From d438d3c34b1da1736f040556944c53b510eb90d1 Mon Sep 17 00:00:00 2001
From: Franck Dary <franck.dary@lis-lab.fr>
Date: Thu, 2 Apr 2020 22:44:23 +0200
Subject: [PATCH] added transition split
---
decoder/src/Decoder.cpp | 2 ++
reading_machine/include/Action.hpp | 1 +
reading_machine/include/Config.hpp | 6 +++++
reading_machine/include/ReadingMachine.hpp | 3 +++
reading_machine/include/Transition.hpp | 1 +
reading_machine/include/TransitionSet.hpp | 1 +
reading_machine/src/Action.cpp | 29 ++++++++++++++++++++++
reading_machine/src/Config.cpp | 10 ++++++++
reading_machine/src/ReadingMachine.cpp | 10 ++++++++
reading_machine/src/Transition.cpp | 17 +++++++++++++
reading_machine/src/TransitionSet.cpp | 12 ++++++++-
torch_modules/include/LSTMNetwork.hpp | 1 +
torch_modules/src/LSTMNetwork.cpp | 24 +++++++++++++++---
trainer/src/Trainer.cpp | 2 ++
14 files changed, 114 insertions(+), 5 deletions(-)
diff --git a/decoder/src/Decoder.cpp b/decoder/src/Decoder.cpp
index a24f6a2..7d563c8 100644
--- a/decoder/src/Decoder.cpp
+++ b/decoder/src/Decoder.cpp
@@ -24,6 +24,8 @@ void Decoder::decode(BaseConfig & config, std::size_t beamSize, bool debug, bool
if (debug)
config.printForDebug(stderr);
+ config.setAppliableSplitTransitions(machine.getSplitWordTransitionSet().getNAppliableTransitions(config, Config::maxNbAppliableSplitTransitions));
+
auto dictState = machine.getDict(config.getState()).getState();
auto context = machine.getClassifier()->getNN()->extractContext(config, machine.getDict(config.getState())).back();
machine.getDict(config.getState()).setState(dictState);
diff --git a/reading_machine/include/Action.hpp b/reading_machine/include/Action.hpp
index 012af61..5561a81 100644
--- a/reading_machine/include/Action.hpp
+++ b/reading_machine/include/Action.hpp
@@ -64,6 +64,7 @@ class Action
static Action ignoreCurrentCharacter();
static Action consumeCharacterIndex(util::utf8string consumed);
static Action setMultiwordIds(int multiwordSize);
+ static Action split(int index);
};
#endif
diff --git a/reading_machine/include/Config.hpp b/reading_machine/include/Config.hpp
index 1e6833e..0bc46f5 100644
--- a/reading_machine/include/Config.hpp
+++ b/reading_machine/include/Config.hpp
@@ -9,6 +9,8 @@
#include "util.hpp"
#include "Dict.hpp"
+class Transition;
+
class Config
{
public :
@@ -20,6 +22,7 @@ class Config
static constexpr const char * deprelColName = "DEPREL";
static constexpr const char * idColName = "ID";
static constexpr int nbHypothesesMax = 1;
+ static constexpr int maxNbAppliableSplitTransitions = 3;
public :
@@ -34,6 +37,7 @@ class Config
std::set<std::string> predicted;
int lastPoppedStack{-1};
int currentWordId{0};
+ std::vector<Transition *> appliableSplitTransitions;
protected :
@@ -124,6 +128,8 @@ class Config
void setCurrentWordId(int currentWordId);
void addMissingColumns();
void addComment();
+ void setAppliableSplitTransitions(const std::vector<Transition *> & appliableSplitTransitions);
+ const std::vector<Transition *> & getAppliableSplitTransitions() const;
};
#endif
diff --git a/reading_machine/include/ReadingMachine.hpp b/reading_machine/include/ReadingMachine.hpp
index cc56ec2..dd0098d 100644
--- a/reading_machine/include/ReadingMachine.hpp
+++ b/reading_machine/include/ReadingMachine.hpp
@@ -26,6 +26,8 @@ class ReadingMachine
std::map<std::string, Dict> dicts;
std::set<std::string> predicted;
+ std::unique_ptr<TransitionSet> splitWordTransitionSet;
+
private :
void readFromFile(std::filesystem::path path);
@@ -36,6 +38,7 @@ class ReadingMachine
ReadingMachine(std::filesystem::path path);
ReadingMachine(std::filesystem::path path, std::vector<std::filesystem::path> models, std::vector<std::filesystem::path> dicts);
