KevinBaselinesw
commented
3 years ago Below is the tensordot function implementation. You can see the error being thrown. This algorithm was ported right from the real numpy code. Assuming I didn't make a mistake in porting, the logic is what python/numpy uses.
I suspect what is happening is that the code does not like your axes definition. I wonder if you can drop this code into your app and then step through it? It may lead to an idea of what is wrong.
If you want to send me a small code sample that I can plug into a unit test, I can try to debug it and provide some more information.
Also, have you tried to get your code to work in python/numpy?
public static ndarray tensordot(object a, object b, (npy_intp[], npy_intp[]) axes)
{
npy_intp[] axes_a = axes.Item1;
npy_intp[] axes_b = axes.Item2;
int na = axes_a.Length;
int nb = axes_b.Length;
var aa = asanyarray(a);
var bb = asanyarray(b);
var as_ = aa.shape;
var nda = aa.ndim;
var bs = bb.shape;
var ndb = bb.ndim;
bool equal = true;
if (na != nb)
{
equal = false;
}
else
{
for (int k = 0; k < na; k++)
{
long asindex = axes_a[k];
asindex = asindex >= 0 ? asindex : asindex += as_.iDims.Length;
long bsindex = axes_b[k];
bsindex = bsindex >= 0 ? bsindex : bsindex += bs.iDims.Length;
if (as_.iDims[asindex] != bs.iDims[bsindex])
{
equal = false;
break;
}
if (axes_a[k] < 0)
{
axes_a[k] += nda;
}
if (axes_b[k] < 0)
{
axes_b[k] += ndb;
}
}
}
if (!equal)
{
throw new ValueError("shape-mismatch for sum");
}
/**********************************/
List<int> notin = new List<int>();
for (int k = 0; k < nda; k++)
{
if (!axes_a.Contains(k))
{
notin.Add(k);
}
}
List<npy_intp> newaxes_a = new List<npy_intp>();
foreach (var k in notin)
newaxes_a.Add(k);
foreach (var k in axes_a)
newaxes_a.Add(k);
npy_intp N2 = 1;
foreach (var axis in axes_a)
{
N2 *= as_.iDims[axis];
}
List<npy_intp> asax = new List<npy_intp>();
foreach (var ax in notin)
{
asax.Add(as_.iDims[ax]);
}
var multreduce = ufunc.reduce(UFuncOperation.multiply, asanyarray(asax.ToArray()));
var newshape_a = new shape((npy_intp)multreduce.GetItem(0), N2);
List<npy_intp> olda = new List<npy_intp>();
foreach (var axis in notin)
{
olda.Add(as_.iDims[axis]);
}
/**********************************/
notin = new List<int>();
for (int k = 0; k < ndb; k++)
{
if (!axes_b.Contains(k))
{
notin.Add(k);
}
}
List<npy_intp> newaxes_b = new List<npy_intp>();
foreach (var k in axes_b)
newaxes_b.Add(k);
foreach (var k in notin)
newaxes_b.Add(k);
N2 = 1;
foreach (var axis in axes_b)
{
N2 *= bs.iDims[axis];
}
List<npy_intp> bsax = new List<npy_intp>();
foreach (var ax in notin)
{
bsax.Add(bs.iDims[ax]);
}
multreduce = ufunc.reduce(UFuncOperation.multiply, asanyarray(bsax.ToArray()));
var newshape_b = new shape(N2, (npy_intp)multreduce.GetItem(0));
List<npy_intp> oldb = new List<npy_intp>();
foreach (var axis in notin)
{
oldb.Add(bs.iDims[axis]);
}
var at = aa.Transpose(newaxes_a.ToArray()).reshape(newshape_a);
var bt = bb.Transpose(newaxes_b.ToArray()).reshape(newshape_b);
var res = np.dot(at, bt);
olda.AddRange(oldb);
return res.reshape(olda);
}
AsiaMartini
I'm trying to convert the python function einsum (commented line below) but I'm getting this error:
Am I passing the axis in the wrong way? In SciSharp/Numpy.NET I can do this without problems:
Could you please help me? Thanks.