mmengine/docs/zh_cn/migration/param_scheduler.md

564 lines
9.8 KiB
Markdown
Raw Normal View History

# 迁移 MMCV 参数调度器到 MMEngine
MMCV 1.x 版本使用 [LrUpdaterHook](https://mmcv.readthedocs.io/zh_CN/v1.6.0/api.html#mmcv.runner.LrUpdaterHook) 和 [MomentumUpdaterHook](https://mmcv.readthedocs.io/zh_CN/v1.6.0/api.html#mmcv.runner.MomentumUpdaterHook) 来调整学习率和动量。
但随着深度学习算法训练方式的不断发展,使用 Hook 修改学习率已经难以满足更加丰富的自定义需求,因此 MMEngine 提供了参数调度器ParamScheduler
一方面,参数调度器的接口与 PyTroch 的学习率调度器LRScheduler对齐另一方面参数调度器提供了更丰富的功能详细请参考[参数调度器使用指南](../tutorials/param_scheduler.md)。
## 学习率调度器LrUpdater迁移
MMEngine 中使用 LRScheduler 替代 LrUpdaterHook配置文件中的字段从原本的 `lr_config` 修改为 `param_scheduler`
MMCV 中的学习率配置与 MMEngine 中的参数调度器配置对应关系如下:
### 学习率预热Warmup迁移
由于 MMEngine 中的学习率调度器在实现时增加了 begin 和 end 参数指定了调度器的生效区间所以可以通过调度器组合的方式实现学习率预热。MMCV 中有 3 种学习率预热方式,分别是 `'constant'`, `'linear'`, `'exp'`,在 MMEngine 中对应的配置应修改为:
#### 常数预热(constant)
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
warmup='constant',
warmup_ratio=0.1,
warmup_iters=500,
warmup_by_epoch=False
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='ConstantLR',
factor=0.1,
begin=0,
end=500,
by_epoch=False),
dict(...) # 主学习率调度器配置
]
```
</td>
</tr>
</thead>
</table>
#### 线性预热(linear)
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
warmup='linear',
warmup_ratio=0.1,
warmup_iters=500,
warmup_by_epoch=False
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='LinearLR',
start_factor=0.1,
begin=0,
end=500,
by_epoch=False),
dict(...) # 主学习率调度器配置
]
```
</td>
</tr>
</thead>
</table>
#### 指数预热(exp)
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
warmup='exp',
warmup_ratio=0.1,
warmup_iters=500,
warmup_by_epoch=False
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='ExponentialLR',
gamma=0.1,
begin=0,
end=500,
by_epoch=False),
dict(...) # 主学习率调度器配置
]
```
</td>
</tr>
</thead>
</table>
### fixed 学习率FixedLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(policy='fixed')
```
</td>
<td>
```python
param_scheduler = [
dict(type='ConstantLR', factor=1)
]
```
</td>
</tr>
</thead>
</table>
### step 学习率StepLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='step',
step=[8, 11],
gamma=0.1,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='MultiStepLR',
milestone=[8, 11],
gamma=0.1,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### poly 学习率PolyLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='poly',
power=0.7,
min_lr=0.001,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='PolyLR',
power=0.7,
eta_min=0.001,
begin=0,
end=num_epochs,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### exp 学习率ExpLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='exp',
power=0.5,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='ExponentialLR',
gamma=0.5,
begin=0,
end=num_epochs,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### CosineAnnealing 学习率CosineAnnealingLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='CosineAnnealing',
min_lr=0.5,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='CosineAnnealingLR',
eta_min=0.5,
T_max=num_epochs,
begin=0,
end=num_epochs,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### FlatCosineAnnealing 学习率FlatCosineAnnealingLrUpdaterHook迁移
