[Draft] Enable training with optional bias

ONE-DCO-1.0-Signed-off-by: ragmani <[email protected]>

runtime/onert/backend/train/ops/FullyConnectedLayer.cc

@@ -186,13 +186,11 @@ void FullyConnectedLayer::backwardFloat32()
     getShape(_grad_weights), getBuffer<float>(_grad_weights));
 
   // Compute gradient for bias
-  if (_bias)
-  {
-    assert(_grad_bias);
-    nnfw::cker::train::FullyConnectedBiasGrad(getShape(backprop_act),
-                                              getBuffer<float>(backprop_act), getShape(_grad_bias),
-                                              getBuffer<float>(_grad_bias));
-  }
+  // Bias tensor must exist on training even if bias is optinal input
+  assert(_bias);
+  assert(_grad_bias);
+  nnfw::cker::train::FullyConnectedBiasGrad(getShape(backprop_act), getBuffer<float>(backprop_act),
+                                            getShape(_grad_bias), getBuffer<float>(_grad_bias));
 }
 
 } // namespace ops

runtime/onert/core/include/ir/IOperation.h

@@ -40,6 +40,7 @@ struct IOperation
 
   virtual void replaceInputs(const OperandIndex &from, const OperandIndex &to) = 0;
   virtual void replaceOutputs(const OperandIndex &from, const OperandIndex &to) = 0;
+  virtual void replaceInput(size_t pos, const OperandIndex &index) = 0;
   virtual const OperandIndexSequence &getInputs() const = 0;
   virtual const OperandIndexSequence &getOutputs() const = 0;
 };

runtime/onert/core/include/ir/OperandIndexSequence.h

@@ -51,6 +51,7 @@ class OperandIndexSequence
   const OperandIndex &at(uint32_t index) const { return _vec.at(index); }
   bool contains(const OperandIndex &index) const;
   void replace(const OperandIndex &from, const OperandIndex &to);
+  void replace(size_t pos, const OperandIndex &index);
   OperandIndexSequence operator|(ir::Remove filter) const
   {
     switch (filter)

runtime/onert/core/include/ir/Operation.h

@@ -50,6 +50,7 @@ class Operation : virtual public IOperation
 public:
   void replaceInputs(const OperandIndex &from, const OperandIndex &to) override;
   void replaceOutputs(const OperandIndex &from, const OperandIndex &to) override;
+  void replaceInput(size_t pos, const OperandIndex &index) override;
   OperandIndexSequence &getInputs() { return _inputs; }
   const OperandIndexSequence &getInputs() const override { return _inputs; }
   const OperandIndexSequence &getOutputs() const override { return _outputs; }

runtime/onert/core/include/ir/train/TrainableGraph.h

@@ -123,6 +123,7 @@ class TrainableGraph : public IGraph
   const Operands &operands() const override { return _graph.operands(); }
   Operands &operands() { return _graph.operands(); } // TODO Remove this non-const accessor
   const Operations &operations() const override { return _graph.operations(); }
+  Operations &operations() { return _graph.operations(); }
   const Operands &backward_operands() const { return _backward_operands; }
   OperandIndex getLossIndex(const IOIndex &pred_io_ind) const;
   const Graph &graph() const { return _graph; }

runtime/onert/core/src/compiler/train/TrainingCompiler.cc

@@ -18,6 +18,7 @@
 
 #include "StaticBackwardShapeInferer.h"
 #include "TrainableOperationConverter.h"
+#include "pass/BiasInsertionPass.h"
 #include "pass/LossInsertionPass.h"
 #include "../CompilerHelpers.h"
 #include "../ExecutorFactory.h"
@@ -86,6 +87,7 @@ std::shared_ptr<CompilerArtifact> TrainingCompiler::compile(void)
     compiler::pass::PassRunner{}
       .append(std::make_unique<compiler::pass::ConstantOutputPass>(subg))
       .append(std::make_unique<compiler::pass::OddOutputPass>(subg))
+      .append(std::make_unique<train::pass::BiasInsertionPass>(subg))
       .run();
 
