diff --git a/include/aidge/operator/PTQMetaOps.hpp b/include/aidge/operator/PTQMetaOps.hpp
index b9bad0d18f099e94d4c52254b08629c7f947db6a..9ca76fbd40b9366aa82c6521fba931d284da137a 100644
--- a/include/aidge/operator/PTQMetaOps.hpp
+++ b/include/aidge/operator/PTQMetaOps.hpp
@@ -29,14 +29,6 @@ namespace Aidge {
/// @return A shared pointer to an instance of the meta-operator node.
std::shared_ptr<Aidge::Node> Quantizer(double scalingFactor, double clipMin, double clipMax, const std::string& name);
-/// @brief The purpose of Scaling is to encapsulate the Mul operator and tag it as a PTQ node rather than a regular Mul operator.
-/// Therefore, this meta-operator consists solely of a [Mul] operation.
-///
-/// @param scalingFactor The scaling factor to apply to the input (a scalar to multiply the input with).
-/// @param name The name of the meta-operator node created.
-/// @return A shared pointer to an instance of the scaling node.
-std::shared_ptr<Aidge::Node> Scaling(double scalingFactor, const std::string& name = "");
-
/// @brief Updates the scaling factor of a PTQ meta-operator node, allowing for dynamic adjustment of the scaling parameter.
/// This function sets a new scaling factor for a specified meta-operator node, modifying the scalar applied in the [Mul] operation.
/// The meta-operator node must be a PTQ-specific operator, such as a Quantizer or Scaling node.
diff --git a/include/aidge/quantization/PTQ/PTQ.hpp b/include/aidge/quantization/PTQ/PTQ.hpp
index bfe671e3556c3af2c367ce7f86708f01c8e3d3b5..1d1b71ba7501580ea99103d351eafac9a7f793d2 100644
--- a/include/aidge/quantization/PTQ/PTQ.hpp
+++ b/include/aidge/quantization/PTQ/PTQ.hpp
@@ -69,6 +69,26 @@ namespace Aidge {
* @return The scheduled vector of nodes
*/
std::vector<std::shared_ptr<Node>> retrieveNodeVector(std::shared_ptr<GraphView> graphView, bool newSchedule = true, bool verbose = false);
+
+ /**
+ * @brief Inserts a scaling node below the given producer node in the graph view.
+ * If the node is already a producer scaling node, it accumulates the scaling factor by multiplyins its value directly.
+ *
+ * @param node A shared pointer to the producer node where the scaling node will be inserted (below).
+ * @param scalingFactor The scaling factor to apply.
+ * @param graphView A shared pointer to the graph view in which the nodes are located.
+ * @return True if the scaling node was successfully inserted or the scaling factor was accumulated; False otherwise.
+ */
+ bool insertScalingBelowProducer(std::shared_ptr<Node> node, double scalingFactor, std::shared_ptr<GraphView> graphView);
+
+ /**
+ * @brief Inserts a rounding node below the given producer (also below its ows producerScaling) node in the graph view.
+ *
+ * @param node A shared pointer to the producer node where the rounding node will be inserted.
+ * @param graphView A shared pointer to the graph view in which the nodes are located.
+ * @return True if the rounding node was successfully inserted; False otherwise.
+ */
+ bool insertRoundBelowProducer(std::shared_ptr<Node> node, std::shared_ptr<GraphView> graphView);
/**
* @brief Determine whether an input GraphView can be quantized or not.
@@ -77,6 +97,14 @@ namespace Aidge {
*/
bool checkArchitecture(std::shared_ptr<GraphView> graphView);
+ /**
+ * @brief This function multiplies the existing scaling factor by a given coefficient. It verifies that the node is of the correct type ("Mul")
+ * and has the `isScaling` attribute. If these conditions are not met, a warning is logged.
+ * @param node A shared pointer to an `Aidge::Node` object representing the node to modify.
+ * @param coeff A double representing the multiplication coefficient to apply to the scaling factor.
+ */
+ void multiplyScalingFactor(std::shared_ptr<Aidge::Node> node, double coeff);
+
void prepareNetwork(std::shared_ptr<GraphView> graphView);
diff --git a/include/aidge/quantization/QAT/QAT_LSQ.hpp b/include/aidge/quantization/QAT/QAT_LSQ.hpp
index 922187abca915daa1c00f3949d0d791b0d3e1c39..4970be07fae8737a1c2863600757bb81ff3a65f9 100644
--- a/include/aidge/quantization/QAT/QAT_LSQ.hpp
+++ b/include/aidge/quantization/QAT/QAT_LSQ.hpp
@@ -9,29 +9,36 @@
*
********************************************************************************/
-#ifndef AIDGE_QUANTIZATION_QUANTIZATION_QAT_LSQ_H_
-#define AIDGE_QUANTIZATION_QUANTIZATION_QAT_LSQ_H_
+#ifndef AIDGE_QUANTIZATION_QAT_LSQ_H_
+#define AIDGE_QUANTIZATION_QAT_LSQ_H_
-#include <cstddef> // std::size_t
-#include <memory>
-
-#include "aidge/data/Tensor.hpp"
+#include "aidge/graph/Node.hpp"
#include "aidge/graph/GraphView.hpp"
+#include "aidge/data/Tensor.hpp"
namespace Aidge {
namespace QuantLSQ {
/**
- * @brief Given a GraphView with parameters properly initialized, insert
- * the LSQ quantizer nodes, and setup the adjustment their step-sizes.
- * @param graphView The GraphView containing the network to quantize.
+ * @brief Insert the LSQ quantizer nodes in a given GraphView
+ * @param graphView The GraphView containing the graph to quantize.
* @param nbBits Number of quantization bits.
+ * @param span Fixed output span of the quantizers.
*/
+void insertQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits, float step_size);
-void setupQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits);
+/**
+ * @brief Given a GraphView with parameters properly initialized and some calibration data,
+ * insert the LSQ quantizer nodes, and adjust their step-sizes.
+ * @param graphView The GraphView containing the graph to quantize.
+ * @param nbBits Number of quantization bits.
+ * @param calibrationData Calibration data used to adjust the spans.
+ * @param scale Multiplicative constant applied to the spans.
+ */
+void insertAndInitQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits, std::shared_ptr<Tensor> calibrationData);
-} // namespace QuantLSQ
-} // namespace Aidge
+}
+}
-#endif /* AIDGE_QUANTIZATION_QUANTIZATION_QAT_LSQ_H_ */
+#endif /* AIDGE_QUANTIZATION_QAT_LSQ_H_ */
diff --git a/python_binding/pybind_PTQ.cpp b/python_binding/pybind_PTQ.cpp
index 1de797693468273814f4c5e82a161991648d06ff..ae0a0def28a861e2fc207adbc27c6af47dc0ded8 100644
--- a/python_binding/pybind_PTQ.cpp
+++ b/python_binding/pybind_PTQ.cpp
@@ -213,13 +213,6 @@ void init_PTQ(py::module &m) {
:type network: :py:class:`aidge_core.GraphView`
)mydelimiter");
- m.def("dev_ptq", &devPTQ, py::arg("network"),
- R"mydelimiter(
- Developement and test routine.
- :param network: The GraphView under test.