像 FlatCosineAnnealing 这种由多个学习率策略拼接而成的学习率,原本需要重写 Hook 来实现,而在 MMEngine 中只需将两个参数调度器组合即可
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='FlatCosineAnnealing',
start_percent=0.5,
min_lr=0.005,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='ConstantLR', factor=1, begin=0, end=num_epochs * 0.75)
dict(type='CosineAnnealingLR',
eta_min=0.005,
begin=num_epochs * 0.75,
end=num_epochs,
T_max=num_epochs * 0.25,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### CosineRestart 学习率CosineRestartLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(policy='CosineRestart',
periods=[5, 10, 15],
restart_weights=[1, 0.7, 0.3],
min_lr=0.001,
by_epoch=True)
```
</td>
<td>
```python
param_scheduler = [
dict(type='CosineRestartLR',
periods=[5, 10, 15],
restart_weights=[1, 0.7, 0.3],
eta_min=0.001,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
### OneCycle 学习率OneCycleLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(policy='OneCycle',
max_lr=0.02,
total_steps=90000,
pct_start=0.3,
anneal_strategy='cos',
div_factor=25,
final_div_factor=1e4,
three_phase=True,
by_epoch=False)
```
</td>
<td>
```python
param_scheduler = [
dict(type='OneCycleLR',
eta_max=0.02,
total_steps=90000,
pct_start=0.3,
anneal_strategy='cos',
div_factor=25,
final_div_factor=1e4,
three_phase=True,
by_epoch=False)
]
```
</td>
</tr>
</thead>
</table>
需要注意的是 `by_epoch` 参数 MMCV 默认是 `False`, MMEngine 默认是 `True`
### LinearAnnealing 学习率LinearAnnealingLrUpdaterHook迁移
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='LinearAnnealing',
min_lr_ratio=0.01,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
dict(type='LinearLR',
start_factor=1,
end_factor=0.01,
begin=0,
end=num_epochs,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
## 动量调度器MomentumUpdater迁移
MMCV 使用 `momentum_config` 字段和 MomentumUpdateHook 调整动量。 MMEngine 中动量同样由参数调度器控制。用户可以简单将学习率调度器后的 `LR` 修改为 `Momentum`,即可使用同样的策略来调整动量。动量调度器只需要和学习率调度器一样添加进 `param_scheduler` 列表中即可。举一个简单的例子:
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(...)
momentum_config = dict(
policy='CosineAnnealing',
min_momentum=0.1,
by_epoch=True
)
```
</td>
<td>
```python
param_scheduler = [
# 学习率调度器配置
dict(...),
# 动量调度器配置
dict(type='CosineAnnealingMomentum',
eta_min=0.1,
T_max=num_epochs,
begin=0,
end=num_epochs,
by_epoch=True)
]
```
</td>
</tr>
</thead>
</table>
## 参数更新频率相关配置迁移
如果在使用 epoch-based 训练循环且配置文件中按 epoch 设置生效区间(`begin``end`)或周期(`T_max`)等变量的同时希望参数率按 iteration 更新,在 MMCV 中需要将 `by_epoch` 设置为 False。而在 MMEngine 中需要注意,配置中的 `by_epoch` 仍需设置为 True通过在配置中添加 `convert_to_iter_based=True` 来构建按 iteration 更新的参数调度器,关于此配置详见[参数调度器教程](../tutorials/param_scheduler.md)。
以迁移CosineAnnealing为例
<table class="docutils">
<thead>
<tr>
<th>MMCV-1.x</th>
<th>MMEngine</th>
<tbody>
<tr>
<td>
```python
lr_config = dict(
policy='CosineAnnealing',
min_lr=0.5,
by_epoch=False
)
```
</td>
<td>
```python
param_scheduler = [
dict(
type='CosineAnnealingLR',
eta_min=0.5,
T_max=num_epochs,
by_epoch=True, # 注意by_epoch 需要设置为 True
convert_to_iter_based=True # 转换为按 iter 更新参数
)
]
```
</td>
</tr>
</thead>
</table>
你可能还想阅读[参数调度器的教程](../tutorials/param_scheduler.md)或者[参数调度器的 API 文档](mmengine.optim.scheduler)。