diff --git a/scikits/learn/logistic.py b/scikits/learn/logistic.py
index b202109d945002d5d281537b5d09fdcfe8259be5..d78a2929541ec32324e1029b570b97416925ab0d 100644
--- a/scikits/learn/logistic.py
+++ b/scikits/learn/logistic.py
@@ -96,13 +96,12 @@ class LogisticRegression(object):
def intercept_(self):
if self.bias_ > 0:
return self.raw_coef_[:,-1]
- else:
- raise ValueError('intercept_ not estimated')
+ return 0.0
+
@property
def coef_(self):
if self.bias_ > 0:
return self.raw_coef_[:,:-1]
- else:
- return self.raw_coef_
+ return self.raw_coef_
diff --git a/scikits/learn/src/liblinear_helper.c b/scikits/learn/src/liblinear_helper.c
index cf67880e4343c3072351ed018e161263a4d72918..714832bc9f019572ad12a27b82b798c891b2b172 100644
--- a/scikits/learn/src/liblinear_helper.c
+++ b/scikits/learn/src/liblinear_helper.c
@@ -13,8 +13,9 @@
* Special care must be taken with indices, since libsvm indices start
* at 1 and not at 0.
*
+ * If bias is > 0, we append an item at the end.
*/
-struct feature_node **dense_to_sparse (double *x, npy_intp *dims)
+struct feature_node **dense_to_sparse (double *x, npy_intp *dims, double bias)
{
struct feature_node **sparse;
register int i, j; /* number of nonzero elements in row i */
@@ -40,53 +41,15 @@ struct feature_node **dense_to_sparse (double *x, npy_intp *dims)
}
/* set bias element */
- T->value = 1.0;
- T->index = j;
- ++ T;
-
- /* set sentinel */
- T->index = -1;
- ++ T;
-
- /* allocate memory and copy collected items*/
- count = T - temp;
- sparse[i] = (struct feature_node *) malloc(count * sizeof(struct feature_node));
- if (sparse[i] == NULL) return NULL;
- memcpy(sparse[i], temp, count * sizeof(struct feature_node));
- }
-
- free(temp);
- return sparse;
-}
-
-struct feature_node **dense_to_sparse_nobias (double *x, npy_intp *dims)
-{
- struct feature_node **sparse;
- register int i, j; /* number of nonzero elements in row i */
- struct feature_node *temp; /* stack for nonzero elements */
- struct feature_node *T; /* pointer to the top of the stack */
- int count;
-
- sparse = (struct feature_node **) malloc (dims[0] * sizeof(struct feature_node *));
- temp = (struct feature_node *) malloc ((dims[1]+1) * sizeof(struct feature_node));
-
- if (sparse == NULL || temp == NULL) return NULL;
-
- for (i=0; i<dims[0]; ++i) {
- T = temp; /* reset stack pointer */
-
- for (j=1; j<=dims[1]; ++j) {
- T->value = *x;
- if (T->value != 0) {
+ if (bias > 0) {
+ T->value = 1.0;
T->index = j;
- ++T;
+ ++ T;
}
- ++x; /* go to next element */
- }
/* set sentinel */
T->index = -1;
- ++T;
+ ++ T;
/* allocate memory and copy collected items*/
count = T - temp;
@@ -99,6 +62,7 @@ struct feature_node **dense_to_sparse_nobias (double *x, npy_intp *dims)
return sparse;
}
+
struct problem * set_problem(char *X,char *Y, npy_intp *dims, double bias)
{
struct problem *problem;
@@ -113,7 +77,7 @@ struct problem * set_problem(char *X,char *Y, npy_intp *dims, double bias)
problem->n = (int) dims[1];
}
problem->y = (int *) Y;
- problem->x = dense_to_sparse((double *) X, dims); /* TODO: free */
+ problem->x = dense_to_sparse((double *) X, dims, bias); /* TODO: free */
problem->bias = bias;
if (problem->x == NULL) {
free(problem);
@@ -191,7 +155,7 @@ int copy_predict(char *train, struct model *model, npy_intp *train_dims,
register int i, n;
struct feature_node **train_nodes;
n = train_dims[0];
- train_nodes = dense_to_sparse((double *) train, train_dims);
+ train_nodes = dense_to_sparse((double *) train, train_dims, -1.0);
if (train_nodes == NULL)
return -1;
for(i=0; i<n; ++i) {
@@ -214,7 +178,7 @@ int copy_prob_predict(char *predict, struct model *model, npy_intp *predict_dims
int n, m;
n = predict_dims[0];
m = model->nr_class;
- predict_nodes = dense_to_sparse((double *) predict, predict_dims);
+ predict_nodes = dense_to_sparse((double *) predict, predict_dims, -1.0);
if (predict_nodes == NULL)
return -1;
for(i=0; i<n; ++i) {
diff --git a/scikits/learn/svm.py b/scikits/learn/svm.py
index 32647428b039efb09adbb6c3285901e49645a6a5..849a2c2b9946f79b0dcc1e70f8caac72c82a24db 100644
--- a/scikits/learn/svm.py
+++ b/scikits/learn/svm.py
@@ -355,9 +355,13 @@ class LinearSVC(object):
@property
def intercept_(self):
- return self.raw_coef[:,-1]
+ if self.bias_ > 0:
+ return self.raw_coef[:,-1]
+ return 0.0
@property
def coef_(self):
- return self.raw_coef[:,:-1]
+ if self.bias_ > 0:
+ return self.raw_coef[:,:-1]
+ return self.raw_coef_
diff --git a/scikits/learn/tests/test_logistic.py b/scikits/learn/tests/test_logistic.py
index 84da42798091680bd078ed1e2ca6bead16d9975a..93fb2040af39455d05759ade4d8ee30fa5c656fb 100644
--- a/scikits/learn/tests/test_logistic.py
+++ b/scikits/learn/tests/test_logistic.py
@@ -8,11 +8,19 @@ X = [[0, 0], [0, 1], [1, 1]]
Y1 = [0,1,1]
Y2 = [0,1,2]
+
def test_predict_2_classes():
clf = logistic.LogisticRegression()
clf.fit(X, Y1)
+ assert_array_almost_equal(clf.coef_, [[-0.27501564, -0.60803562]])
+ assert_array_almost_equal(clf.intercept_, [-0.08642295])
assert_array_equal(clf.predict([[-1, -1], [0, 1],]), [0, 1])
+ clf = logistic.LogisticRegression(intercept=False)
+ clf.fit(X, Y1)
+ assert_array_almost_equal(clf.coef_, [[-0.28540916, -0.63236105]])
+ assert_array_almost_equal(clf.intercept_, [0])
+
def test_predict_3_classes():
clf = logistic.LogisticRegression()
clf.fit(X, Y2)
@@ -22,7 +30,7 @@ def test_predict_proba():
clf = logistic.LogisticRegression()
clf.fit(X, Y2)
assert_array_almost_equal(clf.predict_proba([[1, 1]]),
- [[ 0.23148573, 0.31760051, 0.45091376]])
+ [[ 0.21490268, 0.32639437, 0.45870294]])
clf = logistic.LogisticRegression(penalty='l1')
clf.fit(X, Y2)