- :type network: :py:class:`aidge_core.GraphView`
- )mydelimiter");
-
m.def("prepare_network", &prepareNetwork, py::arg("network"), "prepare the network for the PTQ");
}
diff --git a/python_binding/pybind_QAT_LSQ.cpp b/python_binding/pybind_QAT_LSQ.cpp
index 4bba3b6baa5eda41a024399eb1be1402c74b2c1b..206985efe4558a84ce1ed67a1264bd6902213764 100644
--- a/python_binding/pybind_QAT_LSQ.cpp
+++ b/python_binding/pybind_QAT_LSQ.cpp
@@ -23,6 +23,8 @@ void init_QAT_LSQ(py::module &m) {
auto mQuantLSQ = m.def_submodule("lsq");
- mQuantLSQ.def("setup_quantizers", &QuantLSQ::setupQuantizers, py::arg("network"), py::arg("nb_bits"));
+ mQuantLSQ.def("insert_quantizers", &QuantLSQ::insertQuantizers, py::arg("network"), py::arg("nb_bits"), py::arg("step_size"));
+
+ mQuantLSQ.def("insert_and_init_quantizers", &QuantLSQ::insertAndInitQuantizers, py::arg("network"), py::arg("nb_bits"), py::arg("calibration_data"));
}
} // namespace Aidge
diff --git a/src/PTQ/CLE.cpp b/src/PTQ/CLE.cpp
index 28858d0e3c693a7620bc32806008523e0602faa9..2738f8a92154368962e9162fba62c41b7622d07c 100644
--- a/src/PTQ/CLE.cpp
+++ b/src/PTQ/CLE.cpp
@@ -20,7 +20,10 @@
#include "aidge/quantization/PTQ/PTQ.hpp" // retrieveNodeVector
#include "aidge/graph/GraphView.hpp"
-#include "aidge/graph/Node.hpp"
+
+#include "aidge/scheduler/SequentialScheduler.hpp"
+#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/utils/Log.hpp"
#include "aidge/operator/OperatorTensor.hpp"
#include "aidge/utils/Log.hpp"
@@ -30,6 +33,12 @@
#include "aidge/operator/Reshape.hpp"
#include "aidge/operator/Round.hpp"
+#include "aidge/operator/Mul.hpp"
+#include "aidge/operator/ArgMax.hpp"
+#include "aidge/operator/Abs.hpp"
+#include "aidge/operator/Reshape.hpp"
+#include "aidge/operator/Round.hpp"
+
namespace Aidge
{
@@ -49,7 +58,7 @@ static void rescaleTensor(std::shared_ptr<Tensor> tensor, double scaling)
mulOp.setDataType(tensor->dataType());
mulOp.setBackend(tensor->backend());
- std::shared_ptr<Aidge::Tensor> scalingTensor = std::make_shared<Aidge::Tensor>(scaling);
+ std::shared_ptr<Aidge::Tensor> scalingTensor = std::make_shared<Aidge::Tensor>(Aidge::Array1D<double, 1> {scaling});
scalingTensor->setDataType(tensor->dataType());
scalingTensor->setBackend(tensor->backend());
@@ -67,7 +76,7 @@ static void rescaleTensor(std::shared_ptr<Tensor> tensor, double scaling)
static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
{
// get the abs tensor
-
+ std::shared_ptr<Tensor> fallback; //Fallback tensor for refCastFR
std::shared_ptr<Tensor> absTensor = std::make_shared<Tensor>(tensor->abs());
// flatten the abs tensor
@@ -81,6 +90,7 @@ static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
reshapeOp.associateInput(0, absTensor);
reshapeOp.forward();
std::shared_ptr<Tensor> flatTensor = reshapeOp.getOutput(0);
+ const Tensor& localFlatTensor = flatTensor->refCastFrom(fallback, DataType::Float64, "cpu");
// Get the argmax
@@ -90,13 +100,24 @@ static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
argmaxOp.associateInput(0, flatTensor);
argmaxOp.forward();
- std::shared_ptr<Tensor> argmaxTensor = argmaxOp.getOutput(0);
+
+ const Tensor& argMaxTensor = argmaxOp.getOutput(0)->refCastFrom(fallback, DataType::Float64, "cpu");
// Return the max
- int maxIndex = std::round(argmaxTensor->get<double>(0));
+ int maxIndex = std::round(argMaxTensor.get<double>(0));
- return flatTensor->get<double>(maxIndex);
+ return localFlatTensor.get<double>(maxIndex);
+}
+//Function used to extraxt the local tensor (from a ProducerScalingNode)
+std::shared_ptr<Aidge::Tensor> getLocalTensor(std::shared_ptr<Node> node) {
+ if (node->getParent(1)->attributes()->hasAttr("quantization.ptq.isProducerScaling")) {
+ std::shared_ptr<Aidge::OperatorTensor> operatorTensor = std::static_pointer_cast<OperatorTensor>(node->getParent(1)->getOperator());
+ operatorTensor->forward();// We need the forward pass to compute the scaled value of the Tensor
+ return operatorTensor->getOutput(0);
+ } else {
+ return getWeightTensor(node);
+ }
}
void crossLayerEqualization(std::shared_ptr<GraphView> graphView, double targetDelta)
@@ -131,16 +152,18 @@ void crossLayerEqualization(std::shared_ptr<GraphView> graphView, double targetD
std::shared_ptr<Node> n1 = affineNodeVector[i];
std::shared_ptr<Node> n2 = affineNodeVector[i+1];
- double r1 = getTensorAbsoluteMax(getWeightTensor(n1));
- double r2 = getTensorAbsoluteMax(getWeightTensor(n2));
+ std::shared_ptr<Aidge::Tensor> n1localTensor = getLocalTensor(n1);
+ std::shared_ptr<Aidge::Tensor> n2localTensor = getLocalTensor(n2);
+
+ double r1 = getTensorAbsoluteMax(n1localTensor);
+ double r2 = getTensorAbsoluteMax(n2localTensor);
double s1 = std::sqrt(r1 * r2) / r1;
double s2 = std::sqrt(r1 * r2) / r2;
- rescaleTensor(getWeightTensor(n1), s1);
- rescaleTensor(getWeightTensor(n2), s2);
-
- rescaleTensor(getBiasTensor(n1), s1);
+ insertScalingBelowProducer(n1->getParent(1),s1,graphView);
+ insertScalingBelowProducer(n2->getParent(1),s2,graphView);
+ insertScalingBelowProducer(n1->getParent(2),s1,graphView);
double rangeDelta = std::abs(r1 - r2);
if (rangeDelta > maxRangeDelta)
diff --git a/src/PTQ/Clipping.cpp b/src/PTQ/Clipping.cpp
index 66b0ab36fba7634d7ee350cdccb27895ffa52da1..a4e7fed921604fcf9d18c6e50991220c4785f3bb 100644
--- a/src/PTQ/Clipping.cpp
+++ b/src/PTQ/Clipping.cpp
@@ -222,7 +222,7 @@ std::map<std::string, double> adjustRanges(Clipping clippingMode, std::map<std::
for (std::shared_ptr<Node> node : graphView->getNodes())
{
- if (node->type() == "Scaling")
+ if (node->attributes()->hasAttr("quantization.ptq.isScaling"))
{
std::vector<int> histogram = histograms[node->name()];
diff --git a/src/PTQ/PTQ.cpp b/src/PTQ/PTQ.cpp
index 7c29ee0b9178fbb07f4a2d5edf9f0ad7ac8dcac4..f03fc7bcea039a1939e116cc842f7062f28c5cae 100644
--- a/src/PTQ/PTQ.cpp
+++ b/src/PTQ/PTQ.cpp
@@ -14,7 +14,6 @@
#include "aidge/quantization/PTQ/PTQ.hpp"
#include "aidge/operator/PTQMetaOps.hpp"
-
#include "aidge/data/Tensor.hpp"
#include "aidge/graph/GraphView.hpp"
#include "aidge/graph/Node.hpp"
@@ -22,16 +21,15 @@
#include "aidge/scheduler/Scheduler.hpp"
#include "aidge/utils/Log.hpp"
+
#include "aidge/operator/Producer.hpp"
#include "aidge/operator/Mul.hpp"
+#include "aidge/operator/Round.hpp"
#include "aidge/operator/ReLU.hpp"
#include "aidge/operator/BatchNorm.hpp"
#include "aidge/operator/Conv.hpp"
-
#include "aidge/operator/ArgMax.hpp"
-#include "aidge/operator/Abs.hpp"
#include "aidge/operator/Reshape.hpp"
-#include "aidge/operator/Round.hpp"
#include "aidge/recipes/Recipes.hpp"
@@ -55,59 +53,124 @@ bool isMerging(std::shared_ptr<Node> node)
{
return (mergingNodeTypes.find(node->type()) != mergingNodeTypes.end());
}
+static int getInputIndex(std::shared_ptr<Node> node, std::shared_ptr<Node> parentNode)
+{
+ int index = 0;
+ while (node->getParent(index) != parentNode)
+ index++;
+ return index;
+}
-bool checkArchitecture(std::shared_ptr<GraphView> graphView)
+
+void multiplyScalingFactor(std::shared_ptr<Aidge::Node> node,double coeff)
{
- std::set<std::string> otherNodeTypes({"Flatten", "Softmax", "BatchNorm2D", "ReLU", "Producer"});
+ AIDGE_ASSERT(node->type() == "Mul" && (node->attributes()->hasAttr("quantization.ptq.isProducerScaling") || node->attributes()->hasAttr("quantization.ptq.isScaling")),
+ "Cannot update the scaling factor on Node of type {} with no scaling tag",node->type());
+ auto scalingFactorTensor = std::static_pointer_cast<OperatorTensor>(node->getOperator())->getInput(1);
+ std::shared_ptr<Tensor> fallback;
+ const Tensor& localTensor = scalingFactorTensor->refCastFrom(fallback, DataType::Float64, "cpu");
+ double previousScalingFactor = localTensor.get<double>(0);
