def adagrad_momentum(grads,
params,
learning_rate=1.0,
momentum=0.9,
epsilon=1e-06):
return apply_nesterov_momentum(adagrad(grads, params, learning_rate,
epsilon),
params=params,
momentum=momentum)
def adadelta_momentum(grads,
params,
learning_rate=1.0,
momentum=0.9,
rho=0.95,
epsilon=1e-06):
return apply_nesterov_momentum(adadelta(grads, params, learning_rate, rho,
epsilon),
params=params,
momentum=momentum)
def _create_functions(self):
self._update_func = self._update(self._get_loss_function(),
self._get_all_trainable_params(),
self._learning_rate)
if self._momentum != 0:
self._update_func = apply_nesterov_momentum(
self._update_func, self._get_all_trainable_params(),
self._momentum)
self._loss_func = self._get_loss_function()
self._train_function = theano.function([self.t_input, self.t_label],
self._get_loss_function(),
updates=self._update_func)
def rmsprop_nesterov(loss, params, eta=1e-3, alpha=0.9, **kwargs):
rms = updt.rmsprop(loss, params, learning_rate=eta, **kwargs)
return updt.apply_nesterov_momentum(rms, params, momentum=alpha)
l5=Conv2DLayer(l4,96,(5,5),nonlinearity=very_leaky_rectify,W=GlorotUniform('relu'))
l6=MaxPool2DLayer(l5,(3,3))
l7=DenseLayer(l6,512,nonlinearity=very_leaky_rectify,W=lasagne.init.GlorotNormal())
#l7_5=cyclicpool(l7)
#l7_5=lasagne.layers.DropoutLayer(l7)
l8=DenseLayer(l7,2,nonlinearity=softmax)
rate=theano.shared(.0002)
params = lasagne.layers.get_all_params(l8)
prediction = lasagne.layers.get_output(l8)
loss = lasagne.objectives.categorical_crossentropy(prediction,y1)
loss = loss.mean()
updates_sgd = adagrad(loss, params, learning_rate=rate)
updates = apply_nesterov_momentum(updates_sgd, params, momentum=0.9)
train_model = theano.function([x1,y1],outputs=loss,updates=updates)
pred = theano.function([x1,y1],outputs=lasagne.objectives.categorical_crossentropy(prediction,y1))
# pred=theano.function([x1,y1],outputs=prediction,on_unused_input='ignore')
### begin to train
renewtrain=len(train_x)/batchsize
renewtest=len(test_x)/batchsize
for i in range(15000):
if i>325 and i<3000:
rate.set_value(.001)
elif i>6500 and i<15000:
def _create_functions(self):
self._update_func = self._update(self._get_loss_function(), self._get_all_trainable_params(), self._learning_rate)
if self._momentum != 0:
self._update_func = apply_nesterov_momentum(self._update_func, self._get_all_trainable_params(), self._momentum)
self._loss_func = self._get_loss_function()
self._train_function = theano.function([self.t_input, self.t_label], self._get_loss_function(), updates=self._update_func)
