def __init__(self, model, **kwargs):
super(TaggerTrainerTf, self).__init__()
self.loss = model.create_loss()
self.model = model
span_type = kwargs.get('span_type', 'iob')
self.evaluator = TaggerEvaluatorTf(model, span_type)
self.global_step, self.train_op = optimizer(self.loss, **kwargs)
def __init__(self, model, **kwargs):
super(Seq2SeqTrainerTf, self).__init__()
self.sess = model.sess
self.loss = model.create_loss()
self.test_loss = model.create_test_loss()
self.model = model
self.global_step, self.train_op = optimizer(self.loss, **kwargs)
def __init__(self, model, **kwargs):
super(ClassifyTrainerTf, self).__init__()
self.sess = model.sess
self.loss = model.create_loss()
self.test_loss = model.create_test_loss()
self.model = model
self.global_step, self.train_op = optimizer(
self.loss, colocate_gradients_with_ops=True, **kwargs)
def __init__(self, model, **kwargs):
super(ClassifyTrainerTf, self).__init__()
self.sess = model.sess
self.loss = model.create_loss()
self.test_loss = model.create_test_loss()
self.model = model
self.global_step, train_op = optimizer(
self.loss, colocate_gradients_with_ops=True, **kwargs)
decay = kwargs.get('ema_decay', None)
if decay is not None:
self.ema = True
ema_op, self.ema_load, self.ema_restore = _add_ema(
model, float(decay))
with tf.control_dependencies([ema_op]):
self.train_op = tf.identity(train_op)
else:
self.ema = False
self.train_op = train_op
def __init__(self, model, **kwargs):
super(TaggerTrainerTf, self).__init__()
self.loss = model.create_loss()
self.model = model
self.evaluator = TaggerEvaluatorTf(model)
self.global_step, self.train_op = optimizer(self.loss, **kwargs)
