change default subsampled_target_frac

afsl/acquisition_functions/__init__.py

@@ -346,12 +346,12 @@ class Targeted(ABC):
     def __init__(
         self,
         target: torch.Tensor,
-        subsampled_target_frac: float = 0.5,
+        subsampled_target_frac: float = 1,
         max_target_size: int | None = None,
     ):
         r"""
         :param target: Tensor of prediction targets (shape $m \times d$).
-        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $0.5$.
+        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $1$.
         :param max_target_size: Maximum size of the target to be subsampled in each iteration. Default is `None` in which case the target may be arbitrarily large.
         """
 

afsl/acquisition_functions/bace.py

@@ -118,7 +118,7 @@ def __init__(
         self,
         target: torch.Tensor,
         noise_std=1.0,
-        subsampled_target_frac: float = 0.5,
+        subsampled_target_frac: float = 1,
         max_target_size: int | None = None,
         mini_batch_size=DEFAULT_MINI_BATCH_SIZE,
         embedding_batch_size=DEFAULT_EMBEDDING_BATCH_SIZE,
@@ -129,7 +129,7 @@ def __init__(
         r"""
         :param target: Tensor of prediction targets (shape $m \times d$).
         :param noise_std: Standard deviation of the noise.
-        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $0.5$. Ignored if `target` is `None`.
+        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $1$. Ignored if `target` is `None`.
         :param max_target_size: Maximum size of the target to be subsampled in each iteration. Default is `None` in which case the target may be arbitrarily large. Ignored if `target` is `None`.
         :param mini_batch_size: Size of mini-batch used for computing the acquisition function.
         :param embedding_batch_size: Batch size used for computing the embeddings.

afsl/acquisition_functions/cosine_similarity.py

@@ -38,7 +38,7 @@ class CosineSimilarity(Targeted, BatchAcquisitionFunction):
     def __init__(
         self,
         target: torch.Tensor,
-        subsampled_target_frac: float = 0.5,
+        subsampled_target_frac: float = 1,
         max_target_size: int | None = None,
         mini_batch_size=DEFAULT_MINI_BATCH_SIZE,
         embedding_batch_size=DEFAULT_EMBEDDING_BATCH_SIZE,
@@ -47,7 +47,7 @@ def __init__(
     ):
         r"""
         :param target: Tensor of prediction targets (shape $m \times d$).
-        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $0.5$. Ignored if `target` is `None`.
+        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $1$. Ignored if `target` is `None`.
         :param max_target_size: Maximum size of the target to be subsampled in each iteration. Default is `None` in which case the target may be arbitrarily large. Ignored if `target` is `None`.
         :param mini_batch_size: Size of mini-batch used for computing the acquisition function.
         :param embedding_batch_size: Batch size used for computing the embeddings.

afsl/acquisition_functions/information_density.py

@@ -32,7 +32,7 @@ def __init__(
         self,
         target: torch.Tensor,
         beta=1.0,
-        subsampled_target_frac: float = 0.5,
+        subsampled_target_frac: float = 1,
         max_target_size: int | None = None,
         mini_batch_size=DEFAULT_MINI_BATCH_SIZE,
         embedding_batch_size=DEFAULT_EMBEDDING_BATCH_SIZE,

afsl/acquisition_functions/max_dist.py

@@ -45,7 +45,7 @@ class MaxDist(SequentialAcquisitionFunction[ModelWithEmbeddingOrKernel, Distance
 
     This acquisition function rests on the assumption that the model induces a distance $d(\vx,\vxp)$ between points $\vx$ and $\vxp$, either via an embedding or a kernel.
 
-    - **Embeddings** $\vphi(\cdot)$ induce the (euclidean) *embedding distance* \\[ d_\vphi(\vx,\vxp) \defeq \norm{\vphi(\vx) - \vphi(\vxp)}_2. \\]
+    - **Embeddings** $\vphi(\cdot)$ induce the (Euclidean) *embedding distance* \\[ d_\vphi(\vx,\vxp) \defeq \norm{\vphi(\vx) - \vphi(\vxp)}_2. \\]
     - A **kernel** $k$ induces the *kernel distance* \\[ d_k(\vx,\vxp) \defeq = \sqrt{k(\vx,\vx) + k(\vxp,\vxp) - 2 k(\vx,\vxp)}. \\]
 
     It is straightforward to see that if $k(\vx,\vxp) = \vphi(\vx)^\top \vphi(\vxp)$ then embedding and kernel distances coincide, i.e., $d_\vphi(\vx,\vxp) = d_k(\vx,\vxp)$.

afsl/active_data_loader.py

@@ -36,7 +36,7 @@ class ActiveDataLoader(Generic[M]):
     r"""Acquisition function to be used for data selection."""
 
     subsampled_target_frac: float
-    r"""Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $0.5$."""
+    r"""Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $1$."""
 
     max_target_size: int | None
     r"""
@@ -71,7 +71,7 @@ def initialize(
         dataset: Dataset,
         target: torch.Tensor | None,
         batch_size: int,
-        subsampled_target_frac: float = 0.5,
+        subsampled_target_frac: float = 1,
         max_target_size: int | None = None,
         mini_batch_size: int = DEFAULT_MINI_BATCH_SIZE,
         embedding_batch_size: int = DEFAULT_EMBEDDING_BATCH_SIZE,
@@ -84,7 +84,7 @@ def initialize(
         :param dataset: Inputs (shape $n \times d$) to be selected from.
         :param target: Tensor of prediction targets (shape $m \times d$) or `None`.
         :param batch_size: Size of the batch to be selected.
-        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $0.5$. Ignored if `target` is `None`.
+        :param subsampled_target_frac: Fraction of the target to be subsampled in each iteration. Must be in $(0,1]$. Default is $1$. Ignored if `target` is `None`.
         :param max_target_size: Maximum size of the target to be subsampled in each iteration. Default is `None` in which case the target may be arbitrarily large. Ignored if `target` is `None`.
         :param mini_batch_size: Size of mini batches used for computing the acquisition function.
         :param embedding_batch_size: Batch size used for computing the embeddings.

examples/fine_tuning/cifar_100/experiment.py

@@ -29,7 +29,7 @@
 
 MINI_BATCH_SIZE = 1_000
 NUM_WORKERS = 4
-NUM_ROUNDS = 100
+NUM_ROUNDS = 101
 
 DEFAULT_NOISE_STD = 1.0
 DEFAULT_QUERY_BATCH_SIZE = 10

examples/fine_tuning/mnist/experiment.py

@@ -32,7 +32,7 @@
 
 MINI_BATCH_SIZE = 1_000
 NUM_WORKERS = 4
-NUM_ROUNDS = 100
+NUM_ROUNDS = 101
 
 DEFAULT_NOISE_STD = 0.01
 DEFAULT_QUERY_BATCH_SIZE = 1