Shortcuts

Better performance optimizers

This document provides some third-party optimizers supported by MMEngine, which may bring faster convergence speed or higher performance.

D-Adaptation

D-Adaptation provides DAdaptAdaGrad, DAdaptAdam and DAdaptSGD optimizers.

Note

If you use the optimizer provided by D-Adaptation, you need to upgrade mmengine to 0.6.0.

  • Installation

pip install dadaptation
  • Usage

Take the DAdaptAdaGrad as an example.

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader_cfg,
    # To view the input parameters for DAdaptAdaGrad, you can refer to
    # https://github.com/facebookresearch/dadaptation/blob/main/dadaptation/dadapt_adagrad.py
    optim_wrapper=dict(optimizer=dict(type='DAdaptAdaGrad', lr=0.001, momentum=0.9)),
    train_cfg=dict(by_epoch=True, max_epochs=3),
)
runner.train()

Lion-Pytorch

lion-pytorch provides the Lion optimizer.

Note

If you use the optimizer provided by Lion-Pytorch, you need to upgrade mmengine to 0.6.0.

  • Installation

pip install lion-pytorch
  • Usage

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader_cfg,
    # To view the input parameters for Lion, you can refer to
    # https://github.com/lucidrains/lion-pytorch/blob/main/lion_pytorch/lion_pytorch.py
    optim_wrapper=dict(optimizer=dict(type='Lion', lr=1e-4, weight_decay=1e-2)),
    train_cfg=dict(by_epoch=True, max_epochs=3),
)
runner.train()

Sophia

Sophia provides Sophia, SophiaG, DecoupledSophia and Sophia2 optimizers.

Note

If you use the optimizer provided by Sophia, you need to upgrade mmengine to 0.7.4.

  • Installation

pip install Sophia-Optimizer
  • Usage

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader_cfg,
    # To view the input parameters for SophiaG, you can refer to
    # https://github.com/kyegomez/Sophia/blob/main/Sophia/Sophia.py
    optim_wrapper=dict(optimizer=dict(type='SophiaG', lr=2e-4, betas=(0.965, 0.99), rho = 0.01, weight_decay=1e-1)),
    train_cfg=dict(by_epoch=True, max_epochs=3),
)
runner.train()

bitsandbytes

bitsandbytes provides AdamW8bit, Adam8bit, Adagrad8bit, PagedAdam8bit, PagedAdamW8bit, LAMB8bit, LARS8bit, RMSprop8bit, Lion8bit, PagedLion8bit and SGD8bit optimizers.

Note

If you use the optimizer provided by bitsandbytes, you need to upgrade mmengine to 0.9.0.

  • Installation

pip install bitsandbytes
  • Usage

Take the AdamW8bit as an example.

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader_cfg,
    # To view the input parameters for AdamW8bit, you can refer to
    # https://github.com/TimDettmers/bitsandbytes/blob/main/bitsandbytes/optim/adamw.py
    optim_wrapper=dict(optimizer=dict(type='AdamW8bit', lr=1e-4, weight_decay=1e-2)),
    train_cfg=dict(by_epoch=True, max_epochs=3),
)
runner.train()

transformers

transformers provides Adafactor optimzier.

Note

If you use the optimizer provided by transformers, you need to upgrade mmengine to 0.9.0.

  • Installation

pip install transformers
  • Usage

Take the Adafactor as an example.

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    train_dataloader=train_dataloader_cfg,
    # To view the input parameters for Adafactor, you can refer to
    # https://github.com/huggingface/transformers/blob/v4.33.2/src/transformers/optimization.py#L492
    optim_wrapper=dict(optimizer=dict(type='Adafactor', lr=1e-5,
        weight_decay=1e-2, scale_parameter=False, relative_step=False)),
    train_cfg=dict(by_epoch=True, max_epochs=3),
)
runner.train()
Read the Docs v: v0.10.4
Versions
latest
stable
v0.10.4
v0.10.3
v0.10.2
v0.10.1
v0.10.0
v0.9.1
v0.9.0
v0.8.5
v0.8.4
v0.8.3
v0.8.2
v0.8.1
v0.8.0
v0.7.4
v0.7.3
v0.7.2
v0.7.1
v0.7.0
v0.6.0
v0.5.0
v0.4.0
v0.3.0
v0.2.0
Downloads
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.