From 64fbc46a258fcb5689f46d20da7c542bec9be1da Mon Sep 17 00:00:00 2001
From: cmoineau <cyril.moineau@cea.fr>
Date: Mon, 23 Oct 2023 07:58:05 +0000
Subject: [PATCH] Update hook, add comment and fix minor compilation issues.
---
aidge_quantization/__init__.py | 1 +
aidge_quantization/unit_tests/test_ptq.py | 20 +-
include/aidge/QuantPTQ.hpp | 1 -
python_binding/pybind_QuantPTQ.cpp | 17 +-
src/QuantPTQ.cpp | 324 +++++-----------------
unit_tests/hook/Test_execTime.cpp | 8 +-
unit_tests/hook/Test_outputRange.cpp | 8 +-
7 files changed, 84 insertions(+), 295 deletions(-)
create mode 100644 aidge_quantization/__init__.py
diff --git a/aidge_quantization/__init__.py b/aidge_quantization/__init__.py
new file mode 100644
index 0000000..b00fae1
--- /dev/null
+++ b/aidge_quantization/__init__.py
@@ -0,0 +1 @@
+from aidge_quantization.aidge_quantization import * # import so generated by PyBind
diff --git a/aidge_quantization/unit_tests/test_ptq.py b/aidge_quantization/unit_tests/test_ptq.py
index e2e5180..519fb10 100644
--- a/aidge_quantization/unit_tests/test_ptq.py
+++ b/aidge_quantization/unit_tests/test_ptq.py
@@ -2,12 +2,9 @@
import aidge_core
import aidge_backend_cpu
import aidge_onnx
+import aidge_quantization
import numpy as np
import matplotlib.pyplot as plt
-from jinja2 import Environment, FileSystemLoader
-import os
-import pathlib
-import onnx
model = aidge_onnx.load_onnx("MNIST_model/LeNet_MNIST.onnx")
@@ -29,18 +26,7 @@ input_node.get_operator().set_backend("cpu")
## Link Producer to the Graph
input_node.add_child(model)
-nodesRegexes = {}
-nodesRegexes["Flatten"] = aidge_core.NodeRegex("Flatten")
-
-# Graph Regex
-graphRegex = ["Flatten;"]
-
-graphMatching = aidge_core.GRegex(nodesRegexes, graphRegex)
-all_match = graphMatching.match(model)
-print('Number of match : ', all_match.get_nb_match())
-
-for mn in all_match.get_match_nodes():
- aidge_core.remove_flatten(mn)
+aidge_core.remove_flatten(model)
model.save("my_supported_LeNet")
@@ -73,4 +59,4 @@ for outNode in model.get_output_nodes():
#add PTQ part for testing
#created ordered graph, as in cpp and do
-#quantizeNetwork(ordered_graph_view, 8, verbose);
\ No newline at end of file
+aidge_quantization.quantizeNetwork(scheduler.get_static_scheduling(), 8, True)
diff --git a/include/aidge/QuantPTQ.hpp b/include/aidge/QuantPTQ.hpp
index e1efa9a..21c78f6 100644
--- a/include/aidge/QuantPTQ.hpp
+++ b/include/aidge/QuantPTQ.hpp
@@ -32,7 +32,6 @@ namespace Aidge{
float getCellThreshold(std::shared_ptr<Node> node);
float getMaxParentsScaling(std::shared_ptr<Node> node);
void rescaleParentsToScaling(std::shared_ptr<Node> node, std::unordered_map<std::string, long double>& scalingForCells, long double scaling);
-
long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, std::unordered_map<std::string, long double> biasScalings, bool verbose);
long double quantizeActivation(std::shared_ptr<Node> node, std::size_t nbBits, std::unordered_map<std::string, long double> biasScalings, std::unordered_map<std::string, long double> activationScalings, bool verbose);
void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::size_t nbBits, bool verbose);
diff --git a/python_binding/pybind_QuantPTQ.cpp b/python_binding/pybind_QuantPTQ.cpp
index c89605e..3722a86 100644
--- a/python_binding/pybind_QuantPTQ.cpp
+++ b/python_binding/pybind_QuantPTQ.cpp
@@ -15,20 +15,23 @@
#include <string>
#include "aidge/QuantPTQ.hpp"
+#include "aidge/hook/hook.hpp"
namespace py = pybind11;
namespace Aidge {
void init_QuantPTQ(py::module &m) {
m.def("quantize_network", &quantizeNetwork, py::arg("ordered_graph_view"), py::arg("nb_bits")=8, py::arg("verbose")=false, R"mydelimiter(
- void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::size_t nbBits, bool verbose){
- Parameters
- ----------
- std::vector<std::shared_ptr<Node>> orderedGraphView: Ordered Graph View,
- std::size_t nbBits : number of bits
- bool verbose : print debug messages
+ Recipie to Fuse MatMul and Add operators into an :py:class:`aidge_core.FC` operator.
