Switch Training Time Config Hook （修改训练配置钩子的设计）

[Ezra-Yu]

In some training configurations, it is necessary to modify the training-related parameters in the training, such as the scale in RandomResizeCrop in transform, the magnitude_level of Random Augment, the alpha in MixUp, and so on. Such as :

efficientnetv2
mobileone
Data Augmentation Revisited: Rethinking the Distribution Gap between Clean and Augmented Data

在一个算法中，需要在训练时修改一些训练配置，比如 ’RandomResizeCrop 中 ‘scale‘’， Random Augment 中 ‘‘magnitude_level ’’,等等。以上的一些算法使用这些训练技巧

We are considering three ways to implement it.

Option1 a unified interface based on specific switch timing (基于变化节点)

 base_train_pipeline = [                                                                                                                                                                           
     dict(type='LoadImageFromFile'),                                                                                                                                                               
     dict(type='RandomResizedCrop', scale=224, backend='pillow'),                                                                                                                                  
     dict(type='RandomFlip', prob=0.5, direction='horizontal'),                                                                                                                                    
     dict(type='PackClsInputs')                                                                                                                                                                    
 ]                                                                                                                                                                                                 
                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        
 import copy                                                                                                                                                                                       
 # modify 37 epoch's RandomResizedCrop.scale to 192                                                                                                                                                 
 train_pipeline_37e = copy.deepcopy(base_train_pipeline)                                                                                                                                           
 train_pipeline_37e[1]['scale'] = 192                                                                                                                                                              
                                                                                                                                                                                                   
 # modify 112 epoch's RandomResizedCrop.scale to 224                                                                                                                                               
 train_pipeline_112e = copy.deepcopy(base_train_pipeline)                                                                                                                                          
 train_pipeline_112e[1]['scale'] = 224                                                                                                                                                             
                                                                                                                                                                                                   
                                                                                                                                                                                                   
 custom_hooks = [                                                                                                                                                                                  
     dict(                                                                                                                                                                                         
         type='SwitchTrainAugHook',                                                                                                                                                                 
         action_epoch=37,                                                                                                                                                                          
         pipeline=train_pipeline_37e),   # unchanged the batch augments and the loss                                                                                                                                        
     dict(                                                                                                                                                                                         
         type='SwitchTrainAugHook',                                                                                                                                                                 
         action_epoch=112,
         train_augments=None,            # stop all the batch augments
         loss=dict(type='CrossEntropyLoss', loss_weight=1.0),    # switch the loss                                                                                                                                                                        
         pipeline=train_pipeline_112e),                                                                                                                                                                                                                                                                                                                             
 ] 

Option2 Three interfaces based on the specific configuration (基于特定数据变换，3个hook)

custom_hooks = dict(
    dict(type='SwitchTrainPipelineHook',
         target='RandomResizedCrop',
         schedule=[dict(epoch=37, scale=192),
                   dict(epoch=112, scale=224)]),
    dict(type='SwitchBatchAugmentHook',
         schedule=[dict(epoch=112, augments=None)]),
    dict(type='SwitchLossHook',
         schedule=[dict(epoch=112, type='CrossEntropy', loss_weight=1.0)]),
)

Option3 a unified interface based on the specific configuration with ‘’target‘’ (基于特定数据变换，一个统一hook)

custom_hooks = dict(
    dict(type='SwitchTrainConfigHook',
         target='RandomResizedCrop',
         schedule=[dict(epoch=37, scale=192),
                   dict(epoch=112, scale=224)]),
    dict(type='SwitchTrainConfigHook',
         target='batch_augments',
         schedule=[dict(epoch=112, augments=None)]),
    dict(type='SwitchTrainConfigHook',
         target='loss',
         schedule=[dict(epoch=112, type='CrossEntropy', loss_weight=1.0)]),
)

prons and cons

1. Option1 Support for any requirement，uses only one interface, but needs copy;
2. Option2 Shorter and simpler， use three interfaces, but can not insert or del some data transform
3. Option3 Shorter and simpler， use three interfaces, but can not insert or del some data transforms.

优点及缺点

1. Option1 功能齐全，一个接口，但需要繁琐的复制
2. Option2 更短更简洁，维护三个接口，无法实现插入一个新的data transform，和关闭部分data transform
3. Option3 更短更简洁，维护一个接口，使用‘‘traget’’指定，无法实现插入一个新的data transform，和关闭部分data transform

[okotaku]

@Ezra-Yu If I want to combine it with other custom hooks, I need to put it in a list format, right?

# option1
 custom_hooks = [
     dict(type='EMAHook', momentum=4e-5, priority='ABOVE_NORMAL'),                                                                                                                                                                                  
     dict(                                                                                                                                                                                         
         type='SwitchTrainAugHook',                                                                                                                                                                 
         action_epoch=37,                                                                                                                                                                          
         pipeline=train_pipeline_37e),   # unchanged the batch augments and the loss                                                                                                                                        
     dict(                                                                                                                                                                                         
         type='SwitchTrainAugHook',                                                                                                                                                                 
         action_epoch=112,
         train_augments=None,            # stop all the batch augments
         loss=dict(type='CrossEntropyLoss', loss_weight=1.0),    # switch the loss                                                                                                                                                                        
         pipeline=train_pipeline_112e),                                                                                                                                                                                                                                                                                                                             
 ] 

# option2
custom_hooks = [
     dict(type='EMAHook', momentum=4e-5, priority='ABOVE_NORMAL'),                                                                                                                                                                                  
     dict(type='SwitchTrainPipelineHook',
         target='RandomResizedCrop',
         schedule=[dict(epoch=37, scale=192),
                   dict(epoch=112, scale=224)]),
    dict(type='SwitchBatchAugmentHook',
         schedule=[dict(epoch=112, augments=None)]),
    dict(type='SwitchLossHook',
         schedule=[dict(epoch=112, type='CrossEntropy', loss_weight=1.0)]),                                                                                                                                                                                                                                                                                                                 
 ] 

https://github.com/open-mmlab/mmclassification/blob/v1.0.0rc2/configs/convnext/convnext-base_32xb128_in1k.py#L18

[Ezra-Yu]

If I want to combine it with other custom hooks, I need to put it in a list format, right?

Yes.