Skip to content

class  BOTHOptimizer

Description

Wrapper for performing bi-level optimization and gradient-based initialization optimization

BOTHOptimizer is the wrapper of Bi-Level Optimization(BLO) and Initialization Optimization(Initialization-based EGBR) process which builds LL, UL and Initialization problem solver with corresponding method modules and uses in training phase. The optimization process could also be done by using methods packages directly.

Parameters

  • method: str Define basic method for following training process, it should be included in ['Initial', 'Feature']. 'Initial' type refers to meta-learning optimization strategy, including methods like 'MAML, FOMAML, TNet, WarpGrad, L2F'; 'Feature' type refers to bi-level optimization strategy, includes methods like 'BDA, RHG, Truncated RHG, Onestage, BVFIM, IAPTT-GM, LS, NS, GN, BVFIM'.

  • ll_method: str, default=None method chosen for solving LL problem, including ['Dynamic' | 'Implicit' | 'BVFIM'].

  • ul_method: str, default=None Method chosen for solving UL problem, including ['Recurrence','Onestage' | 'LS','NS', 'GN' | 'BVFIM'].

  • ll_objective: callable, default=None An optimization problem which is considered as the constraint of UL problem.

    Callable with signature callable(state). Defined based on modeling of the specific problem that need to be solved. Computing the loss of ul problem. The state object contains the following:

    • "data" Data used in the LL optimization phase.
    • "target" Target used in the LL optimization phase.
    • "ul_model" UL model of the bi-level model structure.
    • "ll_model" LL model of the bi-level model structure.

  • ul_objective: callable, default=None The main optimization problem in a hierarchical optimization problem.

    Callable with signature callable(state). Defined based on modeling of the specific problem that need to be solved. Computing the loss of ul problem. The state object contains the following:

    • "data" Data used in the UL optimization phase.
    • "target" Target used in the UL optimization phase.
    • "ul_model" Ul model of the bi-level model structure.
    • "ll_model" LL model of the bi-level model structure.

  • inner_objective: callable, default=None The inner loop optimization objective.

    Callable with signature callable(state). Defined based on modeling of the specific problem that need to be solved. Computing the loss of inner objective. The state object contains the following:

    • "data" Data used in inner optimization phase.
    • "target" Target used in inner optimization phase.
    • "model" Meta model to be updated.
    • "updated_weights" Weights of model updated in inner-loop, will be used for forward propagation.

  • outer_objective: callable, default=None The outer optimization objective.

    Callable with signature callable(state). Defined based on modeling of the specific problem that need to be solved. Computing the loss of outer objective. The state object contains the following:

    • "data" Data used in outer optimization phase.
    • "target" Target used in outer optimization phase.
    • "model" Meta model to be updated.
    • "updated_weights" Weights of model updated in inner-loop, will be used for forward propagation.

  • ll_model: Module, default=None The model whose parameters will be updated during ul-level optimization.

  • ul_model: Module, default=None ul model in a hierarchical model structure whose parameters will be updated with ul objective.

  • meta_model: MetaModel, default=None Model whose initial value will be optimized. If choose MAML method to optimize, any user-defined torch nn.Module could be used as long as the definition of forward() function meets the standard; but if choose other derived methods, internally defined both.utils.model.meta_model should be used for related additional modules.

  • total_iters: int, default=60000 Total iterations of the experiment, used to set weight decay.

Methods

build_ll_problem_solver()

Build LL-problem solver with both.ll_optimizer module, which will optimize ll model for further using in ul optimization procedure. Setting the value of parameters according to the selected method.

Parameters:

  • lower_loop(int) - The total number of iterations for ll gradient descent optimization. It is also regarded as z_loop in BVFIM method.

  • ll_optimizer(Optimizer) - Optimizer of ll model, defined outside this module and will be used in LL optimization procedure.

  • truncate_iter(int, default=0) - Specific parameter for Truncated Reverse AD method, defining number of iterations to truncate in the back propagation process during ll optimizing.

  • acquire_max_loss_iter(bool, default=False) - Specific parameter for IAPTT-GM method,if set True then will use PTT method to truncate reverse trajectory during ll optimization.

  • alpha_init(float, default=0.0) - Specify parameter for BDA method. The aggregation parameter for Bi-level descent aggregation method, where alpha ∈ (0, 1) denotes the ratio of ll objective to ul objective during ll optimizing.

  • alpha_decay(float, default=0.0) - Specify parameter for BDA method. Weight decay coefficient of aggregation parameter alpha. The decay rate will accumulate with ll optimization procedure.

