n, dim = x.shape
if not L:
L = np.eye(dim)
Lx = np.dot(x, L)
Lpredx = np.dot(predx, L)
print 'Predicting ...'
predict = []
for ind, ele in enumerate(Lpredx):
dist = ((ele - Lx)**2).sum(1)
pred = y[dist.argsort()[1:K+1]]
bincount = np.bincount(pred)
maxcount = bincount.max()
candidates = [predy for predy in pred if bincount[predy] == maxcount]
predict.append( np.random.choice(candidates) )
return np.array(predict)
if __name__ == '__main__':
train_data, test_data = load_mnist(percentage=0.1, skip_valid=True)
train_x, train_y = train_data
test_x, test_y = test_data
predict_y = KNN_predict(train_x, train_y, test_x, K=5)
correct = (predict_y==test_y).sum()
print 'acc = {}/{} = {}%'.format(correct,
predict_y.shape[0],
float(correct)/predict_y.shape[0]*100
)
# ------------------------------------------------------------------------------
# TRAINING
learning_rate = 0.001 # my experience shows that larger learning rates lead to divergence
# number of training epochs, i.e., passes over training set.
# there are 50.000 samples in training set. 200 epochs means training over 10.000.000 samples. In the paper, they
# keep training for much longer (around 100.000.000)
epoch_count = 10
# report log_likelihood bound on training after this many training samples
report_interval = 20000
# load the training data
(tx, ty), (vx, vy), (_, _) = load_mnist(path='../datasets')
train_n = tx.shape[0]
val_n = vx.shape[0]
# we load the data into shared variables, this is recommended if you do training on GPU.
train_x = theano.shared(np.asarray(tx, dtype=np.float32))
val_x = theano.shared(np.asarray(vx, dtype=np.float32))
# index of the first sample in batch
batch_start = T.lscalar()
# Theano function for training
# This function feeds the current batch to model and applies simple gradient descent updates on model parameters
train_model = theano.function(
[batch_start],
ll_bound,
updates=((w1, w1 - learning_rate * dw1), (b1,
Adventures in Deep Learning
Multilayer Perceptron for MNIST Handwritten Digit Recognition
Goker Erdogan
https://github.com/gokererdogan
"""
import numpy as np
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.utils.np_utils import to_categorical
from common import load_mnist
# load data
(x_train, y_train), (x_val, y_val), (x_test, y_test) = load_mnist()
y_train = to_categorical(y_train)
y_val = to_categorical(y_val)
# build model
model = Sequential()
model.add(Dense(input_dim=784, output_dim=400))
model.add(Activation("sigmoid"))
model.add(Dense(input_dim=400, output_dim=10))
model.add(Activation("softmax"))
model.compile(optimizer="sgd", loss="categorical_crossentropy")
# fit model
model.fit(x=x_train, y=y_train, batch_size=128, nb_epoch=5, verbose=1,
validation_data=(x_val, y_val))
Multilayer Perceptron for MNIST Handwritten Digit Recognition
07 July 2018
Balavivek Sivanantham
https://github.com/bsivanantham
"""
import numpy as np
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.utils.np_utils import to_categorical
from common import load_mnist
# load data
(x_train, y_train), (x_val, y_val), (x_test, y_test) = load_mnist()
y_train = to_categorical(y_train)
y_val = to_categorical(y_val)
# build model
model = Sequential()
model.add(Dense(input_dim=784, output_dim=400))
model.add(Activation("sigmoid"))
model.add(Dense(input_dim=400, output_dim=10))
model.add(Activation("softmax"))
model.compile(optimizer="sgd", loss="categorical_crossentropy")
# fit model
model.fit(x=x_train,
y=y_train,
active_set = new_active_set
violate_set = new_violate_set
gradient = new_gradient
self.M = newM
t += 1
return self
if __name__ == '__main__':
# iris_data = load_iris()
# X = iris_data['data']
# Y = iris_data['target']
# trainx, testx, trainy, testy = train_test_split(X, Y, test_size=0.66)
(trainx, trainy), (testx, testy) = load_mnist(percentage=0.01, skip_valid=True)
pca = PCA(whiten=True)
pca.fit(trainx)
components, variance = 0, 0.0
for components, ele in enumerate(pca.explained_variance_ratio_):
variance += ele
if variance > 0.90: break
components += 1
print 'n_components=%d'%components
pca.set_params(n_components=components)
pca.fit(trainx)
trainx = pca.transform(trainx)
testx = pca.transform(testx)
import numpy as np
from common import load_mnist
from knn import KNN_predict
from sklearn.decomposition import PCA
from sklearn.neighbors import KNeighborsClassifier, DistanceMetric
from metric_learn import LMNN
def knn(train_x, train_y, test_x, test_y):
neigh = KNeighborsClassifier(n_neighbors=5)
neigh.fit(train_x, train_y)
acc = (neigh.predict(test_x) == test_y).sum()
return float(acc))/test_y.shape[0]
if __name__ == '__main__':
(train_x, train_y), (test_x, test_y) = load_mnist(percentage=0.01, skip_valid=True)
pca = PCA(whiten=True)
pca.fit(train_x)
components, variance = 0, 0.0
for components, ele in enumerate(pca.explained_variance_ratio_):
variance += ele
if variance > 0.90: break
components += 1
print 'n_components=%d'%components
pca.set_params(n_components=components)
pca.fit(train_x)
train_x = pca.transform(train_x)
test_x = pca.transform(test_x)
