def cross_validation(y, x, degree, k, k_indices,method, error, feature_augmentation, hyperparams):
""""""
from helpers_data import feature_processing, feat_augmentation, standardize, build_poly
from implementations import ridge_regression, least_squares, least_squares_GD, least_squares_SGD, logistic_regression, reg_logistic_regression
# get k'th subgroup in test, others in train
te_indice = k_indices[k]
tr_indice = k_indices[~(np.arange(k_indices.shape[0]) == k)]
tr_indice = tr_indice.reshape(-1)
y_te = y[te_indice]
y_tr = y[tr_indice]
x_te = x[te_indice]
x_tr = x[tr_indice]
x_tr, y_tr, median = feature_processing (x_tr, y_tr, 'mean', replace_feature = True, suppr_outliers = hyperparams[-1], threshold = 3, ref_median=[])
x_te, y_te, _= feature_processing (x_te, y_te, 'mean', replace_feature = True, suppr_outliers = False, threshold = 3, ref_median=median)
tx_tr_aug = []
tx_te_aug = []
if feature_augmentation:
tx_tr_aug, index = feat_augmentation(x_tr, 0.003)
tx_te_aug, _ = feat_augmentation(x_te, 0.003, False, index)
# form data with polynomial degree
tx_tr = build_poly(x_tr, degree, feature_augmentation, tx_tr_aug)
tx_te = build_poly(x_te, degree, feature_augmentation, tx_te_aug)
tx_tr, mean, std = standardize(tx_tr)
tx_te, _, _ = standardize(tx_te, mean, std)
#print('Mean and std of each feature in train set: {} , {}'.format(tx_tr.mean(axis = 0),tx_tr.std(axis = 0)))
#print('Mean and std of each feature in test set: {} , {}'.format(tx_te.mean(axis = 0),tx_te.std(axis = 0)))
if method == 'rr': w,_ = ridge_regression(y_tr, tx_tr, hyperparams[0]) # ridge regression
elif method == 'ls': w,_ = least_squares(y_tr, tx_tr) # least square
elif method == 'lsGD': w,_ = least_squares_GD(y_tr, tx_tr, hyperparams[0], hyperparams[1], hyperparams[2]) # gradient descent
elif method == 'lsSGD': w,_ = least_squares_SGD(y_tr, tx_tr, hyperparams[0], hyperparams[1], hyperparams[2], hyperparams[3]) # stoch GD
elif method == 'log': w,_ = logistic_regression(y_tr, tx_tr, hyperparams[0], hyperparams[1], hyperparams[2]) # logistic reg
elif method == 'rlog': w,_ =reg_logistic_regression(y_tr, tx_tr, hyperparams[3], np.zeros(tx_tr.shape[1]), hyperparams[1], hyperparams[2]) # regularised logistic reg
else: raise NotImplementedError
if method == 'log':
loss_tr = cal_loglike(y_tr, tx_tr, w)
loss_te = cal_loglike(y_te, tx_te, w)
elif method == 'rlog':
loss_tr = cal_loglike_r(y_tr, tx_tr, w, hyperparams[3])
loss_te = cal_loglike_r(y_te, tx_te, w, hyperparams[3])
else :
# calculate the loss for train and test data
loss_tr = compute_loss(y_tr, tx_tr, w, error)
loss_te = compute_loss(y_te, tx_te, w, error)
y_pred = predict_labels(np.array(w).T, tx_te)
acc = accuracy(y_te,y_pred)
return loss_tr, loss_te, w, acc
def cross_validation(y,
tx,
mlfunction,
split_number=5,
lambda_=1e-6,
gamma=0.001):
'''Performs a ml_function given as parameters using cross validation on the training set split_number folds (5 as default value) '''
# define empty lists to store train/test losses and accuracy
train_loss_ = []
test_loss_ = []
train_accuracy_ = []
test_accuracy_ = []
# get k_indices
k_indices = build_k_indices(len(y), split_number)
for ki in range(len(k_indices)):
# set the k'th indices as test, and others as training set
#train_idx = np.asarray([k_indices[i] for i in np.delete( np.arange(len(k_indices)), ki)]).flatten()
test_idx = np.asarray(k_indices[ki])
train_idx = np.delete(np.arange(len(y)), test_idx)
train_tX = tx[train_idx]
train_y = y[train_idx]
