def _get_optimizer(self):
lr = tf.get_variable('learning_rate',
initializer=0.001,
trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
gradprocs = [
MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.1)),
SummaryGradient()
]
opt = optimizer.apply_grad_processors(opt, gradprocs)
return opt
def optimizer(self, scope_name="master"):
lr = tf.get_variable('learning_rate',
initializer=0.001,
trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
gradprocs = [
MapGradient(lambda grad: tf.clip_by_norm(
grad, 0.1 * tf.cast(tf.size(grad), tf.float32))),
SummaryGradient()
]
opt = optimizer.apply_grad_processors(opt, gradprocs)
return opt
def _get_opt(name, init_lr):
lr = symbf.get_scalar_var('learning_rate/' + name,
init_lr,
summary=True)
opt = tf.train.AdamOptimizer(lr)
logger.info("create opt {}".format(name))
gradprocs = [
# MapGradient(lambda grad: tf.Print(grad, [grad], 'grad {}='.format(grad.op.name), summarize=4)),
MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.1),
regex='^actor/.*'),
MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.05),
regex='^critic/.*'),
# GlobalNormClip(40.),
SummaryGradient(),
]
opt = optimizer.apply_grad_processors(opt, gradprocs)
return opt
def _get_optimizer(self, name):
from tensorpack.tfutils import optimizer
from tensorpack.tfutils.gradproc import SummaryGradient, GlobalNormClip, MapGradient
init_lr = INIT_LEARNING_RATE_A if name == 'actor' else INIT_LEARNING_RATE_C
import tensorpack.tfutils.symbolic_functions as symbf
lr = symbf.get_scalar_var('learning_rate/' + name,
init_lr,
summary=True)
opt = tf.train.AdamOptimizer(lr)
logger.info("create opt {}".format(name))
gradprocs = [
# MapGradient(lambda grad: tf.Print(grad, [grad], 'grad {}='.format(grad.op.name), summarize=4)),
MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.1),
regex='^actor/.*'),
MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.05),
regex='^critic/.*'),
# GlobalNormClip(40.),
SummaryGradient(),
]
opt = optimizer.apply_grad_processors(opt, gradprocs)
return opt
def _get_optimizer(self):
lr = tf.get_variable('learning_rate', initializer=5e-3, trainable=False)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [GlobalNormClip(5), SummaryGradient()])
def get_gradient_processor(self):
return [MapGradient(lambda grad: \
tf.clip_by_global_norm([grad], 5)[0][0]),
SummaryGradient()]
def get_gradient_processor(self):
return [MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.1)),
SummaryGradient()]
def optimizer(self):
lr = tf.get_variable("learning_rate", initializer=self.lr, trainable=False)
opt = tf.train.AdamOptimizer(lr)
return optimizer.apply_grad_processors(
opt, [GlobalNormClip(5), SummaryGradient()]
)