public static void Save(NNCSaveData container, string fileName)
{
string path = string.Format("{0}/{1}.json", GetDirectoryPath(), fileName);
using (StreamWriter sw = new StreamWriter(path))
{
sw.WriteLine(JsonUtility.ToJson(container, true));
sw.Close();
}
}
private void LoadData(NNCSaveData data)
{
m_layerLengths = data.m_layerLengths;
//m_dataFileName = data.m_dataFileName;
m_activisionType = data.m_activisionType;
m_activisionTypeOutput = data.m_activisionTypeOutput;
m_costType = data.m_costType;
m_initializationType = data.m_initializationType;
m_activisionConstant = data.m_activisionConstant;
}
private void LoadContainer(NNCSaveData data)
{
if (data.m_isCorrupted)
{
Debug.Log("Aborted: loading data is corrupted!");
return;
}
LoadData(data);
m_trainingManager.LoadData(data.m_trainingData);
m_sampleManager.LoadData(data.m_sampleData);
m_visualizationNetwork.LoadData(data.m_visuilizationNetworkData);
//m_visualizationSample.LoadData(data.m_visuilizationSampleData);
m_network.LoadData(data.m_networkData);
InitializeContainer(new NeuralNetwork(
data.m_networkData,
m_layerLengths,
m_trainingManager.GetLearnRate(),
m_trainingManager.GetBatchSize(),
m_trainingManager.GetDropoutRate(),
m_trainingManager.GetWeightDecayRate(),
m_activisionType,
m_activisionTypeOutput,
m_costType,
m_initializationType,
m_activisionConstant,
data.m_networkData.m_currentDropoutSeed,
data.m_networkData.m_initDropoutSeed
));
ApplyData();
m_trainingManager.ApplyData();
m_sampleManager.ApplyData();
m_visualizationNetwork.ApplyData();
//m_visualizationSample.ApplyData();
m_network.ApplyData();
}
private void SaveContainer(string fileName)
{
NNCSaveData data = new NNCSaveData {
m_trainingData = m_trainingManager.SaveData(),
m_sampleData = m_sampleManager.SaveData(),
m_visuilizationNetworkData = m_visualizationNetwork.SaveData(),
//m_visuilizationSampleData = m_visualizationSample.SaveData(),
m_networkData = m_network.SaveData(),
m_dropoutSeed = m_initDropoutSeed,
m_layerLengths = m_layerLengths,
m_dataFileName = fileName,
m_activisionType = m_activisionType,
m_activisionTypeOutput = m_activisionTypeOutput,
m_costType = m_costType,
m_initializationType = m_initializationType,
m_activisionConstant = m_activisionConstant,
};
NeuralNetworkData.Save(data, fileName);
}
private void ManageSize()
{
if (!Input.GetKeyDown(m_keyCodeIncreaseSize))
{
return;
}
NNSaveData data = m_network.SaveData();
int oldEnemyWidth = m_sampleManager.GetScreenshotScript().GetInputLayerLengthEnemy(0, 0);
int oldPlayerHeightPixel = m_sampleManager.GetScreenshotScript().GetInputLayerLengthPlayer(0, 0);
m_sampleManager.GetScreenshotScript().SetCaptureHeight(m_sampleManager.GetScreenshotScript().GetCaptureHeight() * 2);
m_sampleManager.GetScreenshotScript().SetCaptureWidth(m_sampleManager.GetScreenshotScript().GetCaptureWidth() * 2);
m_sampleManager.GetScreenshotScript().SetCaptureSize();
m_sampleManager.GetScreenshotScript().SetCaptureSizesPlayer(m_sampleManager.GetScreenshotScript().GetCaptureWidth());
m_layerLengths[0] = m_sampleManager.GetScreenshotScript().GetInputLayerLengthTotal(0, 0);
int newPlayerHeightPixel = m_sampleManager.GetScreenshotScript().GetInputLayerLengthPlayer(0, 0);
int width = m_sampleManager.GetScreenshotScript().GetCaptureWidth();
JaggedArrayContainer[] newWeights = new JaggedArrayContainer[data.m_biases[0].dataFloat.Length];
for (int nodeIndex = 0; nodeIndex < data.m_biases[0].dataFloat.Length; nodeIndex++)
{
bool addIndex = true;
JaggedArrayContainer weights2 = new JaggedArrayContainer(m_layerLengths[0], 0);
int index = 0;
for (int weightIndex = 0; weightIndex < data.m_weights[0].array[0].dataFloat.Length; weightIndex++)
{
if (weightIndex < oldEnemyWidth)
{
if (weightIndex % (width / 2) == 0 && weightIndex != 0)
{
index += width + 2;
}
else if (weightIndex != 0)
{
index += 2;
}
int[] indices = { index, index + 1, index + width, index + width + 1 };
foreach (int i in indices)
{
weights2.dataFloat[i] = data.m_weights[0].array[nodeIndex].dataFloat[weightIndex] * 0.25f;
}
}
else
{
if (2 * oldPlayerHeightPixel > newPlayerHeightPixel)
{
if (weightIndex % (width / 2) == 0 && weightIndex != 0)
{
index += width + 2;
}
else if (weightIndex != 0)
{
index = index + 2;
}
int[] indices = { index, index + 1, index + width, index + width + 1 };
foreach (int i in indices)
{
weights2.dataFloat[i] = data.m_weights[0].array[nodeIndex].dataFloat[weightIndex] * 0.25f;
}
}
else
{
if (addIndex)
{
index += width + 2;
addIndex = false;
}
//Debug.Log(weightIndex + ": (" + index + ")");
weights2.dataFloat[index] = data.m_weights[0].array[nodeIndex].dataFloat[weightIndex] * 0.5f;
weights2.dataFloat[index + 1] = data.m_weights[0].array[nodeIndex].dataFloat[weightIndex] * 0.5f;
index += 2;
}
//else
//Debug.Log("Warning!");
newWeights[nodeIndex] = weights2;
}
}
}
data.m_weights[0].array = newWeights;
NNCSaveData containerData = new NNCSaveData
{
m_trainingData = m_trainingManager.SaveData(),
m_sampleData = m_sampleManager.SaveData(),
m_visuilizationNetworkData = m_visualizationNetwork.SaveData(),
//m_visuilizationSampleData = m_visualizationSample.SaveData(),
m_networkData = data,
m_layerLengths = m_layerLengths,
m_dataFileName = m_dataFileName,
m_activisionType = m_activisionType,
m_activisionTypeOutput = m_activisionTypeOutput,
m_costType = m_costType,
m_initializationType = m_initializationType,
m_activisionConstant = m_activisionConstant
};
LoadContainer(containerData);
}