n_iters_ in SVR is always saved as 10,000 about scikit-learn-intelex

[Alexsandruss]

This is expected behavior for now. Warning for n_iter_ property might be added.

scikit-learn-intelex/onedal/svm/svm.py

Lines 218 to 221 in bfa470b

	max_iter = 10000 if self.max_iter == -1 else self.max_iter
	# TODO: remove this workaround
	# when oneDAL SVM starts support of 'n_iterations' result
	self.n_iter_ = 1 if max_iter < 1 else max_iter

[caspimoshe]

This is expected behavior for now. Warning for n_iter_ property might be added.

scikit-learn-intelex/onedal/svm/svm.py

Lines 218 to 221 in bfa470b

	max_iter = 10000 if self.max_iter == -1 else self.max_iter
	# TODO: remove this workaround
	# when oneDAL SVM starts support of 'n_iterations' result
	self.n_iter_ = 1 if max_iter < 1 else max_iter

Thanks for the fast response, Alex.
My main issue is that the max_iter parameter seem to have no effect.

Attaching code that shows it:
With patch the code ran for -0.84 seconds and managed to fit the data
`import time

import numpy as np
from sklearnex import patch_sklearn

patch_sklearn()
from sklearn.svm import SVR

patch_svr = SVR(max_iter=2)
np.random.seed(0)
X = np.random.randn(int(1e5), 5)
y = np.mean(X, axis=1)

tic = time.time()
patch_svr.fit(X, y)
print("Time to fit SVR with patch:", time.time() - tic)

score = patch_svr.score(X, y)
print("Score:", score)`

Without the patch the model terminated early, after 0.34 seconds and did not fit the data
`import time

import numpy as np
from sklearnex import patch_sklearn

patch_sklearn()
from sklearn.svm import SVR

patch_svr = SVR(max_iter=2)
np.random.seed(0)
X = np.random.randn(int(1e5), 5)
y = np.mean(X, axis=1)

tic = time.time()
patch_svr.fit(X, y)
print("Time to fit SVR with patch:", time.time() - tic)

score = patch_svr.score(X, y)
print("Score:", score)`

[Alexsandruss]

sklearn and sklearnex use different implementations of SVM, thus, different behavior on same number of iterations it expected.
SVM algorithm stopping is controlled by two parameters: maximum number of iterations (max_iter) and threshold/tolerance for stopping criterion (tol). sklearnex SVM might stop in same point while max_iters (>= stopping point) are widely different due to trigger of stopping criterion. Because of previously mentioned workaround it's impossible to know how many iterations were performed. If tol is significantly close to 0 (for example, 1e-32), stopping criterion would be unreachable and SVM training time becomes strictly proportional to max_iter.