+
+ :param ordered_graph_view: Graph view on which we want to apply the recipie
+ :type ordered_graph_view: List[:py:class:`aidge_core.Node`]
+ :param nb_bits: Number of bits to quantize on, default=8
+ :type nb_bits: int, optional
+ :param verbose: if True, log informations, default=False
+ :type verbose: bool, optional
)mydelimiter");
-
+
}
PYBIND11_MODULE(aidge_quantization, m) {
diff --git a/src/QuantPTQ.cpp b/src/QuantPTQ.cpp
index 7704183..528f2b5 100644
--- a/src/QuantPTQ.cpp
+++ b/src/QuantPTQ.cpp
@@ -18,6 +18,7 @@
#include <cmath>
#include <cstdint>
#include <unordered_map>
+#include <iostream>
#include "aidge/QuantPTQ.hpp"
//using namespace Aidge;
@@ -29,31 +30,23 @@
//namespace Aidge_HELPER{
namespace Aidge{
-//getMaxParentsScaling
-long double getMaxParentsScaling(std::shared_ptr<Node> node,
+long double getMaxParentsScaling(std::shared_ptr<Node> node,
std::unordered_map<std::string, long double>& scalingForCells){
long double maxParentsScaling = 0.0;
- //std::cout << " inside getMaxParentsScaling " << std::endl;
- const std::vector<std::shared_ptr<Node>> parentsNodes = node->getParents();
- //std::cout << "after parentsNodes " << std::endl;
- //std::cout << "size of parents nodes = " << parentsNodes.size() << std::endl;
+ const std::vector<std::shared_ptr<Node>> parentsNodes = node->getParents();
for(const std::shared_ptr<Node>& parentNode: parentsNodes) {
- //std::cout << " parent node type = " << parentNode->type() << std::endl;
- //std::cout << " parent node name = " << parentNode->name() << std::endl;
- const long double parentScaling = (parentNode->type()=="Producer" || scalingForCells.empty())?1.0:scalingForCells.at(parentNode->name());
- //std::cout << " parentScaling = " << parentScaling << std::endl;
+ const long double parentScaling = (
+ parentNode->type()=="Producer" || scalingForCells.empty()) ? 1.0 : scalingForCells.at(parentNode->name());
maxParentsScaling = std::max(maxParentsScaling, parentScaling);
- //std::cout << " maxParentsScaling = " << maxParentsScaling << std::endl;
assert(parentScaling > 0.0);
}
-
return maxParentsScaling;
}
//rescaleParentsToScaling
-void rescaleParentsToScaling(std::shared_ptr<Node> node,
+void rescaleParentsToScaling(std::shared_ptr<Node> node,
std::unordered_map<std::string, long double>& scalingForCells,
long double scaling)
{
@@ -74,10 +67,10 @@ void rescaleParentsToScaling(std::shared_ptr<Node> node,
//std::cout << "parentScaling = " << parentScaling << std::endl;
//std::cout << "scaling = " << scaling << std::endl;
-
+
assert(parentScaling < scaling);
- //IF parentScaling is smaller than scaling
+ //IF parentScaling is smaller than scaling
//create scaling operator and add it to the graphView
/*
std::shared_ptr<Node> scaling_op = Scaling(scaling, parentCell->name()+"rescale_branch");
@@ -85,7 +78,7 @@ void rescaleParentsToScaling(std::shared_ptr<Node> node,
scaling_op->getOperator()->setBackend("cpu");
*/
- //TODO : get all graphviews node is attached to and insert a scaling cell before it.
