diff --git a/decoder/include/Beam.hpp b/decoder/include/Beam.hpp
index 2c34d3f9a81792ddda88506ebb88c4fae944078b..1dd40184acbe7f0adb75a24201939959ca12083d 100644
--- a/decoder/include/Beam.hpp
+++ b/decoder/include/Beam.hpp
@@ -21,6 +21,7 @@ class Beam
int nbTransitions = 0;
double totalProbability{0.0};
bool ended{false};
+ float entropy{0.0};
public :
diff --git a/decoder/src/Beam.cpp b/decoder/src/Beam.cpp
index 16d5a32c17d220c75e5c6ab66ab3ca799dbd0e45..32e7360e02fd07bbb6e8c201463312a9e3fb867b 100644
--- a/decoder/src/Beam.cpp
+++ b/decoder/src/Beam.cpp
@@ -51,7 +51,7 @@ void Beam::update(ReadingMachine & machine, bool debug)
auto neuralInput = torch::from_blob(context.data(), {(long)context.size()}, torch::kLong).clone().to(NeuralNetworkImpl::device);
auto prediction = classifier.isRegression() ? classifier.getNN()(neuralInput).squeeze(0) : torch::softmax(classifier.getNN()(neuralInput).squeeze(0), 0);
-
+ float entropy = classifier.isRegression() ? 0.0 : NeuralNetworkImpl::entropy(prediction);
std::vector<std::pair<float, int>> scoresOfTransitions;
for (unsigned int i = 0; i < prediction.size(0); i++)
{
@@ -80,6 +80,7 @@ void Beam::update(ReadingMachine & machine, bool debug)
elements.back().config.setChosenActionScore(scoresOfTransitions[i].first);
elements.back().nbTransitions++;
elements.back().meanProbability = elements.back().totalProbability / elements.back().nbTransitions;
+ elements.back().entropy = entropy;
}
elements[index].nextTransition = scoresOfTransitions[0].second;
@@ -89,6 +90,7 @@ void Beam::update(ReadingMachine & machine, bool debug)
elements[index].name.push_back("0");
elements[index].meanProbability = 0.0;
elements[index].meanProbability = elements[index].totalProbability / elements[index].nbTransitions;
+ elements[index].entropy = entropy;
if (debug)
{
@@ -127,7 +129,7 @@ void Beam::update(ReadingMachine & machine, bool debug)
auto * transition = machine.getTransitionSet(config.getState()).getTransition(element.nextTransition);
- transition->apply(config);
+ transition->apply(config, element.entropy);
config.addToHistory(transition->getName());
auto movement = config.getStrategy().getMovement(config, transition->getName());
diff --git a/decoder/src/Decoder.cpp b/decoder/src/Decoder.cpp
index 7739b1de6096be6fedb313b8d81ed974ee493e99..22f9689ccb23687aec0ba4b21184c1fc61622404 100644
--- a/decoder/src/Decoder.cpp
+++ b/decoder/src/Decoder.cpp
@@ -43,7 +43,7 @@ void Decoder::decode(BaseConfig & baseConfig, std::size_t beamSize, float beamTh
if (machine.getTransitionSet(baseConfig.getState()).getTransition("EOS b.0") and baseConfig.getLastNotEmptyHypConst(Config::EOSColName, baseConfig.getWordIndex()) != Config::EOSSymbol1)
{
- machine.getTransitionSet(baseConfig.getState()).getTransition("EOS b.0")->apply(baseConfig);
+ machine.getTransitionSet(baseConfig.getState()).getTransition("EOS b.0")->apply(baseConfig, 0.0);
if (debug)
{
fmt::print(stderr, "Forcing EOS transition\n");
diff --git a/reading_machine/include/Action.hpp b/reading_machine/include/Action.hpp
index 4d88e48118f90fe3edadd920b98e46c1be1a3a60..1ad7ab700c5c837ff83edac8043f9b804fecd37f 100644
--- a/reading_machine/include/Action.hpp
+++ b/reading_machine/include/Action.hpp
@@ -43,6 +43,7 @@ class Action
static Action moveCharacterIndex(int movement);
static Action addHypothesis(const std::string & colName, std::size_t lineIndex, const std::string & hypothesis);