test_tX = tx[test_idx]
test_y = y[test_idx]
if (mlfunction == 'ridge_regression'):
w, loss = impl.ridge_regression(train_y, train_tX, lambda_)
elif (mlfunction == 'least_squares'):
w, loss = impl.least_squares(train_y, train_tX)
elif (mlfunction == 'logistic_regression'):
w, loss = impl.logistic_regression(train_y, train_tX)
elif (mlfunction == 'reg_logistic_regression'):
w, loss = impl.reg_logistic_regression(train_y, train_tX, lambda_)
elif (mlfunction == 'least_squares_sgd'):
w, loss = impl.least_squares_SGD(train_y, train_tX, gamma)
elif (mlfunction == 'least_squares_gd'):
w, loss = impl.least_squares_GD(train_y, train_tX, gamma)
else:
print('ERROR: ml_function not recognized')
print(
'least_squares, least_squares_gd, least_squares_sgd, logistic_regression, reg_logistic_regression'
)
return None
# Calculate different losses and accuracy
train_loss_.append(impl.compute_loss_mse(train_y, train_tX, w))
test_loss_.append(impl.compute_loss_mse(test_y, test_tX, w))
train_accuracy_ = impl.compute_accuracy(train_y, train_tX, w)
test_accuracy_ = impl.compute_accuracy(test_y, test_tX, w)
return np.mean(train_loss_), np.mean(test_loss_), np.mean(
train_accuracy_), np.mean(test_accuracy_)
def get_model(model, y, tx, initial_w, max_iters, gamma, lambda_, batch_size):
""" Returns the learned weights 'w' (last weight vector) and
the corresponding loss function by a given model.
Parameters
----------
model: string
The model
y: ndarray
The labels
tx: ndarray
The feature matrix
initial_w: ndarray
The initial weights
max_iters: integer
The number of steps to run
gamma: integer
The step size
lambda_: integer
The regularization parameter
batch_size: integer
The batch size
Returns
-------
tuple
The learned weights
"""
if model == "MSE_GD":
w, _ = least_squares_GD(y, tx, initial_w, max_iters, gamma)
elif model == "MSE_SGD":
w, _ = least_squares_SGD(y, tx, initial_w, batch_size, max_iters, gamma)
elif model == "MSE_OPT":
w, _ = least_squares(y, tx)
elif model == "MSE_OPT_REG":
w, _ = ridge_regression(y, tx, lambda_)
elif model == "LOG_GD":
w, _ = logistic_regression(y, tx, initial_w, max_iters, gamma)
elif model == "LOG_REG_GD":
w, _ = reg_logistic_regression(y, tx, lambda_, initial_w, max_iters, gamma)
elif model == "LOG_REG_L1":
w, _ = reg_logistic_regression_L1(y, tx, lambda_, initial_w, max_iters, gamma)
elif model == "MSE_GD_L1":
w, _ = least_squares_GD_L1(y, tx, lambda_, initial_w, max_iters, gamma)
else:
raise UnknownModel
return w
def cross_validation_ls_SGD(y, x, k_indices, k, gamma, max_iters, w_initial, batch_size):
"""train and test least square SGD model using cross validation"""
x_test = x[k_indices[k]]
x_train = np.delete(x, [k_indices[k]], axis=0)
y_test = y[k_indices[k]]
y_train = np.delete(y, [k_indices[k]], axis=0)
opt_w, mse_tr = imp.least_squares_SGD(y_train, x_train, w_initial, batch_size, max_iters, gamma)
mse_te = imp.compute_mse(y_test, x_test,opt_w)
return mse_te, opt_w
def test_least_squares_SGD(y_train, tx_train, y_test, tx_test):
"""
Tests least_squares_SGD method on the splitted data set and
reports percentage of correct predictions.
Args:
y_train: training labels after the splitting
tx_train: training features after the splitting
y_test: test labels after the splitting
tx_test: test features after the splitting
"""
print('\nTesting least_squares_SGD...')
w, _ = least_squares_SGD(y_train, tx_train, np.zeros(tx_train.shape[1]),
3000, 0.005)
report_prediction_accuracy(y_test, tx_test, w)
print('... testing completed.')