+ //TODO : get all graphviews node is attached to and insert a scaling cell before it.
/*
void Aidge::GraphView::insert(Node & newNode, Node & inNode,
std::initializer_list<Node> outNodes,
@@ -99,11 +92,11 @@ void rescaleParentsToScaling(std::shared_ptr<Node> node,
//N2D2 version
/*
auto scalingCell = Registrar<ScalingCell>::create<Float_T>(getCellModelType(*parentCell))
- (mDeepNet,
- mDeepNet.generateNewCellName(parentCell->getName() + "_rescale_branch"),
- parentCell->getNbOutputs(),
+ (mDeepNet,
+ mDeepNet.generateNewCellName(parentCell->getName() + "_rescale_branch"),
+ parentCell->getNbOutputs(),
Scaling::floatingPointScaling(
- std::vector<Float_T>(parentCell->getNbOutputs(),
+ std::vector<Float_T>(parentCell->getNbOutputs(),
parentScaling/scaling), false, std::vector<Float_T>(0.0f))
);
@@ -112,71 +105,17 @@ void rescaleParentsToScaling(std::shared_ptr<Node> node,
}
}
-//N2D2 version
-/*
-std::unordered_map<std::string, long double> N2D2::DeepNetQuantization::quantizeFreeParemeters(std::size_t nbBits) {
-#ifdef VERBOSE_QUANT
- std::cout << " Quantizing free parameters:" << std::endl;
-#endif
-
- std::unordered_map<std::string, long double> biasScalings;
-
- std::vector<std::vector<std::string>> layers = mDeepNet.getLayers();
- for (auto itLayer = layers.begin() + 1; itLayer != layers.end(); ++itLayer) {
- for (auto itCell = itLayer->begin(); itCell != itLayer->end(); ++itCell) {
- std::shared_ptr<Cell> cell = mDeepNet.getCell(*itCell);
- if(!cell) {
- throw std::runtime_error("Invalid cell.");
- }
-
- long double biasScaling = getMaxParentsScaling(cell, biasScalings);
- rescaleParentsToScaling(cell, biasScalings, biasScaling);
-
-
- const long double wQuantScaling = std::pow(2, nbBits - 1) - 1;
- const long double bQuantScaling = DeepNetExport::isCellInputsUnsigned(*cell, mDeepNet)?