     // Optimizations

runtime/onert/core/src/compiler/train/pass/BiasInsertionPass.cc

@@ -0,0 +1,83 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. 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.
+ */
+
+#include "BiasInsertionPass.h"
+
+#include "ir/Graph.h"
+#include "util/logging.h"
+
+namespace onert
+{
+namespace compiler
+{
+namespace train
+{
+namespace pass
+{
+
+void BiasInsertionPass::run()
+{
+  _graph.operations().iterate([&](const ir::OperationIndex &op_index, const ir::IOperation &node) {
+    _current_op_index = op_index;
+    node.accept(*this);
+  });
+}
+
+void BiasInsertionPass::visit(const ir::operation::Conv2D &) {}
+
+void BiasInsertionPass::visit(const ir::operation::DepthwiseConv2D &) {}
+
+void BiasInsertionPass::visit(const ir::operation::FullyConnected &node)
+{
+  const auto &bias_index = node.getInputs().at(ir::operation::Conv2D::Input::BIAS);
+
+  // Insert bias if it is optional
+  if (!bias_index.valid())
+  {
+    const auto &output_index = node.getOutputs().at(0);
+    const auto &output = _graph.operands().at(output_index);
+    const auto &output_shape = output.shape();
+    const auto bias_shape = ir::Shape{output_shape.dim(output_shape.rank() - 1)};
+
+    auto bias_typeinfo = output.typeInfo();
+    if (bias_typeinfo.type() != ir::DataType::FLOAT32)
+      throw std::runtime_error("BiasInsertionPass: Only FLOAT32 is supported for now");
+
+    const auto new_bias_index = _graph.addOperand(bias_shape, output.typeInfo());
+
+    // TODO Replace data with sparse data to reduce memory usage
+    const auto bias_size = bias_shape.num_elements() * ir::sizeOfDataType(bias_typeinfo.type());
+    std::vector<uint8_t> data_vec(bias_size, 0);
+    auto data_obj = std::make_shared<ir::CachedData>(data_vec.data(), bias_size);
+    _graph.setOperandValue(new_bias_index, std::move(data_obj));
+
+    auto &bias = _graph.operands().at(new_bias_index);
+    bias.insertUse(_current_op_index);
+    bias.isConstant();
+
+    _graph.operations()
+      .at(_current_op_index)
+      .replaceInput(ir::operation::Conv2D::Input::BIAS, new_bias_index);
+
+    VERBOSE(BiasInsertionPass) << "Optional bias is inserted for training, bias index : "
+                               << bias_index << std::endl;
+  }
+}
+
+} // namespace pass
+} // namespace train
+} // namespace compiler
+} // namespace onert

runtime/onert/core/src/compiler/train/pass/BiasInsertionPass.h

@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. 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.
+ */
+
+#ifndef __ONERT_COMPILER_TRAIN_PASS_BIAS_INSERTION_PASS_H__
+#define __ONERT_COMPILER_TRAIN_PASS_BIAS_INSERTION_PASS_H__
+
+#include "../../pass/Pass.h"
+#include "ir/OperationVisitor.h"
+
+namespace onert
+{
+namespace compiler
+{
+namespace train
+{
+namespace pass
+{
+
+class BiasInsertionPass final : public compiler::pass::Pass, public ir::OperationVisitor
+{
+public:
+  BiasInsertionPass(ir::Graph &graph) : compiler::pass::Pass{graph} {}
+
+public:
+  std::string id() final { return "BiasInsertionPass"; }
+  void run() final;
+
+public:
+  void visit(const ir::operation::Conv2D &node) override;
+  void visit(const ir::operation::DepthwiseConv2D &node) override;
+  void visit(const ir::operation::FullyConnected &node) override;
+
+private:
+  ir::OperationIndex _current_op_index;
+};
+
+} // namespace pass
+} // namespace train
+} // namespace compiler
+} // namespace onert
+
+#endif // __ONERT_COMPILER_TRAIN_PASS_BIAS_INSERTION_PASS_H__

runtime/onert/core/src/ir/OperandIndexSequence.cc

@@ -17,6 +17,7 @@
 #include "ir/OperandIndexSequence.h"
 
 #include <algorithm>
+#include <cassert>
 #include <sstream>
 
 namespace onert
@@ -55,6 +56,12 @@ void OperandIndexSequence::replace(const OperandIndex &from, const OperandIndex
   std::replace(_vec.begin(), _vec.end(), from, to);
 }
 
