Why am I getting 32 bit response for confusion_matrix when all the inputs are 64 bit?

[simonm3]

I have a sample of 1m rows and the cohen_kappa_score showed as -11.3 when it should be -1 to +1. Further investigation showed that the confusion matrix returns a 32 bit result and in cohen_kappa calculation the outer product overflows.

Here is an example with a small amount of data. If I put np.int64 around the confusion matrix then it works.

import sys
from sklearn.metrics import confusion_matrix
y1 = np.int64([1,0,0,1])
y2 = np.int64([0,0,1,1])
confusion = confusion_matrix(y1, y2)
sys.version, type(y1[0]), type(y2[0]), type(confusion[0,0])

('3.5.2 |Anaconda 4.1.1 (64-bit)| (default, Jul 5 2016, 11:41:13) [MSC v.1900 64 bit (AMD64)]',
numpy.int64,
numpy.int64,
numpy.int32)

[jnothman]

Well, the type of the classes should not technically matter, but if the data set is large, you're right, we should be counting in a larger dtype. Does changing np.int to np.int64 in https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/metrics/classification.py#L252 fix your issue?

…

On 24 November 2016 at 11:00, simon ***@***.***> wrote: I have a sample of 1m rows and the cohen_kappa_score showed as -11.3 when it should be -1 to +1. Further investigation showed that the confusion matrix returns a 32 bit result and in cohen_kappa calculation the outer product overflows. Here is an example with a small amount of data. If I put np.int64 around the confusion matrix then it works. import sys from sklearn.metrics import confusion_matrix y1 = np.int64([1,0,0,1]) y2 = np.int64([0,0,1,1]) confusion = confusion_matrix(y1, y2) sys.version, type(y1[0]), type(y2[0]), type(confusion[0,0]) ('3.5.2 |Anaconda 4.1.1 (64-bit)| (default, Jul 5 2016, 11:41:13) [MSC v.1900 64 bit (AMD64)]', numpy.int64, numpy.int64, numpy.int32) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#7929>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6-xwTCcnRKY28cqW2YQN9M88rJHSks5rBNOEgaJpZM4K7KG5> .

[simonm3]

Yes that works! However I note that np.iinfo(int) = 32 bit and np.finfo(float) = 64 bit. Why is the default int 32 bit when I am running a 64 bit machine?

…

On 24 November 2016 at 00:09, Joel Nothman ***@***.***> wrote: Well, the type of the classes should not technically matter, but if the data set is large, you're right, we should be counting in a larger dtype. Does changing np.int to np.int64 in https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/metrics/ classification.py#L252 fix your issue? On 24 November 2016 at 11:00, simon ***@***.***> wrote: > I have a sample of 1m rows and the cohen_kappa_score showed as -11.3 when > it should be -1 to +1. Further investigation showed that the confusion > matrix returns a 32 bit result and in cohen_kappa calculation the outer > product overflows. > > Here is an example with a small amount of data. If I put np.int64 around > the confusion matrix then it works. > > import sys > from sklearn.metrics import confusion_matrix > y1 = np.int64([1,0,0,1]) > y2 = np.int64([0,0,1,1]) > confusion = confusion_matrix(y1, y2) > sys.version, type(y1[0]), type(y2[0]), type(confusion[0,0]) > > ('3.5.2 |Anaconda 4.1.1 (64-bit)| (default, Jul 5 2016, 11:41:13) [MSC > v.1900 64 bit (AMD64)]', > numpy.int64, > numpy.int64, > numpy.int32) > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#7929>, or mute the > thread > <https://github.com/notifications/unsubscribe-auth/AAEz6- xwTCcnRKY28cqW2YQN9M88rJHSks5rBNOEgaJpZM4K7KG5> > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#7929 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABJN6b0eTtdmm37bhFvG550bqbGDrpO3ks5rBNWcgaJpZM4K7KG5> .

[jnothman]

Could you please submit a PR, ideally including a test, and we'll see if we can get that change merged, or if we want it to be conditional on the size of the data, etc.

…

On 24 November 2016 at 22:40, simon ***@***.***> wrote: Yes that works! However I note that np.iinfo(int) = 32 bit and np.finfo(float) = 64 bit. Why is the default int 32 bit when I am running a 64 bit machine? On 24 November 2016 at 00:09, Joel Nothman ***@***.***> wrote: > Well, the type of the classes should not technically matter, but if the > data set is large, you're right, we should be counting in a larger dtype. > > Does changing np.int to np.int64 in > https://github.com/scikit-learn/scikit-learn/blob/ master/sklearn/metrics/ > classification.py#L252 > fix your issue? > > On 24 November 2016 at 11:00, simon ***@***.***> wrote: > > > I have a sample of 1m rows and the cohen_kappa_score showed as -11.3 when > > it should be -1 to +1. Further investigation showed that the confusion > > matrix returns a 32 bit result and in cohen_kappa calculation the outer > > product overflows. > > > > Here is an example with a small amount of data. If I put np.int64 around > > the confusion matrix then it works. > > > > import sys > > from sklearn.metrics import confusion_matrix > > y1 = np.int64([1,0,0,1]) > > y2 = np.int64([0,0,1,1]) > > confusion = confusion_matrix(y1, y2) > > sys.version, type(y1[0]), type(y2[0]), type(confusion[0,0]) > > > > ('3.5.2 |Anaconda 4.1.1 (64-bit)| (default, Jul 5 2016, 11:41:13) [MSC > > v.1900 64 bit (AMD64)]', > > numpy.int64, > > numpy.int64, > > numpy.int32) > > > > — > > You are receiving this because you are subscribed to this thread. > > Reply to this email directly, view it on GitHub > > <#7929>, or mute the > > thread > > <https://github.com/notifications/unsubscribe-auth/AAEz6- > xwTCcnRKY28cqW2YQN9M88rJHSks5rBNOEgaJpZM4K7KG5> > > . > > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <https://github.com/scikit-learn/scikit-learn/issues/ 7929#issuecomment-262659982>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ ABJN6b0eTtdmm37bhFvG550bqbGDrpO3ks5rBNWcgaJpZM4K7KG5> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#7929 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6w6s8icYec_QbWL7BXFG1ysPmMkXks5rBXergaJpZM4K7KG5> .

[lesteve]

Why is the default int 32 bit when I am running a 64 bit machine?

For the record this happens only on Windows, on my Linux box, np.iinfo(int) is int64. I always wondered what is the point of using np.int to be honest instead of an explicit np.int32 or np.int64. @jnothman I am interested if you have any hints about that.

[amueller]

@jnothman is this still open?

[jnothman]

Waiting for you or someone to review #8094