public void Fit(Tensor x, Tensor y, uint batchSize = 32, uint epochs = 10, uint verbose = 1)
{
KerasProto kerasProto = new KerasProto();
kerasProto.Graph = _graph.ToString(Formatting.None);
kerasProto.BatchSize = batchSize;
kerasProto.Epochs = epochs;
kerasProto.Verbose = verbose;
kerasProto.Inputs.Add(x.GetProto());
kerasProto.Inputs.Add(y.GetProto());
kerasProto.Command = KerasCommand.Fit;
_state = new ProgressCallbackState(new ProgressWriter(epochs, (uint)x.Sizes[0]));
_callback = new ProgressCallback(_state.Callback);
kerasProto.ProgressCallback = (ulong)Marshal.GetFunctionPointerForDelegate(_callback);
using (var stream = new MemoryStream())
{
kerasProto.WriteTo(stream);
var bytes = stream.ToArray();
IntPtr outData = IntPtr.Zero;
uint outLen = 0;
ulong outPtr = 0;
IntPtr exceptionData = IntPtr.Zero;
uint exceptionLen = 0;
ulong exceptionPtr = 0;
KerasFitModel(bytes, (uint)bytes.Length, ref outData, ref outLen, ref outPtr, ref exceptionData, ref exceptionLen, ref exceptionPtr);
if (exceptionLen == 0)
{
var resultBytes = new byte[outLen];
Marshal.Copy(outData, resultBytes, 0, (int)outLen);
KerasDeletePointer(outPtr);
var resultProto = KerasProto.Parser.ParseFrom(resultBytes);
_model = resultProto.Model.ToArray();
}
else
{
var outBytes = new byte[exceptionLen];
Marshal.Copy(exceptionData, outBytes, 0, (int)exceptionLen);
var exception = new KerasException(Encoding.ASCII.GetString(outBytes));
KerasDeletePointer(exceptionPtr);
throw exception;
}
}
}
public void Fit(string path, uint nsamples, uint nfeatures, uint nlabels, uint batchSize, uint epochs, uint verbose = 1)
{
KerasProto kerasProto = new KerasProto();
kerasProto.Graph = _graph.ToString(Formatting.None);
kerasProto.Path = path;
kerasProto.Nsamples = nsamples;
kerasProto.Nfeatures = nfeatures;
kerasProto.Nlabels = nlabels;
kerasProto.BatchSize = batchSize;
kerasProto.Epochs = epochs;
kerasProto.Verbose = verbose;
using (var stream = new MemoryStream())
{
kerasProto.WriteTo(stream);
var bytes = stream.ToArray();
IntPtr outData = IntPtr.Zero;
uint outLen = 0;
ulong outPtr = 0;
IntPtr exceptionData = IntPtr.Zero;
uint exceptionLen = 0;
ulong exceptionPtr = 0;
KerasFitModel(bytes, (uint)bytes.Length, ref outData, ref outLen, ref outPtr, ref exceptionData, ref exceptionLen, ref exceptionPtr);
// KerasCntkDll.KerasFitModel(bytes, (uint)bytes.Length, ref outData, ref outLen, ref outPtr, ref exceptionData, ref exceptionLen, ref exceptionPtr);
if (exceptionLen == 0)
{
_model = new byte[outLen];
Marshal.Copy(outData, _model, 0, (int)outLen);
KerasDeletePointer(outPtr);
}
else
{
var outBytes = new byte[exceptionLen];
Marshal.Copy(exceptionData, outBytes, 0, (int)exceptionLen);
var exception = new KerasException(Encoding.ASCII.GetString(outBytes));
KerasDeletePointer(exceptionPtr);
throw exception;
}
}
}
public Tensor Predict(Tensor x, uint batchSize = 32, uint verbose = 1, bool cache = true)
{
KerasProto kerasProto = new KerasProto();
kerasProto.BatchSize = batchSize;
kerasProto.Verbose = verbose;
var jobj = new JObject()
{
["cache"] = cache
};
kerasProto.PredictParams = jobj.ToString(Formatting.None);
// TODO Consider not copying the model if we have a uuid
if (_model != null)
{
kerasProto.Model = ByteString.CopyFrom(_model);
}
kerasProto.ModelUuid = _uuid;
kerasProto.ModelPath = _path;
kerasProto.Inputs.Add(x.GetProto());
kerasProto.Command = KerasCommand.Predict;
using (var stream = new MemoryStream())
{
kerasProto.WriteTo(stream);
var bytes = stream.ToArray();
IntPtr outData = IntPtr.Zero;
uint outLen = 0;
ulong outPtr = 0;
IntPtr exceptionData = IntPtr.Zero;
uint exceptionLen = 0;
ulong exceptionPtr = 0;
KerasFitModel(bytes, (uint)bytes.Length, ref outData, ref outLen, ref outPtr, ref exceptionData, ref exceptionLen, ref exceptionPtr);
if (exceptionLen == 0)
{
var resultBytes = new byte[outLen];
Marshal.Copy(outData, resultBytes, 0, (int)outLen);
KerasDeletePointer(outPtr);
var resultProto = KerasProto.Parser.ParseFrom(resultBytes);
_uuid = resultProto.ModelUuid;
return(TensorUtils.Deserialize(resultProto.Outputs[0]));
}
else
{
var outBytes = new byte[exceptionLen];
Marshal.Copy(exceptionData, outBytes, 0, (int)exceptionLen);
var exception = new KerasException(Encoding.ASCII.GetString(outBytes));
KerasDeletePointer(exceptionPtr);
throw exception;
}
}
}
