pass value_scale to the LearnerND's loss_per_simplex function

[basnijholt]

Currently, the LearnerND cannot access the unscaled value in the loss_per_simplex.

This PR fixes this by passing the scale. This fixes a regression introduced when we added scaling.

Now the following code:

import numpy as np
import adaptive
adaptive.notebook_extension()

def isoline_loss_function(y_iso, sigma, priority=1):
    from adaptive.learner.learnerND import default_loss

    def gaussian(x, mu, sigma):
        return np.exp(-(x - mu) ** 2 / sigma ** 2)

    def loss(simplex, values, value_scale):
        distance = np.mean([abs(y_iso * value_scale - y) for y in values])
        return priority * gaussian(distance, 0, sigma) + default_loss(simplex, values, value_scale)

    return loss

level = 0.5
loss = isoline_loss_function(y_iso=level, sigma=1, priority=0.3)

def f(xy):
    x, y = xy
    return x ** 2 + y ** 2 / 2 + y ** 3

learner = adaptive.LearnerND(
    f,
    bounds=[(-1, 1), (-1, 1)],
    loss_per_simplex=loss,
)

adaptive.runner.simple(learner, goal=lambda l: l.npoints > 1000)
%opts Image [colorbar=True]
learner.plot_isoline(level, tri_alpha=0.3)

produces

instead of

[jbweston]

Seems ok, don't know why tests are failing

[jbweston]

@basnijholt the test failures are genuine.

>      loss_curvature = triangle_loss(simplex, values, neighbors, neighbor_values)
E       TypeError: triangle_loss() missing 1 required positional argument: 'neighbor_values'

Threre may be others

[basnijholt]

Thanks! I've already pushed a fix.

[basnijholt]

@jbweston merge if you are OK with this 👍