+ std::shared_ptr<Tensor> finalTensor = std::make_shared<Tensor>(Array1D<double, 1> {previousScalingFactor * coeff});
+ node->input(1).first->getOperator()->setOutput(0, finalTensor);
+}
+/* Util function to insert a node below another one already connected */
+void insertNodeBetween(std::shared_ptr<Node> parent,
+ std::shared_ptr<Node> newNode,
+ std::shared_ptr<GraphView> graphView)
+{
+ // Checking the parents always have at least 1 children
+ AIDGE_ASSERT(parent->getChildren().size() > 0, "The parent node must have at least one child to insert a new node.");
+
+ // Retrieve children connection indexes
+ std::vector<std::shared_ptr<Node>> nextNodes = parent->getChildren(0);
+ std::vector<int> inputIndices(nextNodes.size());
+ for (std::size_t i = 0; i < nextNodes.size(); i++) {
+ inputIndices[i] = getInputIndex(nextNodes[i], parent);
+ }
- for (std::shared_ptr<Node> node : graphView->getNodes())
- {
- bool isOther = otherNodeTypes.find(node->type()) != otherNodeTypes.end();
- if (!isOther && !isAffine(node) && !isSeamless(node) && !isMerging(node)) {
- Log::warn(" GraphView can't be quantized : node type {} is not supported !", node->type());
- return false;
- }
+ // Disconnect childs from parent
+ for (std::shared_ptr<Node> nextNode : nextNodes) {
+ parent->removeChild(nextNode, 0);
}
- return true;
+ // Insert the new node between the child and the parent
+ parent->addChild(newNode, 0, 0);
+ for (std::size_t i = 0; i < nextNodes.size(); i++) {
+ newNode->addChild(nextNodes[i], 0, inputIndices[i]);
+ }
+
+ graphView->add(newNode);
}
-static void rescaleTensor(std::shared_ptr<Tensor> tensor, double scaling)
+bool insertRoundBelowProducer(std::shared_ptr<Node> node,std::shared_ptr<GraphView> graphView)
{
- auto mulOp = Mul_Op();
- mulOp.setDataType(tensor->dataType());
- mulOp.setBackend(tensor->backend());
+ if(node->attributes()->hasAttr("quantization.ptq.isProducerScaling") && node->type() != "Round")
+ {
+ std::shared_ptr<Aidge::Node> roundNode = Round(node->name() + "_Round");
+ roundNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
+ roundNode->getOperator()->setBackend(node->getOperator()->backend());
- std::shared_ptr<Aidge::Tensor> scalingTensor = std::make_shared<Aidge::Tensor>(scaling);
- scalingTensor->setDataType(tensor->dataType());
- scalingTensor->setBackend(tensor->backend());
+ insertNodeBetween(node,roundNode,graphView);
+ roundNode->attributes()->addAttr("quantization.ptq.isProducerRounding",0.0);
+ return true;
+ }
+ return false;
+}
+bool insertScalingBelowProducer(std::shared_ptr<Node> node,double scalingFactor, std::shared_ptr<GraphView> graphView)
+{
+ if(node->attributes()->hasAttr("quantization.ptq.isProducerRounding"))
+ {
+ //In this case we 'bump' the node to the one above him (an actual ProducerScaling)
+ // because the round node is not usable (only used when SSA is enabled)
+ node = node->getParent(0);
+ }
+ if(node->attributes()->hasAttr("quantization.ptq.isProducerScaling"))
+ {
+ // We accumulate the multiples scaling factors by multiplying the SF of the ProducerScaling node
+ // (adding new nodes each time would make the graph unusable)
+ multiplyScalingFactor(node,scalingFactor);
+ return true;
+ }
+ AIDGE_ASSERT(node->type() == "Producer","Cannot apply a scaling factor on node of type: {} which is not a producer", node->type());
+ std::string scalingNodeName = makeUniqueName(node->name() + "_ProducerScaling", graphView);
+
+ std::shared_ptr<Aidge::Node> scalingNode = Mul(scalingNodeName);
+ scalingNode->attributes()->addAttr("quantization.ptq.isProducerScaling",0.0);
+
+ std::shared_ptr<Tensor> scalingFactorTensor = std::make_shared<Tensor>(Array1D<double, 1> {scalingFactor});
+ std::shared_ptr<Node> scalingFactorProducer = addProducer(scalingNode, 1, {1}, "Factor");
+ scalingFactorProducer->getOperator()->setOutput(0, scalingFactorTensor);
+ graphView->add(scalingFactorProducer);
+
+ scalingNode->getOperator()->setDataType(DataType::Float64);
+ std::string producerBackend = std::static_pointer_cast<OperatorTensor>(node->getOperator())->getOutput(0)->backend();
+ scalingNode->getOperator()->setBackend(producerBackend);
- mulOp.associateInput(0, tensor);
- mulOp.associateInput(1, scalingTensor);
+ insertNodeBetween(node, scalingNode, graphView);
- mulOp.forward();
-
- auto outTensor = mulOp.getOutput(0);
- *tensor = *outTensor;
+ return true;
}
-static void roundTensor(std::shared_ptr<Tensor> tensor)
+bool checkArchitecture(std::shared_ptr<GraphView> graphView)
{
- auto roundOp = Round_Op();
- roundOp.setDataType(tensor->dataType());
- roundOp.setBackend(tensor->backend());
+ std::set<std::string> otherNodeTypes({"Flatten", "Softmax", "BatchNorm2D", "ReLU", "Producer"});
- roundOp.associateInput(0, tensor);
- roundOp.forward();
-
- auto outTensor = roundOp.getOutput(0);
- *tensor = *outTensor;
+ for (std::shared_ptr<Node> node : graphView->getNodes())
+ {
+ bool isOther = otherNodeTypes.find(node->type()) != otherNodeTypes.end();
+ if (!isOther && !isAffine(node) && !isSeamless(node) && !isMerging(node)) {
+ Log::warn(" GraphView can't be quantized : node type {} is not supported !", node->type());
+ return false;
+ }
+ }
+
+ return true;
}
-// TODO : make the retreival of argmax values backend independant (refCastFrom)
static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
{
// get the abs tensor
+ std::shared_ptr<Tensor> fallback; //Fallback tensor for refCastFR
std::shared_ptr<Tensor> absTensor = std::make_shared<Tensor>(tensor->abs());
@@ -122,6 +185,7 @@ static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
reshapeOp.associateInput(0, absTensor);
reshapeOp.forward();
std::shared_ptr<Tensor> flatTensor = reshapeOp.getOutput(0);
+ const Tensor& localFlatTensor = flatTensor->refCastFrom(fallback, DataType::Float64, "cpu");
// Get the argmax
@@ -131,13 +195,13 @@ static double getTensorAbsoluteMax(std::shared_ptr<Tensor> tensor)
argmaxOp.associateInput(0, flatTensor);
argmaxOp.forward();
- std::shared_ptr<Tensor> argmaxTensor = argmaxOp.getOutput(0);
+ const Tensor& argMaxTensor = argmaxOp.getOutput(0)->refCastFrom(fallback, DataType::Float64, "cpu");
// Return the max
- int maxIndex = std::round(argmaxTensor->get<double>(0));
+ int maxIndex = std::round(argMaxTensor.get<double>(0));
- return flatTensor->get<double>(maxIndex);
+ return localFlatTensor.get<double>(maxIndex);
}
@@ -151,6 +215,15 @@ static std::vector<std::shared_ptr<Node>> removeMatchingNodes(std::vector<std::s
return remainingNodes;
}
+static std::vector<std::shared_ptr<Node>> removeProdScalingNodes(std::vector<std::shared_ptr<Node>> nodeVector)
+{
+ std::vector<std::shared_ptr<Node>> remainingNodes;
+ for (std::shared_ptr<Node> node : nodeVector)
+ if (!node->attributes()->hasAttr("quantization.ptq.isProducerScaling"))
+ remainingNodes.push_back(node);
+
+ return remainingNodes;
+}
static void fixScheduling(std::vector<std::shared_ptr<Node>>& nodeVector) {
@@ -195,6 +268,7 @@ std::vector<std::shared_ptr<Node>> retrieveNodeVector(std::shared_ptr<GraphView>
fixScheduling(nodeVector);
nodeVector = removeMatchingNodes(nodeVector, "Producer");
+ nodeVector = removeProdScalingNodes(nodeVector);
if (verbose)
{
@@ -214,7 +288,6 @@ static std::shared_ptr<Node> getFirstNode(std::shared_ptr<GraphView> graphView)
void prepareNetwork(std::shared_ptr<GraphView> graphView)
{
removeFlatten(graphView);
-
sanitizeNodeNames(graphView);
bool containsBatchNorm = false;
@@ -264,25 +337,26 @@ void insertResidualNodes(std::shared_ptr<GraphView> graphView)
Log::info(" ### inserting multiplicative node ...");
std::string residualNodeName = makeUniqueName(parentNode->name() + "_Res", graphView);
- std::shared_ptr<Node> residualNode = Scaling(1.0, residualNodeName);
