Save, reopen and resume¶

A model definition and a trained run are different files. Save the one you need; use both if you want an editable setup and a recorded training snapshot.

After training, File provides separate model and run/checkpoint actions. Open full size.

Reopen a setup¶

Choose File → Open model…. It restores the graph and settings. It does not load trained weights or recreate old predictions. Loading another setup uses the normal Save / Discard / Cancel check for your current unsaved work.

Continue a trained run¶

Choose File → Open saved run…. Result views stay closed until requested. Use Run → Resume checkpoint… to add more training updates. The archive carries the state required by its method, including SA masks or LRA balancing history.

Resume validates architecture, optimiser settings and input-data identity. Changing imported files, the architecture or numerical settings is not silently accepted as the same run. Copied inputs are included in new run archives; linked CSV/NPZ files must remain available for actual resume. Archived predictions can still be viewed if external input files moved.

What undo and layout remember¶

Undo/redo changes model definitions; it does not certify old results after a physics change. Numerical edits clear stale results/checkpoints. Moving blocks or renaming outputs preserves the trained result. Desktop layout/recovery preferences are stored separately from your model; no result view opens from layout recovery.

For exact formats and validation rules, see desktop persistence and validation scope.

Next: Export predictions.

Adaptive sampler runs also retain active points, importance weights and the sampling schedule. Resume verifies the original data and settings; it does not restart the sampler. See adaptive sampling and resume.

Project-owned files¶

Unsaved models have a private temporary working folder, created only when a component needs files. First Save copies these files beside the model into <name>_files; Save As copies them for the new model and preserves the original. Keep the model and its companion folder together when moving a project. Ordinary models without owned files do not create an empty companion folder. A failed save retains the temporary working files so saving can be retried. Discarding or closing the model removes its temporary working folder.

Jorvik geometry documents and their NetCDF snapshots use this owned folder. Their paths are stored relative to the project workspace. CSV/NPZ files can also be copied there when importing; existing links remain external unless you explicitly change File handling. Desktop recovery copies owned files alongside the recovery model; it pauses during external editing and resumes when the editor closes. Close Jorvik before saving, switching projects or closing Margam.

Choose copied or linked input data¶

Every CSV/NPZ mapping dialog now has File handling:

New sources default to Copy into project. Existing linked sources keep Link existing file when reopened, so upgrading or previewing an old model does not silently copy or replace its inputs. Choose Copy explicitly to convert one. Preview the mapping and click OK to apply it. Cancel and failed validation leave the mapping unchanged and do not install a new dataset.

The file field identifies a project copy by name; no temporary folder path needs to be managed by the user. The mapping and data readers are unchanged.

The choice applies independently to interior points, each named boundary set, evaluation points, Reference data and the older Reference block's imported values. CSV and NPZ contents are copied intact; coordinate/normal/value/ID mappings are preserved. The inspector shows Project copy or Linked file, with the linked location shown explicitly. To replace a copied snapshot with newer data, browse the desired original file and apply Copy again. To create a link from a copied source, choose an external file and select Link; Margam will not create a fragile absolute link into its own temporary folder.

Save a copied-data model and move both <name>.margam.json and <name>_files. Save As copies the owned data for the new model and retains the originals. It preserves trained weights for evaluation without extra optimizer updates. Opening a run archive restores copied inputs into a private workspace, allowing resume without the original model folder. Linked inputs remain external and are not silently bundled. Existing model/archive formats remain readable.

Each imported file is limited to 64 MiB, including the existing expanded-NPZ checks. Run archives retain their 256 MiB/4096-file limits. Copies retained for undo/redo are kept in the workspace; automatic pruning is not implemented. Copy/paste of blocks containing owned files remains disabled. CLI configuration continues to use explicit ordinary file paths; this copy/link choice belongs to the desktop project workflow.

Reproduce the portability check¶

.venv/bin/python -m examples.portable_data --device cpu
.venv/bin/python -m examples.portable_data --device cuda

Each command leaves a fresh run-* folder below build/portable-data and prints its location. Open portable.margam.json with its companion folder to inspect the finished model. The check copies six training/evaluation/reference sources, deletes the originals, verifies recovery, saves twice without changing weights, relocates and trains the model, then restores/resumes a run archive after removing the relocated model's inputs. Continuation matches uninterrupted training exactly on the same device. It also checks reference alignment, original point IDs/order, and CSV/VTU export. See the CPU record and CUDA record. This is a portability check, not a solution-accuracy benchmark.

Inspect the project's input data¶

Open View → Project data overview… before training or sharing a model. The window lists file mappings held by the blocks, including disconnected source blocks and individual named boundary sets. Generated built-in samples, analytical reference expressions and geometry documents are not imported datasets.

Actual triangle example: one copied interior source, three boundary mappings and an evaluation source. One linked boundary file was removed and another changed since its accepted import, detected after reopening the saved project. The overview leaves the model and datasets unchanged.

Select a row to inspect its location and coordinate, normal, value and ID mappings. Reference mappings also show the model's field order. A shared file can appear more than once because each set or block may map different arrays/columns.

• Refresh checks the current files. It does not watch the disk continuously, replace data, accept changed bytes, rewrite mappings or clear trained results. The ordinary baseline stays fixed on subsequent refreshes, including after Save As.
• Select block selects and centres the owner in the graph and opens its inspector. Named boundary rows select that boundary set too. Double-clicking a row does this.
• Map data… opens the existing mapping dialog. For Jorvik NetCDF inputs, Review Jorvik datasets… uses the geometry block's normal explicit update flow. These editing actions are disabled during training/external editing, and until Refresh when graph/settings changes make the table stale.

Newly accepted CSV/NPZ mappings record an import fingerprint, saved with the project and run archive. A changed linked file can therefore be identified after closing and reopening Margam, including on the first check. This applies to training, each boundary, evaluation and reference mappings, for both copies and links.

Existing models without a recorded fingerprint retain the session-only fallback. Their selected-row details explicitly say there is no saved import fingerprint; Margam does not invent a baseline for bytes it never observed. To establish one, open the mapping dialog, preview the file and click OK. The fingerprint describes that accepted version, not an unverifiable earlier import.

To accept newer bytes deliberately, reopen the existing mapping dialog, preview and apply it again. Mapping and fingerprint are one undoable edit. Refresh, Validate, Save and reopening an archive do not accept changed bytes. Save As and moving owned project files preserve the baseline. A fingerprint is bound to the exact mapping; changing its file/columns through other APIs does not reuse an old fingerprint for a different declaration.

A byte-change warning may reflect formatting-only changes; it does not establish that coordinates or fields changed numerically. Valid changed files produce a warning and can be used for a fresh training run. Strict checkpoint continuation still compares actual numerical data independently. Reaccepting a training input clears cached training/prediction state, even when its external pathname stays the same. Jorvik retains its stricter applied-snapshot fingerprint checks. Missing copied files are reported without recreating an empty working folder.

Use Validate for graph compatibility, boundary operators and reference alignment. Resume retains its separate strict checkpoint/data identity checks. The overview is a file and mapping diagnostic; it does not train, evaluate a network, open a result view or automatically repair dependencies.