def weights(self):
'''Get model parameters
Returns:
a numpy array of model parameters
'''
return getParameters(self.net)
def params(self):
'''Get model parameter vector
Returns:
params: A vector of parameters
'''
return param.getParameters(self)
def permuteNet(self, goodIdx, badIdx):
goodNet = self.model.net.net[goodIdx]
badNet = self.model.net.net[badIdx]
goodParams = getParameters(goodNet)
noise = self.config.PERMVAL * np.random.randn(len(goodParams))
goodParams = np.array(goodParams) + noise
setParameters(badNet, goodParams)
def __init__(self, ann, config):
self.saver = save.Saver(config.MODELDIR,
'models',
'bests',
resetTol=256)
self.config = config
print('Initializing new model...')
self.net = ann(config)
self.parameters = Parameter(
torch.Tensor(np.array(getParameters(self.net))))
def __init__(self, ann, config):
self.saver = save.Saver(config.MODELDIR,
'models',
'bests',
resetTol=256)
self.config = config
print('Initializing new model...')
self.net = ann(config)
self.parameters = Parameter(
torch.Tensor(np.array(getParameters(self.net))))
#Have been experimenting with population based
#training. Nothing stable yet -- advise avoiding
if config.POPOPT:
self.opt = PopulationOptimizer(self, config)
else:
self.opt = GradientOptimizer(self, config)
if config.LOAD or config.BEST:
self.load(self.opt, config.BEST)
def initModel(self):
return getParameters(trinity.ANN(self.config))
def params(self):
return param.getParameters(self)
def model(self):
return [getParameters(ann) for ann in self.anns], [getParameters(self.lawmaker)]
def initModel(self):
return getParameters(ANN(self.config, self.args))
