Merge pull request #114 from julianofoleiss/skgpuid

Added gpu_id parameter to the Scikit interface.

python/README.md

@@ -122,6 +122,9 @@ The usage of thundersvm scikit interface is similar to sklearn.svm.
 *max_mem_size*: int, optional (default=-1)\
 	set the maximum memory size (MB) that thundersvm uses, or -1 for no limit.
 
+*gpu_id*: int, optional (default=0)\
+	set which gpu to use for training.
+
 *decision_function_shape*: ‘ovo’, default=’ovo’, not supported yet for 'ovr'\
     only for classifier. Whether to return a one-vs-rest (‘ovr’) decision function of shape (n_samples, n_classes) as all other classifiers, or the original one-vs-one (‘ovo’) decision function of libsvm which has shape (n_samples, n_classes * (n_classes - 1) / 2).
 

python/thundersvmScikit.py

@@ -52,7 +52,7 @@ def __init__(self, kernel, degree,
                  gamma, coef0, C, nu, epsilon,
                  tol, probability, class_weight,
                  shrinking, cache_size, verbose,
-                 max_iter, n_jobs, max_mem_size, random_state):
+                 max_iter, n_jobs, max_mem_size, random_state,gpu_id):
         self.kernel = kernel
         self.degree = degree
         self.gamma = gamma
@@ -72,6 +72,7 @@ def __init__(self, kernel, degree,
         self.max_mem_size = max_mem_size
         thundersvm.model_new.restype = c_void_p
         self.model = thundersvm.model_new(SVM_TYPE.index(self._impl))
+        self.gpu_id = gpu_id
         if self.max_mem_size != -1:
             thundersvm.set_memory_size(c_void_p(self.model), self.max_mem_size)
 
@@ -171,6 +172,7 @@ def _dense_fit(self, X, y, solver_type, kernel):
             c_float(self.C), c_float(self.nu), c_float(self.epsilon), c_float(self.tol),
             self.probability, weight_size, weight_label, weight,
             self.verbose, self.max_iter, self.n_jobs, self.max_mem_size,
+            self.gpu_id,
             n_features, n_classes, self._train_succeed, c_void_p(self.model))
         self.n_features = n_features[0]
         self.n_classes = n_classes[0]
@@ -213,6 +215,7 @@ def _sparse_fit(self, X, y, solver_type, kernel):
                 c_float(self.C), c_float(self.nu), c_float(self.epsilon), c_float(self.tol),
                 self.probability, weight_size, weight_label, weight,
                 self.verbose, self.max_iter, self.n_jobs, self.max_mem_size,
+                self.gpu_id,
                 n_features, n_classes, self._train_succeed, c_void_p(self.model))
         self.n_features = n_features[0]
         self.n_classes = n_classes[0]
@@ -394,15 +397,15 @@ def __init__(self, kernel = 'rbf', degree = 3,
                   gamma = 'auto', coef0 = 0.0, C = 1.0,
                   tol = 0.001, probability = False, class_weight = None,
                   shrinking = False, cache_size = None, verbose = False,
-                  max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None, decison_function_shape = 'ovo'):
+                  max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None, decison_function_shape = 'ovo', gpu_id=0):
          self.decison_function_shape = decison_function_shape
          super(SVC, self).__init__(
              kernel=kernel, degree=degree, gamma=gamma,
              coef0=coef0, C=C, nu=0., epsilon=0.,
              tol=tol, probability=probability,
              class_weight=class_weight, shrinking = shrinking,
              cache_size = cache_size, verbose = verbose,
-             max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state)
+             max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state, gpu_id=gpu_id)
 
 
 