+ std::shared_ptr<Node> residualNode = Mul(residualNodeName);
+ residualNode->attributes()->addAttr("quantization.ptq.isScaling", 0.0);
+ residualNode->attributes()->addAttr("quantization.ptq.isResidual", 0.0);
+
+ //Adding the SF as a producer of the node
+ std::shared_ptr<Tensor> scalingFactorTensor = std::make_shared<Tensor>(Array1D<double, 1> {1.0});
+ std::shared_ptr<Node> scalingFactorProducer = addProducer(residualNode, 1, {1}, "ScalingFactor");
+ scalingFactorProducer->getOperator()->setOutput(0, scalingFactorTensor);
- residualNode->getOperator()->setDataType(DataType::Float64); //getDataType(parentNode)
- residualNode->getOperator()->setBackend("cpu");
+ residualNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
+ residualNode->getOperator()->setBackend(parentNode->getOperator()->backend());
graphView->insertParent(node, residualNode, i, 0, 0);
+ graphView->add(scalingFactorProducer);
}
}
}
}
}
-static int getInputIndex(std::shared_ptr<Node> node, std::shared_ptr<Node> parentNode)
-{
- int index = 0;
- while (node->getParent(index) != parentNode)
- index++;
- return index;
-}
void insertScalingNodes(std::shared_ptr<GraphView> graphView)
{
@@ -295,37 +369,30 @@ void insertScalingNodes(std::shared_ptr<GraphView> graphView)
if (isAffine(parentNode) || isMerging(parentNode))
{
std::string scalingNodeName = makeUniqueName(parentNode->name() + "_Scaling", graphView);
- std::shared_ptr<Node> scalingNode = Scaling(1.0, scalingNodeName);
+ //std::shared_ptr<Node> scalingNode = Scaling(1.0, scalingNodeName);
+
+ //Adding Mul operator with tag "quantization.ptq.isScaling"
+ std::shared_ptr<Aidge::Node> scalingNode = Mul(scalingNodeName);
+ scalingNode->attributes()->addAttr("quantization.ptq.isScaling",0.0);
+
+ //Adding the SF as a producer of the node
+ std::shared_ptr<Tensor> scalingFactorTensor = std::make_shared<Tensor>(Array1D<double, 1> {1.0});
+ std::shared_ptr<Node> scalingFactorProducer = addProducer(scalingNode, 1, {1}, "ScalingFactor");
+ scalingFactorProducer->getOperator()->setOutput(0, scalingFactorTensor);
scalingNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
- scalingNode->getOperator()->setBackend("cpu");
+ scalingNode->getOperator()->setBackend(parentNode->getOperator()->backend());
if (parentNode->getChildren().size() > 0)
{
- // SCALING NODE INSERTION
-
- // We always have one output from Affine and Add nodes, but possibly multiple childs
- std::vector<std::shared_ptr<Node>> nextNodes = parentNode->getChildren(0);
-
- // For each node in nextNodes store the connexion index
- std::vector<int> inputIndices(nextNodes.size());
- for (std::size_t i = 0; i < nextNodes.size(); i++)
- inputIndices[i] = getInputIndex(nextNodes[i], parentNode);
-
- for (std::shared_ptr<Node> nextNode : nextNodes)
- parentNode->removeChild(nextNode, 0);
-
- parentNode->addChild(scalingNode, 0, 0);
-
- for (std::size_t i = 0; i < nextNodes.size(); i++)
- scalingNode->addChild(nextNodes[i], 0, inputIndices[i]);
-
- graphView->add(scalingNode);
+ insertNodeBetween(parentNode,scalingNode,graphView);
+ graphView->add(scalingFactorProducer);
}
else
{
// Log::info(" last node reached ! ");
parentNode->addChild(scalingNode, 0, 0);
+ graphView->add(scalingFactorProducer);
graphView->add(scalingNode);
}
}
@@ -335,7 +402,7 @@ void insertScalingNodes(std::shared_ptr<GraphView> graphView)
static std::shared_ptr<Node> getPreviousScalingNode(std::shared_ptr<Node> mergingNode)
{
std::shared_ptr<Node> currNode = mergingNode;
- while(currNode->type() != "Scaling")
+ while(!currNode->attributes()->hasAttr("quantization.ptq.isScaling"))
{
if (currNode->getParents().size() == 0)
{
@@ -378,7 +445,7 @@ void normalizeParameters(std::shared_ptr<GraphView> graphView)
for (std::shared_ptr<Node> node : nodeVector)
{
// Scaling nodes still have a ratio of 1, so they are seamless ...
- if (node->type() == "ReLU" || node->type() == "Scaling" || isSeamless(node))
+ if (node->type() == "ReLU" || node->attributes()->hasAttr("quantization.ptq.isScaling") || isSeamless(node))
{
if (node != firstNode)
{
@@ -394,7 +461,8 @@ void normalizeParameters(std::shared_ptr<GraphView> graphView)
std::shared_ptr<Tensor> weightTensor = getWeightTensor(node);
double scaling = getTensorAbsoluteMax(weightTensor);
double ratio = 1.0 / scaling;
- rescaleTensor(weightTensor, ratio);
+ //rescaleTensor(weightTensor, ratio);
+ insertScalingBelowProducer(node->getParent(1),ratio,graphView);
// Accumulate the ratio
if (node == firstNode)
@@ -412,7 +480,8 @@ void normalizeParameters(std::shared_ptr<GraphView> graphView)
if (nodeHasBias(node))
{
std::shared_ptr<Tensor> biasTensor = getBiasTensor(node);
- rescaleTensor(biasTensor, accumulatedRatios[node->name()] );
+ //rescaleTensor(biasTensor, accumulatedRatios[node->name()] );
+ insertScalingBelowProducer(node->getParent(2),accumulatedRatios[node->name()],graphView);
}
}
@@ -439,8 +508,7 @@ void normalizeParameters(std::shared_ptr<GraphView> graphView)
std::shared_ptr<Node> scalingNode = getPreviousScalingNode(mergingNode);
- double currScalingFactor = getScalingFactor(scalingNode);
- updateScalingFactor(scalingNode, currScalingFactor / rescaling);
+ multiplyScalingFactor(scalingNode,1/rescaling);
accumulatedRatios[mergingNode->name()] /= rescaling; // optional ...
}
@@ -465,7 +533,7 @@ std::map<std::string, double> computeRanges(std::shared_ptr<GraphView> graphView
std::set<std::shared_ptr<Node>> nodeSet = graphView->getNodes();
for (std::shared_ptr<Node> node : nodeSet)
{
- if ((scalingNodesOnly && (node->type() == "Scaling")) || (!scalingNodesOnly && (node->type() != "Producer")))
+ if ((scalingNodesOnly && (node->attributes()->hasAttr("quantization.ptq.isScaling"))) || (!scalingNodesOnly && (node->type() != "Producer")))
{
std::shared_ptr<Operator> nodeOperator = node->getOperator();
std::shared_ptr<Tensor> valueTensor = std::static_pointer_cast<Tensor> (nodeOperator->getRawOutput(0));
@@ -487,7 +555,7 @@ std::map<std::string, double> computeRanges(std::shared_ptr<GraphView> graphView
// std::shared_ptr<Node> inputNode = getFirstNode(graphView);
for (std::shared_ptr<Node> node : nodeSet)
- if ((scalingNodesOnly && (node->type() == "Scaling")) || (!scalingNodesOnly && (node->type() != "Producer")))
+ if ((scalingNodesOnly && (node->attributes()->hasAttr("quantization.ptq.isScaling"))) || (!scalingNodesOnly && (node->type() != "Producer")))
valueRanges.insert(std::make_pair(node->name(), 0));
if (useCuda)
@@ -514,7 +582,7 @@ std::map<std::string, double> computeRanges(std::shared_ptr<GraphView> graphView
std::map<std::string, double> sampleRanges;
for (std::shared_ptr<Node> node : nodeSet)
{
- if ((scalingNodesOnly && (node->type() == "Scaling")) || (!scalingNodesOnly && (node->type() != "Producer")))
+ if ((scalingNodesOnly && (node->attributes()->hasAttr("quantization.ptq.isScaling"))) || (!scalingNodesOnly && (node->type() != "Producer")))
{
std::shared_ptr<Operator> nodeOperator = node->getOperator();
std::shared_ptr<Tensor> valueTensor = std::static_pointer_cast<Tensor> (nodeOperator->getRawOutput(0));
@@ -536,7 +604,7 @@ std::map<std::string, double> computeRanges(std::shared_ptr<GraphView> graphView
for (std::shared_ptr<Node> node : nodeSet)
{
- if ((scalingNodesOnly && (node->type() == "Scaling")) || (!scalingNodesOnly && (node->type() != "Producer")))
+ if ((scalingNodesOnly && (node->attributes()->hasAttr("quantization.ptq.isScaling"))) || (!scalingNodesOnly && (node->type() != "Producer")))
{
std::string nodeName = node->name();
if (sampleRanges[nodeName] > valueRanges[nodeName])
@@ -572,7 +640,7 @@ void normalizeActivations(std::shared_ptr<GraphView> graphView, std::map<std::st
for (std::shared_ptr<Node> node : nodeVector)
{
// Seamless scaling factor propagation ...
