fixed bounds scaling for deepLS reconstruction

[lfreinberger]

No description provided.

[deepsource-io]

Here's the code health analysis summary for commits 0786f4f..82ea298. View details on DeepSource ↗.

Analysis Summary

Analyzer	Status	Summary	Link
Python	✅ Success	🎯 2 occurences resolved	View Check ↗

---

💡 If you’re a repository administrator, you can configure the quality gates from the settings.

[mkofler96]

What about using Newton:

def reconstruct_from_samples_newton(
    sdf: SDFBase,
    sdfSample: SampledSDF,
    num_iterations=200,          # LBFGS needs far fewer steps
    lr=1.0,                      # this is a step size, not Adam lr
    loss_fn="ClampedL1",
    batch_size=None,             # full batch is strongly recommended
):

    gt_dist = sdfSample.distances

    if loss_fn == "L1":
        Loss = torch.nn.L1Loss()
    elif loss_fn == "ClampedL1":
        Loss = ClampedL1Loss(clamp_val=0.1)
    elif loss_fn == "MSE":
        Loss = torch.nn.MSELoss()
    else:
        raise NotImplementedError(f"Loss function {loss_fn} not available.")

    dataset = TensorDataset(sdfSample.samples, gt_dist)

 
    if batch_size is None:
        dataloader = DataLoader(dataset, batch_size=len(dataset), shuffle=False)
    else:
        dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)

    optimizer = torch.optim.LBFGS(
        sdf.parameters(),
        lr=lr,
        max_iter=20,              # inner Newton iterations
        history_size=100,         # curvature memory
        line_search_fn="strong_wolfe",
    )

    loss_history = []
    pbar = trange(num_iterations, desc="Reconstructing SDF (L-BFGS)", leave=True)

    for _ in pbar:
        for queries, gt_batch in dataloader:

            def closure():
                optimizer.zero_grad()
                pred_dist = sdf(queries)
                loss = Loss(pred_dist, gt_batch)
                loss.backward()
                return loss

            loss = optimizer.step(closure)
            loss_val = loss.item()

            pbar.set_postfix({"loss": f"{loss_val:.5f}"})
            loss_history.append(loss_val)

    return list(sdf.parameters())