Ranges when condition gives a zero length tensor

[Sprinzl]

Ok this one is a nice one: You can do something like this: where h is an tensor and target is a double...

h[h > target] = 1

... but this only works if the resulting vector is bigger than 0.

Is where any help or alternative, because thsi is a very ellegant solution. thanks in advance Michael

[NiklasGustafsson]

Hmmm. It works as I had expected it to do.

Here's my repro:

        [Fact]
        public void Validate_1187()
        {
            var target = 1.5f;
            using var h = torch.rand(10);
            using var expected = h.clone();

            // This should do nothing, including blow up.
            h[h > target] = 1.0f;

            Assert.Equal(expected, h);
        }

What was your expected result?

[Sprinzl]

Ok Niklas got the solution: h[h > target] = h[h > target] .copy_(1); The problem arises when the h[h > target] has zero length or is null. (Think......)

Thanks Niklas

The goal of the hole script us to get the first non-zero of the row and when hot-encode it. Whats for labeling for stock market data. I have an LLM Encoder with next step prediction self-subervised training. I want fine-tune as follows profit - neutral - stop loss. Something like this: from: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 2 0 0 0 0 0 0 0 0 0 1 1 1 2 0 0 0 0 0 0 0 0 0 0 1 1 2 0 0 0 0 0 0 0 0 0 0 1 1 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 1 0 0 0 0 2 2 0 0 0 2 2 2 1 1 0 0 0 2 2 0 0 0 2 2 2 2 2 1 0 0 0 0 0 0 0 0 0 2 2 0 2 2 to: 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 1 I use Rhino3d as visualization for tensors.....

[NiklasGustafsson]

Is the behavior different from PyTorch? If the assignment is supposed to do what your workaround does, can you give me a simple example in Python that illustrates it?

[Sprinzl]

It is a custom made code. All c#. So i cannot say. I have also a runtime-compiler plugin in the CAD-System which is compiling a vs-project and serializes it before. (So you do not to need to restart the programm - something what if you attach an dll in VS-Code does not work) So where can be a series of problems. I changed from
h[h > target] = 1.0f; to h[h > target] = h[h > target].copy_(1); All tensors in cuda. Where is the spagetti-monster code for what task:

private Tensor LabelEncoder(Tensor bar, double target, int steps) 

{ //Risk-Reward 1:1 //bar Sequence -- Open|H|L|C|WAP|Volume|Ratio|YearMonthDay|HourMinutesSeconds Func<Tensor, int, int,Tensor> hotselect = (x, dim, label) => { var max = torch.max(x != 0, dim: dim).indexes; max = max + arange(x.numel() / x.size(dim), dtype: ScalarType.Int64, device: x.device) x.size(dim); max = torch.nn.functional.one_hot(x.flatten()[max].to(ScalarType.Int64), label); var _n0 = max[.., 0]; var _n1 = max[.., 1]; max[.., 0] = _n1; max[.., 1] = n0; return max; }; int high = 1; int low = 2; int close = 3; var v = bar.clone(); v = v.unsqueeze(0); v = v.unfold(dimension: 1, size: steps, step: 1)[0, .., .., ..].permute(0, 2, 1); //[1269674x15x9] var h = (v[.., 1.., high].permute(1, 0) / v[.., 0, close]).permute(1, 0) - 1.0; h[h > target] = h[h > target].copy(1); h[h != 1] = h[h != 1].copy_(0); var l = ((v[.., 1.., low].permute(1, 0) / v[.., 0, close]).permute(1, 0) - 1.0) -1.0; l[l > target] = l[l > target].copy( 1); l[l != 1] = l[l != 1].copy(0); var mix = torch.zeroslike(h, device: bar.device); mix[(l == 1)] = mix[(l == 1)].copy(1); mix[(h == 1)] = mix[(h == 1)].copy_(2); var sum = mix.sum(1); var idx = hotselect(mix, 1, 3); var extend = torch.zeros(new long[] { steps - 1, 3 }, dtype: idx.dtype, device: idx.device); extend[.., 1] = 1.0; idx = torch.cat(new List() { idx, extend }, dim: 0L); return idx; }

[Sprinzl]

I cannot reproduce the error... no idea. Sorry. The mistake was some where else.