def get_training_datas(method, all=True, replace=False):
"""
Construct training, testing data, and kernels.
:param: method: string, method used for computing kernels
:param: replace: Boolean, whether or not replace the existing files in the repo
:return:
- X_train: pd.DataFrame, training sequences
- X_val: pd.DataFrame, validation sequences
- y_train: pd.DataFrame, training labels
- y_val: pd.DataFrame, validation labels
- X_test: pd.DataFrame, testing sequences
- K: np.array, kernel
- ID: np.array, Ids
"""
file = 'training_data_'+method+'.pkl'
if not all:
X_train, y_train, X_val, y_val, X_test = train_test_split()
X_test.loc[:, 'Id'] = -(X_test.loc[:, 'Id'] + 1)
X = pd.concat((X_train, X_val, X_test), axis=0)
ID = X.loc[:, 'Id']
else:
if trainInRepo(file) and not replace:
X_train, y_train, X_val, y_val, X_test, K, ID = pkl.load(open(os.path.join(path, file), 'rb'))
else:
X_train, y_train, X_val, y_val, X_test = train_test_split()
X_test.loc[:, 'Id'] = -(X_test.loc[:, 'Id']+1)
X = pd.concat((X_train, X_val, X_test), axis=0)
ID = np.array(X.loc[:, 'Id'])
K = km.select_method(X, method)
file = 'training_data_'+method+'.pkl'
pkl.dump([X_train, y_train, X_val, y_val, X_test, K, ID], open(os.path.join(path, file), 'wb'))
return X_train, y_train, X_val, y_val, X_test, K, ID
def grid_search(label,
X,
y,
svm_params,
methods,
train_size=0.75,
graph=False):
"""
Implementation of the cross validation
Parameters:
- kernel: function, kernel function
- label: int (0, 1 or 2), label of the set of data
- X: array, observations
- y: array, labels
- svm_params: array, parameters of the SVM classifier
- kernel_params: array, parameters of the kernel function
- train_size: float (between 0 and 1), proportion of data for the train part
- graph: bool, plot the evolution of the accuracy wrt log(svm_params) or not
Returns the best SVM classifier
"""
Xtr, ytr, Xte, yte = train_test_split(X, y, train_size)
best_score = 0
best_clf = None
for method in methods:
kernel, kernel_param = kernels.select_method(method)
print()
scores = []
for c in svm_params:
print('Parameters : ' + str([method, c]))
gram_file = "../gram_matrix/gramMat_" + str(
label) + "_" + method + ".p"
clf = SVM.SupportVectorMachine(kernel=kernel,
C=c,
kernel_params=kernel_param)
clf.fit(Xtr, ytr, gram_file)
score = accuracy(clf.predict(Xte), yte)
if score > best_score:
best_score = score
best_clf = clf
print("Accuracy score = " + str(score) + '\n')
scores.append(score)
if graph:
plt.plot(np.log10(svm_params),
scores,
label='kernel_param = ' + str(kernel_param))
if graph:
plt.title('Evolution of the accuracy wrt log(C)')
plt.legend()
plt.savefig('../res/cross_val' + str(label) + '.png')
plt.show()
return best_clf
def create_models(labels, params, train_size=0.75):
"""
Create the model for each label and save the prediction on the test set
on a csv file
Parameters:
- labels: list, list of the labels
- params: dict, dictionnary containing the kernel method and
the SVM parameter
- train_size: float, proportion of the train part in the data set
"""
ytes = np.array([])
for label in labels:
print("\n*******Treating group " + str(label) + "*******\n")
Xtr, ytr = get_train(label)
Xte = get_test(label)
Xtr, ytr, Xv, yv = train_test_split(Xtr, ytr, train_size)
method, c = params[label]
kernel, kernel_param = kernels.select_method(method)
gram_file = "../gram_matrix/gramMat_" + str(
label) + "_" + method + ".p"
clf = SVM.SupportVectorMachine(kernel=kernel,
C=c,
kernel_params=kernel_param)
clf.fit(Xtr, ytr, gram_file)
score = accuracy(clf.predict(Xv), yv)
print("Accuracy score = " + str(score) + '\n')
ytes = np.concatenate((ytes, clf.predict(Xte)), axis=None)
results = pd.DataFrame({'Id': list(range(3000)), 'Bound': ytes})
results['Bound'] = [int((val + 1) / 2) for val in results['Bound']]
results.to_csv('../predictions/Yte.csv', index=False)
def import_data_to_train(method, all=True, replace=False):
#preparing kernels in this section but first we had to Construct training and testing data
file = 'training_data_' + method + '.pkl'
if not all:
X_train, y_train, X_val, y_val, X_test = spli_test()
X_test.loc[:, 'Id'] = -(X_test.loc[:, 'Id'] + 1)
X = pd.concat((X_train, X_val, X_test), axis=0)
ID = X.loc[:, 'Id']
else:
if trainInRepo(file) and not replace:
X_train, y_train, X_val, y_val, X_test, K, ID = pkl.load(
open(os.path.join(path, file), 'rb'))
else:
X_train, y_train, X_val, y_val, X_test = spli_test()
X_test.loc[:, 'Id'] = -(X_test.loc[:, 'Id'] + 1)
X = pd.concat((X_train, X_val, X_test), axis=0)
ID = np.array(X.loc[:, 'Id'])
K = km.select_method(X, method)
file = 'training_data_' + method + '.pkl'
pkl.dump([X_train, y_train, X_val, y_val, X_test, K, ID],
open(os.path.join(path, file), 'wb'))
return X_train, y_train, X_val, y_val, X_test, K, ID