Partial shadow inheritance

[henrikfo]

Description

Summary of changes
Updated the shadow model inheritance logic.

• Partial shadow_model config inheritance was implemented so missing shadow settings now fall back to the target model setup instead of requiring a fully duplicated shadow config. That includes inherited model path/class, init params, optimizer, criterion, epochs, and batch size, while still allowing shadow-specific overrides.

Furthermore, silent faults or possibility of running leakpro with unintended setups have been reduced.

• Shadow-model cache reuse was fixed so existing shadow models are no longer reused just because they share num_train and model_class. Reuse now checks the full effective training setup, including optimizer settings, criterion settings, init params, batch size, online/offline mode
• Shadow_model.batch_size now actually affects training. Before the field could be present in config but be silently ignored during dataloader creation
• Shadow metadata was expanded so saved shadow runs now record the fields needed for safe reuse decisions, including optimizer params, criterion params, batch size, and model module path
• Tests were added to cover the new behavior: one verifies partial inheritance from the target setup, one checks partial optimizer override behavior, one ensures stale cached shadow metadata is rejected when configs differ, and one verifies that the configured shadow batch size is really used

Resolved Issues

• fixes Partial shadow model inheritance #313
• fixas Expand hashing functionality #347

How Has This Been Tested?

This was tested by running the celebA DP example with both target model and shadow models without DPsgd enabled. Then another time, setting only
shadow_model:
init_params:
dpsgd: True

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB