[Feature] Support Sophia optimizers (#1170)

docs/en/common_usage/better_optimizers.md

@@ -0,0 +1,92 @@
+# 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](https://github.com/facebookresearch/dadaptation) provides `DAdaptAdaGrad`, `DAdaptAdam` and `DAdaptSGD` optimziers。
+
+```{note}
+If you use the optimizer provided by D-Adaptation, you need to upgrade mmengine to `0.6.0`.
+```
+
+- Installation
+
+```bash
+pip install dadaptation
+```
+
+- Usage
+
+Take the `DAdaptAdaGrad` as an example.
+
+```python
+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](https://github.com/lucidrains/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
+
+```bash
+pip install lion-pytorch
+```
+
+- Usage
+
+```python
+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](https://github.com/kyegomez/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
+
+```bash
+pip install Sophia-Optimizer
+```
+
+- Usage
+
+```python
+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()
+```

docs/en/index.rst

@@ -24,6 +24,7 @@ You can switch between Chinese and English documents in the lower-left corner of
    common_usage/distributed_training.md
    common_usage/speed_up_training.md
    common_usage/save_gpu_memory.md
+   common_usage/better_optimizers.md
    common_usage/visualize_training_log.md
    common_usage/set_random_seed.md
    common_usage/debug_tricks.md

docs/en/tutorials/optim_wrapper.md

@@ -243,7 +243,7 @@ As shown in the above example, `OptimWrapperDict` exports learning rates and mom
 
 ### Configure the OptimWapper in [Runner](runner.md)
 
-We first need to configure the `optimizer` for the OptimWrapper. MMEngine automatically adds all optimizers in PyTorch to the `OPTIMIZERS` registry, and users can specify the optimizers they need in the form of a `dict`. All supported optimizers in PyTorch are listed [here](https://pytorch.org/docs/stable/optim.html#algorithms). In addition, `DAdaptAdaGrad`, `DAdaptAdam`, and `DAdaptSGD` can be used by installing [dadaptation](https://github.com/facebookresearch/dadaptation). `Lion` optimizer can used by install [lion-pytorch](https://github.com/lucidrains/lion-pytorch).
+We first need to configure the `optimizer` for the OptimWrapper. MMEngine automatically adds all optimizers in PyTorch to the `OPTIMIZERS` registry, and users can specify the optimizers they need in the form of a `dict`. All supported optimizers in PyTorch are listed [here](https://pytorch.org/docs/stable/optim.html#algorithms).
 
 Now we take setting up a SGD OptimWrapper as an example.
 

docs/zh_cn/common_usage/better_optimizers.md

@@ -0,0 +1,92 @@
+# 性能更优的优化器
+
+本文档提供了一些 MMEngine 支持的第三方优化器，它们可能会带来更快的收敛速度或者更高的性能。
+
+## D-Adaptation
+
+[D-Adaptation](https://github.com/facebookresearch/dadaptation) 提供了 `DAdaptAdaGrad`、`DAdaptAdam` 和 `DAdaptSGD` 优化器。
+
+```{note}
+如使用 D-Adaptation 提供的优化器，需将 mmengine 升级至 `0.6.0`。
+```
+
+- 安装
+
+```bash
+pip install dadaptation
+```
+
+- 使用
+
+以使用 `DAdaptAdaGrad` 为例。
+
+```python
+runner = Runner(
+    model=ResNet18(),
+    work_dir='./work_dir',
+    train_dataloader=train_dataloader_cfg,
+    # 如需查看 DAdaptAdaGrad 的输入参数，可查看
+    # 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
+
+[lion-pytorch](https://github.com/lucidrains/lion-pytorch) 提供了 `Lion` 优化器。
+
+```{note}
+如使用 Lion 提供的优化器，需将 mmengine 升级至 `0.6.0`。
+```
+
+- 安装
+
+```bash
+pip install lion-pytorch
+```
+
+- 使用
+
+```python
+runner = Runner(
+    model=ResNet18(),
+    work_dir='./work_dir',
+    train_dataloader=train_dataloader_cfg,
+    # 如需查看 Lion 的输入参数，可查看
+    # 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](https://github.com/kyegomez/Sophia) 提供了 `Sophia`、`SophiaG`、`DecoupledSophia` 和 `Sophia2` 优化器。
+
+```{note}
+如使用 Sophia 提供的优化器，需将 mmengine 升级至 `0.7.4`。
+```
+
+- 安装
+
+```bash
+pip install Sophia-Optimizer
+```
+
+- 使用
+
+```python
+runner = Runner(
+    model=ResNet18(),
+    work_dir='./work_dir',
+    train_dataloader=train_dataloader_cfg,
+    # 如需查看 SophiaG 的输入参数，可查看
+    # 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()
+```

docs/zh_cn/index.rst

@@ -24,6 +24,7 @@
    common_usage/distributed_training.md
    common_usage/speed_up_training.md
    common_usage/save_gpu_memory.md
+   common_usage/better_optimizers.md
    common_usage/visualize_training_log.md
    common_usage/set_random_seed.md
    common_usage/debug_tricks.md

docs/zh_cn/tutorials/optim_wrapper.md

@@ -243,7 +243,7 @@ print(optim_dict.get_momentum())  # {'gen.momentum': [0], 'disc.momentum': [0]}
 
 ### 在[执行器](./runner.md)中配置优化器封装
 
-优化器封装需要接受 `optimizer` 参数，因此我们首先需要为优化器封装配置 `optimizer`。MMEngine 会自动将 PyTorch 中的所有优化器都添加进 `OPTIMIZERS` 注册表中，用户可以用字典的形式来指定优化器，所有支持的优化器见 [PyTorch 优化器列表](https://pytorch.org/docs/stable/optim.html#algorithms)。另外，可以通过安装 [dadaptation](https://github.com/facebookresearch/dadaptation) 使用 `DAdaptAdaGrad`、`DAdaptAdam` 和 `DAdaptSGD` 3 个优化器。也可以通过安装 [lion-pytorch](https://github.com/lucidrains/lion-pytorch) 使用 `Lion` 优化器。
+优化器封装需要接受 `optimizer` 参数，因此我们首先需要为优化器封装配置 `optimizer`。MMEngine 会自动将 PyTorch 中的所有优化器都添加进 `OPTIMIZERS` 注册表中，用户可以用字典的形式来指定优化器，所有支持的优化器见 [PyTorch 优化器列表](https://pytorch.org/docs/stable/optim.html#algorithms)。
 
 以配置一个 SGD 优化器封装为例：
 

mmengine/optim/optimizer/builder.py

@@ -105,6 +105,30 @@ def register_lion_optimizers() -> List[str]:
 LION_OPTIMIZERS = register_lion_optimizers()
 
 
+def register_sophia_optimizers() -> List[str]:
+    """Register Sophia optimizer to the ``OPTIMIZERS`` registry.
+
+    Returns:
+        List[str]: A list of registered optimizers' name.
+    """
+    optimizers = []
+    try:
+        import Sophia
+    except ImportError:
+        pass
+    else:
+        for module_name in dir(Sophia):
+            _optim = getattr(Sophia, module_name)
+            if inspect.isclass(_optim) and issubclass(_optim,
+                                                      torch.optim.Optimizer):
+                OPTIMIZERS.register_module(module=_optim)
+                optimizers.append(module_name)
+    return optimizers
+
+
+SOPHIA_OPTIMIZERS = register_sophia_optimizers()
+
+
 def build_optim_wrapper(model: nn.Module,
                         cfg: Union[dict, Config, ConfigDict]) -> OptimWrapper:
     """Build function of OptimWrapper.