-
+
if (isAffine(node) || isSeamless(node) || node->type() == "ReLU")
{
if (node == firstNode)
@@ -586,11 +654,13 @@ void normalizeActivations(std::shared_ptr<GraphView> graphView, std::map<std::st
}
}
+
// Here prevNode is either a 'Affine' or a 'Merging'
// => do not split the cases, just handle the bias ...
- if (node->type() == "Scaling")
+ if (node->attributes()->hasAttr("quantization.ptq.isScaling"))
{
+
// retrieve the previous scaling factor ...
std::shared_ptr<Node> prevNode = node->getParent(0);
double prevScalingFactor = scalingFactors[prevNode->name()];
@@ -598,8 +668,7 @@ void normalizeActivations(std::shared_ptr<GraphView> graphView, std::map<std::st
// ValueRanges must contains all the scaling nodes !!!
double scalingFactor = valueRanges[node->name()];
- double currScalingFactor = getScalingFactor(node);
- updateScalingFactor(node, currScalingFactor / (scalingFactor / prevScalingFactor));
+ multiplyScalingFactor(node,1/(scalingFactor / prevScalingFactor));
scalingFactors[node->name()] = scalingFactor;
@@ -607,11 +676,13 @@ void normalizeActivations(std::shared_ptr<GraphView> graphView, std::map<std::st
if (isAffine(prevNode))
{
+
bool prevNodeHasBias = nodeHasBias(prevNode);
if (prevNodeHasBias)
- {
+ {
std::shared_ptr<Tensor> biasTensor = getBiasTensor(prevNode);
- rescaleTensor(biasTensor, 1.0 / prevScalingFactor);
+ //rescaleTensor(biasTensor, 1.0 / prevScalingFactor);
+ insertScalingBelowProducer(prevNode->getParent(2),1.0 / prevScalingFactor,graphView);
}
}
}
@@ -641,9 +712,8 @@ void normalizeActivations(std::shared_ptr<GraphView> graphView, std::map<std::st
std::shared_ptr<Node> scalingNode = getPreviousScalingNode(mergingNode);
//Log::info(" SCALING NODE : {} {}", scalingNode->type(), scalingNode->name());
-
- double currScalingFactor = getScalingFactor(scalingNode);
- updateScalingFactor(scalingNode, currScalingFactor * rescaling);
+
+ multiplyScalingFactor(scalingNode,rescaling) ;
}
}
}
@@ -679,7 +749,7 @@ std::map<std::string, std::pair<bool, bool>> computeSignMap(std::shared_ptr<Grap
signMap[node->name()].second = false;
}
- if (node->type() == "Scaling")
+ if (node->attributes()->hasAttr("quantization.ptq.isScaling"))
{
signMap[node->name()].second = false;
@@ -726,7 +796,7 @@ std::map<std::string, std::pair<bool, bool>> computeSignMap(std::shared_ptr<Grap
// Arbitration : Signed type wins !
for(std::shared_ptr<Node> parent : parentNodes)
{
- while (parent->type() != "Scaling")
+ while (!parent->attributes()->hasAttr("quantization.ptq.isScaling"))
{
signMap[parent->name()] = std::make_pair(false, false);
// We are on a branch so nodes always have 1 parent ...
@@ -808,26 +878,23 @@ void quantizeNormalizedNetwork(std::shared_ptr<GraphView> graphView, std::uint8_
if (isAffine(node))
{
// Rescale the weight tensor
-
std::shared_ptr<Tensor> weightTensor = getWeightTensor(node);
- rescaleTensor(weightTensor, signedMax);
+ insertScalingBelowProducer(node->getParent(1),signedMax,graphView);
if (!noQuant)
- roundTensor(weightTensor);
+ insertRoundBelowProducer(node->getParent(1),graphView);
// Rescale the bias tensor
-
if (nodeHasBias(node))
{
bool inputIsUnsigned = signMap[node->name()].first;
double rescaling = inputIsUnsigned ? unsignedMax * signedMax : signedMax * signedMax;
-
-
+
std::shared_ptr<Tensor> biasTensor = getBiasTensor(node);
- rescaleTensor(biasTensor, rescaling);
+ insertScalingBelowProducer(node->getParent(2),rescaling,graphView);
if (!noQuant)
- roundTensor(biasTensor);
+ insertRoundBelowProducer(node->getParent(2),graphView);
}
// Compensate the rescaling using the next Scaling node
@@ -842,8 +909,7 @@ void quantizeNormalizedNetwork(std::shared_ptr<GraphView> graphView, std::uint8_
std::shared_ptr<Node> scalingNode = *(node->getChildren().begin()); // Assert if scalingNode is a Scaling ...
- double currScalingFactor = getScalingFactor(scalingNode);
- updateScalingFactor(scalingNode, currScalingFactor * rescaling);
+ multiplyScalingFactor(scalingNode,rescaling) ;
}
if (isMerging(node))
@@ -858,23 +924,25 @@ void quantizeNormalizedNetwork(std::shared_ptr<GraphView> graphView, std::uint8_
std::shared_ptr<Node> scalingNode = *(node->getChildren().begin()); // Assert if scalingNode is a Scaling ...
- double currScalingFactor = getScalingFactor(scalingNode); // XXX bad naming
- updateScalingFactor(scalingNode, currScalingFactor * rescaling);
+ multiplyScalingFactor(scalingNode,rescaling) ;
}
// Handle the Scaling Nodes ...
- if (node->type() == "Scaling")
+ if (node->attributes()->hasAttr("quantization.ptq.isScaling"))
{
if (!noQuant)
{
// Replace the Scaling Node by Quantizer
+ auto scalingFactorTensor = std::static_pointer_cast<OperatorTensor>(node->getOperator())->getInput(1);
+ std::shared_ptr<Tensor> fallback;
+ const Tensor& localTensor = scalingFactorTensor->refCastFrom(fallback, DataType::Float64, "cpu");
+ double old_sf = localTensor.get<double>(0);//!\\
- std::shared_ptr<Node> quantizerNode = Quantizer(getScalingFactor(node), -(signedMax + 1), signedMax, node->name());
+ std::shared_ptr<Node> quantizerNode = Quantizer(old_sf, -(signedMax + 1), signedMax, node->name());
quantizerNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
- quantizerNode->getOperator()->setBackend("cpu");
-
- graphView->replace({node}, {quantizerNode});
+ quantizerNode->getOperator()->setBackend(node->getOperator()->backend());
+ graphView->replace({node,node->getParent(1)}, {quantizerNode});
if (optimizeSigns)
{
@@ -888,6 +956,7 @@ void quantizeNormalizedNetwork(std::shared_ptr<GraphView> graphView, std::uint8_
double currScalingFactor = getScalingFactor(quantizerNode);
updateScalingFactor(quantizerNode, currScalingFactor * rescaling);
+
if(outputIsUnsigned)
{
@@ -910,41 +979,37 @@ static void insertCompensationNodes(std::shared_ptr<GraphView> graphView, std::u
{
// A merging node is always followed by a Quantizer node at this point
- if (node->type() == "Quantizer")
+ if (node->type() == "Quantizer" && (node->attributes()->hasAttr("quantization.ptq.isResidual") || !isAffine(node->getParent(0))))
{
- // check if the Quantizer is a residual one, and insert a compensation node if so ...
- bool prevNodeIsForking = ((node->getParent(0))->getChildren().size() > 1);
- bool prevNodeIsAffine = isAffine(node->getParent(0));
- bool insertNode = prevNodeIsForking || !prevNodeIsAffine;
-
- if (insertNode)
- {
- // create and insert the multplicative node before the Quantizer
+ // check if the Quantizer is a residual one, and insert a compensation node if so ...
