thsolver.lr_scheduler

Learning-rate scheduler factory functions used by thsolver.solver.Solver.config_lr_scheduler().

multi_step(optimizer, flags)

Builds a multi-step learning-rate scheduler.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

cos(optimizer, flags)

Builds a cosine annealing learning-rate scheduler.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

poly(optimizer, flags)

Builds a polynomial-decay learning-rate scheduler.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

constant(optimizer, flags)

Builds a constant learning-rate scheduler.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

cos_warmup(optimizer, flags)

Builds a cosine schedule with linear warmup.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

poly_warmup(optimizer, flags)

Builds a polynomial schedule with linear warmup.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

step_warmup(optimizer, flags)

Builds a step schedule with linear warmup.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node.

get_lr_scheduler(optimizer, flags)

Builds the configured learning-rate scheduler.

Parameters:

• optimizer – The optimizer to schedule.

• flags – The solver config node containing lr_type.

Returns:

The configured scheduler.

Return type:

torch.optim.lr_scheduler._LRScheduler