static Function <T> CreateFunction <T>(NodeProto node, long version, List <TensorProto> initializers, int[] inputShape, ref int initilizerIndex, out int[] outputShape) where T : unmanaged, IComparable <T>
{
switch (node.OpType)
{
case "BatchNormalization":
if (version >= 9)
{
TensorProto bn_scale = initializers[initilizerIndex++];
TensorProto bn_b = initializers[initilizerIndex++];
TensorProto bn_mean = initializers[initilizerIndex++];
TensorProto bn_var = initializers[initilizerIndex++];
BatchNormalization <T> batchNormalization = new BatchNormalization <T>(
channelSize: bn_scale.FloatDatas.Length,
useGamma: true,
useBeta: true,
eps: (TVal <T>)node.GetAttribute("epsilon").F,
name: node.Name,
inputNames: new[] { node.Inputs[0] },
outputNames: new[] { node.Outputs[0] },
decay: (TVal <T>)node.GetAttribute("momentum").F
);
Array.Copy(bn_scale.FloatDatas, batchNormalization.Gamma.Data, bn_scale.FloatDatas.Length);
Array.Copy(bn_b.FloatDatas, batchNormalization.Beta.Data, bn_b.FloatDatas.Length);
Array.Copy(bn_mean.FloatDatas, batchNormalization.AvgMean.Data, bn_mean.FloatDatas.Length);
Array.Copy(bn_var.FloatDatas, batchNormalization.AvgVar.Data, bn_var.FloatDatas.Length);
outputShape = inputShape;
return(batchNormalization);
}
else if (version >= 7)
{
TensorProto bn_scale = initializers[initilizerIndex++];
TensorProto bn_b = initializers[initilizerIndex++];
TensorProto bn_mean = initializers[initilizerIndex++];
TensorProto bn_var = initializers[initilizerIndex++];
//[spatial]
// If true, compute the mean and variance across per activation.
// If false, compute the mean and variance across per feature over each mini - batch.
// 真の場合は、活性化ごとに平均と分散を計算します。
// falseの場合は，ミニバッチごとに特徴量ごとの平均と分散を計算します．
//将来Axis対応することがあればコメントアウトを外す
//int[] axis = {0};
//if (node.GetAttribute("spatial").I != 1)
//{
// List<int> tmp = new List<int>();
// tmp.Add(0); //ここの次元指定はミニバッチ数に当たる
// tmp.AddRange(Enumerable.Range(2, inputShape.Length - 2));
// axis = tmp.ToArray();
//}
BatchNormalization <T> batchNormalization = new BatchNormalization <T>(
channelSize: bn_scale.FloatDatas.Length,
eps: (TVal <T>)node.GetAttribute("epsilon").F,
name: node.Name,
inputNames: new[] { node.Inputs[0] },
outputNames: new[] { node.Outputs[0] },
decay: (TVal <T>)node.GetAttribute("momentum").F
//axis: axis
);
Array.Copy(bn_scale.FloatDatas, batchNormalization.Gamma.Data, bn_scale.FloatDatas.Length);
Array.Copy(bn_b.FloatDatas, batchNormalization.Beta.Data, bn_b.FloatDatas.Length);
Array.Copy(bn_mean.FloatDatas, batchNormalization.AvgMean.Data, bn_mean.FloatDatas.Length);
Array.Copy(bn_var.FloatDatas, batchNormalization.AvgVar.Data, bn_var.FloatDatas.Length);
outputShape = inputShape;
return(batchNormalization);
}
else if (version >= 6)
{
//[spatial]
//If true, compute the mean and variance across all spatial elements.
//If false, compute the mean and variance across per feature.