- wQuantScaling*(std::pow(2, nbBits) - 1):
- wQuantScaling*(std::pow(2, nbBits - 1) - 1);
-
-
- const std::pair<Float_T, Float_T> wMinMax
- = cell->getFreeParametersRange(Cell::Multiplicative);
- const Float_T wScalingCell = Utils::max_abs(wMinMax.first, wMinMax.second);
- if(wScalingCell != 0.0) {
- cell->processFreeParameters([&](Float_T w) { return w*(wQuantScaling/wScalingCell); },
- Cell::Multiplicative);
-
- biasScaling *= wScalingCell;
- }
-
- cell->processFreeParameters([&](Float_T b) { return b*(bQuantScaling/biasScaling); },
- Cell::Additive);
- biasScalings[cell->getName()] = biasScaling;
-
-#ifdef VERBOSE_QUANT
- std::cout << " - " << cell->getName() << ": " << biasScaling
- << std::endl;
-#endif
- }
- }
-
- fuseScalingCells();
-
- return biasScalings;
-}
-*/
-
long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, std::unordered_map<std::string, long double> biasScalings, bool verbose){
-
+
long double biasScaling = getMaxParentsScaling(node, biasScalings);
rescaleParentsToScaling(node, biasScalings, biasScaling);
//weights are input 1
std::shared_ptr<Tensor> weight_tensor = node->getOperator()->getInput(1);
- if(verbose){
+ if(verbose){
printf("Weight init :\n");
- weight_tensor->print();
+ weight_tensor->print();
}
float max_value = 0.;
@@ -188,7 +127,7 @@ long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, s
}
const long double wQuantScaling = std::pow(2, nbBits - 1) - 1;
- //find out if the input to the cell is signed/unsigned, n2d2 example :
+ //TODO :find out if the input to the cell is signed/unsigned, n2d2 example :
//return cellFrame.getOutputsRange().first >= 0.0;
/*
virtual std::pair<double, double> getOutputsRange() const {
@@ -210,15 +149,15 @@ long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, s
static_cast<float *>(quant_weight_tensor->getImpl()->rawPtr())[i] = std::round(static_cast<float *>(weight_tensor->getImpl()->rawPtr())[i]*wQuantScaling/max_value);
}
- if(verbose){
+ if(verbose){
printf("Weight quantized :\n");
- quant_weight_tensor->print();
+ quant_weight_tensor->print();
}
std::shared_ptr<Tensor> bias_tensor = node->getOperator()->getInput(2);
- if(verbose){
+ if(verbose){
printf("Bias init :\n");
- bias_tensor->print();
+ bias_tensor->print();
}
std::shared_ptr<Tensor> quant_bias_tensor = std::static_pointer_cast<Tensor>(bias_tensor);
@@ -226,9 +165,9 @@ long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, s
static_cast<float *>(quant_bias_tensor->getImpl()->rawPtr())[i] = static_cast<float *>(bias_tensor->getImpl()->rawPtr())[i]*bQuantScaling/biasScaling;
}
- if(verbose){
+ if(verbose){
printf("Bias quantized :\n");
- quant_bias_tensor->print();
+ quant_bias_tensor->print();
}
//replace weights in the node
@@ -239,155 +178,14 @@ long double quantizeFreeParams(std::shared_ptr<Node> node, std::size_t nbBits, s
return biasScaling;
}
-//N2D2 version with Histogram
-/*
- double N2D2::DeepNetQuantization::getCellThreshold(const std::string& cellName,
- const std::unordered_map<std::string, Histogram>& outputsHistogram,
- const std::unordered_map<std::string, RangeStats>& outputsRange,
- std::size_t nbBits, ClippingMode actClippingMode, double quantileValue)
- {
- switch(actClippingMode) {
- case ClippingMode::KL_DIVERGENCE:
- return outputsHistogram.at(cellName).calibrateKLDivergence(nbBits);
- case ClippingMode::MSE:
- return outputsHistogram.at(cellName).calibrateMSE(nbBits);
- case ClippingMode::QUANTILE:
- return outputsHistogram.at(cellName).getQuantileValue(quantileValue);
- default: {
- const auto& range = outputsRange.at(cellName);
- return Utils::max_abs(range.minVal(), range.maxVal());
- }
- }
- }
-*/
+
float getCellThreshold(std::shared_ptr<Node> node){
float max_output = std::static_pointer_cast<OutputRange>(node->getOperator()->getHook("output_range"))->getOutput(0);
return max_output;
}
-//N2D2 quantizeActivation method
/*
-void N2D2::DeepNetQuantization::quantizeActivations(
- const std::unordered_map<std::string, Histogram>& outputsHistogram,
- const std::unordered_map<std::string, RangeStats>& outputsRange,
