def __call__(self, x):
regularization = 0.
if self.l1:
regularization += K.sum(self.l1 * K.abs(x))
if self.l2:
regularization += K.sum(self.l2 * K.square(x))
return regularization
def get_gradients(self, loss, params):
"""Returns gradients of `loss` with respect to `params`.
Arguments:
loss: Loss tensor.
params: List of variables.
Returns:
List of gradient tensors.
Raises:
ValueError: In case any gradient cannot be computed (e.g. if gradient
function not implemented).
"""
grads = K.gradients(loss, params)
if None in grads:
raise ValueError('An operation has `None` for gradient. '
'Please make sure that all of your ops have a '
'gradient defined (i.e. are differentiable). '
'Common ops without gradient: '
'K.argmax, K.round, K.eval.')
if hasattr(self, 'clipnorm') and self.clipnorm > 0:
norm = K.sqrt(sum([K.sum(K.square(g)) for g in grads]))
grads = [clip_norm(g, self.clipnorm, norm) for g in grads]
if hasattr(self, 'clipvalue') and self.clipvalue > 0:
grads = [K.clip(g, -self.clipvalue, self.clipvalue) for g in grads]
return grads
def __call__(self, w):
norms = K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True))
desired = (
self.rate * K.clip(norms, self.min_value, self.max_value) +
(1 - self.rate) * norms)
w *= (desired / (K.epsilon() + norms))
return w
def get_gradients(self, loss, params):
grads = K.gradients(loss, params)
if hasattr(self, 'clipnorm') and self.clipnorm > 0:
norm = K.sqrt(sum([K.sum(K.square(g)) for g in grads]))
grads = [clip_norm(g, self.clipnorm, norm) for g in grads]
if hasattr(self, 'clipvalue') and self.clipvalue > 0:
grads = [K.clip(g, -self.clipvalue, self.clipvalue) for g in grads]
return grads
def get_initial_state(self, inputs):
# build an all-zero tensor of shape (samples, output_dim)
initial_state = K.zeros_like(inputs) # (samples, timesteps, input_dim)
initial_state = K.sum(initial_state, axis=(1, 2)) # (samples,)
initial_state = K.expand_dims(initial_state) # (samples, 1)
initial_state = K.tile(initial_state, [1,
self.units]) # (samples, output_dim)
initial_state = [initial_state for _ in range(len(self.states))]
return initial_state
def get_initial_state(self, inputs):
# (samples, timesteps, rows, cols, filters)
initial_state = K.zeros_like(inputs)
# (samples, rows, cols, filters)
initial_state = K.sum(initial_state, axis=1)
shape = list(self.kernel_shape)
shape[-1] = self.filters
initial_state = self.input_conv(
initial_state, K.zeros(tuple(shape)), padding=self.padding)
initial_states = [initial_state for _ in range(2)]
return initial_states
def get_constants(self, inputs, training=None):
constants = []
if self.implementation == 0 and 0 < self.dropout < 1:
ones = K.zeros_like(inputs)
ones = K.sum(ones, axis=1)
ones += 1
def dropped_inputs():
return K.dropout(ones, self.dropout)
dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(4)
]
constants.append(dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(4)])
if 0 < self.recurrent_dropout < 1:
shape = list(self.kernel_shape)
shape[-1] = self.filters
ones = K.zeros_like(inputs)
ones = K.sum(ones, axis=1)
ones = self.input_conv(ones, K.zeros(shape), padding=self.padding)
ones += 1.
def dropped_inputs(): # pylint: disable=function-redefined
return K.dropout(ones, self.recurrent_dropout)
rec_dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(4)
]
constants.append(rec_dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(4)])
return constants
def get_initial_state(self, inputs):
# (samples, timesteps, rows, cols, filters)
initial_state = K.zeros_like(inputs)
# (samples, rows, cols, filters)
initial_state = K.sum(initial_state, axis=1)
shape = list(self.cell.kernel_shape)
shape[-1] = self.cell.filters
initial_state = self.cell.input_conv(initial_state,
K.zeros(tuple(shape)),
padding=self.cell.padding)
if hasattr(self.cell.state_size, '__len__'):
return [initial_state for _ in self.cell.state_size]
else:
return [initial_state]
def softmax(x, axis=-1):
"""Softmax activation function.
Arguments:
x : Tensor.
axis: Integer, axis along which the softmax normalization is applied.
Returns:
Tensor, output of softmax transformation.
Raises:
ValueError: In case `dim(x) == 1`.
"""
ndim = K.ndim(x)
if ndim == 2:
return K.softmax(x)
elif ndim > 2:
e = K.exp(x - K.max(x, axis=axis, keepdims=True))
s = K.sum(e, axis=axis, keepdims=True)
return e / s
else:
raise ValueError('Cannot apply softmax to a tensor that is 1D')
def __call__(self, w): norms = K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True)) desired = K.clip(norms, 0, self.max_value) w *= (desired / (K.epsilon() + norms)) return w
def __call__(self, w):
return w / (
K.epsilon() + K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True)))
def cosine_proximity(y_true, y_pred): y_true = K.l2_normalize(y_true, axis=-1) y_pred = K.l2_normalize(y_pred, axis=-1) return -K.sum(y_true * y_pred, axis=-1)
def cosine_proximity(y_true, y_pred): y_true = K.l2_normalize(y_true, axis=-1) y_pred = K.l2_normalize(y_pred, axis=-1) return -K.sum(y_true * y_pred, axis=-1)
def kullback_leibler_divergence(y_true, y_pred): y_true = K.clip(y_true, K.epsilon(), 1) y_pred = K.clip(y_pred, K.epsilon(), 1) return K.sum(y_true * K.log(y_true / y_pred), axis=-1)
def categorical_hinge(y_true, y_pred): pos = K.sum(y_true * y_pred, axis=-1) neg = K.max((1. - y_true) * y_pred, axis=-1) return K.maximum(neg - pos + 1., 0.)
def kullback_leibler_divergence(y_true, y_pred): y_true = K.clip(y_true, K.epsilon(), 1) y_pred = K.clip(y_pred, K.epsilon(), 1) return K.sum(y_true * K.log(y_true / y_pred), axis=-1)
def __call__(self, w):
return w / (K.epsilon() +
K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True)))
def categorical_hinge(y_true, y_pred):
pos = K.sum(y_true * y_pred, axis=-1)
neg = K.max((1. - y_true) * y_pred, axis=-1)
return K.maximum(neg - pos + 1., 0.)
def __call__(self, w):
norms = K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True))
desired = K.clip(norms, 0, self.max_value)
w *= (desired / (K.epsilon() + norms))
return w
def __call__(self, w):
norms = K.sqrt(K.sum(K.square(w), axis=self.axis, keepdims=True))
desired = (self.rate * K.clip(norms, self.min_value, self.max_value) +
(1 - self.rate) * norms)
w *= (desired / (K.epsilon() + norms))
return w