Graph Fusion With Post Ops Fix
- Fusing ConvolutionBatchNormalization Nodes with post ops (activation
or element wise ops)
Resolves: COMPMID-4982
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I5b2d32cad00f710fd744cb5aa2d59fd7e5c97e0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6766
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
diff --git a/arm_compute/graph/DataLayerVisitor.h b/arm_compute/graph/DataLayerVisitor.h
index 670b9f0..ac7f1c8 100644
--- a/arm_compute/graph/DataLayerVisitor.h
+++ b/arm_compute/graph/DataLayerVisitor.h
@@ -48,6 +48,7 @@
void visit(ConvolutionLayerNode &n) override;
void visit(DepthwiseConvolutionLayerNode &n) override;
void visit(FusedConvolutionBatchNormalizationNode &n) override;
+ void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
void visit(OutputNode &n) override;
diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h
index 4cb601f..97e9533 100644
--- a/arm_compute/graph/INodeVisitor.h
+++ b/arm_compute/graph/INodeVisitor.h
@@ -106,6 +106,11 @@
* @param[in] n Node to visit.
*/
virtual void visit(FusedConvolutionBatchNormalizationNode &n) = 0;
+ /** Visit FusedConvolutionBatchNormalizationWithPostOpsNode.
+ *
+ * @param[in] n Node to visit.
+ */
+ virtual void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) = 0;
/** Visit FusedConvolutionWithPostOpNode.
*
* @param[in] n Node to visit.
@@ -210,6 +215,7 @@
virtual void visit(FlattenLayerNode &n) override;
virtual void visit(FullyConnectedLayerNode &n) override;
virtual void visit(FusedConvolutionBatchNormalizationNode &n) override;
+ virtual void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
virtual void visit(FusedConvolutionWithPostOpNode &n) override;
virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
virtual void visit(InputNode &n) override;
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index a8a20c9..8f97bbf 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -116,6 +116,9 @@
case NodeType::FusedConvolutionBatchNormalizationLayer:
os << "FusedConvolutionBatchNormalizationLayer";
break;
+ case NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer:
+ os << "FusedConvolutionBatchNormalizationLayerWithPostOpsLayer";
+ break;
case NodeType::FusedConvolutionWithPostOp:
os << "FusedConvolutionWithPostOp";
break;
diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h
index e802e9d..ff33d50 100644
--- a/arm_compute/graph/Types.h
+++ b/arm_compute/graph/Types.h
@@ -216,6 +216,7 @@
FullyConnectedLayer,
FusedConvolutionBatchNormalizationLayer,
FusedConvolutionWithPostOp,
+ FusedConvolutionBatchNormalizationLayerWithPostOpsLayer,
FusedDepthwiseConvolutionBatchNormalizationLayer,
GenerateProposalsLayer,
L2NormalizeLayer,
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 1e420a8..a7e52d4 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -32,6 +32,7 @@
#include "arm_compute/graph/Types.h"
#include "arm_compute/graph/Utils.h"
#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
+#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h"
#include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h"
#include "arm_compute/graph/backends/Utils.h"
#include "arm_compute/graph/nodes/Nodes.h"
@@ -540,7 +541,7 @@
return std::move(func);
}
-/** Create a backend convolution layer function with post opreator
+/** Create a backend convolution layer function with post operator
*
* @tparam ConvolutionLayerFunctions Backend convolution functions
* @tparam TargetInfo Target-specific information
@@ -629,6 +630,91 @@
<< " Output shape: " << output->info()->tensor_shape()
<< qss.str()
<< (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+ << " Post ops" << post_ops;
+ << std::endl);
+ return std::move(func);
+}
+
+/** Create a backend convolution batch normalization layer function with post operator
+ *
+ * @tparam FusedLayerTypes Backend convolution functions
+ * @tparam TargetInfo Target-specific information
+ *
+ * @param[in] node Node to create the backend function for
+ * @param[in] ctx Graph context
+ *
+ * @return Backend fused convolution with batch normalization layer function
+ */
+template <typename FusedLayerTypes, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_with_post_op(FusedConvolutionBatchNormalizationWithPostOpsNode &node, GraphContext &ctx)
+{
+ validate_node<TargetInfo>(node, 8 /* expected inputs */, 1 /* expected outputs */);
+
+ // Extract IO and info
+ typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
+ typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
+ typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
+ typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(3));
+ typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(4));
+ typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(5));
+ typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(6));
+
+ typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
+
+ const PadStrideInfo conv_info = node.convolution_info();
+ const unsigned int num_groups = node.num_groups();
+ const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled;
+ const float epsilon = node.epsilon();
+
+ experimental::PostOpList<typename TargetInfo::TensorType *> post_ops;
+
+ auto &post_op_info_list = node.post_op_info_list();
+ for(const auto &post_op_info : post_op_info_list)
+ {
+ switch(post_op_info->type())
+ {
+ case PostOpType::Activation:
+ {
+ const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get());