static Action addToHypothesis(const std::string & colName, std::size_t lineIndex, const std::string & addition);
+ static Action sumToHypothesis(const std::string & colName, std::size_t lineIndex, float addition);
static Action addHypothesisRelative(const std::string & colName, Config::Object object, int relativeIndex, const std::string & hypothesis);
static Action addHypothesisRelativeRelaxed(const std::string & colName, Config::Object object, int relativeIndex, const std::string & hypothesis);
static Action addToHypothesisRelative(const std::string & colName, Config::Object object, int relativeIndex, const std::string & addition);
diff --git a/reading_machine/include/Transition.hpp b/reading_machine/include/Transition.hpp
index 3e76c724245800586f78482e0f974689f93b8fb6..75f9c40643ae3487ba54ddddd14349e65cb581f5 100644
--- a/reading_machine/include/Transition.hpp
+++ b/reading_machine/include/Transition.hpp
@@ -61,6 +61,7 @@ class Transition
public :
Transition(const std::string & name);
+ void apply(Config & config, float entropy);
void apply(Config & config);
bool appliable(const Config & config) const;
int getCostDynamic(const Config & config) const;
diff --git a/reading_machine/src/Action.cpp b/reading_machine/src/Action.cpp
index cd4c20cf66d641eb8eea2063cac1ab79c00d44b4..a269968ca658db6f1d2cdaa7aadbcbd173ca8404 100644
--- a/reading_machine/src/Action.cpp
+++ b/reading_machine/src/Action.cpp
@@ -184,6 +184,55 @@ Action Action::addToHypothesis(const std::string & colName, std::size_t lineInde
return {Type::Write, apply, undo, appliable};
}
+Action Action::sumToHypothesis(const std::string & colName, std::size_t lineIndex, float addition)
+{
+ auto apply = [colName, lineIndex, addition](Config & config, Action &)
+ {
+ // Knuth’s algorithm for online mean
+ auto curStr = config.getLastNotEmptyHypConst(colName, lineIndex).get();
+ auto splited = util::split(curStr, '|');
+ float curVal = 0.0;
+ int curNb = 0;
+ if (splited.size() == 2)
+ {
+ curVal = std::stof(splited[0]);
+ curNb = std::stoi(splited[1]);
+ }
+
+ curNb += 1;
+ float delta = addition - curVal;
+ curVal += delta / curNb;
+
+ config.getLastNotEmptyHyp(colName, lineIndex) = fmt::format("{}|{}", curVal, curNb);
+ };
+
+ auto undo = [colName, lineIndex, addition](Config & config, Action &)
+ {
+ // Knuth’s algorithm for online mean
+ auto curStr = config.getLastNotEmptyHypConst(colName, lineIndex).get();
+ auto splited = util::split(curStr, '|');
+ float curVal = 0.0;
+ int curNb = 0;
+ if (splited.size() == 2)
+ {
+ curVal = std::stof(splited[0]);
+ curNb = std::stoi(splited[1]);
+ }
+
+ curNb -= 1;
+ curVal = (curNb*curVal - addition) / (curNb - 1);
+
+ config.getLastNotEmptyHyp(colName, lineIndex) = fmt::format("{}|{}", curVal, curNb);
+ };
+
+ auto appliable = [colName, lineIndex, addition](const Config & config, const Action &)
+ {
+ return config.has(colName, lineIndex, 0);
+ };
+
+ return {Type::Write, apply, undo, appliable};
+}
+
Action Action::addToHypothesisRelative(const std::string & colName, Config::Object object, int relativeIndex, const std::string & addition)
{
auto apply = [colName, object, relativeIndex, addition](Config & config, Action & a)
diff --git a/reading_machine/src/Transition.cpp b/reading_machine/src/Transition.cpp
index 29dcb6cba418d988ca4a9bc1c4a7776995c83131..879c074f8e8de9797097ba849194b8226e07d955 100644
