chore: sort evaluation results (#45)

Now we can get a consistent natural ordering of the output when
performing evaluation. Previously we converted sets directly into lists
which caused a different ordering on each run.

clip_eval/evaluation/evaluator.py

@@ -3,6 +3,7 @@
 from typing import Literal
 
 import numpy as np
+from natsort import natsorted
 from tabulate import tabulate
 
 from clip_eval.common.data_models import EmbeddingDefinition, Embeddings
@@ -21,8 +22,8 @@ def print_evaluation_results(
     classifier_column: (Literal["linear_probe"] | Literal["zero_shot"] | Literal["wKNN"]) = "linear_probe",
 ):
     defs = list(results.keys())
-    model_names = list(set(map(lambda d: d.model, defs)))
-    dataset_names = list(set(map(lambda d: d.dataset, defs)))
+    model_names = natsorted(set(map(lambda d: d.model, defs)))
+    dataset_names = natsorted(set(map(lambda d: d.dataset, defs)))
 
     table: list[list[float | str]] = [
         ["Model/Dataset"] + dataset_names,

pyproject.toml

@@ -31,6 +31,7 @@ tabulate = "^0.9.0"
 sentencepiece = "^0.2.0"
 protobuf = "^4.25.3"
 open-clip-torch = "^2.24.0"
+natsort = "^8.4.0"
 
 [tool.poetry.group.dev.dependencies]
 mypy = "^1.8.0"