PyTorch and autograd

Install the optional backend with uv sync --extra torch or python -m pip install -e ".[torch]" from your checkout.

The explicit orivex.torch namespace keeps tensors on their existing device, preserves their floating dtype, and returns scalar tensors connected to the autograd graph. The differentiable profile contains distribution skewness and kurtosis, the ELA meta-model adjusted R-squared and linear-intercept features, and all six fitness-distance statistics:

import torch

from orivex.torch import TensorLandscapeSample, compute

device = "cuda" if torch.cuda.is_available() else "cpu"
x = torch.rand(200, 2, device=device, dtype=torch.float32, requires_grad=True)
y = torch.sum(x**2, dim=1)
bounds = torch.full((2,), 5.0, device=x.device, dtype=x.dtype)
sample = TensorLandscapeSample(x, y, lower=-bounds, upper=bounds)

result = compute(sample, "ela_distr.skewness")
feature = result.values["ela_distr.skewness"].value
assert feature is not None
feature.backward()

There is no implicit fallback to NumPy: requesting a known feature that has no Torch calculator raises UnsupportedFeatureError. Use orivex.torch.list_features() and orivex.torch.list_capabilities() to discover the implemented profile and its declared device, dtype, and autograd support.

Input and execution requirements

x and y must be tensors with matching float32 or float64 dtype and device. Tensor bounds must match them; array-like bounds are converted to the input dtype and device. Samples clone inputs while preserving autograd history and detect in-place mutation through tensor version counters.

Sample construction computes a fingerprint using a detached CPU snapshot; feature computation operates on the sample device. Account for construction separately when benchmarking GPU work.

The Torch compute function accepts y_normalization and options; it has no workers or rng arguments. Check backend capabilities for device support before requesting features. IC and NBC currently have no Torch calculators.

Preprocessing and gradients

Min-max normalization is the default and introduces piecewise differentiability at extrema. Use list_capabilities(y_normalization=None) to inspect the raw pipeline. Read normalization and fitness-distance selection for details on ties and gradient conventions.