77b4670b6a
COSMIT use np.clip in SGD
2012-10-29 00:47:00 +01:00
f65efcf8eb
More use sklearn.utils.testing.
...
My model is training :)
2012-10-25 21:21:22 +09:00
66fc97708a
ENH don't use deprecated interface in PPCA & cosmit
2012-10-18 14:09:48 +02:00
e64cea0342
ENH add random seed to logistic regression
2012-10-18 13:48:10 +02:00
60ae168464
COSMIT rm unused import
2012-10-16 14:36:49 +02:00
a06c367ba2
ENH sparse matrix support in randomized logistic regression
2012-10-16 14:34:48 +02:00
8d3c75feff
Store X as an ordered array.
...
This simplifies and makes more efficient the `fit` and transform
method (re-ordering in transform is no longer needed).
2012-10-16 14:18:43 +02:00
04251e155d
Better tests + cosmetic changes.
2012-10-16 14:18:43 +02:00
342c5be9f1
ENH: Consider order in X for IsotonicRegression.
...
Before, IsotonicRegression assumed X was an increasing sequence. This
allows to shuffle both X, y (see the test) and obtain the same
result. This makes IsotonicRegression behave more like a standard
Regression (which is what I would expect).
2012-10-16 14:18:43 +02:00
53b220358c
COSMIT correct error msgs in SGD and make them more consistent
2012-10-14 12:09:59 +02:00
97ed66151a
BUG + DOC l1_ratio in SGD and CD
...
SGD{Classifier,Regressor} treated l1_ratio as previously rho. Docstrings
stated that l1_ratio must be >0, which means pure L2 regularization would
not be possible.
2012-10-14 11:43:26 +02:00
72e71bb439
ENH address @agramfort's comments, fix some doctests
2012-10-11 20:56:03 +01:00
b39609af4e
ENH rename rho in SGD
2012-10-11 20:56:03 +01:00
16aabc7c3f
ENH rename rho to l1_ratio in ElasticNet and friends
2012-10-11 20:56:03 +01:00
9423e9c808
DOC various typos and other minor stuff
2012-10-11 11:26:22 +02:00
57001002bb
COSMIT pep8
2012-10-10 09:38:10 +02:00
03b8829a83
extend eps_ description
2012-10-08 19:35:39 +02:00
9701bb81ac
add dual_gap_ and eps_ to Enet and Lasso docstring
2012-10-08 17:04:59 +02:00
55ef1d617a
DOC: protect `classes_` for valid rst
2012-10-07 16:39:40 +02:00
0984a2a59f
cd_fast: use square norm directly
...
Previously the ℓ₂ norm was computed (which uses a sqrt), then squared.
Also, ℓ₂ norm is computed using np.dot.
2012-10-03 19:12:47 -07:00
244e8a68fa
Merge pull request #1187 from ogrisel/bugfix-logistic-ovr-probabilities
...
MRG: FIX: wrong probabilities for OvR LogisticRegression
2012-09-30 06:09:51 -07:00
db37872f5e
DOC: Enet alpha=0 => advice to use LinearRegression
2012-09-29 18:50:34 +02:00
0c2b07f492
Fix warning message.
2012-09-29 02:16:06 +09:00
2a7ebd955f
Cosmit: not a fan of defining a function in a loop :)
2012-09-29 02:16:06 +09:00
7575410adc
PEP8!
2012-09-29 02:16:06 +09:00
71fd662f63
Better documentation on the choice of solver.
2012-09-29 02:16:06 +09:00
15ce0557b2
lsqr is not available in old scipy versions...
2012-09-29 02:16:06 +09:00
1e117305b6
Fall back to dense_cholesky if sample_weight is given.
2012-09-29 02:16:05 +09:00
c5892aabb8
Fix damp parameter.
2012-09-29 02:16:05 +09:00
ec07896195
Add minimalistic test for each solver.
2012-09-29 02:16:05 +09:00
abee624bf6
Fix docstrings and doctests.
2012-09-29 02:16:05 +09:00
047109233f
BUG: transmit parameters correctly from Ridge to ridge_regression.
2012-09-29 02:16:05 +09:00
107abb6f3e
Add lsqr solver.
2012-09-29 02:16:05 +09:00
2cae34b059
Move solver option to constructor.
2012-09-29 02:16:05 +09:00
fa93e209ff
FIX: wrong probabilities for OvR LogisticRegression
2012-09-28 19:13:46 +02:00
e0d45a99f8
BUG: remove leftout debug prints
...
Bad bad bad @GaelVaroquaux
2012-09-28 09:41:31 +02:00
7842748cf7
BUG: copy and keep ordering
...
This is due to a change in behavior in numpy that now does
not keep the ordering by default when copying.
Fixes #1186
2012-09-28 09:32:03 +02:00
3941a4b34a
COSMIT pep8
2012-09-23 18:03:07 +01:00
c0a057843c
ENH rename Scaler to StandardScaler everywhere
2012-09-23 17:31:15 +01:00
b5339a9f24
Added the paper I used to implement isotonic_regression.
2012-09-22 22:06:55 +02:00
a164d57e74
TST writable coef_ and intercept_ on LogisticRegression
...
Solves issue #470 .
2012-09-20 23:04:52 +02:00
3ca3e95b72
DOC rm references to deprecated stuff
2012-09-20 22:15:39 +02:00
77e69edcec
COSMIT rm deprecated stuff -- lots of it
2012-09-20 21:40:33 +02:00
46f4c55787
Merge pull request #1169 from larsmans/ridge-cg
...
Ridge CG performance improvements
2012-09-20 12:01:27 -07:00
1313f25cda
COSMIT pep8
2012-09-20 14:49:14 +01:00
d93b33ed1a
ENH add max_iter argument to Ridge estimators
2012-09-19 14:08:52 +02:00
33efa17665
ENH Improve Ridge's conjugate gradient descent
...
Reorder terms in the algorithm so that X'X is not explicitly computed and
add LinearOperator as accepted types. This improves precedent code in two ways:
* Speed: I get speedups ranging from 2.5x to more than 10x.
Benchmark suite can be seen here: https://gist.github.com/1322711
where I print the time spent for various sizes and densities.
* Interface. By allowing a LinearOperator type, we can use the
conjugate gradient method almost for free on dense matrices with
special structure, pytables, etc., about anything that
implements a fast matrix-vector product.
2012-09-19 14:08:52 +02:00
35704c8351
COSMIT rm deprecated class_weight from fit in Ridge
2012-09-19 14:08:41 +02:00
4183cb3b34
COSMIT use LinearClassifierMixin in RidgeClassifier
...
Had to change a test due to different tie breaking.
2012-09-17 23:22:16 +02:00
88a0e5dccf
TST improve Ridge test
2012-09-17 23:22:16 +02:00
Lars Buitinck