public void SpeedTest()
{
int inwidth = 32, inheight = 32, inchannels = 33, outchannels = 33, ksize = 3, stride = 2;
int outwidth = (inwidth - ksize) / stride + 1, outheight = (inheight - ksize) / stride + 1;
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map2D(outchannels, outwidth, outheight));
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel2D(inchannels / 3 * 4, outchannels / 3, ksize, ksize));
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map2D(inchannels, inwidth, inheight));
TrivectorDeconvolution2D ope = new TrivectorDeconvolution2D(inwidth, inheight, outchannels, inchannels, ksize, ksize, stride);
ope.Execute(y_tensor, w_tensor, x_tensor);
Stopwatch sw = new Stopwatch();
sw.Start();
ope.Execute(y_tensor, w_tensor, x_tensor);
ope.Execute(y_tensor, w_tensor, x_tensor);
ope.Execute(y_tensor, w_tensor, x_tensor);
ope.Execute(y_tensor, w_tensor, x_tensor);
sw.Stop();
Console.WriteLine($"{sw.ElapsedMilliseconds / 4} msec");
}
public void OverflowTest()
{
foreach (bool gradmode in new bool[] { false, true })
{
foreach (int batch in new int[] { 1, 2, 3 })
{
foreach (int inchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int outchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int kheight in new int[] { 1, 3, 5 })
{
foreach (int kwidth in new int[] { 1, 3, 5 })
{
foreach (int stride in new int[] { 1, 2, 3 })
{
foreach (int inwidth in new int[] { 8, 9, 13, 17 })
{
foreach (int inheight in new int[] { 8, 9, 19, 23 })
{
int outwidth = (inwidth - kwidth) / stride + 1, outheight = (inheight - kheight) / stride + 1;
float[] yval = (new float[outwidth * outheight * outchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[kwidth * kheight * inchannels * outchannels / 9 * 4]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map2D(outchannels, outwidth, outheight, batch), yval);
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel2D(inchannels / 3 * 4, outchannels / 3, kwidth, kheight), wval);
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map2D(inchannels, inwidth, inheight, batch));
TrivectorDeconvolution2D ope = new TrivectorDeconvolution2D(inwidth, inheight, outchannels, inchannels, kwidth, kheight, stride, gradmode, batch);
ope.Execute(y_tensor, w_tensor, x_tensor);
CollectionAssert.AreEqual(yval, y_tensor.State);
CollectionAssert.AreEqual(wval, w_tensor.State);
x_tensor.CheckOverflow();
Console.WriteLine($"pass: {inchannels},{outchannels},{kwidth},{kheight},{stride},{inwidth},{inheight},{batch},{gradmode}");
}
}
}
}
}
}
}
}
}
}
public void ExecuteTest()
{
int inchannels = 9, outchannels = 12, inwidth = 13, inheight = 17, kwidth = 3, kheight = 5, stride = 2, batch = 7;
VariableField x = new Tensor(Shape.Map2D(inchannels, inwidth, inheight, batch));
Layer layer = new TrivectorDeconvolution2D(inchannels, outchannels, kwidth, kheight, stride, use_bias: true, pad_mode: PaddingMode.Edge, "conv");
Field y = layer.Forward(x);
(Flow flow, Parameters parameters) = Flow.Optimize(y);
flow.Execute();
Assert.AreEqual(2, parameters.Count);
Assert.AreEqual(inchannels, layer.InChannels);
Assert.AreEqual(outchannels, layer.OutChannels);
Assert.AreEqual(kwidth, layer.Width);
Assert.AreEqual(kheight, layer.Height);
}
public void ExecuteTest()
{
float max_err = 0;
foreach (int batch in new int[] { 1, 2, 3 })
{
foreach (int inchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int outchannels in new int[] { 3, 6, 9, 12 })
{
foreach (int kheight in new int[] { 1, 3, 5 })
{
foreach (int kwidth in new int[] { 1, 3, 5 })
{
foreach (int stride in new int[] { 1, 2, 3 })
{
foreach (int inwidth in new int[] { 8, 9, 13, 17 })
{
foreach (int inheight in new int[] { 8, 9, 19, 23 })
{
int outwidth = (inwidth - kwidth) / stride + 1, outheight = (inheight - kheight) / stride + 1;
float[] yval = (new float[outwidth * outheight * outchannels * batch]).Select((_, idx) => idx * 1e-3f).ToArray();
float[] wval = (new float[kwidth * kheight * inchannels * outchannels / 9 * 4]).Select((_, idx) => idx * 1e-3f).Reverse().ToArray();
Trivector[] ycval = (new Trivector[yval.Length / 3])
.Select((_, idx) => new Trivector(yval[idx * 3], yval[idx * 3 + 1], yval[idx * 3 + 2])).ToArray();
Quaternion.Quaternion[] wcval = (new Quaternion.Quaternion[wval.Length / 4])
.Select((_, idx) => new Quaternion.Quaternion(wval[idx * 4], wval[idx * 4 + 1], wval[idx * 4 + 2], wval[idx * 4 + 3])).ToArray();
TrivectorMap2D y = new TrivectorMap2D(outchannels / 3, outwidth, outheight, batch, ycval);
Quaternion.QuaternionFilter2D w = new Quaternion.QuaternionFilter2D(inchannels / 3, outchannels / 3, kwidth, kheight, wcval);
TrivectorMap2D x = Reference(y, w, inwidth, inheight, kwidth, kheight, stride);
OverflowCheckedTensor y_tensor = new OverflowCheckedTensor(Shape.Map2D(outchannels, outwidth, outheight, batch), yval);
OverflowCheckedTensor w_tensor = new OverflowCheckedTensor(Shape.Kernel2D(inchannels / 3 * 4, outchannels / 3, kwidth, kheight), wval);
OverflowCheckedTensor x_tensor = new OverflowCheckedTensor(Shape.Map2D(inchannels, inwidth, inheight, batch));
TrivectorDeconvolution2D ope = new TrivectorDeconvolution2D(inwidth, inheight, outchannels, inchannels, kwidth, kheight, stride, gradmode: false, batch);
ope.Execute(y_tensor, w_tensor, x_tensor);
float[] x_expect = x.ToArray();
float[] x_actual = x_tensor.State;
CollectionAssert.AreEqual(yval, y_tensor.State);
CollectionAssert.AreEqual(wval, w_tensor.State);
AssertError.Tolerance(x_expect, x_actual, 1e-7f, 1e-5f, ref max_err, $"mismatch value {inchannels},{outchannels},{kwidth},{kheight},{stride},{inwidth},{inheight},{batch}");
Console.WriteLine($"pass: {inchannels},{outchannels},{kwidth},{kheight},{stride},{inwidth},{inheight},{batch}");
}
}
}
}
}
}
}
}
Console.WriteLine($"maxerr:{max_err}");
}