# Cross validation over gamma avg_test_accuracy_LR = cross_validation_LR(X_train, y_train, k_fold=4, seed=1) # Cross validation over both gamma and lambda g, l, avg_test_accuracy_RLR = cross_validation_RLR(X_train, y_train, k_fold=4, seed=1) #%% Testing functions #np.random.seed(42) gamma = 0.2 lambda_ = 4E-5 w, loss = least_squares(y = y_train, tx = X_train) # w, loss = least_squares_SGD(y = y_train, tx = X_train, initial_w = np.random.random(size=num_features)*0.01, max_iters = 200000, gamma = gamma) # w, loss = ridge_regression(y = y_train, tx = X_train, lambda_ = lambda_) # w, loss = logistic_regression(y = y_train, tx = X_train, initial_w = np.random.random(size=num_features)*10, max_iters = 125000, gamma = gamma) # w, loss = reg_logistic_regression(y = y_train, tx = X_train, lambda_ = lambda_, initial_w = np.random.random(size=num_features)*0.01, max_iters = 200000, gamma = gamma) plt.plot(w) #%% Predictive step y_test = X_test @ w plt.hist(y_test, bins=200) y_pred = predict_labels(w, X_test)
def cross_validation(x,
y,
k,
mode,
gamma=None,
lambda_=None,
max_iters=None,
initial_w=None):
"""
INPUT:
@x : input data, dimensions (NxD)
@y : target labels, (Nx1) array
@k : number of folds
OUTPUT:
"""
D = x.shape[1]
#randomly permute data maybe?
x_split = np.array_split(x, k, axis=0)
y_split = np.array_split(y, k, axis=0)
#initialize weights and metrics
weights = list()
acc = list()
tpr = list()
fpr = list()
losses = list()
#loop over folds
for fold in range(k):
#create model
#train_ind = [i for i in range(k) if i!=fold]
#val_ind = [i for i in range(k) if i==fold]
#pdb.set_trace()
x_train = [x_split[i] for i in range(k) if i != fold]
y_train = [y_split[i] for i in range(k) if i != fold]
x_train = np.concatenate(x_train, axis=0)
y_train = np.concatenate(y_train, axis=0)
x_val = x_split[fold]
y_val = y_split[fold]
#model = Proj1_Model(x_train, y_train, mode)
#train model for fold
#weights[k] = model.train()
"""here the choice of method"""
if mode == 'linear_regression_eq':
update, loss = imp.least_squares(y_train, x_train)
predictions = np.dot(x_val, update)
pr_bool = predictions >= np.mean(predictions)
elif mode == 'ridge_regression_eq':
update, loss = imp.ridge_regression(y_train, x_train, lambda_)
predictions = np.dot(x_val, update)
pr_bool = predictions >= np.mean(predictions)
elif mode == 'linear_regression_GD':
update, loss = imp.least_squares_GD(y_train, x_train, initial_w,
max_iters, gamma)
predictions = np.dot(x_val, update)
pr_bool = predictions >= np.mean(predictions)
elif mode == 'linear_regression_SGD':
update, loss = imp.least_squares_SGD(y_train, x_train, initial_w,
max_iters, gamma)
predictions = np.dot(x_val, update)
pr_bool = predictions >= np.mean(predictions)
elif mode == 'logistic_regression':
update, loss = imp.logistic_regression(y_train, x_train, initial_w,
max_iters, gamma)
predictions = np.dot(x_val, update)
predicted_prob = H.sigmoid(predictions)
#pdb.set_trace()
pr_bool = predicted_prob > 0.5
elif mode == 'reg_logistic_regression':
update, loss = imp.reg_logistic_regression(y_train, x_train,
initial_w, max_iters,
gamma)
predictions = np.dot(x_val, update)
predicted_prob = H.sigmoid(predictions)
#pdb.set_trace()
pr_bool = predicted_prob > 0.5
weights.append(update)
losses.append(loss)
pr_bool = predictions >= np.mean(predictions)
y_bool = y_val == 1
correct = pr_bool == y_bool
tp = np.logical_and(correct, y_bool)
fp = np.logical_and(np.logical_not(correct), pr_bool)
#tp = [i for i in range(len(pr_bool)) if (pr_bool[i] == True and y_bool[i] == True)]
#all_p = [i for i in range(len(pr_bool)) if y_bool == True]
