Merge pull request #68 from Strife-AI/nn-service

Remove tinycc and finish neural network service

Author: BrendanMulcahy

src/CMakeLists.txt

@@ -161,7 +161,7 @@ add_library(Strife.Engine STATIC
 		Net/FileTransfer.cpp
         Net/SyncVar.cpp
 		Resource/ResourceManager.hpp Resource/SpriteResource.hpp Resource/SpriteResource.cpp Resource/ResourceManager.cpp Resource/TilemapResource.hpp Resource/TilemapResource.cpp Resource/SpriteFontResource.hpp Resource/SpriteFontResource.cpp
-		Resource/ShaderResource.cpp Resource/ShaderResource.hpp Components/ParticleSystemComponent.hpp Components/ParticleSystemComponent.cpp Scene/Isometric.hpp Scene/Isometric.cpp Components/IsometricSpriteComponent.hpp Components/IsometricSpriteComponent.cpp Resource/FileResource.hpp Resource/FileResource.cpp ML/UtilityAI.hpp Resource/ScriptResource.hpp Resource/ScriptResource.cpp ML/GridSensor.hpp Renderer/SpriteEffect.hpp Renderer/SpriteEffect.cpp Renderer/Stage/RenderPipeline.hpp Renderer/Stage/RenderPipeline.cpp Resource/SpriteAtlasResource.hpp Resource/SpriteAtlasResource.cpp Resource/ResourceSettings.hpp Components/AnimatorComponent.hpp Components/AnimatorComponent.cpp ML/NeuralNetworkService.hpp)
+		Resource/ShaderResource.cpp Resource/ShaderResource.hpp Components/ParticleSystemComponent.hpp Components/ParticleSystemComponent.cpp Scene/Isometric.hpp Scene/Isometric.cpp Components/IsometricSpriteComponent.hpp Components/IsometricSpriteComponent.cpp Resource/FileResource.hpp Resource/FileResource.cpp ML/UtilityAI.hpp   ML/GridSensor.hpp Renderer/SpriteEffect.hpp Renderer/SpriteEffect.cpp Renderer/Stage/RenderPipeline.hpp Renderer/Stage/RenderPipeline.cpp Resource/SpriteAtlasResource.hpp Resource/SpriteAtlasResource.cpp Resource/ResourceSettings.hpp Components/AnimatorComponent.hpp Components/AnimatorComponent.cpp ML/NeuralNetworkService.hpp)
 
 message("SDL DIRS: ${SDL2_INCLUDE_DIRS}")
 

src/Engine.cpp

@@ -19,7 +19,6 @@
 #include "Tools/MetricsManager.hpp"
 #include "Sound/SoundManager.hpp"
 #include "UI/UI.hpp"
-#include "Scripting/ScriptCompiler.hpp"
 #include "Net/ServerGame.hpp"
 
 using namespace std::chrono;
@@ -39,18 +38,13 @@ void ExecuteOnGameThread(const std::function<void()>& function)
     g_workQueue.Enqueue(function);
 }
 
-void RegisterScriptFunctions();
-
 static void StrifeMlLog(const char* message)
 {
     Log("%s", message);
 }
 
 Engine::Engine(const EngineConfig& config)
 {
-    StrifeML::SetLogFunction(StrifeMlLog);
-    RegisterScriptFunctions();
-
     _config = config;
     _defaultBlockAllocator = std::make_unique<BlockAllocator>(config.blockAllocatorSizeBytes);
 
@@ -197,8 +191,6 @@ void Engine::RunFrame()
     AccurateSleepFor(timeUntilUpdate);
     nextGameToRun->RunFrame(GetTimeSeconds());
     nextGameToRun->nextUpdateTime = nextGameToRun->nextUpdateTime + 1.0f / nextGameToRun->targetTickRate;
-
-    ScriptCompiler::GetInstance()->Update();
 }
 
 void Engine::PauseGame()

src/ML/ML.hpp

@@ -18,141 +18,6 @@ struct StrifeML::Serializer<Vector2>
     }
 };
 