+ // create and insert the multplicative node before the Quantizer
- std::string mulNodeName = makeUniqueName(node->name() + "_Mul", graphView);
- std::shared_ptr<Node> mulNode = Mul(mulNodeName);
- mulNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
- mulNode->getOperator()->setBackend("cpu");
+ std::string mulNodeName = makeUniqueName(node->name() + "_Mul", graphView);
+ std::shared_ptr<Node> mulNode = Mul(mulNodeName);
+
+ mulNode->attributes()->addAttr("quantization.ptq.isCompensation",0.0);
+ mulNode->getOperator()->setDataType(DataType::Float64); // getDataType(parentNode)
+ mulNode->getOperator()->setBackend(node->getOperator()->backend());
- graphView->insertParent(node, mulNode, 0, 0, 0);
+ graphView->insertParent(node, mulNode, 0, 0, 0);
- // Add the coeff producer to the multiplier node
+ // Add the coeff producer to the multiplier node
- std::shared_ptr<Node> coeffProducer = addProducer(mulNode, 1, {1}, "");
- std::shared_ptr<Tensor> coeffTensor = std::make_shared<Tensor>(signedMax);
- coeffProducer->getOperator()->setOutput(0, coeffTensor);
+ std::shared_ptr<Node> coeffProducer = addProducer(mulNode, 1, {1}, "");
+ std::shared_ptr<Tensor> coeffTensor = std::make_shared<Tensor>(Array1D<double, 1> {signedMax});
+ coeffProducer->getOperator()->setOutput(0, coeffTensor);
- coeffProducer->getOperator()->setDataType(DataType::Float64);
- coeffProducer->getOperator()->setBackend("cpu");
+ coeffProducer->getOperator()->setDataType(DataType::Float64);
+ coeffProducer->getOperator()->setBackend(node->getOperator()->backend());
- graphView->add(coeffProducer); // needed ?
+ graphView->add(coeffProducer); // needed ?
- // Adapt the scaling factor value accordingly
+ // Adapt the scaling factor value accordingly
- double currScalingFactor = getScalingFactor(node);
- updateScalingFactor(node, currScalingFactor / signedMax);
- }
+ double currScalingFactor = getScalingFactor(node);
+ updateScalingFactor(node, currScalingFactor / signedMax);
+
}
}
}
@@ -955,9 +1020,7 @@ void performSingleShiftApproximation(std::shared_ptr<GraphView> graphView, bool
for (std::shared_ptr<Node> node : nodeVector)
{
- // TODO : use Compensation nodes instead of Mul nodes
-
- if (isAffine(node) || (node->type() == "Mul"))
+ if (isAffine(node) || (node->type() == "Mul" && node->attributes()->hasAttr("quantization.ptq.isCompensation")))
{
std::shared_ptr<Node> scalingNode = (*node->getChildren().begin());
@@ -965,21 +1028,20 @@ void performSingleShiftApproximation(std::shared_ptr<GraphView> graphView, bool
double approx = std::pow(2, std::ceil(std::log2(base)));
- updateScalingFactor(scalingNode, approx);
+ updateScalingFactor(scalingNode,approx);
double ratio = base / approx;
- std::shared_ptr<Tensor> weightTensor = getWeightTensor(node);
- rescaleTensor(weightTensor, ratio);
+ insertScalingBelowProducer(node->getParent(1),ratio,graphView);
if (!noQuant)
- roundTensor(weightTensor);
+ insertRoundBelowProducer(node->getParent(1),graphView);
if (nodeHasBias(node))
{
- std::shared_ptr<Tensor> biasTensor = getBiasTensor(node);
- rescaleTensor(biasTensor, ratio);
+ insertScalingBelowProducer(node->getParent(2),ratio,graphView);
+
if (!noQuant)
- roundTensor(biasTensor);
+ insertRoundBelowProducer(node->getParent(2),graphView);
}
}
}
@@ -988,7 +1050,7 @@ void performSingleShiftApproximation(std::shared_ptr<GraphView> graphView, bool
static void printScalingFactors(std::shared_ptr<GraphView> graphView)
{
for (auto node : retrieveNodeVector(graphView))
- if (node->type() == "Scaling" || node->type() == "Quantizer")
+ if (node->attributes()->hasAttr("quantization.ptq.isScaling") || node->type() == "Quantizer")
{
double scalingFactor = getScalingFactor(node);
Log::info(" {:.6f} ({})", scalingFactor, node->name());
@@ -1010,18 +1072,6 @@ static void setupDataType(std::shared_ptr<GraphView> graphView, std::vector<std:
tensor->setDataType(dataType);
}
-static void printRanges(std::shared_ptr<GraphView> graphView, std::map<std::string, double> valueRanges)
-{
- SequentialScheduler scheduler(graphView);
- scheduler.resetScheduling();
- scheduler.generateScheduling();
-
- auto scheduling = scheduler.getStaticScheduling();
- for (auto node : scheduling)
- if (node->type() == "Scaling")
- Log::info(" {} range = {} ", node->name(), valueRanges[node->name()]);
-}
-
void quantizeNetwork(std::shared_ptr<GraphView> graphView, std::uint8_t nbBits, std::vector<std::shared_ptr<Tensor>> inputDataSet, Clipping clippingMode, bool noQuant, bool optimizeSigns, bool singleShift, bool useCuda, bool verbose)
{
Log::info(" === QUANT PTQ 0.2.21 === ");
@@ -1041,7 +1091,6 @@ void quantizeNetwork(std::shared_ptr<GraphView> graphView, std::uint8_t nbBits,
insertScalingNodes(graphView);
crossLayerEqualization(graphView);
-
Log::info(" Normalizing the parameters ...");
normalizeParameters(graphView);
@@ -1049,14 +1098,12 @@ void quantizeNetwork(std::shared_ptr<GraphView> graphView, std::uint8_t nbBits,
std::map<std::string, double> valueRanges = computeRanges(graphView, inputDataSet, true, useCuda);
//Log::info(" === RANGES (BEFORE ADJUST) ===");
- //printRanges(graphView, valueRanges);
Log::info(" Optimizing the clipping values ...");
valueRanges = adjustRanges(clippingMode, valueRanges, nbBits, graphView, inputDataSet, useCuda, verbose);
- //Log::info(" === RANGES (AFTER ADJUST) ===");
+ //Log:debug("=== RANGES (AFTER ADJUST) ===");
//printRanges(graphView, valueRanges);
-
Log::info(" Normalizing the activations ...");
normalizeActivations(graphView, valueRanges);
@@ -1075,17 +1122,9 @@ void quantizeNetwork(std::shared_ptr<GraphView> graphView, std::uint8_t nbBits,
if (verbose)
printScalingFactors(graphView);
- //Log::info(" === SCALINGS (BEFORE CAST) ===");
- //printScalingFactors(graphView);
-
- setupDataType(graphView, inputDataSet, initialDataType);
-
if (useCuda)
graphView->setBackend("cuda");
- //Log::info(" === SCALINGS (AFTER CAST) ===");
- //printScalingFactors(graphView);
-
Log::info(" Reseting the scheduler ...");
SequentialScheduler scheduler(graphView);
scheduler.resetScheduling();
@@ -1115,15 +1154,9 @@ void clearBiases(std::shared_ptr<GraphView> graphView)
for (std::shared_ptr<Node> node : graphView->getNodes()) {
if (node->type() == "FC" || node->type() == "Conv2D") {
std::shared_ptr<Tensor> biasTensor = std::static_pointer_cast<OperatorTensor>(node->getOperator())->getInput(2);
- rescaleTensor(biasTensor, 0);
+ //rescaleTensor(biasTensor, 0);
+ insertScalingBelowProducer(node->getParent(2),0,graphView);
}
}
}
-
-void devPTQ(std::shared_ptr<GraphView> graphView)
-{
- for (std::shared_ptr<Node> node : graphView->getNodes())
- Log::info(" UUU : {}", node->name());
-}
-
}
diff --git a/src/QAT/QAT_LSQ.cpp b/src/QAT/QAT_LSQ.cpp
index 6eae077b060027eb4029f6b59f55376a1674df70..9b51e846df498a9303b7373ae1c86d4b007a96f0 100644
--- a/src/QAT/QAT_LSQ.cpp
+++ b/src/QAT/QAT_LSQ.cpp
@@ -21,152 +21,193 @@
#include "aidge/graph/Matching.hpp"
#include "aidge/recipes/QuantRecipes.hpp"
+namespace Aidge {
-namespace Aidge
+void QuantLSQ::insertQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits, float stepSize)
{
+ const auto matches = SinglePassGraphMatching(graphView).match("(Conv2D#|FC#)");
-static float getTensorAbsMean(std::shared_ptr<Tensor> tensor)
-{
- auto valueTensor = (*tensor).abs().mean();
- std::shared_ptr<Tensor> fallback;
- const Tensor& localTensor = valueTensor.refCastFrom(fallback, DataType::Float32, "cpu");
- return localTensor.get<float>(0);
-}
+ for (const auto& match : matches)
+ {
+ auto linearNode = match.graph->rootNode();
-static float getTensorStd(std::shared_ptr<Tensor> tensor)
-{
- auto valueTensor = (*tensor);
-
- auto skewedTensor = valueTensor - valueTensor.mean();
- auto squaredTensor = skewedTensor * skewedTensor;
- auto varianceTensor = squaredTensor.mean();
+ std::pair<int, int> signedRange = {-std::pow(2, nbBits - 1), std::pow(2, nbBits - 1) - 1};
+ std::pair<int, int> unsignedRange = {0, std::pow(2, nbBits) - 1};
- std::shared_ptr<Tensor> fallback;
- auto localTensor = varianceTensor.refCastFrom(fallback, DataType::Float32, "cpu");
-
- float variance = localTensor.get<float>(0);
- return std::sqrt(variance);
-}
+ // INPUT QUANTIZERS INSERTION
+ // TODO : double check this, and use createUniqueName()
+ auto inputQuantizerName = makeUniqueName(linearNode->name() + "_lsq_i", graphView);
+ auto inputQuantizerNode = LSQ(signedRange, inputQuantizerName);
-// INIT THE STEP SIZE OF A QUANTIZER NODE
+ // Set the step size
-static bool initStepSize(std::shared_ptr<Node> quantizer)
-{
- const auto quantizerOp = std::static_pointer_cast<LSQ_Op>(quantizer->getOperator());
+ auto inputStepSizeOp = inputQuantizerNode->getParent(1)->getOperator();
+ auto inputStepSizeTensor = std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}}));
+ inputStepSizeOp->setOutput(0, inputStepSizeTensor);
- // This formula is the one proposed in the paper ...
