class ModelFrame:
def __init__(self):
self.tparams = OrderedDict()
self.flag_training = Flag(False)
def build(self, options):
raise NotImplementedError
def load_value(self, name, value):
self.tparams[name].set_value(value)
def get_params(self):
'''
get value from tparams to an OrderedDict for saving
'''
params = OrderedDict()
for name, value in self.tparams.iteritems():
params[name] = value.get_value()
return params
def update_tparams(self, params):
'''
set values for tparams using params
'''
for name, value in self.tparams.iteritems():
if not name in params:
raise Warning('param %s not found' % (name))
self.tparams[name].set_value(params[name])
def predict(self, *inputs):
return self.f_pred(*inputs)
def generate(self, inputs, output, pred, cost, optimizer):
if not isinstance(inputs, list):
inputs = [
inputs,
]
outputs = [
output,
] # for possible extension
self.f_pred = theano.function(inputs, pred, name='f_pred_proba')
#self.f_pred_proba = theano.function(inputs, pred, name='f_pred_proba')
#self.f_pred = theano.function(inputs, pred.argmax(axis=1), name='f_pred')
self.f_cost = theano.function(inputs + outputs, cost, name='f_cost')
grads = T.grad(cost, wrt=self.tparams.values())
self.learning_rate = T.scalar(name='lr')
self.f_grad_shared, self.f_update = optimizer.build(
self.learning_rate, self.tparams, grads, inputs, output, cost)
def train(self, inputs, output, learning_rate):
if isinstance(inputs, tuple):
inputs = list(inputs)
if not isinstance(inputs, list):
inputs = [
inputs,
]
outputs = [
output,
] # for possible extension
self.flag_training.on()
cost = self.f_grad_shared(*(inputs + outputs))
self.f_update(learning_rate)
self.flag_training.off()
return cost
