def _batch_norm(self, name, x, add_ops=True):
with tf.variable_scope(name):
n_out = x.get_shape()[-1]
try:
n_out = int(n_out)
shape = [n_out]
except:
shape = None
beta = self._weight_variable(shape,
init_method="constant",
init_param={"val": 0.0},
name="beta",
dtype=self.dtype)
gamma = self._weight_variable(shape,
init_method="constant",
init_param={"val": 1.},
name="gamma",
dtype=self.dtype)
normed, ops = batch_norm(x,
self.is_training,
gamma=gamma,
beta=beta,
axes=[0, 1, 2],
eps=1e-3,
name="bn_out")
if add_ops:
if ops is not None:
self._bn_update_ops.extend(ops)
return normed
def _batch_norm(self, name, x):
"""
Batch normalization.
Only active if specified in config
"""
if self.config.batch_norm:
with tf.variable_scope(name):
n_out = x.get_shape()[-1]
beta = nn.weight_variable([n_out],
init_method="constant",
init_param={"val": 0.0},
name="beta")
gamma = nn.weight_variable([n_out],
init_method="constant",
init_param={"val": 1.0},
name="gamma")
return nn.batch_norm(x,
self.is_training,
gamma=gamma,
beta=beta,
axes=[0, 1, 2],
eps=1e-3,
scope="bn",
name="bn_out")
else:
return x