TransitionSet & getTransitionSet();
+ TransitionSet & getSplitWordTransitionSet();
Strategy & getStrategy();
Dict & getDict(const std::string & state);
std::map<std::string, Dict> & getDicts();
diff --git a/reading_machine/include/Transition.hpp b/reading_machine/include/Transition.hpp
index 0c1ccd7..87531db 100644
--- a/reading_machine/include/Transition.hpp
+++ b/reading_machine/include/Transition.hpp
@@ -28,6 +28,7 @@ class Transition
void initEndWord();
void initAddCharToWord();
void initSplitWord(std::vector<std::string> words);
+ void initSplit(int index);
public :
diff --git a/reading_machine/include/TransitionSet.hpp b/reading_machine/include/TransitionSet.hpp
index 2daaa2e..a1bc2c1 100644
--- a/reading_machine/include/TransitionSet.hpp
+++ b/reading_machine/include/TransitionSet.hpp
@@ -17,6 +17,7 @@ class TransitionSet
TransitionSet(const std::string & filename);
std::vector<std::pair<Transition*, int>> getAppliableTransitionsCosts(const Config & c);
Transition * getBestAppliableTransition(const Config & c);
+ std::vector<Transition *> getNAppliableTransitions(const Config & c, int n);
std::size_t getTransitionIndex(const Transition * transition) const;
Transition * getTransition(std::size_t index);
Transition * getTransition(const std::string & name);
diff --git a/reading_machine/src/Action.cpp b/reading_machine/src/Action.cpp
index e425740..889aaa7 100644
--- a/reading_machine/src/Action.cpp
+++ b/reading_machine/src/Action.cpp
@@ -1,4 +1,5 @@
#include "Action.hpp"
+#include "Transition.hpp"
Action::Action(Action::Type type, std::function<void(Config & config, Action & action)> apply, std::function<void(Config & config, Action & action)> undo, std::function<bool(const Config & config, const Action & action)> appliable)
{
@@ -619,6 +620,33 @@ Action Action::attach(Object governorObject, int governorIndex, Object dependent
return {Type::Write, apply, undo, appliable};
}
+Action Action::split(int index)
+{
+ auto apply = [index](Config & config, Action &)
+ {
+ Transition * t = config.getAppliableSplitTransitions()[index];
+ t->apply(config);
+ };
+
+ auto undo = [](Config &, Action &)
+ {
+ //TODO : undo this
+ };
+
+ auto appliable = [index](const Config & config, const Action &)
+ {
+ auto & transitions = config.getAppliableSplitTransitions();
+
+ if (index < 0 or index >= (int)transitions.size())
+ return false;
+
+ Transition * t = transitions[index];
+ return t->appliable(config);
+ };
+
+ return {Type::Write, apply, undo, appliable};
+}
+
Action::Object Action::str2object(const std::string & s)
{
if (s == "b")
@@ -629,3 +657,4 @@ Action::Object Action::str2object(const std::string & s)
util::myThrow(fmt::format("Invalid object '{}'", s));
return Object::Buffer;
}
+
diff --git a/reading_machine/src/Config.cpp b/reading_machine/src/Config.cpp
index 6aec0fb..bb2366c 100644
--- a/reading_machine/src/Config.cpp
+++ b/reading_machine/src/Config.cpp
@@ -601,3 +601,13 @@ long Config::getRelativeWordIndex(int relativeIndex) const
return -1;
}
+void Config::setAppliableSplitTransitions(const std::vector<Transition *> & appliableSplitTransitions)
+{
+ this->appliableSplitTransitions = appliableSplitTransitions;
+}
+
+const std::vector<Transition *> & Config::getAppliableSplitTransitions() const
+{
+ return appliableSplitTransitions;
+}
+
diff --git a/reading_machine/src/ReadingMachine.cpp b/reading_machine/src/ReadingMachine.cpp
index 50ce655..96b5e71 100644
--- a/reading_machine/src/ReadingMachine.cpp
+++ b/reading_machine/src/ReadingMachine.cpp
@@ -63,6 +63,11 @@ void ReadingMachine::readFromFile(std::filesystem::path path)
--curLine;