- std::unordered_map<std::string, long double>& biasScalings,
- std::size_t nbBits, ClippingMode actClippingMode, double quantileValue)
-{
-#ifdef VERBOSE_QUANT
- std::cout << " Quantizing activations:" << std::endl;
-#endif
-
- std::unordered_map<std::string, long double> activationScalings;
-
- std::vector<std::vector<std::string>> layers = mDeepNet.getLayers();
- for (auto itLayer = layers.begin() + 1; itLayer != layers.end(); ++itLayer) {
- for (auto itCell = itLayer->begin(); itCell != itLayer->end(); ++itCell) {
- std::shared_ptr<Cell> cell = mDeepNet.getCell(*itCell);
- std::shared_ptr<Cell_Frame_Top> cellFrame = std::dynamic_pointer_cast<Cell_Frame_Top>(cell);
- if(!cell || !cellFrame) {
- throw std::runtime_error("Invalid cell.");
- }
-
- const long double prevActivationScaling = getMaxParentsScaling(cell, activationScalings);
- rescaleParentsToScaling(cell, activationScalings, prevActivationScaling);
-
-
- long double activationScaling;
-
- const std::shared_ptr<Activation>& activation = cellFrame->getActivation();
- if(cell->getType() == ElemWiseCell::Type) {
- activationScaling = getCellThreshold(cell->getName(),
- outputsHistogram, outputsRange,
- nbBits, ClippingMode::NONE);
- }
- else if(cell->getType() == PaddingCell::Type ||
- cell->getType() == PoolCell::Type ||
- cell->getType() == ResizeCell::Type ||
- cell->getType() == ScalingCell::Type ||
- cell->getType() == SoftmaxCell::Type ||
- cell->getType() == TransposeCell::Type ||
- cell->getType() == ReshapeCell::Type)
- {
- activationScalings[cell->getName()] = prevActivationScaling;
- continue;
- }
- else if(activation) {
- const bool clip = cell->getNbOutputs() > 2 &&
- (activation->getType() == RectifierActivation::Type ||
- activation->getType() == LinearActivation::Type ||
- activation->getType() == SaturationActivation::Type);
-
-
- auto childrenCells = mDeepNet.getChildCells(cell->getName());
- const bool isNextCellMaxPool = childrenCells.size() == 1 &&
- childrenCells[0]->getType() == PoolCell::Type &&
- dynamic_cast<const PoolCell&>(*childrenCells[0]).getPooling() == PoolCell::Max;
-
-
- const std::string cellStatsName = clip && isNextCellMaxPool?childrenCells[0]->getName():
- cell->getName();
- activationScaling = getCellThreshold(cellStatsName,
- outputsHistogram, outputsRange,
- nbBits, clip?actClippingMode:ClippingMode::NONE, quantileValue);
- }
- else {
- throw std::runtime_error("Quantization of cell '" + cell->getName() + "' of type '" +
- cell->getType() + "' is not supported yet.");
- }
-
- const long double biasScaling = biasScalings.at(cell->getName());
-
-#ifdef VERBOSE_QUANT
- std::cout << " - " << cell->getName() << ": "
- << "prev=" << prevActivationScaling
- << ", act=" << activationScaling
- << ", bias=" << biasScaling << std::endl;
-#endif
-
- activationScaling /= biasScaling;
- activationScaling = (activationScaling == 0.0)?1.0:activationScaling;
-
- activationScalings[cell->getName()] = activationScaling;
-
- cell->processFreeParameters([&](Float_T d) { return d/prevActivationScaling; },
- Cell::Additive);
-
-
- const long double actQuantScaling = getActivationQuantizationScaling(*cell, nbBits);
- auto scalingCell = Registrar<ScalingCell>::create<Float_T>(getCellModelType(*cell))
- (mDeepNet,
- mDeepNet.generateNewCellName(cell->getName() + "_rescale_act"),
- cell->getNbOutputs(),
- Scaling::floatingPointScaling(
- std::vector<Float_T>(cell->getNbOutputs(),
- (prevActivationScaling/activationScaling)/actQuantScaling
- ),
- false,
- std::vector<Float_T>(0.0f)
- )
- );
-
- mDeepNet.addCellAfter(scalingCell, cell);
-
- activationScalings[scalingCell->getName()] = activationScalings[cell->getName()];
- biasScalings[scalingCell->getName()] = biasScaling;
-
-#ifdef VERBOSE_QUANT
- std::cout << " quant=" << actQuantScaling
- << ", global scaling=" << Utils::cnotice << activationScaling
- << Utils::cdef << " -> cell scaling=" << Utils::cwarning
- << ((prevActivationScaling/activationScaling)
- /actQuantScaling)
- << Utils::cdef << std::endl;
-#endif
- }
- }
-
- fuseScalingCells();
-}
-
-
//TODO: add this !!!