+void OperandIndexSequence::replace(size_t pos, const OperandIndex &index)
+{
+  assert(pos < _vec.size() && "OperandIndexSequence: Out of range");
+  _vec.at(pos) = index;
+}
+
 bool OperandIndexSequence::operator==(const OperandIndexSequence &other) const
 {
   return _vec == other._vec;

runtime/onert/core/src/ir/Operation.cc

@@ -62,5 +62,7 @@ void Operation::replaceOutputs(const OperandIndex &from, const OperandIndex &to)
   _outputs.replace(from, to);
 }
 
+void Operation::replaceInput(size_t pos, const OperandIndex &index) { _inputs.replace(pos, index); }
+
 } // namespace ir
 } // namespace onert

runtime/onert/core/src/ir/train/UseDefGenerator.cc

@@ -211,6 +211,23 @@ void UseDefGenerator::visit(const train::operation::FullyConnected &node)
   const auto weights_forwarding_index = TrainingOperandIndex{weights_index, true};
   insertUse(weights_forwarding_index, backwarding_op_index);
 
+  // // Set def and use for bias if bias is optional since bias is required on training
+  // const auto forwarding_op_index = TrainingOperationIndex{op_index, true};
+  // const auto &bias_index = node.getInputs().at(ir::operation::Conv2D::Input::BIAS);
+  // const auto bias_forwarding_index = TrainingOperandIndex{bias_index, true};
+  // if (!bias_index.valid())
+  // {
+  //   // If bias is not optional, the use for bias is already inserted before
+  //   insertUse(bias_forwarding_index, forwarding_op_index);
+
+  //   // The forwarding bias is not used in backwarding
+  // }
+  // else
+  // {
+  //   [[maybe_unused]] const auto &usedef_chain = _training_usedefs.at(bias_forwarding_index);
+  //   assert(usedef_chain.getTrainingUses().count(forwarding_op_index) == 0);
+  // }
+
   // Insert uses of forwarding output
   if (node.param().activation != ir::Activation::NONE)
   {
@@ -220,18 +237,19 @@ void UseDefGenerator::visit(const train::operation::FullyConnected &node)
   }
 
   // Set def of backwarding inputs
+  // The uses of backwarding inputs has already been inserted before
   const auto outgoing_index = TrainingOperandIndex{in_index, false};
   insertBackPropDef(outgoing_index, backwarding_op_index);
 
   const auto weights_gradient_index = TrainingOperandIndex{weights_index, false};
   insertDef(weights_gradient_index, backwarding_op_index);
 
+  // Set def for gradient bias even if it is optional bias since bias is required
+  // on training.
   const auto &bias_index = node.getInputs().at(ir::operation::Conv2D::Input::BIAS);
-  if (bias_index.valid())
-  {
-    const auto bias_gradient_index = TrainingOperandIndex{bias_index, false};
-    insertDef(bias_gradient_index, backwarding_op_index);
-  }
+  assert(bias_index.valid());
+  const auto bias_gradient_index = TrainingOperandIndex{bias_index, false};
+  insertDef(bias_gradient_index, backwarding_op_index);
 }
 
 void UseDefGenerator::visit(const train::operation::Loss &node)
@@ -427,7 +445,10 @@ void UseDefGenerator::initForForwardingNodes()
     assert(_training_usedefs.at(forwarding_operand_index).getTrainingUses().size() == 0);
     const auto uses = operand.getUses();
     for (const auto &use : uses)
+    {
+      // if (use.valid())
       insertUse(forwarding_operand_index, TrainingOperationIndex{use, is_forward});
+    }
   });
 }
 

tests/nnfw_api/src/GenModelTests/one_op_trains/FullyConnected.test.cc

@@ -74,7 +74,7 @@ TEST_F(GenModelTrain, OneOp_FullyConnected_OptionalBias)
   _context->addTrainCase(
     uniformTCD<float>({{{1, 3, 2, 1}}},                                            // inputs
                       {{{2, 1, 5, 5, 2, 1, 5, 5, 2, 1, 5, 5, 2, 1, 5, 6}}},        // expected
-                      {{14.4375f}, {13.9950f}, {13.5668f}, {13.1523f}, {12.7512f}} // loss
+                      {{14.4375f}, {13.9242f}, {13.4298f}, {12.9537f}, {12.4952f}} // loss
                       ));
 
   _context->setBackends({"train"});