def __init__(self,
units,
activation=None,
use_bias=False,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=unit_norm(),
bias_constraint=None,
**kwargs):
"""
Initialize like Dense.
*****************************
"""
# explicit call to parent constructor
Dense.__init__(self,
units,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
kernel_constraint=kernel_constraint,
bias_constraint=bias_constraint,
**kwargs)
def __init__(self,
alpha=1,
**kwargs):
# Weight decay (regularizer_l2) is important because we want to keep the trace of the W matrix
# (the matrix of the weight of this layer, i.e. the label-flip confusion matrix) low.
# This because in the paper is proven that the tr(Q*) <= tr(Q), where Q* is the Q that best represents
# the label-flip noise.
Dense.__init__(self, output_dim=-1, bias=False, trainable=False, init='identity', **kwargs)
if alpha >= 1:
raise ValueError("OutlierNoise Layer: alpha must be < 1 "
"(theorically alpha = (outlaiers labelled as outlaier i.e. class K+1) / (total outlaiers)")
self.alpha = alpha
def __init__(self, alpha=1, **kwargs):
Dense.__init__(self,
output_dim=-1,
bias=False,
trainable=False,
init='identity',
**kwargs)
if alpha >= 1:
raise ValueError(
"OutlierNoise Layer: alpha must be < 1 "
"(theorically alpha = (outlaiers labelled as outlaier i.e. class K+1) / (total outlaiers)"
)
self.alpha = alpha
def __init__(self,
weight_decay=0.1,
W_learning_rate_multiplier=None,
b_learning_rate_multiplier=None,
**kwargs):
# Weight decay (regularizer_l2) is important because we want to keep the trace of the W matrix
# (the matrix of the weight of this layer, i.e. the label-flip confusion matrix) low.
# This because in the paper is proven that the tr(Q*) <= tr(Q), where Q* is the Q that best represents
# the label-flip noise.
Dense.__init__(self, output_dim=-1, bias=False, b_learning_rate_multiplier=None,
# W_learning_rate_multiplier=W_learning_rate_multiplier,
W_regularizer=l2(weight_decay),
W_constraint=stochastic2(),
init='identity',
**kwargs)
def __init__(
self,
task_features,
use_task_bias,
use_task_gain,
units,
activation=None,
use_bias=False, # default false, since can be achieved by task-specific gains
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
self.current_task_bias = None
self.current_task = None
self.current_task_gain = None
if 'input_shape' not in kwargs and 'input_dim' in kwargs:
kwargs['input_shape'] = (kwargs.pop('input_dim'), )
Dense.__init__(self,
units,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
kernel_constraint=kernel_constraint,
bias_constraint=bias_constraint,
**kwargs)
self.task_features = task_features
self.use_task_bias = use_task_bias
self.use_task_gain = use_task_gain
def __init__(self, units, prob=0.5, drop_bias=False, drop_noise_shape=None, **kwargs):
DropConnect.__init__(self, prob=prob, drop_bias=drop_bias, drop_noise_shape=drop_noise_shape)
Dense.__init__(self, units, **kwargs)
if self.needs_drop:
self.uses_learning_phase = True
