UTILS

Classes

class models.tak_utils.utils.FisherLoader(fisher_cache, dataset_name, device, fp_precision='fp32')

Bases: object

load_diff_ekfac(task_id, only_counts=False)

Return type:

tuple[dict[str, Tensor], dict[str, Tensor], dict[str, Tensor], int, int]

load_ekfac(task_id, only_counts=False)

Return type:

tuple[dict[str, Tensor], dict[str, Tensor], dict[str, Tensor], int, int]

load_kfac(task_id: int, only_counts: Literal[True]) → tuple[int, int]

load_kfac(task_id: int, only_counts: Literal[False] = False) → tuple[dict[str, Tensor], dict[str, Tensor], dict[str, Tensor], int, int]

store_kfac(task_id, ggT, aaT, ffT, num_ggT, num_aaT)

class models.tak_utils.utils.OptimizerBuilder(cmd_args)

Bases: object

build_opt_and_sched(all_params, num_batches)

build_opt_and_sched_multiple_lr(params_group_1, params_group_2, num_batches)

Functions

models.tak_utils.utils.add_clip_args(parser)

models.tak_utils.utils.assign_learning_rate(param_group, new_lr)

models.tak_utils.utils.compute_acc_on_last_task(model, dataset)

models.tak_utils.utils.cosine_lr(optimizer, base_lrs, warmup_length, steps, min_lr)

models.tak_utils.utils.get_delta_w_backbone(named_params, delta_w, delta_w_names, training_type, device)

models.tak_utils.utils.get_delta_w_parameterlist(named_params, delta_w, delta_w_names, peft_type, device)

models.tak_utils.utils.get_parameter(shape, device, type_init='orto', transpose=False, requires_grad=True)

models.tak_utils.utils.get_params(net, features=True, classifier=False, offset_1=-1, offset_2=-1)

Return type:

Tensor

models.tak_utils.utils.make_psd(x, to64=False)

models.tak_utils.utils.replace_non_dynamically_quantizable_linear(module)

Recursively replace all NonDynamicallyQuantizableLinear layers with Linear layers in a model.

models.tak_utils.utils.set_params(net, new_params, features=True, classifier=False, offset_1=-1, offset_2=-1)

models.tak_utils.utils.set_requires_grad_to(model, namevars, mode)

models.tak_utils.utils.step_lr_decay(optimizer, base_lrs, warmup_length, steps)

models.tak_utils.utils.add_clip_args(parser)

models.tak_utils.utils.assign_learning_rate(param_group, new_lr)

models.tak_utils.utils.compute_acc_on_last_task(model, dataset)

models.tak_utils.utils.cosine_lr(optimizer, base_lrs, warmup_length, steps, min_lr)

models.tak_utils.utils.get_delta_w_backbone(named_params, delta_w, delta_w_names, training_type, device)

models.tak_utils.utils.get_delta_w_parameterlist(named_params, delta_w, delta_w_names, peft_type, device)

models.tak_utils.utils.get_parameter(shape, device, type_init='orto', transpose=False, requires_grad=True)

models.tak_utils.utils.get_params(net, features=True, classifier=False, offset_1=-1, offset_2=-1)

Return type:

Tensor

models.tak_utils.utils.make_psd(x, to64=False)

models.tak_utils.utils.replace_non_dynamically_quantizable_linear(module)

Recursively replace all NonDynamicallyQuantizableLinear layers with Linear layers in a model.

models.tak_utils.utils.set_params(net, new_params, features=True, classifier=False, offset_1=-1, offset_2=-1)

models.tak_utils.utils.set_requires_grad_to(model, namevars, mode)

models.tak_utils.utils.step_lr_decay(optimizer, base_lrs, warmup_length, steps)