-enum class NeuralNetworkMode
-{
-    Disabled,
-    Deciding,
-    CollectingSamples,
-	ReinforcementLearning,
-};
-
-template<typename TNeuralNetwork>
-struct NeuralNetworkComponent : ComponentTemplate<NeuralNetworkComponent<TNeuralNetwork>>
-{
-    using InputType = typename TNeuralNetwork::InputType;
-    using OutputType = typename TNeuralNetwork::OutputType;
-    using NetworkType = TNeuralNetwork;
-    using SampleType = StrifeML::Sample<InputType, OutputType>;
-
-    explicit NeuralNetworkComponent(float decisionsPerSecond_ = 1)
-        : decisionInput(StrifeML::MlUtil::SharedArray<InputType>(TNeuralNetwork::SequenceLength)),
-          decisionsPerSecond(decisionsPerSecond_)
-    {
-
-    }
-
-    virtual ~NeuralNetworkComponent() = default;
-
-    void Update(float deltaTime) override;
-
-    void SetNetwork(const char* name);
-
-    bool StartMakingDecision();
-
-    StrifeML::NetworkContext<NetworkType>* networkContext = nullptr;
-    std::shared_ptr<StrifeML::MakeDecisionWorkItem<TNeuralNetwork>> decisionInProgress;
-    StrifeML::MlUtil::SharedArray<InputType> decisionInput;
-    float makeDecisionsTimer = 0;
-    float decisionsPerSecond;
-
-    std::shared_ptr<StrifeML::RunTrainingBatchWorkItem<TNeuralNetwork>> trainingInProgress;
-
-	int inputsCollected = 0;
-	InputType previousInput;
-    std::function<void(InputType& input)> collectInput;
-    std::function<void(OutputType& decision)> collectDecision;
-
-    std::function<void(OutputType& decision)> receiveDecision;
-
-    NeuralNetworkMode mode = NeuralNetworkMode::Disabled;
-};
-
-template <typename TNeuralNetwork>
-void NeuralNetworkComponent<TNeuralNetwork>::Update(float deltaTime)
-{
-    if (mode == NeuralNetworkMode::Disabled)
-    {
-        return;
-    }
-    else if (mode == NeuralNetworkMode::Deciding || mode == NeuralNetworkMode::ReinforcementLearning)
-    {
-        OutputType output;
-        if (decisionInProgress != nullptr && decisionInProgress->TryGetResult(output))
-        {
-            if (receiveDecision != nullptr)
-            {
-                receiveDecision(output);
-            }
-
-            decisionInProgress = nullptr;
-        }
-
-        makeDecisionsTimer -= deltaTime;
-
-        if (makeDecisionsTimer <= 0)
-        {      	
-            makeDecisionsTimer = 1.0f / decisionsPerSecond;
-            auto wasStarted = StartMakingDecision();
-
-        	if (wasStarted && mode == NeuralNetworkMode::ReinforcementLearning && inputsCollected > TNeuralNetwork::SequenceLength)
-	        {
-		        SampleType sample;
-	        	sample.input = previousInput;
-		        collectDecision(sample.output);
-		        networkContext->trainer->AddSample(sample);
-	        }
-        }
-    }
-    else if (mode == NeuralNetworkMode::CollectingSamples)
-    {
-        SampleType sample;
-        collectInput(sample.input);
-        collectDecision(sample.output);
-        networkContext->trainer->AddSample(sample);
-    }
-}
-
-template <typename TNeuralNetwork>
-bool NeuralNetworkComponent<TNeuralNetwork>::StartMakingDecision()
-{
-    // Don't allow making more than one decision at a time
-    if (decisionInProgress != nullptr
-        && !decisionInProgress->IsComplete())
-    {
-        return false;
-    }
-
-    if (collectInput == nullptr
-        || networkContext == nullptr
-        || networkContext->decider == nullptr)
-    {
-        return false;
-    }
-
-	if (mode == NeuralNetworkMode::ReinforcementLearning && inputsCollected > TNeuralNetwork::SequenceLength)
-	{
-		previousInput = decisionInput.data.get()[TNeuralNetwork::SequenceLength - 1];
-	}
-
-    // Expire oldest input
-    for (int i = 0; i < TNeuralNetwork::SequenceLength - 1; ++i)
-    {
-        decisionInput.data.get()[i] = std::move(decisionInput.data.get()[i + 1]);
-    }
-
-    // Collect new input
-    collectInput(decisionInput.data.get()[TNeuralNetwork::SequenceLength - 1]);
-	inputsCollected++;
-
-    // Start making decision if we have enough data
-    if (inputsCollected >= TNeuralNetwork::SequenceLength)
-    {
-	    decisionInProgress = networkContext->decider->MakeDecision(decisionInput, TNeuralNetwork::SequenceLength);
-    }
-	
-	return true;
-}
-
 struct SensorObjectDefinition
 {
     SensorObjectDefinition()
@@ -253,7 +118,7 @@ struct NeuralNetworkManager
     }
 
