def init_lambda_network(self):
assert hasattr(self, 'lambda_network')
if self.use_cuda:
use_cuda(True, self._opt['device_id'])
self.lambda_network.cuda()
self.optimizer = use_optimizer(self.lambda_network,
self.lambda_network_opt)
self.scheduler = ExponentialLR(
self.optimizer, gamma=self.lambda_network_opt['lr_exp_decay'])
def init_episode(self):
opt = self.opt
self.model = MF(opt)
self._train_step_idx = 0
if self.use_cuda:
use_cuda(True, opt['device_id'])
self.model.cuda()
self.optimizer = use_optimizer(self.model, opt)
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt['lr_exp_decay'])
self.param = [p.data.clone() for p in self.model.parameters()]
def __init__(self, opt):
super(MFBPRFactorizer, self).__init__(opt)
self.model = MF(opt)
self.opt = opt
if self.use_cuda:
use_cuda(True, opt['device_id'])
self.model.cuda()
self.optimizer = use_optimizer(self.model, opt)
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt['lr_exp_decay'])
def copy(self, new_factorizer):
"""Return a new copy of factorizer
# Note: directly using deepcopy wont copy factorizer.scheduler correctly
the gradient of self.model is not copied!
"""
self.train_step_idx = new_factorizer.train_step_idx
self.param = new_factorizer.param
self.model.load_state_dict(new_factorizer.model.state_dict())
self.optimizer = use_optimizer(self.model, self.opt)
self.optimizer.load_state_dict(new_factorizer.optimizer.state_dict())
self.scheduler = ExponentialLR(self.optimizer,
gamma=self.opt['lr_exp_decay'],
last_epoch=self.scheduler.last_epoch)
def init_episode(self):
opt = self.opt
if opt['model'] == 'linear':
self.model = LR(opt)
elif opt['model'] == 'fm':
self.model = FM(opt)
elif opt['model'] == 'deepfm':
self.model = DeepFM(opt)
elif opt['model'] == 'autoint':
self.model = AutoInt(opt)
else:
raise ValueError("Invalid FM model type: {}".format(opt['model']))
self._train_step_idx = 0
if self.use_cuda:
use_cuda(True, opt['device_id'])
self.model.cuda()
self.optimizer = use_optimizer(self.model, opt)
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt['lr_exp_decay'])
def __init__(self, opt):
super(FMFactorizer, self).__init__(opt)
self.opt = opt
if opt['model'] == 'linear':
self.model = LR(opt)
elif opt['model'] == 'fm':
self.model = FM(opt)
elif opt['model'] == 'deepfm':
self.model = DeepFM(opt)
elif opt['model'] == 'autoint':
self.model = AutoInt(opt)
else:
raise ValueError("Invalid FM model type: {}".format(opt['model']))
if self.use_cuda:
use_cuda(True, opt['device_id'])
self.model.cuda()
self.optimizer = use_optimizer(self.model, opt)
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt['lr_exp_decay'])
