I assume it fails on alpha=1000 ?

2/3 examples fail when alpha = 0.001, so not exclusively, no.

Failures:

ValueError('Floating-point under-/overflow occurred at epoch #2. Scaling input data with StandardScaler or MinMaxScaler might help.',)
{u'alpha': 1000.0,
 u'class_weight': None,
 u'eta0': 0.1,
 u'fit_intercept': 0.0,
 u'l1_ratio': 0.5,
 u'learning_rate': 'constant',
 u'loss': 'modified_huber',
 u'n_iter': 25.0,
 u'penalty': 'l2',
 u'shuffle': True}

ValueError('Floating-point under-/overflow occurred at epoch #3. Scaling input data with StandardScaler or MinMaxScaler might help.',)
{u'alpha': 0.001,
 u'class_weight': None,
 u'eta0': 0.1,
 u'fit_intercept': 0.0,
 u'l1_ratio': 0.5,
 u'learning_rate': 'optimal',
 u'loss': 'squared_hinge',
 u'n_iter': 25.0,
 u'penalty': 'elasticnet',
 u'shuffle': True}

ValueError('Floating-point under-/overflow occurred at epoch #3. Scaling input data with StandardScaler or MinMaxScaler might help.',)
{u'alpha': 0.001,
 u'class_weight': None,
 u'eta0': 0.001,
 u'fit_intercept': 0.0,
 u'l1_ratio': 0.5,
 u'learning_rate': 'optimal',
 u'loss': 'squared_hinge',
 u'n_iter': 100.0,
 u'penalty': 'l2',
 u'shuffle': True}

MinMaxScaler on the (0,1) range does help, but not always:

from sklearn.linear_model import SGDClassifier
from sklearn.datasets import load_iris
from sklearn import cross_validation
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, MinMaxScaler

iris = load_iris()

hyperparameter_choices = [ 

    # some examples, by no means exhausive
    {u'loss': 'modified_huber', u'shuffle': True, u'n_iter': 25.0, 
    u'l1_ratio': 0.5, u'learning_rate': 'constant', u'fit_intercept': 0.0, 
    u'penalty': 'l2', u'alpha': 1000.0, u'eta0': 0.1, u'class_weight': None},

    {u'loss': 'squared_hinge', u'shuffle': True, u'n_iter': 25.0, u'l1_ratio': 0.5, 
    u'learning_rate': 'optimal', u'fit_intercept': 0.0, u'penalty': 'elasticnet', 
    u'alpha': 0.001, u'eta0': 0.1, u'class_weight': None},

    {u'loss': 'squared_hinge', u'shuffle': True, u'n_iter': 100.0, u'l1_ratio': 0.5, 
    u'learning_rate': 'optimal', u'fit_intercept': 0.0, u'penalty': 'l2', u'alpha': 0.001, 
    u'eta0': 0.001, u'class_weight': None}
]

print "\nWith Standard scaling..."
for params in hyperparameter_choices:
    try:
        pipeline = Pipeline([
            ("standard_scaler", StandardScaler()),
            ("sgd", SGDClassifier(**params))
        ])
        scores = cross_validation.cross_val_score(pipeline, iris.data, iris.target, cv=5)
    except ValueError as ve:
        print "ValueError: %s" % ve

print "\nWith MinMax scaling..."     
for params in hyperparameter_choices:
    try:
        pipeline = Pipeline([
            ("minmax_scaler", MinMaxScaler()),
            ("sgd", SGDClassifier(**params))
        ])
        scores = cross_validation.cross_val_score(pipeline, iris.data, iris.target, cv=5)
        print "Success with: %s" % params
    except ValueError as ve:
        print "ValueError: %s" % ve

I just did some testing with this example and I think the parameters are just not good. With too high a learning rate, gradient descent will overshoot its target; that's an inherent risk with this algorithm. Scaling with StandardScaler and using the optimal learning rate fixes the first problem.

What remains are the other two. Here, the squared hinge loss goes off into infinity and its dloss becomes huge; so do the updates. Regularizing more or using the vanilla hinge loss solve this issue.

I think many SGD practitioners from the deep learning community clip the gradient norms (or the norm of the weights) to e.g. [-100, 100] to avoid such numerical stability issues in practice. That might be worth trying.

true -- we should experiment with this -- its a major annoyance during grid
search

2014-11-24 14:59 GMT+01:00 Olivier Grisel [email protected]:

I think many SGD practitioners from the deep learning community clip the
gradient norms or coefficients to [-100, 100] to avoid such numerical
stability issues in practice. That might be worth trying.

—
Reply to this email directly or view it on GitHub
#3040 (comment)
.

Peter Prettenhofer

we could add a param max_grad ? or clip_grad ?

true -- we should experiment with this -- its a major annoyance during grid search

I am hacking my copy of sgd_fast to clip dloss to [-100, 100], it helps for some cases on @worldveil's script but not all: the huber loss case is still unstable. Needs more investigation.

Note that the problem disappears when clipping dloss to [-100, 100 ] and preventing alpha to be larger than 10..

Do you need to check the gradient in every step then? Doesn't that impact performance quite a bit?

true -- we should experiment with this -- its a major annoyance during
grid
search

But at least in dev version it's possible to ask grid search to catch the
error and return 0 score...

On 25 November 2014 at 09:13, Andreas Mueller [email protected]
wrote:

Do you need to check the gradient in every step then? Doesn't that
impact performance quite a bit?

—
Reply to this email directly or view it on GitHub
#3040 (comment)
.

But at least in dev version it's possible to ask grid search to catch the
error and return 0 score...

completely missed that -- that's great!

2014-11-24 23:23 GMT+01:00 jnothman [email protected]:

true -- we should experiment with this -- its a major annoyance during
grid
search

But at least in dev version it's possible to ask grid search to catch the
error and return 0 score...

On 25 November 2014 at 09:13, Andreas Mueller [email protected]
wrote:

Do you need to check the gradient in every step then? Doesn't that
impact performance quite a bit?

—
Reply to this email directly or view it on GitHub
<
https://github.com/scikit-learn/scikit-learn/issues/3040#issuecomment-64274690>

.

—
Reply to this email directly or view it on GitHub
#3040 (comment)
.

Peter Prettenhofer

#3883 has an implementations that seems to work. Please have a look. I have not yet run benchmarks to see the computational overhead vs master but I have to go now so I cannot run them right now.

@worldveil #3883 seems to fix all the problems you reported. Please feel free to test on more cases on your own data and report any remaining issues.

Should be fixed by f5e0ea0, closing.