COMPMID-2273: Fuse Batch Normalization with Depthwise Convolution layer at graph level (only for CL)
Change-Id: I1d941c6e66722f39583bf68148c980bb28ff89a1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1423
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h
index be43b57..5c5b777 100644
--- a/arm_compute/graph/INodeVisitor.h
+++ b/arm_compute/graph/INodeVisitor.h
@@ -96,6 +96,11 @@
* @param[in] n Node to visit.
*/
virtual void visit(FusedConvolutionBatchNormalizationNode &n) = 0;
+ /** Visit FusedDepthwiseConvolutionBatchNormalizationNode.
+ *
+ * @param[in] n Node to visit.
+ */
+ virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) = 0;
/** Visit InputNode.
*
* @param[in] n Node to visit.
@@ -214,6 +219,10 @@
{
default_visit();
}
+ virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override
+ {
+ default_visit();
+ }
virtual void visit(InputNode &n) override
{
default_visit();
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index 4fb5b73..9da0e61 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -101,6 +101,9 @@
case NodeType::FusedConvolutionBatchNormalizationLayer:
os << "FusedConvolutionBatchNormalizationLayer";
break;
+ case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer:
+ os << "FusedDepthwiseConvolutionBatchNormalizationLayer";
+ break;
case NodeType::GenerateProposalsLayer:
os << "GenerateProposalsLayer";
break;
diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h
index 2f09abb..9f96242 100644
--- a/arm_compute/graph/Types.h
+++ b/arm_compute/graph/Types.h
@@ -141,6 +141,7 @@
FlattenLayer,
FullyConnectedLayer,
FusedConvolutionBatchNormalizationLayer,
+ FusedDepthwiseConvolutionBatchNormalizationLayer,
GenerateProposalsLayer,
NormalizationLayer,
NormalizePlanarYUVLayer,
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 5ac4fda..ed5b35c 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -30,6 +30,7 @@
#include "arm_compute/graph/Types.h"
#include "arm_compute/graph/Utils.h"
#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
+#include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h"
#include "arm_compute/graph/backends/Utils.h"
#include "arm_compute/graph/nodes/Nodes.h"
@@ -197,12 +198,6 @@
const ActivationLayerInfo fused_act = node.fused_activation();
const float epsilon = node.epsilon();
- const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
- if(is_quantized && biases != nullptr)
- {
- biases->info()->set_data_type(DataType::S32);
- }
-
// Create and configure function
auto func = support::cpp14::make_unique<FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>();
func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act);
@@ -210,7 +205,55 @@
// Log info
ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
<< node.name()
- << " Type: " << 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()) : "")
+ << std::endl);
+ return std::move(func);
+}
+
+/** Create a backend fused depthwise convolution batch normalization layer function
+ *
+ * @tparam FusedLayerTypes Fused layer types
+ * @tparam TargetInfo Target-specific information
+ *
+ * @param[in] node Node to create the backend function for
+ *
+ * @return Backend fused depthwise convolution batch normalization layer function
+ */
+template <typename FusedLayerTypes, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node)
+{
+ validate_node<TargetInfo>(node, 7 /* 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 depth_multiplier = node.depth_multiplier();
+ const ActivationLayerInfo fused_act = node.fused_activation();
+ const float epsilon = node.epsilon();
+
+ // Create and configure function
+ auto func = support::cpp14::make_unique<FusedDepthwiseConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>();
+ func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, depth_multiplier, fused_act);
+
+ // 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()
diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
index 92af17b..a6da76b 100644
--- a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
+++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
@@ -54,7 +54,7 @@
* 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, except for input of QASYMM8 type where biases should be of S32 type.
+ * 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
diff --git a/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
new file mode 100644
index 0000000..6f70d3c
--- /dev/null
+++ b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2019 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_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+#define __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+
+#include "arm_compute/core/Types.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}DepthwiseConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedDepthwiseConvolutionBatchNormalizationFunction : public IFunction
+{
+public:
+ using TensorType = typename TargetInfo::TensorType;
+ using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+ FusedDepthwiseConvolutionBatchNormalizationFunction()
+ : _depth_conv_layer(), _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: F16/F32.
+ * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+ * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [IFM].
+ * 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] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] fused_act Activation layer information in case of a fused activation.