+ util::doIfNameMatch(std::regex("Splitwords : (.+)"), lines[curLine++], [this,path](auto sm)
+ {
+ this->splitWordTransitionSet.reset(new TransitionSet(path.parent_path() / sm.str(1)));
+ });
+
if (!util::doIfNameMatch(std::regex("Predictions : (.+)"), lines[curLine++], [this](auto sm)
{
auto predictions = sm.str(1);
@@ -84,6 +89,11 @@ TransitionSet & ReadingMachine::getTransitionSet()
return classifier->getTransitionSet();
}
+TransitionSet & ReadingMachine::getSplitWordTransitionSet()
+{
+ return *splitWordTransitionSet;
+}
+
Strategy & ReadingMachine::getStrategy()
{
return *strategy;
diff --git a/reading_machine/src/Transition.cpp b/reading_machine/src/Transition.cpp
index 0a9bb89..d45a794 100644
--- a/reading_machine/src/Transition.cpp
+++ b/reading_machine/src/Transition.cpp
@@ -27,6 +27,8 @@ Transition::Transition(const std::string & name)
[this](auto){initEndWord();}},
{std::regex("ADDCHARTOWORD"),
[this](auto){initAddCharToWord();}},
+ {std::regex("SPLIT (.+)"),
+ [this](auto sm){(initSplit(std::stoi(sm.str(1))));}},
{std::regex("SPLITWORD ([^@]+)(:?(:?@[^@]+)+)"),
[this](auto sm)
{
@@ -210,6 +212,21 @@ void Transition::initSplitWord(std::vector<std::string> words)
};
}
+void Transition::initSplit(int index)
+{
+ sequence.emplace_back(Action::split(index));
+
+ cost = [index](const Config & config)
+ {
+ auto & transitions = config.getAppliableSplitTransitions();
+
+ if (index < 0 or index >= (int)transitions.size())
+ return std::numeric_limits<int>::max();
+
+ return transitions[index]->getCost(config);
+ };
+}
+
void Transition::initShift()
{
sequence.emplace_back(Action::pushWordIndexOnStack());
diff --git a/reading_machine/src/TransitionSet.cpp b/reading_machine/src/TransitionSet.cpp
index d5b9716..8fbb6a5 100644
--- a/reading_machine/src/TransitionSet.cpp
+++ b/reading_machine/src/TransitionSet.cpp
@@ -42,6 +42,17 @@ std::vector<std::pair<Transition*, int>> TransitionSet::getAppliableTransitionsC
return appliableTransitions;
}
+std::vector<Transition *> TransitionSet::getNAppliableTransitions(const Config & c, int n)
+{
+ std::vector<Transition *> result;
+
+ for (unsigned int i = 0; i < transitions.size() && result.size() < n; i++)
+ if (transitions[i].appliable(c))
+ result.emplace_back(&transitions[i]);
+
+ return result;
+}
+
Transition * TransitionSet::getBestAppliableTransition(const Config & c)
{
Transition * result = nullptr;
@@ -94,4 +105,3 @@ Transition * TransitionSet::getTransition(const std::string & name)
return nullptr;
}
-
diff --git a/torch_modules/include/LSTMNetwork.hpp b/torch_modules/include/LSTMNetwork.hpp
index 0e1ad1c..8c14e76 100644
--- a/torch_modules/include/LSTMNetwork.hpp
+++ b/torch_modules/include/LSTMNetwork.hpp
@@ -22,6 +22,7 @@ class LSTMNetworkImpl : public NeuralNetworkImpl
torch::nn::Linear linear2{nullptr};
torch::nn::LSTM contextLSTM{nullptr};
torch::nn::LSTM rawInputLSTM{nullptr};
+ torch::nn::LSTM splitTransLSTM{nullptr};
std::vector<torch::nn::LSTM> lstms;
public :
diff --git a/torch_modules/src/LSTMNetwork.cpp b/torch_modules/src/LSTMNetwork.cpp
index 0ae4cd1..201feea 100644
--- a/torch_modules/src/LSTMNetwork.cpp
+++ b/torch_modules/src/LSTMNetwork.cpp
@@ -1,4 +1,5 @@
#include "LSTMNetwork.hpp"
+#include "Transition.hpp"
LSTMNetworkImpl::LSTMNetworkImpl(int nbOutputs, int unknownValueThreshold, std::vector<int> bufferContext, std::vector<int> stackContext, std::vector<std::string> columns, std::vector<int> focusedBufferIndexes, std::vector<int> focusedStackIndexes, std::vector<std::string> focusedColumns, std::vector<int> maxNbElements, int leftWindowRawInput, int rightWindowRawInput) : unknownValueThreshold(unknownValueThreshold), focusedColumns(focusedColumns), maxNbElements(maxNbElements), leftWindowRawInput(leftWindowRawInput), rightWindowRawInput(rightWindowRawInput)
