def pretrain_begin(self, begin, end):
TrainFlow.clear(self)
TrainFlow.start_training(self, begin, end)
Sampler.pretrain_begin(self, begin, end)
self._sequence.extend([None] * self.pretrain_size)
def __init__(self,
ds,
pretrain_size=10,
embdim=16,
beta=None,
lr=0.1,
batchsize=None,
sampling_args=None):
if beta is None:
beta = [0.1, 0.1]
if sampling_args is None:
sampling_args = {}
self.__dataset = ds
TrainFlow.__init__(self,
embdim=embdim,
beta=beta,
trainmod=self.name,
datasetmod=ds.name)
Sampler.__init__(self, **sampling_args)
self.pretrain_size = pretrain_size
self.lr = lr
self.batchsize = batchsize
self.__pretrain = None
self.__online = None
def online_begin(self, begin, end):
TrainFlow.start_training(self, begin, end)
Sampler.online_begin(self, begin, end)
initv = np.random.uniform(
0, 1,
(self.dataset.nsize, self.flowargs['embdim'])).astype('float32')
K.set_value(self.online['vars'][0], initv)
self._sequence.append(None)
def pretrain_begin_iteration(self):
Sampler.pretrain_begin_iteration(self)
if self.cur_train_end > self.init_train_begin + self.pretrain_size: # online phase
return
# compute EM coefficients here
neg1_int, neg1_float = self.__emcoef(self._neg[1])
self._neg = self._neg[:1] + [neg1_int, neg1_float]
def make_online_input(self, batch):
ret = Sampler.make_online_input(self, batch)
for d in ret[0]: # data
d[0] -= self.init_train_begin
ret.append(self._sequence[self.cur_train_begin - 1])
return ret
def make_pretrain_input(self, batch):
ret = Sampler.make_pretrain_input(self, batch)
# (data, weight, triad)
# because embedding variable starts from index 0
for d in ret[0]: # data
d[0] -= self.init_train_begin
for d in ret[2]:
d[0] -= self.init_train_begin
return ret
def online_end(self):
Sampler.online_end(self)
assert self.cur_train_end == self.cur_train_begin + 1, "{} {}".format(
self.cur_train_end, self.cur_train_begin)
self.save_model()
TrainFlow.stop_training(self)
def pretrain_end(self):
Sampler.pretrain_end(self)
TrainFlow.stop_training(self)
def pretrain_end_iteration(self):
Sampler.pretrain_end_iteration(self)
self.save_model()