protected override void ConfirmField()
{
TemporaryOutput = new BufferField(GPU, new OpenCvSharp.Size(outW, outH), outCh);
TemporarySigma = new BufferField(GPU, new OpenCvSharp.Size(outW, outH), outCh);
Kernel = new KernelField(inCh, outCh, KernelSize, optimizer);
Kernel.Randmize();
}
public void Process(Gpu gpu, BufferField input, KernelField kernel, int dilation, int expand, ref BufferField output)
{
var iw = input.Width;
var ih = input.Height;
var ic = input.Channels;
var ibuf = input.Buffer;
var ow = output.Width;
var oh = output.Height;
var oc = output.Channels;
var obuf = output.Buffer;
var ks = kernel.Size;
var kbias = kernel.Bias;
var kbuf = kernel.Buffer;
gpu.For(0, output.Length, n =>
{
int c = (int)(n / (ow * oh));
int l = n - c * (ow * oh);
int y = (int)(l / ow);
int x = l - y * ow;
double v = kbias[c];
for (int _c = 0; _c < ic; _c++)
{
for (int s = -ks; s <= ks; s++)
{
for (int t = -ks; t <= ks; t++)
{
int i = x + s * dilation;
int j = y + t * dilation;
if (i % expand == 0 && j % expand == 0)
{
int _i = i / expand;
int _j = j / expand;
if (_i >= 0 && _i < iw && _j >= 0 && _j < ih)
{
v += ibuf[_c][_i, _j] * kbuf[_c][c][s + ks, t + ks];
}
}
}
}
}
obuf[c][x, y] = v;
});
}
protected override void InitializeInnerField()
{
M = Kernel.Congruence();
V = Kernel.Congruence();
}
public void Initialize(KernelField kernel)
{
Kernel = kernel;
InitializeInnerField();
}
public void Process(Gpu gpu, BufferField input, BufferField sigma, int dilation, int expand, ref BufferField propagater, ref KernelField kernel)
{
var iw = input.Width;
var ih = input.Height;
var ic = input.Channels;
var ibuf = input.Buffer;
var sw = sigma.Width;
var sh = sigma.Height;
var sc = sigma.Channels;
var sbuf = sigma.Buffer;
var ks = kernel.Size;
var kbias = kernel.Bias;
var kbuf = kernel.Buffer;
var dkbias = kernel.dBias;
var dkbuf = kernel.dBuffer;
var pw = propagater.Width;
var ph = propagater.Height;
var pc = propagater.Channels;
var pbuf = propagater.Buffer;
#region Update Kernel
gpu.For(0, sc, c =>
{
double db = dkbias[c];
double ddb = 0;
for (int x = 0; x < sw; x++)
{
for (int y = 0; y < sh; y++)
{
ddb += sbuf[c][x, y];
}
}
dkbias[c] = db + (ddb / (sw * sh));
});
gpu.For(0, (2 * ks + 1) * (2 * ks + 1), n =>
{
int s = (int)(n / (2 * ks + 1));
int t = n - s * (2 * ks + 1);
int _s = s - ks;
int _t = t - ks;
for (int c = 0; c < ic; c++)
{
for (int d = 0; d < sc; d++)
{
double dk = dkbuf[c][d][s, t];
double ddk = 0;
int cnk = 0;
for (int x = 0; x < sw; x++)
{
for (int y = 0; y < sh; y++)
{
int _x = x + _s * dilation;
int _y = y + _t * dilation;
if (_x % expand == 0 && _y % expand == 0)
{
int __x = _x / expand;
int __y = _y / expand;
if (__x >= 0 && __x < iw && __y >= 0 && __y < ih)
{
cnk++;
ddk += ibuf[c][__x, __y] * sbuf[d][x, y];
}
}
}
}
dkbuf[c][d][s, t] = dk + ((ddk / (sw * sh * (2 * (ks - 1) + 1))));
}
}
});
#endregion
#region Calculate Propagater
gpu.For(0, propagater.Length, n =>
{
int c = (int)(n / (pw * ph));
int l = n - c * (pw * ph);
int y = (int)(l / pw);
int x = l - y * pw;
double v = 0;
for (int _c = 0; _c < sc; _c++)
{
for (int s = ks; s >= -ks; s--)
{
for (int t = ks; t >= -ks; t--)
{
int i = x + s * dilation;
int j = y + t * dilation;
if (i % expand == 0 && j % expand == 0)
{
int _i = i / expand;
int _j = j / expand;
if (_i >= 0 && _i < sw && _j >= 0 && _j < sh)
{
v += sbuf[_c][_i, _j] * kbuf[c][_c][s + ks, t + ks];
}
}
}
}
}
pbuf[c][x, y] = v;
});
#endregion
}
