added functionality to specify seed in randomisation tests

quantus/helpers/pytorch_model.py

@@ -44,7 +44,7 @@ def state_dict(self):
         """Get a dictionary of the model's learnable parameters."""
         return self.model.state_dict()
 
-    def get_random_layer_generator(self, order: str = "top_down"):
+    def get_random_layer_generator(self, order: str = "top_down", seed: int = 42):
         """
         In every iteration yields a copy of the model with one additional layer's parameters randomized.
         Set order to top_down for cascading randomization.
@@ -65,5 +65,6 @@ def get_random_layer_generator(self, order: str = "top_down"):
         for module in modules:
             if order == "independent":
                 random_layer_model.load_state_dict(original_parameters)
+            torch.manual_seed(seed=seed+1)
             module[1].reset_parameters()
             yield module[0], random_layer_model

quantus/helpers/tf_model.py

@@ -52,7 +52,7 @@ def load_state_dict(self, original_parameters):
         """Set model's learnable parameters."""
         self.model.set_weights(original_parameters)
 
-    def get_random_layer_generator(self, order: str = "top_down"):
+    def get_random_layer_generator(self, order: str = "top_down", seed: int = 42):
         """
         In every iteration yields a copy of the model with one additional layer's parameters randomized.
         Set order to top_down for cascading randomization.
@@ -70,5 +70,6 @@ def get_random_layer_generator(self, order: str = "top_down"):
             if order == "independent":
                 random_layer_model.set_weights(original_parameters)
             weights = layer.get_weights()
+            np.random.seed(seed=seed+1)
             layer.set_weights([np.random.permutation(w) for w in weights])
             yield layer.name, random_layer_model

quantus/metrics/randomisation_metrics.py

@@ -66,6 +66,7 @@ def __init__(self, *args, **kwargs):
         self.display_progressbar = self.kwargs.get("display_progressbar", False)
         self.similarity_func = self.kwargs.get("similarity_func", correlation_spearman)
         self.layer_order = kwargs.get("layer_order", "independent")
+        self.seed = self.kwargs.get("seed", 42)
         self.last_results = {}
         self.all_results = []
 
@@ -183,7 +184,7 @@ def __call__(
             pbar = tqdm(total=n_iterations)
 
         for layer_name, random_layer_model in model.get_random_layer_generator(
-            order=self.layer_order
+            order=self.layer_order, seed=self.seed
         ):
 
             similarity_scores = []
@@ -266,6 +267,7 @@ def __init__(self, *args, **kwargs):
         self.normalise_func = self.kwargs.get("normalise_func", normalise_by_negative)
         self.similarity_func = self.kwargs.get("similarity_func", ssim)
         self.num_classes = self.kwargs.get("num_classes", 1000)
+        self.seed = self.kwargs.get("seed", 42)
         self.last_results = []
         self.all_results = []
 
@@ -384,6 +386,7 @@ def __call__(
                 a = self.normalise_func(a)
 
             # Randomly select off-class labels.
+            random.seed(a=self.seed)
             y_off = np.array(
                 [
                     random.choice(