+ post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act);
+ break;
+ }
+ case PostOpType::Eltwise_Add:
+ {
+ typename TargetInfo::TensorType *add_input = get_backing_tensor<TargetInfo>(node.input(3));
+ const auto eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get());
+ post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy);
+ break;
+ }
+ default:
+ {
+ ARM_COMPUTE_ERROR("Unsupported PostOpType");
+ }
+ }
+ }
+
+ // Create and configure function (we assume that functions have been validated before creation)
+ std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
+ std::unique_ptr<IFunction> func;
+ std::string func_name;
+
+ using FType = FusedConvolutionBatchNormalizationWithPostOpsFunction<TargetInfo, FusedLayerTypes>;
+
+ // Create and configure function
+ std::tie(func, func_name) = create_named_memory_managed_function<FType>(
+ std::string("FusedConvolutionBatchNormalizationLayerWithPostOpsLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, post_ops);
+
+ // Log info
+ ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+ << node.name()
+ << " Type: " << node.type()
+ << " Target: " << TargetInfo::TargetType
+ << " Data Type: " << input->info()->data_type()
+ << " Input shape: " << input->info()->tensor_shape()
+ << " Weights shape: " << weights->info()->tensor_shape()
+ << " Output shape: " << output->info()->tensor_shape()
+ << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+ << " Post Ops:" << post_ops;
<< std::endl);
return std::move(func);
}
diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h
new file mode 100644
index 0000000..10f2e5c
--- /dev/null
+++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H
+#define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/experimental/IPostOp.h"
+#include "arm_compute/runtime/IFunction.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+namespace backends
+{
+/** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}ConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedConvolutionBatchNormalizationWithPostOpsFunction : public IFunction
+{
+public:
+ using TensorType = typename TargetInfo::TensorType;
+ using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+ FusedConvolutionBatchNormalizationWithPostOpsFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr)
+ : _conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false)
+ {
+ }
+
+ /** Set the input and output tensors.
+ *
+ * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+ * while every optional dimension from 4 and above represent a batch of inputs.
+ * Data types supported: QASYMM8/F16/F32.
+ * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
+ * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+ * Data type supported: Should match @p input data type.
+ * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+ * Data types supported: Same as @p input.
+ * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+ * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+ * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input
+ * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input
+ * @param[in] epsilon Small value to avoid division with zero. Default value is 0.001f.
+ * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
+ * @param[in] num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
+ * @param[in] fast_math Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+ * available which may introduce a drop of accuracy as well. Default is false
+ * @param[in] post_ops A sequence of post operations that are performed after the main operation.
+ *
+ */
+ void configure(TensorType *input,
+ TensorType *weights,
+ TensorType *bias,
+ TensorType *output,
+ const TensorType *mean,
+ const TensorType *var,
+ const TensorType *beta,
+ const TensorType *gamma,
+ float epsilon, const PadStrideInfo &conv_info, unsigned int num_groups, bool fast_math,
+ const arm_compute::experimental::PostOpList<TensorType *> &post_ops = experimental::PostOpList<TensorType *> {})
+ {
+ // We don't run any validate, as we assume that the layers have been already validated
+ const bool has_bias = (bias != nullptr);
+ const TensorType *bias_to_use;
+
+ // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one
+ // as batch normalization might end up with a bias != 0
+ if(has_bias)
+ {
+ _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon);
+ bias_to_use = bias;
+ }
+ else
+ {
+ _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon);
+ bias_to_use = &_fused_bias;
+ }
+
+ ActivationLayerInfo fused_act = ActivationLayerInfo(); // Passing an empty ActivationLayerInfo.
+ _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups, post_ops);
+
+ if(!has_bias)
+ {
+ _fused_bias.allocator()->allocate();
+ }
+ }
+
+ // Inherited methods overridden:
+ void run()
+ {
+ prepare();
+ _conv_layer.run();
+ }
+
+ void prepare()
+ {
+ if(!_is_prepared)
+ {
+ _fused_batch_norm_layer.run();
+ _is_prepared = true;
+ }
+ }
+
+private:
+ typename FusedLayerTypes::ConvolutionLayer _conv_layer;
+ typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer;
+ TensorConcreteType _fused_bias;
+ bool _is_prepared;
+};
+} // namespace backends
+} // namespace graph
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H */
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
index b3661c3..b0051b1 100644
--- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019 Arm Limited.