//真の場合、すべての空間要素の平均と分散を計算します。
//偽の場合は，特徴量ごとの平均と分散を計算します。
throw new NotImplementedException();
}
else if (version >= 1)
{
throw new NotImplementedException();
}
break;
case "Conv":
if (version >= 11)
{
throw new NotImplementedException();
}
else if (version >= 1)
{
TensorProto conv_w = initializers[initilizerIndex++];
TensorProto conv_b = null;
if (node.Inputs.Count > 2)
{
conv_b = initializers[initilizerIndex++];
}
outputShape = inputShape;
return(new Convolution2D <T>(
inputChannels: (int)conv_w.Dims[1],
outputChannels: (int)conv_w.Dims[0],
kernelSize: Array.ConvertAll(node.GetAttribute("kernel_shape").Ints, s => (int)s),
stride: Array.ConvertAll(node.GetAttribute("strides").Ints, s => (int)s),
pad: Array.ConvertAll(node.GetAttribute("pads").Ints, s => (int)s), //pads: [x1_begin, x2_begin...x1_end, x2_end,...]で入ってくるので使用するのは前2つ
noBias: node.Inputs.Count < 3,
initialW: conv_w.FloatDatas,
initialb: conv_b?.FloatDatas,
name: node.Name,
inputNames: new[] { node.Inputs[0] },
outputNames: new[] { node.Outputs[0] }));
}
break;
case "Dropout":
if (version >= 12)
{
throw new NotImplementedException();
}
else if (version >= 10)
{
throw new NotImplementedException();
}
else if (version >= 7)
{
outputShape = inputShape;
return(new Dropout <T>((TVal <T>)node.GetAttribute("ratio").F, name: node.Name, inputNames: new[] { node.Inputs[0] }, outputNames: new[] { node.Outputs[0] }));
}
else if (version >= 6)
{
throw new NotImplementedException();
}
else if (version >= 1)
{
throw new NotImplementedException();
}
break;
case "Flatten":
outputShape = inputShape; //厳密には変わるが、関数内で吸収されるため不要
return(null);
case "Gemm":
if (version >= 11)
{
throw new NotImplementedException();
}
else if (version >= 9)
{
throw new NotImplementedException();
}
else if (version >= 7)
{
TensorProto w = initializers[initilizerIndex++];
TensorProto b = initializers[initilizerIndex++];
outputShape = new[]
{
inputShape[0], //バッチカウント
(int)w.Dims[0] //出力数
};
return(new Linear <T>(
inputCount: (int)w.Dims[1],
outputCount: (int)w.Dims[0],
name: node.Name,
inputNames: new[] { node.Inputs[0] },
outputNames: new[] { node.Outputs[0] },
noBias: false,
initialW: w.FloatDatas,
initialb: b.FloatDatas
));
}
else if (version >= 6)
{
throw new NotImplementedException();
}
else if (version >= 1)
{
throw new NotImplementedException();
}
break;
case "MaxPool":
if (version >= 12)
{
throw new NotImplementedException();
}
else if (version >= 11)
{
throw new NotImplementedException();
}
else if (version >= 10)
{
throw new NotImplementedException();
}
else if (version >= 8)
{
int[] kernelSize = Array.ConvertAll(node.GetAttribute("kernel_shape").Ints, s => (int)s);
int[] stride = Array.ConvertAll(node.GetAttribute("strides").Ints, s => (int)s);
int[] pad = Array.ConvertAll(node.GetAttribute("pads").Ints, s => (int)s);
List <int> tmpOutputShape = new List <int>();
tmpOutputShape.Add(inputShape[0]); //ミニバッチカウント
tmpOutputShape.Add(inputShape[1]); //チャンネル
tmpOutputShape.Add((int)Math.Floor((inputShape[2] - kernelSize[1] + pad[1] * 2.0f + stride[1] - 1.0f) / stride[1]) + 1);
tmpOutputShape.Add((int)Math.Floor((inputShape[3] - kernelSize[0] + pad[0] * 2.0f + stride[0] - 1.0f) / stride[0]) + 1);
outputShape = tmpOutputShape.ToArray();
return(new MaxPooling2D <T>(
kernelSize: kernelSize,
stride: stride,
pad: pad,
name: node.Name,
inputNames: new[] { node.Inputs[0] },
outputNames: new[] { node.Outputs[0] }
));
}
else if (version >= 1)
{
throw new NotImplementedException();
}
break;
case "Relu":
if (version >= 6)
{
outputShape = inputShape;
return(new ReLU <T>(name: node.Name, inputNames: new[] { node.Inputs[0] }, outputNames: new[] { node.Outputs[0] }));
}
else if (version >= 1)
{
throw new NotImplementedException();
}
break;
}
Console.WriteLine(node.OpType + "was not implemented.");
throw new NotImplementedException();
}
public static AttributeProto GetAttribute(this NodeProto node, string str)
{
return(node.Attributes.First(o => o.Name == str));
}