Fix replacement option in loss and gradnorm downsampling (#529)

There was an unused replacement option which defaulted to True, which
caused multiple keys to be chosen, which caused the storage to crash
since it cannot deal with multiple keys in its current implementation.
However, it should not have happened in the first place.

modyn/trainer_server/internal/trainer/remote_downsamplers/abstract_remote_downsampling_strategy.py

@@ -50,8 +50,6 @@ def __init__(
         assert "downsampling_ratio" in params_from_selector
         self.downsampling_ratio = params_from_selector["downsampling_ratio"]
 
-        self.replacement = params_from_selector.get("replacement", True)
-
         # The next variable is used to keep a mapping index <-> sample_id
         # This is needed since the data selection policy works on indexes (the policy does not care what the sample_id
         # is, it simply stores its score in a vector/matrix) but for retrieving again the data we need somehow to

modyn/trainer_server/internal/trainer/remote_downsamplers/remote_gradnorm_downsampling.py

@@ -84,7 +84,7 @@ def select_points(self) -> tuple[list[int], torch.Tensor]:
         probabilities = torch.cat(self.probabilities, dim=0)
         probabilities = probabilities / probabilities.sum()
 
-        downsampled_idxs = torch.multinomial(probabilities, target_size, replacement=self.replacement)
+        downsampled_idxs = torch.multinomial(probabilities, target_size, replacement=False)
 
         # lower probability, higher weight to reduce the variance
         weights = 1.0 / (self.number_of_points_seen * probabilities[downsampled_idxs])

modyn/trainer_server/internal/trainer/remote_downsamplers/remote_loss_downsampling.py

@@ -66,7 +66,7 @@ def select_points(self) -> tuple[list[int], torch.Tensor]:
         probabilities = torch.cat(self.probabilities, dim=0)
         probabilities = probabilities / probabilities.sum()
 
-        downsampled_idxs = torch.multinomial(probabilities, target_size, replacement=self.replacement)
+        downsampled_idxs = torch.multinomial(probabilities, target_size, replacement=False)
 
         # lower probability, higher weight to reduce the variance
         weights = 1.0 / (self.number_of_points_seen * probabilities[downsampled_idxs])