{
@@ -29,8 +30,9 @@ LSTMNetworkImpl::LSTMNetworkImpl(int nbOutputs, int unknownValueThreshold, std::
lstmDropout = register_module("lstm_dropout", torch::nn::Dropout(0.3));
hiddenDropout = register_module("hidden_dropout", torch::nn::Dropout(0.3));
contextLSTM = register_module("contextLSTM", torch::nn::LSTM(torch::nn::LSTMOptions(columns.size()*embeddingsSize, contextLSTMSize).batch_first(true).bidirectional(true)));
+ splitTransLSTM = register_module("splitTransLSTM", torch::nn::LSTM(torch::nn::LSTMOptions(embeddingsSize, embeddingsSize).batch_first(true).bidirectional(true)));
- int totalLSTMOutputSize = contextLSTM->options.hidden_size() * (contextLSTM->options.bidirectional() ? 4 : 1) + rawInputLSTMOutputSize;
+ int totalLSTMOutputSize = contextLSTM->options.hidden_size() * (contextLSTM->options.bidirectional() ? 4 : 1) + rawInputLSTMOutputSize + (Config::maxNbAppliableSplitTransitions * splitTransLSTM->options.hidden_size() * (splitTransLSTM->options.bidirectional() ? 2 : 1));
for (auto & col : focusedColumns)
{
@@ -49,21 +51,28 @@ torch::Tensor LSTMNetworkImpl::forward(torch::Tensor input)
auto embeddings = embeddingsDropout(wordEmbeddings(input));
- auto context = embeddings.narrow(1, rawInputSize, getContextSize());
+ auto splitTrans = embeddings.narrow(1, 0, Config::maxNbAppliableSplitTransitions);
+
+ auto context = embeddings.narrow(1, splitTrans.size(1)+rawInputSize, getContextSize());
context = context.view({context.size(0), context.size(1)/(int)columns.size(), (int)columns.size()*(int)wordEmbeddings->options.embedding_dim()});
- auto elementsEmbeddings = embeddings.narrow(1, rawInputSize+context.size(1), input.size(1)-(rawInputSize+context.size(1)));
+ auto elementsEmbeddings = embeddings.narrow(1, splitTrans.size(1)+rawInputSize+context.size(1), input.size(1)-(splitTrans.size(1)+rawInputSize+context.size(1)));
std::vector<torch::Tensor> lstmOutputs;
if (rawInputSize != 0)
{
- auto rawLetters = embeddings.narrow(1, 0, rawInputSize);
+ auto rawLetters = embeddings.narrow(1, splitTrans.size(1), rawInputSize);
auto lstmOut = rawInputLSTM(rawLetters).output;
lstmOutputs.emplace_back(lstmOut.reshape({lstmOut.size(0), -1}));
}
+ {
+ auto lstmOut = splitTransLSTM(splitTrans).output;
+ lstmOutputs.emplace_back(lstmOut.reshape({lstmOut.size(0), -1}));
+ }
+
auto curIndex = 0;
for (unsigned int i = 0; i < focusedColumns.size(); i++)
{
@@ -101,6 +110,13 @@ std::vector<std::vector<long>> LSTMNetworkImpl::extractContext(Config & config,
std::vector<std::vector<long>> context;
context.emplace_back();
+ auto & splitTransitions = config.getAppliableSplitTransitions();
+ for (int i = 0; i < Config::maxNbAppliableSplitTransitions; i++)
+ if (i < (int)splitTransitions.size())
+ context.back().emplace_back(dict.getIndexOrInsert(splitTransitions[i]->getName()));
+ else
+ context.back().emplace_back(dict.getIndexOrInsert(Dict::nullValueStr));
+
if (rawInputSize > 0)
{
for (int i = 0; i < leftWindowRawInput; i++)
diff --git a/trainer/src/Trainer.cpp b/trainer/src/Trainer.cpp
index ccbad4e..621441c 100644
--- a/trainer/src/Trainer.cpp
+++ b/trainer/src/Trainer.cpp
@@ -41,6 +41,8 @@ void Trainer::extractExamples(SubConfig & config, bool debug, std::vector<torch:
if (debug)
config.printForDebug(stderr);
+ config.setAppliableSplitTransitions(machine.getSplitWordTransitionSet().getNAppliableTransitions(config, Config::maxNbAppliableSplitTransitions));
+
auto * transition = machine.getTransitionSet().getBestAppliableTransition(config);
if (!transition)
{
--
GitLab