+ * Copyright (c) 2019, 2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -100,7 +100,7 @@
*/
ConvolutionMethod convolution_method() const;
- /** Sets the fast math fast hint
+ /** Sets the fast math hint
*
* @param[in] hint Hint to use for convolution
*/
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h
new file mode 100644
index 0000000..a42e06d
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h
@@ -0,0 +1,127 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H
+#define ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H
+
+#include "arm_compute/graph/INode.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Batch Normalization node */
+class FusedConvolutionBatchNormalizationWithPostOpsNode final : public INode
+{
+public:
+ /** Constructor
+ *
+ * @param[in] epsilon Epsilon parameter.
+ * @param[in] info Convolution layer attributes.
+ * @param[in] num_groups (Optional) Number of groups (Defaults to 1)
+ * @param[in] method (Optional) Convolution method to use
+ * @param[in] fast_math_hint (Optional) Fast math hint
+ */
+ FusedConvolutionBatchNormalizationWithPostOpsNode(float epsilon, PadStrideInfo info,
+ unsigned int num_groups = 1,
+ ConvolutionMethod method = ConvolutionMethod::Default,
+ FastMathHint fast_math_hint = FastMathHint::Disabled);
+
+ /** Epsilon parameter accessor
+ *
+ * @return Epsilon parameter
+ */
+ float epsilon() const;
+
+ /** Computes convolution output descriptor
+ *
+ * @param[in] input_descriptor Input descriptor
+ * @param[in] weights_descriptor Weights descriptor
+ * @param[in] info Convolution operation attributes
+ *
+ * @return Output descriptor
+ */
+ static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+ const TensorDescriptor &weights_descriptor,
+ const PadStrideInfo &info);
+
+ /** Sets the convolution layer method to use
+ *
+ * @param[in] method Method to use for convolution
+ */
+ void set_convolution_method(ConvolutionMethod method);
+
+ /** Number of groups in convolution accessor
+ *
+ * @return Number of groups in convolution
+ */
+ unsigned int num_groups() const;
+
+ /** Convolution layer method accessor
+ *
+ * @note This is an indication on which convolution layer implementation to use,
+ * if it fails to be created the library's heuristic approach will be used
+ *
+ * @return Convolution layer method to be used by the node
+ */
+ ConvolutionMethod convolution_method() const;
+
+ /** Sets the fast math hint
+ *
+ * @param[in] hint Hint to use for convolution
+ */
+ void set_fast_math_hint(FastMathHint hint);
+
+ /** Fast math hint accessor
+ *
+ * @return Fast math hint to be used by the node
+ */
+ FastMathHint fast_math_hint() const;
+
+ /** Convolution metadata accessor
+ *
+ * @return Convolution information
+ */
+ PadStrideInfo convolution_info() const;
+
+ // Inherited overridden methods:
+ NodeType type() const override;
+ bool forward_descriptors() override;
+ TensorDescriptor configure_output(size_t idx) const override;
+ void accept(INodeVisitor &v) override;
+
+public:
+ static constexpr NodeType node_type = NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer;
+
+private:
+ float _epsilon;
+
+ PadStrideInfo _info;
+ unsigned int _num_groups;
+ ConvolutionMethod _method;
+ FastMathHint _fast_math_hint;
+};
+
+} // namespace graph
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_GRAPH_BATCH_NORMALIZATION_LAYER_NODE_H */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index fb0eb15..3887eae 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -43,6 +43,7 @@
#include "arm_compute/graph/nodes/FlattenLayerNode.h"
#include "arm_compute/graph/nodes/FullyConnectedLayerNode.h"
#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h"
+#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h"
#include "arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h"
#include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
#include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 6393b1d..f1576d6 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -51,6 +51,7 @@
class FusedConvolutionBatchNormalizationNode;
class FusedConvolutionWithPostOpNode;
class FusedDepthwiseConvolutionBatchNormalizationNode;
+class FusedConvolutionBatchNormalizationWithPostOpsNode;
class GenerateProposalsLayerNode;
class InputNode;
class L2NormalizeLayerNode;
diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h
index 42a2067..63b8927 100644
--- a/arm_compute/graph/printers/DotGraphPrinter.h
+++ b/arm_compute/graph/printers/DotGraphPrinter.h
@@ -57,6 +57,7 @@
void visit(DepthwiseConvolutionLayerNode &n) override;
void visit(EltwiseLayerNode &n) override;
void visit(FusedConvolutionBatchNormalizationNode &n) override;
+ void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
void visit(FusedConvolutionWithPostOpNode &n) override;
void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
void visit(NormalizationLayerNode &n) override;