import jittor as jit
from jittor import Module
from typing import List, Callable, Dict
from ..higher_jit.patch import _MonkeyPatchBase
from jboat.utils.op_utils import update_tensor_grads, neumann
from jboat.operation_registry import register_class
from jboat.na_ol.hyper_gradient import HyperGradient
[docs]
@register_class
class NS(HyperGradient):
"""
Calculation of the hyper gradient of the upper-level variables with Neumann Series (NS) [1].
Parameters
----------
ll_objective : Callable
The lower-level objective function of the BLO problem.
ul_objective : Callable
The upper-level objective function of the BLO problem.
ll_model : jittor.Module
The lower-level model of the BLO problem.
ul_model : jittor.Module
The upper-level model of the BLO problem.
ll_var : List[jittor.Var]
List of variables optimized with the lower-level objective.
ul_var : List[jittor.Var]
List of variables optimized with the upper-level objective.
solver_config : Dict[str, Any]
Dictionary containing solver configurations, including:
- `gm_op` (str): Indicates dynamic initialization type (e.g., "DI").
- `lower_level_opt` (Optimizer): Lower-level optimizer configuration.
- `CG` (Dict): Conjugate Gradient-specific parameters:
- `tolerance` (float): Tolerance for convergence.
- `k` (int): Number of iterations for Neumann approximation.
- GDA-specific parameters, such as `alpha_init` and `alpha_decay`.
- `gda_loss` (Callable, optional): Custom loss function for GDA.
References
----------
[1] J. Lorraine, P. Vicol, and D. Duvenaud, "Optimizing millions of hyperparameters
by implicit differentiation," in AISTATS, 2020.
"""
def __init__(
self,
ll_objective: Callable,
ul_objective: Callable,
ll_model: Module,
ul_model: Module,
ll_var: List,
ul_var: List,
solver_config: Dict,
):
super(NS, self).__init__(
ll_objective,
ul_objective,
ul_model,
ll_model,
ll_var,
ul_var,
solver_config,
)
self.dynamic_initialization = "DI" in solver_config["gm_op"]
self.ll_lr = solver_config["lower_level_opt"].defaults["lr"]
self.tolerance = solver_config["CG"]["tolerance"]
self.K = solver_config["CG"]["k"]
self.alpha = solver_config["GDA"]["alpha_init"]
self.alpha_decay = solver_config["GDA"]["alpha_decay"]
self.alpha = solver_config["GDA"]["alpha_init"]
self.alpha_decay = solver_config["GDA"]["alpha_decay"]
self.gda_loss = solver_config.get("gda_loss", None)
[docs]
def compute_gradients(
self,
ll_feed_dict: Dict,
ul_feed_dict: Dict,
auxiliary_model: _MonkeyPatchBase,
max_loss_iter: int = 0,
hyper_gradient_finished: bool = False,
next_operation: str = None,
**kwargs
):
"""
Compute the hyper-gradients of the upper-level variables with the data from feed_dict and patched models.
Parameters
----------
ll_feed_dict : Dict
Dictionary containing the lower-level data used for optimization.
It typically includes training data, targets, and other information required to compute the LL objective.
ul_feed_dict : Dict
Dictionary containing the upper-level data used for optimization.
It typically includes validation data, targets, and other information required to compute the UL objective.
auxiliary_model : _MonkeyPatchBase
A patched lower model wrapped by the `higher` library.
It serves as the lower-level model for optimization.
max_loss_iter : int, optional
The number of iterations used for backpropagation. Defaults to 0.
next_operation : str, optional
The next operator for the calculation of the hypergradient. Defaults to None.
hyper_gradient_finished : bool, optional
A boolean flag indicating whether the hypergradient computation is finished. Defaults to False.
Returns
-------
Dict
A dictionary containing the upper-level objective and the status of hypergradient computation.
"""
assert (
not hyper_gradient_finished
), "CG does not support multiple hypergradient computation"
lower_model_params = kwargs.get(
"lower_model_params", list(auxiliary_model.parameters())
)
hparams = kwargs.get("hparams", list(self.ul_var))
def fp_map(params, loss_f):
lower_grads = list(jit.grad(loss_f, params))
updated_params = []
for i in range(len(params)):
updated_params.append(params[i] - self.ll_lr * lower_grads[i])
return updated_params
if self.gda_loss is not None:
ll_feed_dict["alpha"] = self.alpha * self.alpha_decay**max_loss_iter
lower_loss = self.gda_loss(
ll_feed_dict,
ul_feed_dict,
self.ul_model,
auxiliary_model,
params=lower_model_params,
)
else:
lower_loss = self.ll_objective(
ll_feed_dict, self.ul_model, auxiliary_model, params=lower_model_params
)
upper_loss = self.ul_objective(
ul_feed_dict, self.ul_model, auxiliary_model, params=lower_model_params
)
if self.dynamic_initialization:
grads_lower = jit.grad(
upper_loss, list(auxiliary_model.parameters(time=0)), retain_graph=True
)
update_tensor_grads(self.ll_var, grads_lower)
grads_upper = neumann(
lower_model_params,
hparams,
upper_loss,
lower_loss,
self.K,
fp_map,
self.tolerance,
)
update_tensor_grads(self.ul_var, grads_upper)
return {"upper_loss": upper_loss.item(), "hyper_gradient_finished": True}