  • y_loop(int, default=5) - Specify parameter for BVFIM method. Num of steps to obtain a optimal ll variable under the ll problem value obtained after z_loop, i.e. optimize the updated ll variable with ul problem. Regarded as Regarded as T_y in the paper.

  • ll_l2_reg(float, default=0.1) - Specify parameter for BVFIM method. Weight of L2 regularization term in the value function of the regularizedLL problem. Referring to module both.ul_optimize.bvfim for more details.

  • ul_l2_reg(float, default=0.01) - Specify parameter for BVFIM method. Weight of L2 regularization term in the value function of the regularized UL problem. Referring to module both.ul_optimize.bvfim for more details.

  • ul_ln_reg(float, default=10.) - Specify parameter for BVFIM method. Weight of the log-barrier penalty term in the value function of the regularized UL problem. Referring to module both.ul_optimize.bvfim for more details.

  • reg_decay(bool, default=True) - Specify parameter for BVFIM method. Whether to use weight decay coefficient of L2 regularization term and log-barrier penalty term.

build_ul_problem_solver()

Setting up UL optimization module. Select desired method through given parameters and set related experiment parameters.

Details of parameter settings for each particular method are available in the specific method module of both.ul_optimize.

Parameters:

  • ul_optimizer(Optimizer) - Optimizer of ul model, defined outside this module and will be used in UL optimization procedure.

  • update_ll_model_init(bool, default=False) - Specific parameter for Dynamic method. If set True, the initial value of ll model will be updated after this iteration.

  • k(int, default=10) - Specific parameter for Implicit method. The maximum number of conjugate gradient iterations.

  • tolerance(float, default=1e-10) - Specific parameter for Implicit method. End the method earlier when the norm of the residual is less than tolerance.

  • r(float, default=1e-2) - Parameter for One-stage RAD method and used to adjust scalar epsilon. Value 0.01 of r is recommended for sufficiently accurate in the paper. Referring to module both.ul_optimize.onestage for more details.

  • ll_l2_reg(float, default=0.1) - Specify parameter for BVFIM method. Weight of L2 regularization term in the value function of the regularizedLL problem. Referring to module both.ul_optimize.bvfim for more details.

  • ul_l2_reg(float, default=0.01) - Specify parameter for BVFIM method. Weight of L2 regularization term in the value function of the regularized UL problem. Referring to module both.ul_optimize.bvfim for more details.

  • ul_ln_reg(float, default=10.) - Specify parameter for BVFIM method. Weight of the log-barrier penalty term in the value function of the regularized UL problem. Referring to module both.ul_optimize.bvfim for more details.

build_meta_problem_solver()

Setting up meta-learning optimization module. Select desired method through given parameters and set set related experiment parameters.

Note that among three methods MT-net, Warpgrad and L2F, only one can be used; while First-order and MSGD can be combined with others.

Parameters:

  • meta_optimizer(Optimizer) - The optimizer used to update initial values of meta model after an iteration.

  • inner_loop(int, default=5) - Num of inner optimization steps.

  • inner_learning_rate(float, default=0.01) - Step size for inner optimization.

  • use_second_order(bool, default=True) - Optional argument, whether to calculate precise second-order gradients during inner-loop.

  • learn_lr(bool, default=False) - Optional argument, whether to update inner learning rate during outer optimization, i.e. use MSGD method.

  • use_t(bool, default=False) - Optional argument, whether to using T-layers during optimization,i.e. use MT-net method.

  • use_warp(bool, default=False) - Optional argument, whether to using warp modules during optimization,i.e. use Warp-grad method.

  • use_forget(bool, default=False) - Optional argument, whether to add attenuation to each layers, i.e. use L2F method.

run_iter()

Run an iteration with a data batch and updates the parameters of ul model or meta-model.

Parameters:

  • train_data_batch(Tensor) - A batch of train data,which is used during ll optimizing.

  • train_target_batch(Tensor) - A batch of train target,which is used during ll optimizing.

  • validate_data_batch(Tensor) - A batch of test data,which is used during ul optimizing.

  • validate_target_batch(Tensor) - A batch of test target,which is used during ul optimizing.

  • batch_size(int) - The number of training samples in each batch.

  • current_iter(int) - The num of current iter.

  • forward_with_whole_batch(bool, default=True) - Whether to feed in the whole data batch when doing forward propagation. When setting to False, each single data in the batch will be fed into model during this iteration. This useful for some experiment having special setting, like few-shot learning.

Returns

        validation_loss(Tensor) - The value of validation loss value.