Skip to content

class  MetaModel

Description

Special model used for initialization optimization with MAML and MAML based methods. Containing backbone model(CONV4, for example) and additional modules.

Parameters

  • backbone: Module Backbone model, could

  • learn_lr: bool, default=False Whether to learning inner learning rate during outer optimization, i.e. use Meta-SGD method.

  • meta_lr: float, default=0.1 Learning rate of inner optimization.

  • use_t: bool, default=False Whether to add T-layers, i.e. use MT-net method.

  • use_warp: bool, default=False Whether to add Warp-blocks, i.e. use Warp-grad method.

  • num_warp_layers: int, default=1 Num of conv layers in one warp block.

  • use_forget: bool, default=False Whether to add attenuator, i.e. use Learning-to-Forget method.

  • enable_inner_loop_optimizable_bn_params: bool, default=False When use L2F method, whether to add the attenuation operation to the batch-norm modules.