     template<typename TDecider, typename TTrainer>
-    StrifeML::NetworkContext<typename TDecider::NetworkType>* CreateNetwork(const char* name, TDecider* decider, TTrainer* trainer)
+    StrifeML::NetworkContext<typename TDecider::NetworkType>* CreateNetwork(const char* name, TDecider* decider, TTrainer* trainer, int sequenceLength)
     {
         static_assert(std::is_same_v<typename TDecider::NetworkType, typename TTrainer::NetworkType>, "Trainer and decider must accept the same type of neural network");
 
@@ -263,7 +128,7 @@ struct NeuralNetworkManager
             throw StrifeML::StrifeException("Network already exists: " + std::string(name));
         }
 
-        auto newContext = std::make_shared<StrifeML::NetworkContext<typename TDecider::NetworkType>>(decider, trainer);
+        auto newContext = std::make_shared<StrifeML::NetworkContext<typename TDecider::NetworkType>>(decider, trainer, sequenceLength);
         _networksByName[name] = newContext;
 
         trainer->networkContext = newContext;
@@ -303,13 +168,6 @@ struct NeuralNetworkManager
     std::unordered_map<std::string, std::shared_ptr<StrifeML::INetworkContext>> _networksByName;
 };
 
-template<typename TNeuralNetwork>
-void NeuralNetworkComponent<TNeuralNetwork>::SetNetwork(const char* name)
-{
-    auto nnManager = this->owner->GetEngine()->GetNeuralNetworkManager();
-    networkContext = nnManager->template GetNetwork<TNeuralNetwork>(name);
-}
-
 gsl::span<uint64_t> ReadGridSensorRectangles(
     Scene* scene,
     Vector2 center,

src/ML/NeuralNetworkService.hpp

@@ -99,7 +99,7 @@ void DecisionBatch<TNetwork>::AddToBatch(Entity* entity, DecisionBatch::InputCir
     entitiesInBatch.emplace_back(entity);
 
     int col = 0;
-    for (PlayerInput& sample : buffer)
+    for (InputType& sample : buffer)
     {
         decisionInput.data.get()[row * sequenceLength + col] = sample;
         ++col;
@@ -124,6 +124,7 @@ struct NeuralNetworkService : ISceneService, IEntityObserver
 {
     using InputType = typename TNetwork::InputType;
     using OutputType = typename TNetwork::OutputType;
+	using SampleType = typename TNetwork::SampleType;
     using TrainerType = StrifeML::Trainer<TNetwork>;
 
     using InputCircularBuffer = CircularQueue<InputType>;
@@ -140,10 +141,7 @@ struct NeuralNetworkService : ISceneService, IEntityObserver
     void ReceiveEvent(const IEntityEvent& ev) override;
 
 protected:
-    virtual void CollectInput(TEntity* entity, InputType& input)
-    {
-
-    }
+    virtual void CollectInput(TEntity* entity, InputType& input) = 0;
 
     void ForEachEntity(const std::function<void(TEntity*)>& func);