mkofler96 / DeepSDFStruct / 30428676523

83%

push

github

Merge PR #75 (Improve reconstruction) with conflicts resolved (#76)

* cooking

* new training with mlflow tracking

* Modified parameterization such that kont vectors are defined on actual normalized domain.

* Fix MMA optimizer for multi-constraint problems

Corrects reshape of gradient matrices for n_constraints > 1, uses
parameters.numel() instead of len(parameters), and initializes
state vectors as proper column arrays.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Add quantum decoder with optional pennylane dependency

Adds QuantumDeepSDFDecoder (PennyLane-based QNN) as a drop-in replacement
for classical decoders. Pennylane is now an optional dependency installed
via `uv sync --extra quantum`. The workspace registry loads the quantum
decoder only when pennylane is available.

Removes the unused cell_phase_deep_sdf_decoder. Adds quantum-aware device
selection and DataParallel guards to the training pipeline.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Improve reconstruction pipeline: structured return, CSV export

reconstruct_from_samples now returns a dict with params, loss_history,
final_loss, and num_steps instead of a bare parameter list. Adds
loss_csv_path parameter for CSV export of training loss. Removes debug
mesh export code and duplicate function definition.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Simplify save_points_to_vtp to accept single (N,4) tensor

Replaces the old (filename, neg, pos) signature with (filename, points)
where points is an (N,4) array of [x, y, z, sdf]. Handles torch tensors
and numpy arrays transparently. Fixes the internal call site to match.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Add near-surface SDF error metrics and Chamfer distance

Adds error_metrics module with compute_near_surface_metrics (MAE, RMSE,
median, quantiles) and compute_metrics_from_vtp for comparing VTP files.
Extends mesh_to_analytical with symmetric Chamfer distance.

Co-... (continued)

633 of 747 branches covered (84.74%)

Branch coverage included in aggregate %.

5307 of 6422 relevant lines covered (82.64%)

0.83 hits per line