def __init__(self,
n_classes,
optimizer=None,
n_filters=96,
keep_prob=0.5,
learning_rate=None,
model_dir=None,
config=None,
warm_start_from=None,
multi_gpu=False,
n_examples=1.0,
prior_path=None):
params = {
'n_classes': n_classes,
# If an instance of an optimizer is passed in, this will just
# return it.
'optimizer': (
None if optimizer is None
else get_optimizer_instance(optimizer, learning_rate)),
'n_filters': n_filters,
'n_examples': n_examples,
'prior_path': prior_path
}
_model_fn = model_fn
if multi_gpu:
params['optimizer'] = TowerOptimizer(params['optimizer'])
_model_fn = replicate_model_fn(_model_fn)
super(MeshNetBWN, self).__init__(
model_fn=_model_fn, model_dir=model_dir, params=params,
config=config, warm_start_from=warm_start_from)
def __init__(self,
n_classes,
optimizer=None,
learning_rate=None,
model_dir=None,
config=None,
warm_start_from=None,
multi_gpu=False):
params = {
'n_classes': n_classes,
# If an instance of an optimizer is passed in, this will just
# return it.
'optimizer': (
None if optimizer is None
else get_optimizer_instance(optimizer, learning_rate)),
}
_model_fn = model_fn
if multi_gpu:
params['optimizer'] = TowerOptimizer(params['optimizer'])
_model_fn = replicate_model_fn(_model_fn)
super(HighRes3DNet, self).__init__(
model_fn=_model_fn, model_dir=model_dir, params=params,
config=config, warm_start_from=warm_start_from,
)
def __init__(self,
n_classes,
optimizer,
n_filters=64,
n_examples=1.0,
n_prior_samples=1.0,
learning_rate=None,
model_dir=None,
config=None,
warm_start_from=None,
prior_path=None,
multi_gpu=False,
only_kld=False,
is_mc='True'):
print('Learning Rate: ' + str(learning_rate))
params = {
'n_classes': n_classes,
# If an instance of an optimizer is passed in, this will just
# return it.
'optimizer': get_optimizer_instance(optimizer, learning_rate),
'n_filters': n_filters,
'n_examples': n_examples,
'prior_path': prior_path,
'n_prior_samples': n_prior_samples,
'only_kld': only_kld,
'is_mc': is_mc
}
_model_fn = model_fn
if multi_gpu:
params['optimizer'] = TowerOptimizer(params['optimizer'])
_model_fn = replicate_model_fn(_model_fn)
super(MeshNetCWN, self).__init__(model_fn=_model_fn,
model_dir=model_dir,
params=params,
config=config,
warm_start_from=warm_start_from)
def _create_train_op(loss, learning_rate):
optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate,
decay=0.9)
optimizer = TowerOptimizer(optimizer)
return slim.learning.create_train_op(
loss, optimizer, global_step=tf.train.get_or_create_global_step())