public override void Setup(TensorCollection bottom, TensorCollection top)
{
base.Setup(bottom, top);
this.n = Parameters.Outputs;
this.m = bottom[0].Num;
this.k = bottom[0].Count / bottom[0].Num;
// Reshape the output
top[0].Reshape(m, n, 1, 1);
if (this.weights == null || this.bias == null)
{
// Fill the weights
this.weights = new Tensor(1, 1, n, k);
var weightsFiller = FillerFactory.Create(Parameters.WeightsFiller);
weightsFiller.Fill(weights);
// If necessary, initialize and fill the bias term
if (Parameters.UseBias)
{
this.bias = new Tensor(1, 1, 1, n);
var biasFiller = FillerFactory.Create(Parameters.BiasFiller);
biasFiller.Fill(bias);
}
}
else
{
// LOG we are skipping the parameter initialization
}
if (Parameters.UseBias)
{
this.biasMultiplier = new Tensor(1, 1, 1, m);
using (var biasMultiplierCpu = this.biasMultiplier.OnCpu())
{
biasMultiplierCpu.Data.Map(v => 1, biasMultiplierCpu.Data, Zeros.Include);
}
}
}
public override void Setup(IList <Blob> bottom, IList <Blob> top)
{
base.Setup(bottom, top);
Guard.That(() => this.Parameters.Kernel.Width).IsGreaterThan(0);
Guard.That(() => this.Parameters.Kernel.Height).IsGreaterThan(0);
Guard.That(() => this.Parameters.Kernel.Depth).IsEqual(0);
Guard.That(() => this.Parameters.NumberOfOutputs).IsGreaterThan(0);
int num = bottom[0].Num;
int channels = bottom[0].Channels;
int height = bottom[0].Height;
int width = bottom[0].Width;
foreach (var bottomBlob in bottom)
{
Guard.That(() => bottomBlob.Num).IsEqual(num);
Guard.That(() => bottomBlob.Channels).IsEqual(channels);
Guard.That(() => bottomBlob.Height).IsEqual(height);
Guard.That(() => bottomBlob.Width).IsEqual(width);
}
Size padding = this.Parameters.Padding;
Size stride = this.Parameters.Stride;
Size kernel = this.Parameters.Kernel;
if (this.IsScaleKernel)
{
throw new NotSupportedException("1x1 kernels are not supported yet.");
}
// Number of output should be multiples of group.
Guard.That(() => this.Parameters.NumberOfOutputs % this.Parameters.Groups).IsEqual(0);
// Number of channels should be multiples of group.
Guard.That(() => channels % this.Parameters.Groups).IsEqual(0);
// We are going to work 1 image at a time to avoid overly large memory usage.
int outputHeight = (height + 2 * padding.Height - kernel.Height) / stride.Height + 1;
int outputWidth = (width + 2 * padding.Width - kernel.Width) / stride.Width + 1;
this.imageBuffer = new Blob(1, channels * kernel.Height * kernel.Width, outputHeight, outputWidth);
// Figure out the dimensions for individual multiplications.
this.m = this.Parameters.NumberOfOutputs / this.Parameters.Groups;
this.k = channels * kernel.Height * kernel.Width / this.Parameters.Groups;
this.n = outputHeight * outputWidth;
// Resize the output
foreach (var topBlob in top)
{
topBlob.Reshape(num, this.Parameters.NumberOfOutputs, outputHeight, outputWidth);
}
// Initialize the weight
this.weights = new Blob(this.Parameters.NumberOfOutputs, channels / this.Parameters.Groups, this.Parameters.Kernel.Height, this.Parameters.Kernel.Width);
var weightsFiller = FillerFactory.Create(this.Parameters.WeightsFiller);
weightsFiller.Fill(this.weights);
this.SetPropagateDownForParameter(0, true);
if (this.Parameters.UseBias)
{
// Initialize the bias
this.bias = new Blob(1, 1, 1, this.Parameters.NumberOfOutputs);
var biasFiller = FillerFactory.Create(this.Parameters.BiasFiller);
biasFiller.Fill(this.bias);
this.SetPropagateDownForParameter(1, true);
}
this.biasMultiplier = new Blob(1, 1, 1, this.n);
this.biasMultiplier.InitializeWith(1, 0);
}