@@ -411,28 +414,28 @@ class NuSVC(SvmModel, ClassifierMixin):
     def __init__(self, kernel = 'rbf', degree = 3, gamma = 'auto',
                  coef0 = 0.0, nu = 0.5, tol = 0.001,
                  probability = False, shrinking = False, cache_size = None, verbose = False,
-                 max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None, decison_function_shape = 'ovo'):
+                 max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None, decison_function_shape = 'ovo', gpu_id=0):
         self.decison_function_shape = decison_function_shape
         super(NuSVC, self).__init__(
             kernel = kernel, degree = degree, gamma = gamma,
             coef0 = coef0, C = 0., nu = nu, epsilon= 0.,
             tol = tol, probability = probability, class_weight = None,
             shrinking = shrinking, cache_size = cache_size, verbose = verbose,
-            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state
+            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state, gpu_id=gpu_id
         )
 
 class OneClassSVM(SvmModel):
     _impl = 'one_class'
     def __init__(self, kernel = 'rbf', degree = 3, gamma = 'auto',
                  coef0 = 0.0, nu = 0.5, tol = 0.001,
                  shrinking = False, cache_size = None, verbose = False,
-                 max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None):
+                 max_iter = -1, n_jobs = -1, max_mem_size = -1, random_state = None, gpu_id=0):
         super(OneClassSVM, self).__init__(
             kernel = kernel, degree = degree, gamma = gamma,
             coef0 = coef0, C = 0., nu = nu, epsilon = 0.,
             tol = tol, probability= False, class_weight = None,
             shrinking = shrinking, cache_size = cache_size, verbose = verbose,
-            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state
+            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = random_state, gpu_id=gpu_id
         )
 
     def fit(self, X, y=None):
@@ -444,25 +447,25 @@ def __init__(self, kernel = 'rbf', degree = 3, gamma = 'auto',
                  coef0 = 0.0, C = 1.0, epsilon = 0.1,
                  tol = 0.001, probability = False,
                  shrinking = False, cache_size = None, verbose = False,
-                 max_iter = -1, n_jobs = -1,max_mem_size = -1):
+                 max_iter = -1, n_jobs = -1,max_mem_size = -1, gpu_id=0):
         super(SVR, self).__init__(
             kernel = kernel, degree = degree, gamma = gamma,
             coef0 = coef0, C = C, nu = 0., epsilon = epsilon,
             tol = tol, probability = probability, class_weight = None,
             shrinking = shrinking, cache_size = cache_size, verbose = verbose,
-            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = None
+            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = None, gpu_id=gpu_id
         )
 
 class NuSVR(SvmModel, RegressorMixin):
     _impl = 'nu_svr'
     def __init__(self, kernel = 'rbf', degree = 3, gamma = 'auto',
                coef0 = 0.0, nu = 0.5, C = 1.0, tol = 0.001, probability = False,
                shrinking = False, cache_size = None, verbose = False,
-               max_iter = -1, n_jobs = -1, max_mem_size = -1):
+               max_iter = -1, n_jobs = -1, max_mem_size = -1, gpu_id=0):
         super(NuSVR, self).__init__(
             kernel = kernel, degree = degree, gamma = gamma,
             coef0 = coef0, nu = nu, C = C, epsilon = 0.,
             tol = tol, probability = probability, class_weight = None,
             shrinking = shrinking, cache_size = cache_size, verbose = verbose,
-            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = None
+            max_iter = max_iter, n_jobs = n_jobs, max_mem_size = max_mem_size, random_state = None, gpu_id=gpu_id
         )

src/thundersvm/thundersvm-scikit.cpp

@@ -43,7 +43,12 @@ extern "C" {
                                   float cost, float nu, float epsilon, float tol, int probability,
                                   int weight_size, int* weight_label, float* weight,
                                   int verbose, int max_iter, int n_cores, int max_mem_size,
+                                  int gpu_id,
                                   int* n_features, int* n_classes, int* succeed, SvmModel* model){
+#ifdef USE_CUDA
+        CUDA_CHECK(cudaSetDevice(gpu_id));
+#endif
+
         succeed[0] = 1;
         if(verbose)
             el::Loggers::reconfigureAllLoggers(el::ConfigurationType::Enabled, "true");
@@ -145,7 +150,13 @@ extern "C" {
                                  float cost, float nu, float epsilon, float tol, int probability,
                                  int weight_size, int* weight_label, float* weight,
                                  int verbose, int max_iter, int n_cores, int max_mem_size,
+                                 int gpu_id,
                                  int* n_features, int* n_classes, int* succeed, SvmModel* model){
+
+#ifdef USE_CUDA
+        CUDA_CHECK(cudaSetDevice(gpu_id));
+#endif
+
         succeed[0] = 1;
         if(verbose)
             el::Loggers::reconfigureAllLoggers(el::ConfigurationType::Enabled, "true");