Tinsae/Finetuning
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

reports from optimised linear filtration

Browse Source
This commit is contained in:
Tinsae
2026-07-04 10:34:51 +02:00
parent 61da187012
commit 7f848b0485
16 changed files with 3014 additions and 1227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
Tune_new.py
View File

@@ -133,7 +133,7 @@ def run_finetuning_or_baseline_eval(env_dict, run_training=False, lr_rate=0.0001
# Evaluate original base checkpoint performance
current_mode = "Finetuned"
# Check if weights exist or model was trained before evaluating
# evaluate finetuned model
try:
accuracy, report_dict = model.evaluate(loader=test_loader, mode=current_mode)
Util._log_to_csv(
@@ -178,6 +178,8 @@ def run_unlearning_and_strategy_eval(env_dict, forget_class_idx, strategy, evalu
print("fine tunned model loaded into evaluation sandbox")
# Execute strategic parameter unlearning step
# we are using only training data to unlearn.
# Test data is never touched here.
unlearned = strategy.apply(reloaded.model, train_data)
strategy_in_use = strategy.__class__.__name__
@@ -200,7 +202,7 @@ def run_unlearning_and_strategy_eval(env_dict, forget_class_idx, strategy, evalu
print(domain["label"])
accuracy, report_dict = reloaded.evaluate(loader=domain["loader"], mode=domain["mode"])
Util._log_to_csv(
arch=ARCH.name,#reloaded.__class__.__name__,
arch=ARCH.name,
mode=domain["mode"],
accuracy=accuracy,
report_dict=report_dict,
@@ -211,7 +213,7 @@ def run_unlearning_and_strategy_eval(env_dict, forget_class_idx, strategy, evalu
# entry
if __name__ == "__main__":
outer_loop = 10
outer_loop = 11
inner_loop = CLASS_SIZE
for k in range(outer_loop):
@@ -221,15 +223,17 @@ if __name__ == "__main__":
runtime_environment = prepare_data_and_model_environment()
# Baseline Evaluation
# switch finetuning for tests on strategies only,
# to avoid finetunning every time we test a strategy
finetuning = True
# switch finetuning for tests on strategies only
run_finetuning_or_baseline_eval(runtime_environment, run_training = finetuning)
# scale 16400.0 for ResNet
scale = 20200
scale = 20300
# batch 8 for resNet,
unlearning_batches = 16
# regularis
# strategies
# implementation of Certified Removal for DNNs
certified_unlearning = CertifiedUnlearning(
target_class_index=0, #arch ResNet18 GoogLeNet Inception
l2_reg=0.000002 , # 0.000002 0.00001 0.0
@@ -241,12 +245,13 @@ if __name__ == "__main__":
unlearn_bs=unlearning_batches # 8 32 8
)
# works perfectly
# Normalisation Filtration
linear_filtration = LinearFiltration(
target_class_index=0
target_class_index=0,
num_classes=CLASS_SIZE
)
# WF-Net
weight_filtration = WeightFiltration(
target_class_index=0, #arch ResNet18 GoogLeNet/Inception
epochs=6, #