+ // Absorb the ReLU when possible ...
- // float inputAbsMean = getTensorAbsMean(quantizerOp->getInput(0));
- // float stepSize = 2.0f * (inputAbsMean / std::sqrt(quantizerOp->range().second));
+ // XXX is this safe ???
+ bool nodeHasParent = static_cast<bool> (linearNode->getParents()[0]);
+ // bool nodeHasParent = (linearNode->getParents().size() != 0);
- // .. but this formula seems to work better !!!
+ if (nodeHasParent) {
+ auto parentNode = linearNode->getParents()[0];
+ if (parentNode->type() == "ReLU") {
+ auto inputQuantizerOp = std::static_pointer_cast<LSQ_Op> (inputQuantizerNode->getOperator());
+ inputQuantizerOp->range() = unsignedRange;
+ graphView->replace({parentNode}, {});
+ }
+ }
- float inputStd = getTensorStd(quantizerOp->getInput(0));
- float stepSize = 8.0f * (inputStd / (quantizerOp->range().second));
+ // We need to handle the case where the linear node is the first one ...
- // TODO : use the scalar constructor
- auto stepSizeTensor = std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}}));
+ if (nodeHasParent) {
+ graphView->insertParent(linearNode, inputQuantizerNode, 0, 0, 0);
+ } else {
+ inputQuantizerNode->addChild(graphView);
+ graphView->add(inputQuantizerNode);
+ }
- // XXX Manage backend here ?
- stepSizeTensor->setBackend(quantizerOp->getInput(0)->backend());
- stepSizeTensor->setDataType(quantizerOp->getInput(0)->dataType());
+ // PARAM QUANTIZERS INSERTION
- auto stepSizeProducer = quantizer->getParent(1);
+ // TODO : double check this, and use createUniqueName()
+ auto paramQuantizerName = makeUniqueName(linearNode->name() + "_lsq_p", graphView);
+ auto paramQuantizerNode = LSQ(signedRange, paramQuantizerName);
+ graphView->insertParent(linearNode, paramQuantizerNode, 1, 0, 0);
- stepSizeProducer->getOperator()->setOutput(0, stepSizeTensor);
+ // Set the step size
- Log::notice(" [ INIT STEP SIZE = {} ] ", stepSize);
+ auto paramStepSizeOp = paramQuantizerNode->getParent(1)->getOperator();
+ auto paramStepSizeTensor = std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}}));
+ paramStepSizeOp->setOutput(0, paramStepSizeTensor);
+ }
- return false;
}
-static void setupInputQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits)
+static float getTensorAbsMean(std::shared_ptr<Tensor> tensor)
{
- const auto matches = SinglePassGraphMatching(graphView).match("(Conv2D#|PaddedConv2D#|FC#)");
+ auto backend = tensor->backend();
+ if (backend == "cuda")
+ tensor->setBackend("cpu");
- for (const auto& match : matches)
- {
- auto linearNode = match.graph->rootNode();
+ float acc = 0;
+ float* castedTensor = static_cast<float *> (tensor->getImpl()->rawPtr());
+ for(std::size_t i = 0; i < tensor->size(); i++)
+ acc += std::abs(castedTensor[i]);
+ acc /= static_cast<float> (tensor->size());
- // Log::notice(" SET INPUT QUANTIZER : {} ", linearNode->type());
-
- std::pair<int, int> signedRange = {-std::pow(2, nbBits - 1), std::pow(2, nbBits - 1) - 1};
- std::pair<int, int> unsignedRange = {0, std::pow(2, nbBits) - 1};
+ if (backend == "cuda")
+ tensor->setBackend("cuda");
- // Create the input quantizer node
-
- auto quantizerName = makeUniqueName(linearNode->name() + "_lsq_i", graphView);
- auto quantizerNode = LSQ(signedRange, quantizerName);
+ return acc;
+}
- // Init the step-size using the node call stack
+static std::map<std::string, float> collectInputStats(std::shared_ptr<GraphView> graphView, std::shared_ptr<Tensor> calibrationData, bool useCuda)
+{
+ // Propagate the calibration tensor
- quantizerNode->addBeforeForward([quantizerNode](){ return initStepSize(quantizerNode); });
+ SequentialScheduler scheduler(graphView);
+ scheduler.resetScheduling();
+ scheduler.forward(true, {calibrationData});
- // Absorb the ReLU when possible ...
+ // Store the input tensor statistics
- bool nodeHasParent = static_cast<bool> (linearNode->getParents()[0]); // XXX is this safe ?
+ if (useCuda)
+ graphView->setBackend("cpu");
- if (nodeHasParent)
+ std::map<std::string, float> inputStats;
+ for (auto node : graphView->getNodes())
+ {
+ if (node->type() == "FC" || node->type() == "Conv2D") // TODO: use graph matching !!!
{
- bool allParentsAreReLU = true;
- for (auto parentNode : linearNode->getParents())
- if (parentNode->type() != "ReLU")
- allParentsAreReLU = false;
-
- if (allParentsAreReLU) {
- auto quantizerOp = std::static_pointer_cast<LSQ_Op> (quantizerNode->getOperator());
- quantizerOp->range() = unsignedRange;
- }
-
- // TODO : remove the ReLUs when possible
+ const auto op = std::static_pointer_cast<LSQ_Op>(node->getOperator());
+ float inputAbsMean = getTensorAbsMean(op->getInput(0));
+ inputStats.insert(std::make_pair(node->name(), inputAbsMean));
+ fmt::println("{} -> {}", node->name(), inputAbsMean);
}
+ }
- // Insert the quantizer in the graphView ...
- // (We need to handle the case where the linear node is the first one)
+ if (useCuda)
+ graphView->setBackend("cuda");
- if (nodeHasParent) {
- graphView->insertParent(linearNode, quantizerNode, 0, 0, 0);
- } else {
- quantizerNode->addChild(graphView);
- graphView->add(quantizerNode);
+ return inputStats;
+}
+
+static std::map<std::string, float> collectParamStats(std::shared_ptr<GraphView> graphView, bool useCuda)
+{
+ if (useCuda)
+ graphView->setBackend("cpu");
+
+ std::map<std::string, float> paramStats;
+ for (auto node : graphView->getNodes())
+ {
+ if (node->type() == "FC" || node->type() == "Conv2D") // TODO: use graph matching !!!