double N2D2::DeepNetQuantization::getActivationQuantizationScaling(const Cell& cell, std::size_t nbBits) const {
const double unsignedMax = std::pow(2, nbBits) - 1;
@@ -401,9 +199,9 @@ double N2D2::DeepNetQuantization::getActivationQuantizationScaling(const Cell& c
}
const std::string activationType = activation->getType();
-
- if(activationType == LogisticActivation::Type ||
- activationType == LogisticActivation::TypeWithLoss)
+
+ if(activationType == LogisticActivation::Type ||
+ activationType == LogisticActivation::TypeWithLoss)
{
return 2*(DeepNetExport::isCellInputsUnsigned(cell, mDeepNet)?
signedMax*unsignedMax/signedMax:
@@ -425,63 +223,64 @@ double N2D2::DeepNetQuantization::getActivationQuantizationScaling(const Cell& c
long double quantizeActivation(std::shared_ptr<Node> node, std::size_t /*nbBits*/, std::unordered_map<std::string, long double> biasScalings, std::unordered_map<std::string, long double> activationScalings, bool /*verbose*/){
const long double prevActivationScaling = getMaxParentsScaling(node, activationScalings);
+
rescaleParentsToScaling(node, activationScalings, prevActivationScaling);
long double activationScaling;
long double biasScaling = 1.0;
if(node->type() == "ElemWise") {
- /*
- activationScaling = getCellThreshold(node->name(),
- outputsHistogram, outputsRange,
- nbBits, ClippingMode::NONE);
- */
activationScaling = getCellThreshold(node);
}
- else if(node->type() == "Padding" ||
- node->type() == "Pool" ||
- node->type() == "Resize" ||
- node->type() == "Scaling" ||
- node->type() == "Softmax" ||
+ else if(node->type() == "Padding" ||
+ node->type() == "Pool" ||
+ node->type() == "Resize" ||
+ node->type() == "Scaling" ||
+ node->type() == "Softmax" ||
node->type() == "Transpose" ||
node->type() == "Reshape")
{
return prevActivationScaling;
}
- else if(node->type() == "ReLU"
- || node->type() == "Linear"
+ else if(node->type() == "ReLU"
+ || node->type() == "Linear"
|| node->type() == "Saturation") {
- //TODO :: nbOutputs > 2 is relevant for clip, check it
+ //TODO :: nbOutputs > 2 is relevant for clip, check it
//const bool clip = node->nbOutputs() > 2 && isLinearActivation;
activationScaling = getCellThreshold(node);
}
else {
- throw std::runtime_error("Quantization of cell '" + node->name() + "' of type '" +
+ throw std::runtime_error("Quantization of cell '" + node->name() + "' of type '" +
node->type() + "' is not supported yet.");
}
//get the parent conv/fc bias
- const std::vector<std::shared_ptr<Node>> parentsNodes = node->getParents();
+ const std::vector<std::shared_ptr<Node>> parentsNodes = node->getParents();
for(const std::shared_ptr<Node>& parentNode: parentsNodes) {
if(parentNode->type() == "Conv"){
biasScaling = biasScalings[parentNode->name()];
+ }else{
+ throw std::runtime_error("Quantization of activation with parent: '" + parentNode->type() + "' is not supported yet.");
}
}
activationScaling /= biasScaling;
- activationScaling = (activationScaling == 0.0)?1.0:activationScaling;
+ activationScaling = (activationScaling == 0.0) ? 1.0 : activationScaling;
return activationScaling;
}
+// TODO: Give directly a scheduler object ?
