class SVM:
def __init__(self, db, ratio_split, lbd, kernel="raw_kernel", k=None, C=1, nkp=False):
self.n = int(len(db)*ratio_split)
self.lbd = lbd
self.k = Kernel(self.lbd,k)
self.ker = getattr(self.k, kernel)
self.X, self.y = db[:,0], db[:,1]#shuffle(db[:,0], db[:,1])
self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(db[:,0], db[:,1], train_size=ratio_split, random_state=42, stratify=db[:,1])
self.svc = SVC(kernel='precomputed', C=1)
self.nkp = nkp
def learn(self):
self.kernel_train = self.k.build_gram_matrix(self.X_train, self.ker)
self.svc.fit(self.kernel_train, self.y_train)
return self.k.comp_time
def score(self):
#check is matrice dans bon sens
self.kernel_test = self.k.build_gram_matrix_nonsq(self.X_test, self.X_train.T, self.ker)
self.y_pred = self.svc.predict(self.kernel_test)
return 1-zero_one_loss(self.y_test,self.y_pred)
def cross_val_score(self, k):
start = time()
self.kernel = self.k.build_gram_matrix(self.X, self.ker, nkp=self.nkp)
end = time() - start
score = cross_val_score(self.svc, self.kernel, self.y, cv=k)
return { "accuracy": sum(score)/len(score), "time": end, "stddev": np.std(score)}
def display_heatmap(self, categ = "traintest"):
#revoir labels
if "train" in categ.lower():
plt.figure()
sns.heatmap(self.kernel_train,xticklabels=self.y_train,yticklabels=self.y_train).set_title("Train")
if "test" in categ.lower():
plt.figure()
sns.heatmap(self.kernel_test,xticklabels=self.y_train,yticklabels=self.y_train).set_title("Test")
