refactor: limit usage of `scipy` and `skilearn` dependencies (#2097)

* refactor: move scipy and sklearn module imports to func imports
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>

* refactor: consolidate weighted_f1_score func into lm_eval utils
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>

* lint: allow for utils file to have unused imports

this allows for shared functions to be defined only
once while allowing for the YAML function importing
to continue working
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>

---------
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>

lm_eval/api/metrics.py

@@ -8,7 +8,6 @@ from typing import List
 
 import numpy as np
 import sacrebleu
-import sklearn.metrics
 
 from lm_eval.api.registry import register_aggregation, register_metric
 
@@ -51,21 +50,24 @@ def bits_per_byte(items):
 
 @register_aggregation("f1")
 def f1_score(items):
+    from sklearn.metrics import f1_score
+
     unzipped_list = list(zip(*items))
     golds = unzipped_list[0]
     preds = unzipped_list[1]
-    fscore = sklearn.metrics.f1_score(golds, preds)
+    fscore = f1_score(golds, preds)
 
     return np.max(fscore)
 
 
 @register_aggregation("matthews_corrcoef")
 def matthews_corrcoef(items):
+    from sklearn.metrics import matthews_corrcoef
+
     unzipped_list = list(zip(*items))
     golds = unzipped_list[0]
     preds = unzipped_list[1]
-    # print(preds)
-    return sklearn.metrics.matthews_corrcoef(golds, preds)
+    return matthews_corrcoef(golds, preds)
 
 
 @register_aggregation("bleu")

lm_eval/tasks/afrimmlu/direct/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_choice(doc):
@@ -30,11 +30,3 @@ def doc_to_text(doc):
         choice4=choices[3],
     )
     return text
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrimmlu/translate/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_choice(doc):
@@ -30,11 +30,3 @@ def doc_to_text(doc):
         choice4=choices[3],
     )
     return text
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrimmlu/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_choice(doc):
@@ -30,11 +30,3 @@ def doc_to_text(doc):
         choice4=choices[3],
     )
     return text
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrixnli/anli prompt/en-direct/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_target(doc):
     replacements = {0: "True", 1: "Neither", 2: "False"}
     return replacements[doc["label"]]
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrixnli/anli prompt/native-direct/utils.py

-from sklearn.metrics import f1_score
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore
+from lm_eval.utils import weighted_f1_score

lm_eval/tasks/afrixnli/anli prompt/translate/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_target(doc):
     replacements = {0: "True", 1: "Neither", 2: "False"}
     return replacements[doc["label"]]
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrixnli/lai prompt/direct/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_text(doc):
@@ -17,11 +17,3 @@ def doc_to_text(doc):
 def doc_to_target(doc):
     replacements = {0: "entailment", 1: "neutral", 2: "contradiction"}
     return replacements[doc["label"]]
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/afrixnli/lai prompt/translate/utils.py

-from sklearn.metrics import f1_score
+from lm_eval.utils import weighted_f1_score
 
 
 def doc_to_text(doc):
@@ -17,11 +17,3 @@ def doc_to_text(doc):
 def doc_to_target(doc):
     replacements = {0: "entailment", 1: "neutral", 2: "contradiction"}
     return replacements[doc["label"]]
-
-
-def weighted_f1_score(items):
-    unzipped_list = list(zip(*items))
-    golds = unzipped_list[0]
-    preds = unzipped_list[1]
-    fscore = f1_score(golds, preds, average="weighted")
-    return fscore

lm_eval/tasks/drop/utils.py

@@ -2,7 +2,6 @@ import re
 import string
 
 import numpy as np
-from scipy.optimize import linear_sum_assignment
 
 
 _ARTICLES = re.compile(r"\b(a|an|the)\b", re.UNICODE)
@@ -117,6 +116,8 @@ def _align_bags(predicted, gold):
     Takes gold and predicted answer sets and first finds the optimal 1-1 alignment
     between them and gets maximum metric values over all the answers.
     """
+    from scipy.optimize import linear_sum_assignment
+
     scores = np.zeros([len(gold), len(predicted)])
     for gold_index, gold_item in enumerate(gold):
         for pred_index, pred_item in enumerate(predicted):