--- a/reading_machine/src/Transition.cpp
+++ b/reading_machine/src/Transition.cpp
@@ -97,6 +97,17 @@ Transition::Transition(const std::string & name)
}
+void Transition::apply(Config & config, float entropy)
+{
+ if (config.hasColIndex("ENTROPY"))
+ {
+ auto action = Action::sumToHypothesis("ENTROPY", config.getWordIndex(), entropy);
+ action.apply(config, action);
+ }
+
+ apply(config);
+}
+
void Transition::apply(Config & config)
{
for (Action & action : sequence)
diff --git a/torch_modules/include/NeuralNetwork.hpp b/torch_modules/include/NeuralNetwork.hpp
index ee32d2b2eadc666ef7e38ac70b8ed9f64055d3e4..6058cebf9c39dff266622c656aefff66dd094f7b 100644
--- a/torch_modules/include/NeuralNetwork.hpp
+++ b/torch_modules/include/NeuralNetwork.hpp
@@ -27,6 +27,8 @@ class NeuralNetworkImpl : public torch::nn::Module, public NameHolder, public St
virtual void setDictsState(Dict::State state) = 0;
virtual void setCountOcc(bool countOcc) = 0;
virtual void removeRareDictElements(float rarityThreshold) = 0;
+
+ static float entropy(torch::Tensor probabilities);
};
TORCH_MODULE(NeuralNetwork);
diff --git a/torch_modules/src/NeuralNetwork.cpp b/torch_modules/src/NeuralNetwork.cpp
index 02e8a191bfb4b2bc718b6e815a266bec252fb24b..dbab2ebdb9a8e38ccd0cd660e2b5a56a10c3b829 100644
--- a/torch_modules/src/NeuralNetwork.cpp
+++ b/torch_modules/src/NeuralNetwork.cpp
@@ -2,3 +2,15 @@
torch::Device NeuralNetworkImpl::device(torch::cuda::is_available() ? torch::kCUDA : torch::kCPU);
+float NeuralNetworkImpl::entropy(torch::Tensor probabilities)
+{
+ float res = 0.0;
+ for (unsigned int i = 0; i < probabilities.size(0); i++)
+ {
+ float val = probabilities[i].item<float>();
+ res -= val * log(val);
+ }
+
+ return res;
+}
+
diff --git a/trainer/src/Trainer.cpp b/trainer/src/Trainer.cpp
index bdd5b33515e23c8fa79bbb599666b7c3f10a73ee..0f0ad1a4cfa9fbbb1543d0a58a8943243101d208 100644
--- a/trainer/src/Trainer.cpp
+++ b/trainer/src/Trainer.cpp
@@ -83,11 +83,14 @@ void Trainer::extractExamples(SubConfig & config, bool debug, std::filesystem::p
int nbClasses = machine.getTransitionSet(config.getState()).size();
float bestScore = -std::numeric_limits<float>::max();
+
+ float entropy = 0.0;
if (dynamicOracle and util::choiceWithProbability(1.0) and config.getState() != "tokenizer" and config.getState() != "segmenter")
{
auto neuralInput = torch::from_blob(context[0].data(), {(long)context[0].size()}, torch::kLong).clone().to(NeuralNetworkImpl::device);
auto prediction = torch::softmax(machine.getClassifier(config.getState())->getNN()(neuralInput), -1).squeeze(0);
+ entropy = NeuralNetworkImpl::entropy(prediction);
std::vector<int> candidates;
@@ -123,6 +126,7 @@ void Trainer::extractExamples(SubConfig & config, bool debug, std::filesystem::p
float regressionTarget = 0.0;
if (machine.getClassifier(config.getState())->isRegression())
{
+ entropy = 0.0;
auto errMessage = fmt::format("Invalid regression transition '{}'", transition->getName());
auto splited = util::split(transition->getName(), ' ');
if (splited.size() != 3 or splited[0] != "WRITESCORE")
@@ -154,7 +158,7 @@ void Trainer::extractExamples(SubConfig & config, bool debug, std::filesystem::p
config.setChosenActionScore(bestScore);
- transition->apply(config);
+ transition->apply(config, entropy);
config.addToHistory(transition->getName());
auto movement = config.getStrategy().getMovement(config, transition->getName());