【训练营】学习率调度器实现#113

Open

littleotherut wants to merge 8 commits intoInfiniTensor:masterfrom

littleotherut:lr_scheduler

Commits on Apr 1, 2026

refactor(optimizer): hoist learning_rate_ to Optimizer base and add lr accessors, passthrough SetLearningRate/GetLearningRate, and add initial_learning_rate and it's accessors

kinorw
authored and
littleotherut
committed
feat(lr_scheduler): add LRScheduler abstract base class, ConstantLR, StepLR, LinearLR, LambdaLR and SequentialLR

kinorw
authored and
littleotherut
committed
refactor(lr_scheduler): apply template method pattern to LRScheduler base class, add factory method Create<T>() with two-phase init and update all tests to use Create<T>() factory method.

kinorw
authored and
littleotherut
committed
feat(lr_scheduler): add ChainedScheduler composite strategy, and add closed and chained form, adjust LinearLR、SequentialLR

kinorw
authored and
littleotherut
committed
feat(lr_scheduler): add scheduler factory for CLI integration

kinorw
authored and
littleotherut
committed
refactor(examples): add scheduler placeholder and use runtime lr in logs, and integrate scheduler into training loop

kinorw
authored and
littleotherut
committed
style: apply clang-format to all legacy files
littleotherut
committed
refactor: rename current_lr_ to recover_lr_ and update related methods, add validation tests for learning rate schedulers

kinorw
authored and
littleotherut
committed