lm_eval/tasks/kobest/utils.py

 from datasets import Dataset
-from sklearn.metrics import f1_score
 
 
 def copa_doc_to_text(doc: dict) -> str:
@@ -41,6 +40,8 @@ def hellaswag_process_doc(doc: Dataset) -> Dataset:
 
 
 def macro_f1_score(items):
+    from sklearn.metrics import f1_score
+
     unzipped_list = list(zip(*items))
     golds = unzipped_list[0]
     preds = unzipped_list[1]

lm_eval/tasks/super_glue/cb/aggregate.py

 import numpy as np
-import sklearn
 
 
 def cb_multi_fi(items):
+    from sklearn.metrics import f1_score
+
     preds, golds = zip(*items)
     preds = np.array(preds)
     golds = np.array(golds)
-    f11 = sklearn.metrics.f1_score(y_true=golds == 0, y_pred=preds == 0)
-    f12 = sklearn.metrics.f1_score(y_true=golds == 1, y_pred=preds == 1)
-    f13 = sklearn.metrics.f1_score(y_true=golds == 2, y_pred=preds == 2)
+    f11 = f1_score(y_true=golds == 0, y_pred=preds == 0)
+    f12 = f1_score(y_true=golds == 1, y_pred=preds == 1)
+    f13 = f1_score(y_true=golds == 2, y_pred=preds == 2)
     avg_f1 = np.mean([f11, f12, f13])
     return avg_f1

lm_eval/tasks/super_glue/cb/t5_utils.py

-import sklearn.metrics
-
-
 def mean_3class_f1(predictions, references):  # This is a passthrough function
     string_label = ["entailment", "contradiction", "neutral"]
     predictions = (
@@ -23,6 +20,8 @@ def agg_mean_3class_f1(items):
     }
 
     def _fn(predictions, references):
+        import sklearn.metrics
+
         metric_fn = getattr(sklearn.metrics, metric_str)
         metric_val = metric_fn(references, predictions, **metric_fn_kwargs)
         return metric_val

lm_eval/tasks/super_glue/multirc/t5_utils.py

 import collections
 
 import numpy as np
-import sklearn.metrics
 
 
 def f1(predictions, references):  # This is a passthrough function
@@ -19,10 +18,12 @@ def f1(predictions, references):  # This is a passthrough function
 
 
 def agg_f1(items):
+    from sklearn.metrics import f1_score
+
     predictions, references = zip(*items)
     references, predictions = np.asarray(references), np.asarray(predictions)
 
-    return sklearn.metrics.f1_score(references, predictions)
+    return f1_score(references, predictions)
 
 
 def em(predictions, references):  # This is a passthrough function

lm_eval/utils.py

@@ -487,3 +487,13 @@ def create_iterator(raw_iterator, *, rank=0, world_size=1, limit=None):
     among ranks in multigpu setting or only pulling a sample of documents
     """
     return islice(raw_iterator, rank, limit, world_size)
+
+
+def weighted_f1_score(items):
+    from sklearn.metrics import f1_score
+
+    unzipped_list = list(zip(*items))
+    golds = unzipped_list[0]
+    preds = unzipped_list[1]
+    fscore = f1_score(golds, preds, average="weighted")
+    return fscore

pyproject.toml

@@ -104,3 +104,4 @@ known-first-party = ["lm_eval"]
 
 [tool.ruff.lint.extend-per-file-ignores]
 "__init__.py" = ["F401","F402","F403"]
+"utils.py" = ["F401"]

scripts/model_comparator.py

@@ -4,7 +4,6 @@ from typing import Dict, List, Tuple
 
 import numpy as np
 import pandas as pd
-import scipy.stats
 import torch
 
 import lm_eval.evaluator
@@ -23,11 +22,13 @@ def memory_stats():
 
 
 def calculate_z_value(res1: Dict, res2: Dict) -> Tuple[float, float]:
+    from scipy.stats.norm import sf
+
     acc1, acc2 = res1["acc,none"], res2["acc,none"]
     st_err1, st_err2 = res1["acc_stderr,none"], res2["acc_stderr,none"]
     Z = (acc1 - acc2) / np.sqrt((st_err1**2) + (st_err2**2))
     # Determining the p-value
-    p_value = 2 * scipy.stats.norm.sf(abs(Z))  # two-tailed test
+    p_value = 2 * sf(abs(Z))  # two-tailed test
     return Z, p_value