basecls.solver#
- class basecls.solver.BaseSolver[源代码]#
基类:
object
Base class for solver factory.
A solver factory should return a
Solver
object, which combines anOptimizer
and aGradManager
.
- class basecls.solver.DefaultSolver[源代码]#
基类:
BaseSolver
The default solver factory.
According to
cfg.reduce_mode
, learning rate and weight decay will be scaled automatically following the linear scaling rule, see “Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour” for more details.It supports
"sgd"
,"adam"
and"adamw"
.备注
This linear scaling rule can only work well with SGD. We are still looking for the applicable scaling rule for Adam and AdamW. Thus we recommend keeping default training settings (like learning rate and world size) when using Adam and AdamW.
- class basecls.solver.Solver(optimizer, grad_manager, grad_scaler)#
基类:
tuple
- grad_manager#
Alias for field number 1
- grad_scaler#
Alias for field number 2
- optimizer#
Alias for field number 0