def test_xor_tanh(self):
nnet = Network(
Perceptron(2, 2, activation='tanh'),
Perceptron(2, 1, activation='tanh')
)
teacher = MinibatchSGDTeacher(nnet, test_set, batch_size=1, learning_rate=0.6)
e = Evaluator(nnet)
for _ in xrange(0, 200):
teacher.train_epoch()
self.assertEquals(map(lambda x: round(x[0]), e.evaluate(test_set[0].get_value())), [0, 1, 1, 0])
def test_xor_relu(self):
nnet = Network(
Perceptron(2, 2, activation='relu'),
Perceptron(2, 2, activation='softmax')
)
teacher = MinibatchSGDTeacher(nnet, test_set, batch_size=1, learning_rate=0.05)
e = Evaluator(nnet)
c = Classifier(nnet)
for _ in xrange(0, 400):
teacher.train_epoch()
self.assertTrue(all(p == q for p, q in zip(c.classify(test_set[0].get_value()), [0,1,1,0])))
data = pickle.load(open('../mnist.pkl', 'rb'))
data = [(
theano.shared(np.asarray(das[0], dtype=theano.config.floatX), borrow=True),
theano.shared(np.asarray(das[1], dtype=np.int32), borrow=True)
) for das in data]
nnet = Network(
PerceptronLayer(28*28, 500, activation='relu', dropout=0.5),
PerceptronLayer(500, 300, activation='relu', dropout=0.5),
PerceptronLayer(300, 100, activation='relu', dropout=0.5),
PerceptronLayer(100, 30, activation='relu', dropout=0.5),
PerceptronLayer(30, 10, activation='softmax')
)
teacher = MinibatchSGDTeacher(nnet, data[0], batch_size=100, learning_rate=0.1)
validator = Validator(nnet, data[1])
test_validator = Validator(nnet, data[2])
for i in xrange(0, 100):
error = validator.validate() * 100
print('Accuracy (validation) is %s%%' % (error, ))
# teacher.learning_rate = 1 / (i+2) - uncommen to adapt learning rate
teacher.train_epoch()
test_error = 100 - (test_validator.validate() * 100)
print('Model error rate is %s%%' % (test_error, ))
from nnetsys.ff import Network, Perceptron, MinibatchSGDTeacher, Validator, Classifier
data = pickle.load(open('mnist.pkl', 'rb'))
data = [(
theano.shared(np.asarray(das[0], dtype=theano.config.floatX), borrow=True),
theano.shared(np.asarray(das[1], dtype=np.int32), borrow=True)
) for das in data]
nnet = Network(
Perceptron(28*28, 200, activation='tanh'),
Perceptron(200, 30, activation='tanh'),
Perceptron(30, 10, activation='softmax')
)
teacher = MinibatchSGDTeacher(nnet, data[0], batch_size=20, learning_rate=0.5, l2=0.0001)
validator = Validator(nnet, data[1])
classifier = Classifier(nnet)
#while error > 3: # learn!
for i in xrange(0, 100):
if (i % 10) == 0:
error = validator.validate() * 100
print('Accuracy is %s%%' % (error, ))
teacher.learning_rate = 1 / (i+2)
teacher.train_epoch()