def mn_i(weights, name=None):
"""Applies max-norm regularization to weights."""
try: # TF12
with ops.name_scope(scope, 'maxnorm_i_regularizer', [weights]) as name:
my_scale = ops.convert_to_tensor(scale,
dtype=weights.dtype.base_dtype,
name='scale')
return standard_ops.mul(my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 1)), name=scope)
except: # TF11
with ops.op_scope([weights], name, 'maxnorm_i_regularizer') as scope:
my_scale = ops.convert_to_tensor(scale,
dtype=weights.dtype.base_dtype,
name='scale')
return standard_ops.mul(my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 1)), name=scope)
def mn_i(weights, name=None):
"""Applies max-norm regularization to weights."""
with ops.op_scope([weights], name, 'maxnorm_o_regularizer') as scope:
my_scale = ops.convert_to_tensor(scale,
dtype=weights.dtype.base_dtype,
name='scale')
return standard_ops.mul(my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 1)), name=scope)
def mn_o(weights, name=None):
"""Applies max-norm regularization to weights."""
with ops.op_scope([weights], name, 'maxnorm_o_regularizer') as scope:
my_scale = ops.convert_to_tensor(scale,
dtype=weights.dtype.base_dtype,
name='scale')
return standard_ops.mul(my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 0)), name=scope)
def mn(weights, name='max_regularizer'):
"""Applies max-norm regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
# if tf.__version__ <= '0.12':
# standard_ops_fn = standard_ops.mul
# else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(my_scale, standard_ops.reduce_max(standard_ops.abs(weights)), name=scope)
def mn_i(weights, name='maxnorm_i_regularizer'):
"""Applies max-norm regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
if tf.__version__ <= '0.12':
standard_ops_fn = standard_ops.mul
else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 1)), name=scope)
def _project_log_stochastic_matrix_wrt_kl_divergence(log_matrix):
"""Projects its argument onto the set of log-left-stochastic matrices.
Args:
log_matrix: 2d square tensor, the element-wise logarithm of the matrix to
project.
Returns:
The 2d square tensor that results from projecting exp(`matrix`) onto the set
of left-stochastic matrices w.r.t. the KL-divergence applied column-wise.
"""
# For numerical reasons, make sure that the largest matrix element is zero
# before exponentiating.
log_matrix -= standard_ops.reduce_max(log_matrix, axis=0, keepdims=True)
log_matrix -= standard_ops.log(
standard_ops.reduce_sum(
standard_ops.exp(log_matrix), axis=0, keepdims=True))
return log_matrix
def _project_log_stochastic_matrix_wrt_kl_divergence(log_matrix):
"""Projects its argument onto the set of log-left-stochastic matrices.
Args:
log_matrix: 2d square tensor, the element-wise logarithm of the matrix to
project.
Returns:
The 2d square tensor that results from projecting exp(`matrix`) onto the set
of left-stochastic matrices w.r.t. the KL-divergence applied column-wise.
"""
# For numerical reasons, make sure that the largest matrix element is zero
# before exponentiating.
log_matrix -= standard_ops.reduce_max(log_matrix, axis=0, keep_dims=True)
log_matrix -= standard_ops.log(
standard_ops.reduce_sum(
standard_ops.exp(log_matrix), axis=0, keep_dims=True))
return log_matrix