#fp = [i for i in range(len(pr_bool)) if (pr_bool == True and y_bool == False)]
#all_n = [i for i in range(len(pr_bool)) if y_bool == False]
#print('True signal samples:' + str(sum(y_val)) + ' - Predicted signal samples:' + str(sum(pr_bool)))
acc.append(sum(correct) / float(len(y_val)))
tpr.append(sum(tp) / float(sum(y_bool)))
fpr.append(sum(fp) / float(sum(np.logical_not(y_bool))))
#acc[k] = model.acc()
#tpr[k] = model.tpr()
#fpr[k] = model.fpr()
return acc, tpr, fpr, losses
def cross_validation(y, tX, gamma, method='logistic_regression'):
"""Cross validation for logistic regression
@param gamma: learning rate
@return : the average accuracy over the four fold validations
"""
N, D = tX.shape
# Logistic regression parameters
max_iters = 100
batch_size = N / 100
# Cross validation parameters
seed = 1
k_fold = 4
k_indices = build_k_indices(y, k_fold, seed)
N_fold = N * (k_fold - 1) / k_fold
N_test = N / k_fold
acc = []
for k in range(k_fold):
yTr = np.array([])
xTr = np.zeros((0, D))
for i in range(k_fold):
if i == k:
yTe = y[k_indices[i]]
xTe = tX[k_indices[i]]
else:
yTr = np.append(yTr, y[k_indices[i]], axis=0)
xTr = np.append(xTr, tX[k_indices[i]], axis=0)
initial_w = np.zeros(tX.shape[1])
if method == 'logistic_regression':
initial_w = np.zeros((tX.shape[1], 1))
w, loss = logistic_regression(yTr, xTr, initial_w, max_iters,
gamma)
y_est = sigmoid(np.dot(xTe, w))
y_label = [0 if i < 0.5 else 1 for i in y_est]
elif method == 'reg_logistic_regression':
initial_w = np.zeros((tX.shape[1], 1))
lambda_ = 0.1
w, loss = reg_logistic_regression(yTr, xTr, lambda_, initial_w,
max_iters, gamma)
y_est = sigmoid(np.dot(xTe, w))
y_label = [0 if i < 0.5 else 1 for i in y_est]
elif method == 'least_squares_GD':
w, loss = least_squares_GD(yTr, xTr, initial_w, max_iters, gamma)
y_label = predict_labels(w, xTe)
elif method == 'least_squares_SGD':
w, loss = least_squares_SGD(yTr, xTr, initial_w, max_iters, gamma)
y_label = predict_labels(w, xTe)
elif method == 'least_squares':
w, loss = least_squares(yTr, xTr)
y_label = predict_labels(w, xTe)
elif method == 'ridge_regression':
w, loss = ridge_regression(yTr, xTr, 0.1)
y_label = predict_labels(w, xTe)
else:
raise Exception('Invalid method')
corr = [
True if i == yTe[ind] else False for ind, i in enumerate(y_label)
]
acc.append(sum(corr) / N_test)
# print("Fold: {f}, Accuracy: {acc}, Loss:{loss}".format(f=k, acc=acc[k], loss=loss))
return (sum(acc) / k_fold), acc
import numpy as np
from proj1_helpers import load_csv_data
from implementations import least_squares_SGD
yb, input_data, ids = load_csv_data('all/train.csv', step=50)
losses, w = least_squares_SGD(yb,
input_data,
initial_w=np.zeros(30),
batch_size=1,
max_iters=30,
gamma=0.5)
with open('output.txt', 'w') as fp:
print(yb, input_data, losses, w, file=fp)
initial_w = np.zeros(tx.shape[1])
max_iters = 1000
gamma = 0.01
w_lr, loss_lr = implementations.least_squares(y, tx)
y_pred_lr = tx @ w_lr
print(f"Linear regression eq: {r2_score(y_pred_lr, y)}")
w_lr_gd, loss_lr_gd = implementations.least_squares_GD(y, tx, initial_w,
max_iters, gamma, verbose=False)
y_pred_lr_gd = tx @ w_lr_gd
print(f"Linear regression gd: {r2_score(y_pred_lr_gd, y)}")
w_lr_sgd, loss_lr_sgd = implementations.least_squares_SGD(y, tx, initial_w,
max_iters, gamma, verbose=False)
y_pred_lr_sgd = tx @ w_lr_sgd
print(f"Linear regression sgd: {r2_score(y_pred_lr_sgd, y)}")
reg = LinearRegression().fit(X, y)
y_pred_sk = reg.predict(X)
print(f"Sklearn linear regression: {r2_score(y_pred_sk, y)}")
# Ridge regression
print("Ridge Reression \n -----------------")
lambda_ = 0.01
w_r, loss_r = implementations.ridge_regression(y, tx, lambda_)