V2 features signature

hobj/benchmarks/make_model.py

@@ -0,0 +1,58 @@
+"""
+This module provides an alternative interface for instantiating a linear learning model.
+"""
+from hobj.learning_models.linear import LinearLearner, RepresentationalModel
+import hobj.learning_models.linear.update_rules as update_rules
+from typing import Literal, Dict
+import mref
+import numpy as np
+from typing import List
+
+
+# %%
+def make_linear_learner_from_features(
+        ref_to_features: Dict[mref.ImageRef, np.ndarray],
+        calibration_images: List[mref.ImageRef],
+        update_rule_name: Literal[
+            'Prototype',
+            'Square',
+            'Perceptron',
+            'Hinge',
+            'MAE',
+            'Exponential',
+            'CE',
+            'REINFORCE'
+        ] = 'Square',
+        alpha: float = 1,
+) -> LinearLearner:
+    """
+    Instantiates a linear learning model from precomputed features.
+    :param ref_to_features: Dict[mref.ImageRef, np.ndarray], the features to use.
+    :param calibration_images: List[mref.ImageRef], the images that will be used to calibrate the features (i.e. for mean centering and ensuring they fit within a unit ball).
+    :param update_rule_name: str, the name of the update rule to use.
+    :param alpha: float, the learning rate.
+    :return: LinearLearner
+    """
+
+    f_calibration = np.array([ref_to_features[ref] for ref in calibration_images])
+    mu_calibration = np.mean(f_calibration, axis=0)
+    norms_calibration = np.linalg.norm(f_calibration - mu_calibration, axis=1)
+    norm_cutoff = np.quantile(norms_calibration, 0.999)  # Will clip the rest
+
+    ref_to_calibrated_features = {}
+    for ref in ref_to_features:
+        f = ref_to_features[ref]
+        fc = f - mu_calibration
+        fcn = fc / norm_cutoff
+        norm_cur = np.linalg.norm(fcn)
+        if norm_cur > 1:
+            fcn = fcn / norm_cur
+        ref_to_calibrated_features[ref] = np.array(fcn)
+
+    update_rule_name = getattr(update_rules, update_rule_name)
+    return LinearLearner(
+        representational_model=RepresentationalModel.from_precomputed_features(
+            image_ref_to_features=ref_to_calibrated_features
+        ),
+        update_rule=update_rule_name(alpha=alpha)
+    )

hobj/benchmarks/mut_highvar_benchmark.py

@@ -84,3 +84,7 @@ def __init__(self):
         super().__init__(
             config=config
         )
+
+if __name__ == '__main__':
+    experiment = MutatorHighVarBenchmark()
+    print(sorted(experiment.config.subtask_name_to_data.keys()))

hobj/config.py

@@ -1,3 +1,3 @@
 from pathlib import Path
 
-cachedir: Path = Path.home() / 'hobj_cache2'
+cachedir: Path = Path.home() / 'hobj_cache'

pyproject.toml

@@ -23,7 +23,7 @@ dependencies = [
     "pydantic>=2.10",
     "xarray>=2025.1",
     "pytest>=8.3",
-    #"mref @ git+https://github.com/himjl/mref.git",
+    "mref @ git+https://github.com/himjl/mref.git",
 ]
 
 [project.urls]