void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::size_t nbBits, bool verbose){
//void quantizeNetwork(std::shared_ptr<GraphView> g1, std::size_t nbBits, bool verbose){
//keep all bias scalings here
+ // Map: node name <-> bias scaling
std::unordered_map<std::string, long double> biasScalings;
- //and activations
+ //and activations
+ // Map: node name <-> activation scaling
std::unordered_map<std::string, long double> activationScalings;
//loop on all nodes for weights/bias quantization
@@ -493,28 +292,29 @@ void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::s
std::set<std::shared_ptr<Node>> childrenNodes = nodePtr->getChildren();
for (const std::shared_ptr<Node>& child_node : childrenNodes)
{
- printf("- type %s, and name %s \n", child_node->type().c_str(), child_node->name().c_str());
+ printf("\t- type %s, and name %s \n", child_node->type().c_str(), child_node->name().c_str());
}
const std::vector<std::shared_ptr<Node>> parentsNodes = nodePtr->getParents();
printf("And %zu parents: \n", parentsNodes.size());
for(const std::shared_ptr<Node>& parent_node: parentsNodes) {
- printf("- name %s : \n", parent_node->name().c_str());
+ printf("\t- name %s : \n", parent_node->name().c_str());
}
}
if (nodePtr->type() == "Conv") {
+ // TODO: pass biasScaling by reference and update it in quantizeFreeParams
biasScalings[nodePtr->name()] = quantizeFreeParams(nodePtr, nbBits, biasScalings, verbose);
- if(verbose){
+ if(verbose){
std::cout << "outside quantizeFreeParams :: biasScalings[node->name()] = " << biasScalings[nodePtr->name()] << std::endl;
}
} else if(nodePtr->type() == "ReLU") {
activationScalings[nodePtr->name()] = quantizeActivation(nodePtr, nbBits, biasScalings, activationScalings, verbose);
- if(verbose){
+ if(verbose){
std::cout << "outside quantizeActivation :: activationScalings[node->name()] = " << activationScalings[nodePtr->name()] << std::endl;
}
- //TODO :
- //correct bias for previous activation scaling here ...
+ //TODO :
+ //correct bias for previous activation scaling here ...
/*
cell->processFreeParameters([&](Float_T d) { return d/prevActivationScaling; },
Cell::Additive);
@@ -524,18 +324,18 @@ void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::s
scaling_node->getOperator()->setDatatype(DataType::Float32);
scaling_node->getOperator()->setBackend("cpu");
- for (auto& graphPtr : nodePtr->views()) {
+ for (auto& graphPtr : nodePtr->views()) { // TODO : use maxence/thibualt insert !
graphPtr->addChild(scaling_node);
}
//add scaling cell
/*
const long double actQuantScaling = getActivationQuantizationScaling(*cell, nbBits);
auto scalingCell = Registrar<ScalingCell>::create<Float_T>(getCellModelType(*cell))
- (mDeepNet,
- mDeepNet.generateNewCellName(cell->getName() + "_rescale_act"),
- cell->getNbOutputs(),
+ (mDeepNet,
+ mDeepNet.generateNewCellName(cell->getName() + "_rescale_act"),
+ cell->getNbOutputs(),
Scaling::floatingPointScaling(
- std::vector<Float_T>(cell->getNbOutputs(),
+ std::vector<Float_T>(cell->getNbOutputs(),
(prevActivationScaling/activationScaling)/actQuantScaling
),
false,
@@ -552,13 +352,13 @@ void quantizeNetwork(std::vector<std::shared_ptr<Node>> orderedGraphView, std::s
}
else{
std::cout << "this is not Conv or ReLu, moving on" << std::endl;
- }
+ }
}
-
+ // TODO : add fuseScaling recipes ?
}
}
-//TODO :
-//Scaling cell insert missing - for now it's a problem for activation quantization, as in simple cases there is no need to additional rescaling in rescaleParentsToScaling ...
