Fix ReduceLROnPlateauScheduler ignoring min_lr

[G26karthik]

Fixes #3875

Description:

ReduceLROnPlateauScheduler replaced min_lr with 0 when param_group_index is None, and with param_group_index=k > 0 it put the floor at index k while _reduce_lr reads min_lrs[0] for the single selected group. In both cases the LR decayed past min_lr.

This PR stops rewriting min_lr and passes it to torch's ReduceLROnPlateau as given. torch expands a float to one value per param group, so _reduce_lr reads the right floor whether one group or all groups are scheduled. The TypeError for a non-float min_lr together with param_group_index stays. A list min_lr without param_group_index now reaches torch too, instead of being replaced by 0.

For the test, I parametrized the existing test_reduce_lr_on_plateau_scheduler over param_group_index (None, 0, 1) and raised its min_lr from 1e-7 to 0.3, so the last reduction now hits the floor (0.25 is clamped to 0.3). The old 1e-7 was never reached, which is why the test didn't catch this. On master the None and 1 cases fail (they end at 0.25), and with the fix all 6 cases pass. The rest of tests/ignite/handlers/test_param_scheduler.py passes locally (65 passed; test_simulate_and_plot_values needs matplotlib, which isn't in my env), and ruff check/format are clean.

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[aaishwarymishra]

Could we extend this test with a few additional plateau metrics after the LR reaches 0.3, and assert that it remains at 0.3 on subsequent evaluations? Also, for param_group_index=None, the test currently checks only the first group asserting the full LR history would verify that both groups respect min_lr.?

[G26karthik]

Thanks for the review! Done in d3668ab. The test now runs four more plateau evaluations after the LR reaches 0.3 and checks that it stays there, and it compares the full LR history of every param group the scheduler updates, so with param_group_index=None both groups are checked at every step.

[G26karthik]

Hi @vfdev-5, sorry to bother you! This one was approved by @aaishwarymishra on Oct 7 and CI is green. Could you take a look when you get a chance? #3880 is a related fix for the same scheduler (its state_dict can't be restored from a checkpoint), in case you'd like to review them together. Thanks!

[vfdev-5]

Do we still need to make a copy now?

[G26karthik]

No, good point. It was only there so overwriting min_lr wouldn't touch the caller's dict, and **scheduler_kwargs is already a fresh dict, so setting verbose on it is fine too. Removed it in 9832774.