fix: Adapt torch JIT pass for removeDropout

- Adapt JIT pass to remove dropout to accommodate multiple dropout schemas - Include additional test cases to verify new removal code
pytorch · Jan 24, 2023 · e7a469d · e7a469d
1 parent ae8b569
commit e7a469d
Show file tree

Hide file tree

Showing 2 changed files with 122 additions and 4 deletions.
diff --git a/core/lowering/passes/remove_dropout.cpp b/core/lowering/passes/remove_dropout.cpp
@@ -1,15 +1,57 @@
-#include "torch/csrc/jit/passes/remove_dropout.h"
-#include <torch/csrc/jit/passes/subgraph_rewrite.h>
-
 #include "core/util/prelude.h"
+#include "torch/csrc/jit/passes/dead_code_elimination.h"
 
 namespace torch_tensorrt {
 namespace core {
 namespace lowering {
 namespace passes {
 
+// Schemas for dropout variants
+const std::unordered_set<c10::Symbol> DropoutNodeKinds = {
+    c10::Symbol::fromQualString("aten::dropout"),
+    c10::Symbol::fromQualString("aten::dropout_"),
+    c10::Symbol::fromQualString("aten::feature_dropout"),
+    c10::Symbol::fromQualString("aten::feature_dropout_"),
+    c10::Symbol::fromQualString("aten::feature_alpha_dropout"),
+    c10::Symbol::fromQualString("aten::feature_alpha_dropout_"),
+};
+
+void removeDropoutInBlock(torch::jit::Block* block) {
+  /*
+  Function adapted from:
+  torch/csrc/jit/passes/remove_dropout.cpp
+
+  Modified for conciseness, documentation, and allowing new variants of dropout operators to be quickly added
+  */
+  std::vector<torch::jit::Node*> dropout_nodes_to_remove;
+
+  for (auto node : block->nodes()) {
+    // Remove dropout for each member block within a node
+    for (auto block : node->blocks()) {
+      removeDropoutInBlock(block);
+    }
+
+    // For each node having a dropout-variant Schema, remove the node
+    if (DropoutNodeKinds.find(node->kind()) != DropoutNodeKinds.end()) {
+      // Extract input and output tensors of dropout operator
+      auto input_value = node->inputs()[0];
+      auto output_value = node->outputs()[0];
+
+      output_value->replaceAllUsesWith(input_value);
+      dropout_nodes_to_remove.push_back(node);
+    }
+  }
+
+  // Delete dropout nodes
+  for (auto del_node : dropout_nodes_to_remove) {
+    del_node->destroy();
+  }
+}
+
 void RemoveDropout(std::shared_ptr<torch::jit::Graph>& graph) {
-  torch::jit::removeDropout(graph);
+  // Remove all instances of dropout variants from graph
+  removeDropoutInBlock(graph->block());
+  torch::jit::EliminateDeadCode(graph);
   LOG_GRAPH("Post remove dropout: " << *graph);
 }
 

diff --git a/tests/core/lowering/test_remove_dropout_pass.cpp b/tests/core/lowering/test_remove_dropout_pass.cpp
@@ -132,3 +132,79 @@ TEST(LoweringPasses, RemoveFeatureDropoutInplaceLowersCorrectly) {
 
   ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
 }
+
+TEST(LoweringPasses, RemoveFeatureAlphaDropoutLowersCorrectly) {
+  std::string source_graph = R"IR(
+    graph(%x.1):
+      %3 : float = prim::Constant[value=0.5]()
+      %4 : bool = prim::Constant[value=0]()
+      %y.1 : Tensor = aten::feature_alpha_dropout(%x.1, %3, %4)
+      %11 : Tensor = aten::relu(%y.1)
+      return (%11))IR";
+  std::string target_graph = R"IR(
+    graph(%x.1):
+      %11 : Tensor = aten::relu(%x.1)
+      return (%11))IR";
+
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, sg.get());
+  torch_tensorrt::core::lowering::passes::RemoveDropout(sg);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, tg.get());
+
+  ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
+}
+
+TEST(LoweringPasses, RemoveFeatureAlphaDropoutNestedLowersCorrectly) {
+  std::string source_graph = R"IR(
+    graph(%x.1):
+      %3 : float = prim::Constant[value=0.5]()
+      %4 : bool = prim::Constant[value=0]()
+      %y.1 : Tensor = aten::feature_alpha_dropout(%x.1, %3, %4)
+      %z.1 : Tensor = aten::feature_alpha_dropout(%y.1, %3, %4)
+      %12 : Tensor = aten::relu(%z.1)
+      return (%12))IR";
+  std::string target_graph = R"IR(
+    graph(%x.1):
+      %11 : Tensor = aten::relu(%x.1)
+      return (%11))IR";
+
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, sg.get());
+  torch_tensorrt::core::lowering::passes::RemoveDropout(sg);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, tg.get());
+
+  ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
+}
+
+TEST(LoweringPasses, RemoveFeatureAlphaDropoutInplaceLowersCorrectly) {
+  std::string source_graph = R"IR(
+    graph(%x.1):
+      %3 : float = prim::Constant[value=0.5]()
+      %4 : bool = prim::Constant[value=0]()
+      %y.1 : Tensor = aten::feature_alpha_dropout_(%x.1, %3, %4)
+      %11 : Tensor = aten::relu(%y.1)
+      return (%11))IR";
+  std::string target_graph = R"IR(
+    graph(%x.1):
+      %11 : Tensor = aten::relu(%x.1)
+      return (%11))IR";
+
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, sg.get());
+  torch_tensorrt::core::lowering::passes::RemoveDropout(sg);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, tg.get());
+
+  ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
+}