def optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=args.init_lr,
trainable=False)
opt = tf.train.GradientDescentOptimizer(lr)
return optimizer.apply_grad_processors(opt,
[gradproc.GlobalNormClip(5)])
def _get_optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=1e-3,
trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [gradproc.GlobalNormClip(10),
gradproc.SummaryGradient()])
def _get_optimizer(self):
conf = Config()
lr = tf.get_variable('learning_rate',
initializer=conf.learning_rate,
trainable=False)
opt = tf.train.AdamOptimizer(lr)
tf.summary.scalar('learning_rate', lr)
return optimizer.apply_grad_processors(
opt, [gradproc.GlobalNormClip(conf.max_grad_norm)])
def _get_optimizer(self):
gradprocs = [
FilterGradientVariables('.*net2.*', verbose=False),
gradproc.MapGradient(
lambda grad: tf.clip_by_value(grad, hp.train2.clip_value_min, hp.train2.clip_value_max)),
gradproc.GlobalNormClip(hp.train2.clip_norm),
# gradproc.PrintGradient(),
# gradproc.CheckGradient(),
]
lr = tf.get_variable('learning_rate', initializer=hp.train2.lr, trainable=False)
opt = tf.train.AdamOptimizer(learning_rate=lr)
return optimizer.apply_grad_processors(opt, gradprocs)
def _get_optimizer(self):
gradprocs = [
tensorpack_extension.FilterGradientVariables('.*net2.*', verbose=False),
gradproc.MapGradient(
lambda grad: tf.clip_by_value(grad, hp.train2.clip_value_min, hp.train2.clip_value_max)),
gradproc.GlobalNormClip(hp.train2.clip_norm),
# gradproc.PrintGradient(),
# gradproc.CheckGradient(),
]
global_step = tf.Variable(0, name='global_step',trainable=False)
#self.lr = self.learning_rate_decay(global_step, hp.train2.lr)
#lr = learning_rate_decay(initial_lr = hp.train2.lr, global_step)
lr = tf.get_variable('learning_rate', initializer=hp.train2.lr, trainable=False)
opt = tf.train.AdamOptimizer(learning_rate=lr)
return optimizer.apply_grad_processors(opt, gradprocs)
def _get_optimizer(self):
lr = symbf.get_scalar_var('learning_rate', 1e-3, summary=True)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [gradproc.GlobalNormClip(10),
gradproc.SummaryGradient()])