+ *
+ */
+ 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 depth_multiplier, ActivationLayerInfo const &fused_act)
+ {
+ // 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, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+ bias_to_use = bias;
+ }
+ else
+ {
+ _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+ bias_to_use = &_fused_bias;
+ }
+
+ _depth_conv_layer.configure(input, weights, bias_to_use, output, conv_info, depth_multiplier, fused_act.enabled() ? fused_act : ActivationLayerInfo());
+
+ if(!has_bias)
+ {
+ _fused_bias.allocator()->allocate();
+ }
+ }
+
+ // Inherited methods overridden:
+ void run()
+ {
+ prepare();
+ _depth_conv_layer.run();
+ }
+
+ void prepare()
+ {
+ if(!_is_prepared)
+ {
+ _fused_batch_norm_layer.run();
+ _is_prepared = true;
+ }
+ }
+
+private:
+ typename FusedLayerTypes::DepthwiseConvolutionLayer _depth_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_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__ */
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
index 9b0f5b7..c124c982 100644
--- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
@@ -41,14 +41,13 @@
* @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
- * @param[in] out_quant_info (Optional) Output quantization info
* @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
*/
FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info,
unsigned int num_groups = 1,
ConvolutionMethod method = ConvolutionMethod::Default,
FastMathHint fast_math_hint = FastMathHint::Disabled,
- QuantizationInfo out_quant_info = QuantizationInfo(), ActivationLayerInfo fused_activation = ActivationLayerInfo());
+ ActivationLayerInfo fused_activation = ActivationLayerInfo());
/** Epsilon parameter accessor
*
@@ -135,7 +134,6 @@
unsigned int _num_groups;
ConvolutionMethod _method;
FastMathHint _fast_math_hint;
- QuantizationInfo _out_quant_info;
ActivationLayerInfo _fused_activation;
};
diff --git a/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
new file mode 100644
index 0000000..a2241ef
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2019 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_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+#define __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+
+#include "arm_compute/graph/INode.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Fused Depthwise Convolution Batch Normalization node */
+class FusedDepthwiseConvolutionBatchNormalizationNode final : public INode
+{
+public:
+ /** Constructor
+ *
+ * @param[in] epsilon Epsilon parameter.
+ * @param[in] info Convolution layer attributes.
+ * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] method (Optional) Convolution method to use
+ * @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
+ */
+ FusedDepthwiseConvolutionBatchNormalizationNode(float epsilon,
+ PadStrideInfo info,
+ unsigned int depth_multiplier,
+ DepthwiseConvolutionMethod method,
+ ActivationLayerInfo fused_activation = ActivationLayerInfo());
+
+ /** Sets the depthwise convolution layer method to use
+ *
+ * @param[in] method Method to use for depthwise convolution
+ */
+ void set_depthwise_convolution_method(DepthwiseConvolutionMethod method);
+
+ /** Depthwise convolution layer method accessor
+ *
+ * @note This is an indication on which depthwise convolution layer implementation to use,
+ * if it fails to be created the library's heuristic approach will be used
+ *
+ * @return Depthwise convolution layer method to be used by the node
+ */
+ DepthwiseConvolutionMethod depthwise_convolution_method() const;
+
+ /** Epsilon parameter accessor
+ *
+ * @return Epsilon parameter
+ */
+ float epsilon() const;
+
+ /** Returns fused activation
+ *
+ * @return Fused activation
+ */
+ ActivationLayerInfo fused_activation() const;
+
+ /** Sets fused activation
+ *
+ * @param[in] fused_activation Fused activation to set
+ */
+ void set_fused_activation(ActivationLayerInfo fused_activation);
+
+ /** Computes convolution output descriptor
+ *
+ * @param[in] input_descriptor Input descriptor
+ * @param[in] weights_descriptor Weights descriptor
+ * @param[in] info Convolution operation attributes
+ * @param[in] depth_multiplier Depth multiplier
+ *
+ * @return Output descriptor
+ */
+ static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+ const TensorDescriptor &weights_descriptor,
+ const PadStrideInfo &info,
+ int depth_multiplier);
+
+ /** Sets the convolution layer method to use
+ *
+ * @param[in] method Method to use for convolution
+ */
+ void set_convolution_method(ConvolutionMethod method);
+
+ /** Depth multiplier accessor
+ *
+ * @return Depth multiplier
+ */
+ unsigned int depth_multiplier() 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::FusedDepthwiseConvolutionBatchNormalizationLayer;
+
+private:
+ float _epsilon;
+
+ PadStrideInfo _info;
+ unsigned int _depth_multiplier;
+ DepthwiseConvolutionMethod _method;
+ ActivationLayerInfo _fused_activation;
+};
+
+} // namespace graph
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__ */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index c891bc2..52e2f88 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -39,6 +39,7 @@
#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/FusedDepthwiseConvolutionBatchNormalizationNode.h"
#include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
#include "arm_compute/graph/nodes/InputNode.h"
#include "arm_compute/graph/nodes/NormalizationLayerNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 0f3450b..2c89679 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -45,6 +45,7 @@
class FlattenLayerNode;
class FullyConnectedLayerNode;
class FusedConvolutionBatchNormalizationNode;
+class FusedDepthwiseConvolutionBatchNormalizationNode;
class GenerateProposalsLayerNode;
class InputNode;
class NormalizationLayerNode;
diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h
index 9d2ea46..c28a17b 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(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
void visit(NormalizationLayerNode &n) override;
void visit(PoolingLayerNode &n) override;
void default_visit() override;