input, label, init_hidden, init_cell): init_h = fluid.layers.reshape( diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py index c27fd0b8024a8fa3310a62de34299fb621e2902f..128d18621db8374c6c385dddbefc0d29e760a02f 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py @@ -21,7 +21,6 @@ import paddle import paddle.fluid as fluid from paddle.fluid import core from paddle.fluid.layer_helper import LayerHelper -from paddle.fluid.optimizer import SGDOptimizer from paddle.fluid.imperative.nn import Conv2D, Pool2D, BatchNorm, FC from paddle.fluid.imperative.base import to_variable from test_imperative_base import new_program_scope @@ -173,11 +172,13 @@ class ResNet(fluid.imperative.Layer): for block in range(len(depth)): shortcut = False for i in range(depth[block]): - bottleneck_block = BottleneckBlock( - num_channels=num_channels, - num_filters=num_filters[block], - stride=2 if i == 0 and block != 0 else 1, - shortcut=shortcut) + bottleneck_block = self.add_sublayer( + 'bb_%d_%d' % (block, i), + BottleneckBlock( + num_channels=num_channels, + num_filters=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + shortcut=shortcut)) num_channels = bottleneck_block._num_channels_out self.bottleneck_block_list.append(bottleneck_block) shortcut = True @@ -223,8 +224,7 @@ class TestImperativeResnet(unittest.TestCase): batch_size=batch_size) dy_param_init_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): dy_param_init_value[param.name] = param._numpy() for batch_id, data in enumerate(train_reader()): @@ -247,16 +247,14 @@ class TestImperativeResnet(unittest.TestCase): dy_out = avg_loss._numpy() if batch_id == 0: - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): if param.name not in dy_param_init_value: dy_param_init_value[param.name] = param._numpy() avg_loss._backward() dy_grad_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): if not param.stop_gradient: np_array = np.array(param._ivar._grad_ivar().value() .get_tensor()) @@ -267,8 +265,7 @@ class TestImperativeResnet(unittest.TestCase): resnet.clear_gradients() dy_param_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): dy_param_value[param.name] = param._numpy() with new_program_scope(): @@ -349,6 +346,7 @@ class TestImperativeResnet(unittest.TestCase): self.assertTrue(np.allclose(static_out, dy_out)) self.assertEqual(len(dy_param_init_value), len(static_param_init_value)) + for key, value in six.iteritems(static_param_init_value): self.assertTrue(np.allclose(value, dy_param_init_value[key])) self.assertTrue(np.isfinite(value.all())) diff --git a/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..c0f480e34dcac3351ba3008ad632a29943afdb81 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py @@ -0,0 +1,48 @@ +# 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 os +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core + +os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0" +os.environ[ + 'RECORDIO_FILENAME'] = '/tmp/ir_memory_optimize_transformer.wmt16.recordio' + +from test_parallel_executor_transformer import TestTransformer +from test_parallel_executor_transformer import transformer + + +# NOTE(dzhwinter): test diferent strategy colisions. +# open the eager delete tensor strategy by default. +class TestTransformerWithIR(TestTransformer): + def test_main(self): + if core.is_compiled_with_cuda(): + # check python transpiler + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=True, + use_ir_memory_optimize=False) + # check IR memory optimize + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=False, + use_ir_memory_optimize=True) + + +if __name__ == '__main__': + unittest.main()