public string SerializeCNN(string config, ML.CNN.Model Model, string saveas)
{
CNNModel cnnmodel = new CNNModel();
cnnmodel.config = config;
for (int i = 1; i < Model.Layers.Length; i++)
{
ML.CNN.Model.Unit lyr;
lyr = Model.Layers[i];
bool convtype = lyr.GetType().ToString() == "AI.ML.CNN.Layers.Convolution" ? true : false;
bool conctype = lyr.GetType().ToString() == "AI.ML.CNN.Layers.Connected" ? true : false;
switch (convtype)
{
case true:
ML.CNN.Layers.Convolution convLyr = (ML.CNN.Layers.Convolution)lyr;
AI.ML.CNN.Layers.Convolution.Kernel krn;
CNNModel.CNNWeights cnnweight;
for (int j = 0; j < convLyr.Filters.Count; j++)
{
// j: indexing filters
for (int k = 0; k < ((AI.ML.CNN.Layers.Convolution.Filter)convLyr.Filters[j]).Kernels.Length; k++)
{
//k: indexing kernels
krn = (AI.ML.CNN.Layers.Convolution.Kernel)((AI.ML.CNN.Layers.Convolution.Filter)convLyr.Filters[j]).Kernels[k];
cnnweight = new CNNModel.CNNWeights
{
Weights = krn.Weights
};
cnnmodel.cnnWeights.Add(cnnweight);
}
}
break;
case false:
switch (conctype)
{
case true:
ML.CNN.Layers.Connected connLyr = (ML.CNN.Layers.Connected)lyr;
ML.ANN.Neuron n; ML.ANN.Synapse syn;
CNNModel.ANNWeights annweight;
for (int j = 0; j < connLyr.Neurons.Length; j++)
{
n = connLyr.Neurons[j];
annweight = new CNNModel.ANNWeights();
annweight.Weights = new double?[n.Synapse.Count];
for (int k = 0; k < n.Synapse.Count; k++)
{
syn = n.Synapse[k];
annweight.Weights[k] = syn.W;
}
cnnmodel.annWeights.Add(annweight);
}
break;
case false:
break;
}
break;
default:
break;
}
}
saveas += ".txt";
JsonSerializer serializer = new JsonSerializer();
using (StreamWriter sw = new StreamWriter(saveas))
using (JsonWriter writer = new JsonTextWriter(sw))
{
serializer.Serialize(writer, cnnmodel);
}
return(Path.GetFullPath(saveas));
}
public ML.CNN.Model DeserializeCNN(string filename)
{
CNNModel cnnmodel = JsonConvert.DeserializeObject <CNNModel>(File.ReadAllText(filename));
ML.CNN.Model Model = new AI.ML.CNN.Model();
Model.Configure(cnnmodel.config);
int count = 0;
int anncount = 0;
for (int i = 1; i < Model.Layers.Length; i++)
{
ML.CNN.Model.Unit lyr;
lyr = Model.Layers[i];
bool convtype = lyr.GetType().ToString() == "AI.ML.CNN.Layers.Convolution" ? true : false;
bool conctype = lyr.GetType().ToString() == "AI.ML.CNN.Layers.Connected" ? true : false;
switch (convtype)
{
case true:
ML.CNN.Layers.Convolution convLyr = (ML.CNN.Layers.Convolution)lyr;
AI.ML.CNN.Layers.Convolution.Kernel krn;
for (int j = 0; j < convLyr.Filters.Count; j++)
{
// j: indexing filters
for (int k = 0; k < ((AI.ML.CNN.Layers.Convolution.Filter)convLyr.Filters[j]).Kernels.Length; k++)
{
//k: indexing kernels
krn = (AI.ML.CNN.Layers.Convolution.Kernel)((AI.ML.CNN.Layers.Convolution.Filter)convLyr.Filters[j]).Kernels[k];
krn.Weights = cnnmodel.cnnWeights[count++].Weights;
}
}
break;
case false:
switch (conctype)
{
case true:
ML.CNN.Layers.Connected connLyr = (ML.CNN.Layers.Connected)lyr;
ML.ANN.Neuron n; ML.ANN.Synapse syn;
double?[] weights;
for (int j = 0; j < connLyr.Neurons.Length; j++)
{
n = connLyr.Neurons[j];
weights = cnnmodel.annWeights[anncount++].Weights;
for (int k = 0; k < n.Synapse.Count; k++)
{
syn = n.Synapse[k];
syn.W = weights[k];
}
}
break;
case false:
break;
}
break;
default:
break;
}
}
return(Model);
}
