53 lines
2.0 KiB
Python
53 lines
2.0 KiB
Python
import time
|
|
from pathlib import Path
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.data import DataLoader
|
|
from unlearning.Strategy import Strategy
|
|
from architectures.Model import Model, Architecture
|
|
|
|
class Retrain(Strategy):
|
|
"""
|
|
Implements the Exact Unlearning Baseline by re-instantiating a fresh,
|
|
pre-trained instance of the specific architecture and training it from scratch
|
|
on the retain set using the Model's internal train function.
|
|
"""
|
|
def __init__(self, target_class_index: int, arch: Architecture, size: int,
|
|
lr: float = 0.001, epochs: int = 5):
|
|
super().__init__(target_class_index)
|
|
self.arch = arch
|
|
self.size = size
|
|
self.lr = lr
|
|
self.epochs = epochs
|
|
|
|
def _run(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader) -> nn.Module:
|
|
# 1. Determine the active execution device from the running sandbox
|
|
device = next(model.parameters()).device
|
|
|
|
print(f">> Triggering Exact Unlearning Baseline (Retraining {self.arch.name} from pristine state)...")
|
|
|
|
# a new model with default params is created
|
|
fresh_meat = Model.create(self.arch, device, self.size)
|
|
|
|
# we train it with retain set
|
|
fresh_meat.train(
|
|
epochs=self.epochs,
|
|
loader=retain_loader,
|
|
rate=self.lr,
|
|
mode="retrain"
|
|
)
|
|
|
|
# 4. Extract the trained nn.Module parameter state dict
|
|
# In-place copy onto the existing sandbox model structure to seamlessly retain downstream evaluations
|
|
model.load_state_dict(fresh_meat.model.state_dict())
|
|
|
|
print(">> Retraining pipeline finished. Pristine baseline weights successfully established.")
|
|
return model
|
|
|
|
def _split_data(self, dataset):
|
|
from sets.Data import get_unlearning_loaders
|
|
return get_unlearning_loaders(
|
|
dataset=dataset,
|
|
forget_class_idx=self.target_class_index,
|
|
batch_size=32
|
|
) |