def fit(self, X_train, y_train, mini_batch=100, verbose=10):
old_loss = np.inf
of = 0
for i in range(30):
X_train, y_train = shuffle_data(X_train, y_train)
for b in range(mini_batch, X_train.shape[0] + 1, mini_batch):
x_batch, y_batch = X_train[b -
mini_batch:b], y_train[b -
mini_batch:b]
y_pred = self.predict(x_batch)
self.SGD(y_batch, y_pred, x_batch)
if i % verbose == 0:
loss = self.cost(y_batch, y_pred)
if old_loss - loss >= 0:
of -= 1
self.eta *= 2
else:
of += 1
self.eta /= 3
if of > 5:
self.eta /= 20
if of > 10:
break
#print 'rate: ', self.eta
old_loss = loss
#print 'loss: ', loss
if loss < 0.001:
#print 'enough (loss = 0.01)'
break
def fit(self, X_train, y_train, mini_batch = 100, verbose = 10):
old_loss = np.inf
of = 0
for i in range(30):
X_train, y_train = shuffle_data(X_train, y_train)
for b in range(mini_batch, X_train.shape[0]+1, mini_batch):
x_batch, y_batch = X_train[b-mini_batch:b], y_train[b-mini_batch:b]
y_pred = self.predict(x_batch)
self.SGD(y_batch, y_pred, x_batch)
if i % verbose == 0:
loss = self.cost(y_batch, y_pred)
if old_loss - loss >= 0:
of -= 1
self.eta *= 2
else:
of += 1
self.eta /= 3
if of > 5:
self.eta /= 20
if of > 10:
break
#print 'rate: ', self.eta
old_loss = loss
#print 'loss: ', loss
if loss < 0.001:
#print 'enough (loss = 0.01)'
break
def run_lr_bfgs(X_train, y_train):
dim = X_train.shape[1]
for i in range(10):
lr = LogRegBFGS(dim)
X_train, y_train = shuffle_data(X_train, y_train)
y_true = (y_train == i).astype(int)
lr.fit(X_train, y_true)
y_true_test = (y_test == i).astype(int)
print('test accuracy: ', lr.score(X_test, y_true_test))
def run_lr_bfgs(X_train, y_train):
dim = X_train.shape[1]
for i in range(10):
lr = LogRegBFGS(dim)
X_train, y_train = shuffle_data(X_train, y_train)
y_true = (y_train == i).astype(int)
lr.fit(X_train, y_true)
y_true_test = (y_test == i).astype(int)
print('test accuracy: ', lr.score(X_test, y_true_test))
from time import clock
from Preprocessing import shuffle_data, zscore
from DataLoader import load_mnist
from sklearn.preprocessing import MinMaxScaler, StandardScaler
from scipy.special import expit as sigmoid
from scipy.optimize import fmin_l_bfgs_b
from matplotlib import pyplot as plt
from numba import jit
#%matplotlib inline
X_train, y_train, X_test, y_test = load_mnist()
X_train = zscore(X_train)
X_test = zscore(X_test)
X_train, y_train = shuffle_data(X_train, y_train)
n_classes = len(set(y_train))
def accuracy_score(y_true, y_pred):
return (y_true == y_pred).mean()
class LogisticRegression():
def __init__(self, dim):
self.w = np.zeros(dim)
self.eta = 10e-7
def predict(self, x):
return sigmoid(x.dot(self.w))
from Preprocessing import shuffle_data, zscore
from DataLoader import load_mnist
from sklearn.preprocessing import MinMaxScaler, StandardScaler
from scipy.special import expit as sigmoid
from scipy.optimize import fmin_l_bfgs_b
from matplotlib import pyplot as plt
from numba import jit
#%matplotlib inline
X_train, y_train, X_test, y_test = load_mnist()
X_train = zscore(X_train)
X_test = zscore(X_test)
X_train, y_train = shuffle_data(X_train, y_train)
n_classes = len(set(y_train))
def accuracy_score(y_true, y_pred):
return (y_true == y_pred).mean()
class LogisticRegression():
def __init__(self, dim):
self.w = np.zeros(dim)
self.eta = 10e-7
def predict(self, x):
return sigmoid(x.dot(self.w))
