pdf generator is fixed

examples/mia/celebA_HQ/.gitignore

@@ -1,5 +1,5 @@
 data/*
 !data/.gitkeep
 
-leakprp_output/*
+leakpro_output/*
 target/*

examples/mia/celebA_HQ/audit.yaml

@@ -11,34 +11,6 @@ audit:  # Configurations for auditing
       gamma: 2.0
       offline_a: 0.33 # parameter from which we compute p(x) from p_OUT(x) such that p_IN(x) = a p_OUT(x) + b.
       offline_b: 0.66
-    # qmia:
-    #   training_data_fraction: 1.0  # Fraction of the auxilary dataset (data without train and test indices) to use for training the quantile regressor
-    #   epochs: 5  # Number of training epochs for quantile regression
-    # population:
-    #   attack_data_fraction: 1.0  # Fraction of the auxilary dataset to use for this attack
-    # lira:
-    #   training_data_fraction: 0.3  # Fraction of the auxilary dataset to use for this attack (in each shadow model training)
-    #   num_shadow_models: 2 # Number of shadow models to train
-    #   online: False # perform online or offline attack
-    #   fixed_variance: True # Use a fixed variance for the whole audit
-    #   boosting: True
-    # loss_traj:
-    #   training_distill_data_fraction : 0.7 # Fraction of the auxilary dataset to use for training the distillation models D_s = (1-D_KD)/2
-    #   number_of_traj: 10 # Number of epochs (number of points in the loss trajectory)
-    #   label_only: False # True or False
-    #   mia_classifier_epochs: 100
-    # HSJ:
-    #   attack_data_fraction: 0.01  # Fraction of the auxilary dataset to use for this attack
-    #   target_metadata_path: "./target/model_metadata.pkl"
-    #   num_iterations: 2 # Number of iterations for the optimization
-    #   initial_num_evals: 100 # Number of evaluations for number of random vecotr to estimate the gradient
-    #   max_num_evals: 10000 # Maximum number of evaluations
-    #   stepsize_search: "geometric_progression" # Step size search method
-    #   gamma: 1.0 # Gamma for the optimization
-    #   constraint: 2 
-    #   batch_size: 50 
-    #   verbose: True
-    #   epsilon_threshold: 1e-6
 
   output_dir: "./leakpro_output"
   attack_type: "mia" #mia, gia

examples/mia/celebA_HQ/celebA_HQ_handler.py

@@ -19,9 +19,10 @@ def get_criterion(self) -> torch.nn.Module:
 
     def get_optimizer(self, model: torch.nn.Module) -> optim.Optimizer:
         """Set the optimizer for the model."""
-        learning_rate =0.1  # Default learning rate
+        learning_rate =0.01  # Default learning rate
         momentum =  0.9  # Default momentum
-        return optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum)
+        weight_decay = 0.0001  # Default weight decay
+        return optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum, weight_decay=weight_decay)
 
     def train(
         self,

examples/mia/celebA_HQ/train_config.yaml

@@ -8,7 +8,7 @@ train: # Configuration for training
   optimizer: SGD # String which indicates the optimizer. We support Adam and SGD. For speedyresnet, it uses its own optimizer.
   learning_rate: 0.001 # Float number for indicating learning rate for training the target model. For speedyresnet, it uses its own learning_rate.
   momentum: 0.9
-  weight_decay: 0.1 # Float number for indicating weight decay for training the target model. For speedyresnet, it uses its own weight_decay.
+  weight_decay: 0.001 # Float number for indicating weight decay for training the target model. For speedyresnet, it uses its own weight_decay.
 
 data: # Configuration for data
   dataset: celebA_HQ # String indicates the name of the dataset