@algoGuruZ:

I ran into the exact same issue yesterday and I wanted to comment on your PR. I did something very similar but subtly different. I couldn't think of a single scenario where I'd want to be able to supply independent sample (importance) weights for training (fit) and test (score and metric calculations). It seems to me that unless this serves a particular purpose (backed by a real use case), I would opt for a flag that just enables the sample_weight passed to fit_params, possibly as another fit_params item. This would eliminate silly errors like cut and paste errors.

I also found that for my purposes, I only had to change sklearn/model_selection/_validation.py, but you might be hitting _fit_and_score in _validation.py from a different caller. Here's the example I used to debug into the code in PyCharm. I set break points in _validation.py and that's all I really needed because fit_params passed to _fit_and_score had the sample_weights so I could isolate my changes.

import numpy as np
import sklearn.datasets
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import RandomizedSearchCV

X, y = sklearn.datasets.make_classification(
    n_samples=10000, n_features=10, n_informative=2,
    n_redundant=2, shuffle=True, random_state=0
)

np.random.seed = 5
wt = np.abs(np.random.random(len(y)))
rss = np.random.randint(1, 1000000, 5)
params = { 'random_state': rss}

rscv = RandomizedSearchCV(
    estimator=LogisticRegression(),
    param_distributions=params,
    n_iter=1,
    scoring='roc_auc',
    n_jobs=1,
    refit=True,
    cv=4,
    random_state=2,
    fit_params={'sample_weight': wt}  # maybe add flag here (see comments in PR)
)

rscv.fit(X, y)