-//activation is no more attached to cell, now it's the child of conv/fc - quantize only after Relu/?
+//TODO :
+//Scaling cell insert missing - for now it's a problem for activation quantization, as in simple cases there is no need to additional rescaling in rescaleParentsToScaling ...
+//activation is no more attached to cell, now it's the child of conv/fc - quantize only after Relu/?
//Figure out how to check the range of operator : need to add the function to every possible operator ?
diff --git a/unit_tests/hook/Test_execTime.cpp b/unit_tests/hook/Test_execTime.cpp
index a9fb8bd..c71d69f 100644
--- a/unit_tests/hook/Test_execTime.cpp
+++ b/unit_tests/hook/Test_execTime.cpp
@@ -67,7 +67,7 @@ TEST_CASE("[hook] ExecTime(forward)") {
}
};
Tensor myBias = Array1D<float,4> {{7.,0.,9.,0.}};
- Tensor myInput = Array4D<float,2,3,5,5> {
+ Tensor myInput = Array4D<float,2,3,5,5> {
{
{
{{ 0., 1., 2., 3., 4.},
@@ -109,7 +109,7 @@ TEST_CASE("[hook] ExecTime(forward)") {
}
}
};
- Tensor myOutput = Array4D<float,2,4,3,3> {
+ Tensor myOutput = Array4D<float,2,4,3,3> {
{
{
{{ 15226., 15577., 15928.},
@@ -147,7 +147,7 @@ TEST_CASE("[hook] ExecTime(forward)") {
myConv1->getOperator()->computeOutputDims();
myConv1->getOperator()->addHook("execution_time");
myConv1->forward();
-
+
//std::static_pointer_cast<Tensor>(myConv1->getOperator()->getOutput(0))->print();
assert(*std::static_pointer_cast<Tensor>(myConv1->getOperator()->getOutput(0)) == myOutput);
//std::static_pointer_cast<Tensor>(myConv1->getOperator()->getInput(1))->print();
@@ -156,4 +156,4 @@ TEST_CASE("[hook] ExecTime(forward)") {
const std::time_t t_c = std::chrono::system_clock::to_time_t(time_conv);
//std::cout << "the execution time of the module was " << std::put_time(std::localtime(&t_c), "%F %T.\n") << std::flush;
-}
\ No newline at end of file
+}
diff --git a/unit_tests/hook/Test_outputRange.cpp b/unit_tests/hook/Test_outputRange.cpp
index 18916e9..9a7de72 100644
--- a/unit_tests/hook/Test_outputRange.cpp
+++ b/unit_tests/hook/Test_outputRange.cpp
@@ -66,7 +66,7 @@ TEST_CASE("[hook] OutputRange(forward)") {
}
};
Tensor myBias = Array1D<float,4> {{7.,0.,9.,0.}};
- Tensor myInput = Array4D<float,2,3,5,5> {
+ Tensor myInput = Array4D<float,2,3,5,5> {
{
{
{{ 0., 1., 2., 3., 4.},
@@ -108,7 +108,7 @@ TEST_CASE("[hook] OutputRange(forward)") {
}
}
};
- Tensor myOutput = Array4D<float,2,4,3,3> {
+ Tensor myOutput = Array4D<float,2,4,3,3> {
{
{
{{ 15226., 15577., 15928.},
@@ -146,11 +146,11 @@ TEST_CASE("[hook] OutputRange(forward)") {
myConv1->getOperator()->computeOutputDims();
myConv1->getOperator()->addHook("output_range");
myConv1->forward();
-
+
//std::static_pointer_cast<Tensor>(myConv->getOperator()->getOutput(0))->print();
assert(*std::static_pointer_cast<Tensor>(myConv1->getOperator()->getOutput(0)) == myOutput);
float max_output = std::static_pointer_cast<OutputRange>(myConv1->getOperator()->getHook("output_range"))->getOutput(0);
//std::cout << "the output of the conv was " << max_output << std::flush;
-}
\ No newline at end of file
+}
--
GitLab