+ {
+ const auto op = std::static_pointer_cast<LSQ_Op>(node->getOperator());
+ float paramAbsMean = getTensorAbsMean(op->getInput(1));
+ paramStats.insert(std::make_pair(node->name(), paramAbsMean));
+ fmt::println("{} -> {}", node->name(), paramAbsMean);
}
}
-}
+
+ if (useCuda)
+ graphView->setBackend("cuda");
-// PARAM QUANTIZERS INSERTION
+ return paramStats;
+}
-static void setupParamQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits)
+static void adjustQuantizersStepSizes(std::shared_ptr<GraphView> graphView, std::map<std::string, float> inputStats, std::map<std::string, float> paramStats)
{
- const auto matches = SinglePassGraphMatching(graphView).match("(Conv2D#|PaddedConv2D#|FC#)");
-
- std::pair<int, int> signedRange = {-std::pow(2, nbBits - 1), std::pow(2, nbBits - 1) - 1};
+ const auto matches = SinglePassGraphMatching(graphView).match("(Conv2D#|FC#)");
for (const auto& match : matches)
- {
- auto linearNode = match.graph->rootNode();
+ {
+ auto linearNode = match.graph->rootNode();
- // Log::notice(" SET PARAM QUANTIZER : {} ", linearNode->type());
+ // INPUT QUANTIZERS STEP-SIZES
- // TODO : double check this, and use createUniqueName()
- auto quantizerName = makeUniqueName(linearNode->name() + "_lsq_p", graphView);
- auto quantizerNode = LSQ(signedRange, quantizerName);
+ auto inputQuantNode = linearNode->getParent(0);
+ auto inputQuantOp = std::static_pointer_cast<LSQ_Op>(inputQuantNode->getOperator());
+
+ float absMean = inputStats[linearNode->name()];
+ float stepSize = 2.0f * (absMean / std::sqrt(inputQuantOp->range().second));
- // Init the step-size using the node call stack
+ auto inputStepSizeOp = inputQuantNode->getParent(1)->getOperator();
+ // XXX inputStepSizeOp->setOutput(0, std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}})));
+ auto inputStepSizeTensor = std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}}));
+ inputStepSizeOp->setOutput(0, inputStepSizeTensor);
- quantizerNode->addBeforeForward([quantizerNode](){ return initStepSize(quantizerNode); });
+ // PARAM QUANTIZERS STEP-SIZES
- // Insert the quantizer in the graphView
+ auto paramQuantNode = linearNode->getParent(1);
+ auto paramQuantOp = std::static_pointer_cast<LSQ_Op>(paramQuantNode->getOperator());
- graphView->insertParent(linearNode, quantizerNode, 1, 0, 0);
+ absMean = paramStats[linearNode->name()];
+ stepSize = 2.0f * (absMean / std::sqrt(paramQuantOp->range().second));
+
+ auto paramStepSizeOp = paramQuantNode->getParent(1)->getOperator();
+ // XXX paramStepSizeOp->setOutput(0, std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}})));
+ auto paramStepSizeTensor = std::make_shared<Tensor>(Array1D<float, 1>({{stepSize}}));
+ paramStepSizeOp->setOutput(0, paramStepSizeTensor);
}
}
-void QuantLSQ::setupQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits)
+void QuantLSQ::insertAndInitQuantizers(std::shared_ptr<GraphView> graphView, size_t nbBits, std::shared_ptr<Tensor> calibrationData)
{
- sanitizeNodeNames(graphView);
- setupInputQuantizers(graphView, nbBits);
- setupParamQuantizers(graphView, nbBits);
+ bool useCuda = (calibrationData->backend() == "cuda");
+
+ // Collect the tensor statisics
+ auto inputStats = collectInputStats(graphView, calibrationData, useCuda);
+
+ auto paramStats = collectParamStats(graphView, useCuda);
+
+ // Insert the quantizers
+ insertQuantizers(graphView, nbBits, 1.0);
+
+ // Adjust the quantizers step-sizes
+ adjustQuantizersStepSizes(graphView, inputStats, paramStats);
}
}
\ No newline at end of file
diff --git a/src/operator/PTQMetaOps.cpp b/src/operator/PTQMetaOps.cpp
index 56245da47076d8930ce29ab75e549d97d0d7493d..f86d454245a7fe088edd027732a91f5775cd2acf 100644
--- a/src/operator/PTQMetaOps.cpp
+++ b/src/operator/PTQMetaOps.cpp
@@ -60,23 +60,6 @@ std::shared_ptr<Node> Quantizer(double scalingFactor, double clipMin, double cli
return metaopNode;
}
-std::shared_ptr<Node> Scaling(double scalingFactor, const std::string& name)
-{
- std::shared_ptr<Tensor> scalingFactorTensor = std::make_shared<Tensor>(Array1D<double, 1> {scalingFactor});
-
- std::shared_ptr<Node> mulNode = Mul((!name.empty()) ? name + "_Scaling" : "");
-
- std::shared_ptr<Node> scalingFactorProducer = addProducer<1>(mulNode, 1, {1}, "ScalingFactor");
- scalingFactorProducer->getOperator()->setOutput(0, scalingFactorTensor);
-
- std::shared_ptr<GraphView> graphView = Sequential({mulNode});
- std::shared_ptr<GraphView> connectedGraphView = getConnectedGraphView(mulNode);
-
- NodePtr metaopNode = MetaOperator("Scaling", connectedGraphView, {}, name);
-
- return metaopNode;
-}
-
static std::shared_ptr<Node> getSubNode(std::shared_ptr<GraphView> graphView, std::string nodeType)
{
std::shared_ptr<Node> mulNode = nullptr;
@@ -87,10 +70,12 @@ static std::shared_ptr<Node> getSubNode(std::shared_ptr<GraphView> graphView, st
return mulNode;
}
+
+
void updateScalingFactor(std::shared_ptr<Node> metaOpNode, double scalingFactor)
{
if(metaOpNode->type() != "Scaling" && metaOpNode->type() != "Quantizer")
- Log::warn(" Cannot update the scaling factor on Node of type {}", metaOpNode->type());
+ Log::warn("Cannot update the scaling factor on Node of type {}", metaOpNode->type());
std::shared_ptr<Tensor> scalingFactorTensor = std::make_shared<Tensor>(Array1D<double, 1> {scalingFactor});
@@ -99,7 +84,7 @@ void updateScalingFactor(std::shared_ptr<Node> metaOpNode, double scalingFactor)
std::shared_ptr<Node> mulNode = getSubNode(metaOp->getMicroGraph(), "Mul");
if (!mulNode)
- Log::warn(" Invalid PTQ MetaOperator, no Mul node found inside ! ");
+ Log::warn("Invalid PTQ MetaOperator, no Mul node found inside ! ");
mulNode->input(1).first->getOperator()->setOutput(0, scalingFactorTensor);
}
@@ -107,7 +92,7 @@ void updateScalingFactor(std::shared_ptr<Node> metaOpNode, double scalingFactor)
double getScalingFactor(std::shared_ptr<Node> MetaOpNode)
{
if (MetaOpNode->type() != "Scaling" && MetaOpNode->type() != "Quantizer") {
- Log::warn(" Cannot get the scaling factor on Node of type {}", MetaOpNode->type());
+ Log::warn("Cannot get the scaling factor on Node of type {}", MetaOpNode->type());
return 0;
}
@@ -116,7 +101,7 @@ double getScalingFactor(std::shared_ptr<Node> MetaOpNode)
std::shared_ptr<Node> mulNode = getSubNode(metaOp->getMicroGraph(), "Mul");
if (!mulNode) {
- Log::warn(" Invalid PTQ MetaOperator, no Mul found inside node of type {}", MetaOpNode->type());
+ Log::warn("Invalid PTQ MetaOperator, no Mul found inside node of type {}", MetaOpNode->type());
return 0;
}
@@ -131,7 +116,7 @@ double getScalingFactor(std::shared_ptr<Node> MetaOpNode)
void setClipRange(std::shared_ptr<Node> quantizerNode, double min, double max)
{
if (quantizerNode->type() != "Quantizer") {
- Log::warn(" Cannot set the clipping range on Node of type {}", quantizerNode->type());
+ Log::warn("Cannot set the clipping range on Node of type {}", quantizerNode->type());
return;
}
@@ -140,7 +125,7 @@ void setClipRange(std::shared_ptr<Node> quantizerNode, double min, double max)
std::shared_ptr<Node> clipNode = getSubNode(metaOp->getMicroGraph(), "Clip");
if (!clipNode) {
- Log::warn(" Invalid PTQ MetaOperator, no Clip found inside node of type {}", quantizerNode->type());
+ Log::warn("Invalid PTQ MetaOperator, no Clip found inside node of type {}", quantizerNode->type());
return;
}