import mindspore as ms
from mindspore import nn, ops, numpy as mnp
from typing import Dict, Any, Callable, List
import copy
from boat_ms.operation_registry import register_class
from boat_ms.gm_ol.dynamical_system import DynamicalSystem
[docs]
@register_class
class VFO(DynamicalSystem):
"""
Implements the optimization procedure of Value-function based First-Order Method (VFO) [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 : mindspore.nn.Cell
The lower-level model of the BLO problem.
ul_model : mindspore.nn.Cell
The upper-level model of the BLO problem.
ll_var : List[mindspore.Tensor]
A list of lower-level variables of the BLO problem.
ul_var : List[mindspore.Tensor]
A list of upper-level variables of the BLO problem.
lower_loop : int
The number of iterations for lower-level optimization.
solver_config : Dict[str, Any]
A dictionary containing configurations for the solver. Expected keys include:
- "lower_level_opt" (mindspore.nn.optim.Optimizer): Optimizer for the lower-level model.
- "VFO" (Dict): Configuration for the VFO algorithm:
- "y_hat_lr" (float): Learning rate for optimizing the surrogate variable `y_hat`.
- "eta" (float): Step size for value-function updates.
- "u1" (float): Hyperparameter controlling the penalty in the value function.
- "device" (str): Device on which computations are performed, e.g., "cpu" or "cuda".
References
----------
[1] R. Liu, X. Liu, X. Yuan, S. Zeng and J. Zhang, "A Value-Function-based Interior-point Method for Non-convex Bi-level Optimization," in ICML, 2021.
"""
def __init__(
self,
ll_objective: Callable,
lower_loop: int,
ul_model: nn.Cell,
ul_objective: Callable,
ll_model: nn.Cell,
ll_var: List,
ul_var: List,
solver_config: Dict[str, Any],
):
super(VFO, self).__init__(
ll_objective, ul_objective, lower_loop, ul_model, ll_model, solver_config
)
self.ll_opt = solver_config["lower_level_opt"]
self.ul_opt = solver_config["upper_level_opt"]
self.ll_var = ll_var
self.ul_var = ul_var
self.y_hat_lr = float(solver_config["VFO"]["y_hat_lr"])
self.eta = solver_config["VFO"]["eta"]
self.u1 = solver_config["VFO"]["u1"]
self.device = solver_config["device"]
[docs]
def optimize(self, ll_feed_dict: Dict, ul_feed_dict: Dict, current_iter: int):
"""
Execute the optimization procedure with the data from feed_dict.
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.
current_iter : int
The current iteration number of the optimization process.
Returns
-------
The upper-level loss.
"""
y_hat = copy.deepcopy(self.ll_model)
y_hat_opt = nn.SGD(
y_hat.trainable_params(), learning_rate=self.y_hat_lr, momentum=0.9
)
n_params_y = sum([p.size for p in self.ll_model.trainable_params()])
n_params_x = sum([p.size for p in self.ul_model.trainable_params()])
delta_f = mnp.zeros((n_params_x + n_params_y), dtype=ms.float32)
delta_F = mnp.zeros((n_params_x + n_params_y), dtype=ms.float32)
def g_x_xhat_w(y, y_hat, x):
loss = self.ll_objective(ll_feed_dict, x, y) - self.ll_objective(
ll_feed_dict, x, y_hat
)
grad_y = ops.GradOperation(get_by_list=True)(
self.ll_objective, y.trainable_params()
)(ll_feed_dict, x, y)
grad_x = ops.GradOperation(get_by_list=True)(
self.ll_objective, x.trainable_params()
)(ll_feed_dict, x, y)
return loss, grad_y, grad_x
for y_itr in range(self.lower_loop):
for param in y_hat.trainable_params():
param.set_data(mnp.zeros_like(param.data))
grad_fn = ops.GradOperation(get_by_list=True)(
self.ll_objective, y_hat.trainable_params()
)
grads_hat = grad_fn(ll_feed_dict, self.ul_model, y_hat)
y_hat_opt(grads_hat)
F_y = self.ul_objective(ul_feed_dict, self.ul_model, self.ll_model)
grad_F_y = ops.GradOperation(get_by_list=True)(
self.ul_objective, self.ll_model.trainable_params()
)(ul_feed_dict, self.ul_model, self.ll_model)
grad_F_x = ops.GradOperation(get_by_list=True)(
self.ul_objective, self.ul_model.trainable_params()
)(ul_feed_dict, self.ul_model, self.ll_model)
for param in y_hat.trainable_params():
param.requires_grad = False
loss, gy, gx_minus_gx_k = g_x_xhat_w(self.ll_model, y_hat, self.ul_model)
delta_F[:n_params_y] = mnp.concatenate([p.view(-1) for p in grad_F_y]).astype(
ms.float32
)
delta_f[:n_params_y] = mnp.concatenate([p.view(-1) for p in gy]).astype(
ms.float32
)
delta_F[n_params_y:] = mnp.concatenate([p.view(-1) for p in grad_F_x]).astype(
ms.float32
)
delta_f[n_params_y:] = mnp.concatenate(
[p.view(-1) for p in gx_minus_gx_k]
).astype(ms.float32)
norm_dq = (ops.norm(delta_f) ** 2).astype(ms.float32)
dot = ops.ReduceSum()(delta_F * delta_f)
correction = ops.ReLU()((self.u1 * loss - dot) / (norm_dq + 1e-8))
d = delta_F + correction * delta_f
y_grad = []
x_grad = []
all_numel = 0
for param in self.ll_model.trainable_params():
grad_slice = d[all_numel : all_numel + param.size].reshape(param.shape)
y_grad.append(ms.Tensor(grad_slice, dtype=ms.float32))
all_numel += param.size
for param in self.ul_model.trainable_params():
grad_slice = d[all_numel : all_numel + param.size].reshape(param.shape)
x_grad.append(ms.Tensor(grad_slice, dtype=ms.float32))
all_numel += param.size
for param, grad in zip(self.ll_model.trainable_params(), y_grad):
new_param = param - self.y_hat_lr * grad
param.set_data(new_param)
for param, grad in zip(self.ul_model.trainable_params(), x_grad):
new_param = param - self.y_hat_lr * grad
param.set_data(new_param)
return F_y.item()