def call(self, inputs, reverse=False, ddi=False, **kwargs):
logscale_factor = 3.
x = inputs
reduce_axis = list(range(K.ndim(inputs)))[:-1]
if not reverse:
log_scale = self.log_scale
bias = self.bias
if ddi:
x_var = tf.reduce_mean(x**2, reduce_axis, keepdims=True)
init_scale = tf.log(1. /
(tf.sqrt(x_var) + 1e-6)) / logscale_factor
init_bias = tf.reduce_mean(x, reduce_axis, keepdims=True)
log_scale = K.switch(K.all(K.equal(self.log_scale, 0.)),
init_scale, self.log_scale)
bias = K.switch(K.all(K.equal(self.bias, 0.)), -init_bias,
self.bias)
self.add_update(K.update_add(
self.log_scale,
K.switch(K.all(K.equal(self.log_scale, 0.)), init_scale,
K.zeros_like(init_scale))),
inputs=x)
self.add_update(K.update_add(
self.bias,
K.switch(K.all(K.equal(self.bias, 0.)), -init_bias,
K.zeros_like(init_bias))),
inputs=x)
return (x + bias) * K.exp(log_scale)
else:
return x / K.exp(self.log_scale) - self.bias
def get_initial_state(self, inputs):
initial_states = []
first = True
if self._stackedcells:
for cell in self.cell.cells:
shape = list(cell.kernel_shape)
shape[-1] = cell.filters
if first: # Make m, h, c states
initial_state = K.zeros_like(inputs)
initial_state = K.sum(initial_state, axis=1)
initial_state = cell.input_conv(initial_state,
K.zeros(tuple(shape)),
padding=cell.padding)
initial_states += [initial_state for _ in range(3)]
first = False
else: # if not first make h, c states
initial_state = K.zeros_like(initial_state)
initial_state = cell.input_conv(initial_state,
K.zeros(tuple(shape)),
padding=cell.padding)
initial_states += [initial_state for _ in range(2)]
else: # Single cell
shape = list(self.cell.kernel_shape)
shape[-1] = self.cell.filters
initial_state = K.zeros_like(inputs)
initial_state = self.cell.inputs_conv(initial_state,
K.zeros(tuple(shape)),
padding=self.cell.padding)
initial_states += [initial_state for _ in range(3)]
return initial_states
def _moments(self, inputs, reduction_axes, keep_dims):
mean, variance = nn.moments(inputs, reduction_axes, keep_dims=keep_dims)
# TODO(b/129279393): Support zero batch input in non DistributionStrategy
# code as well.
if self._support_zero_size_input():
inputs_size = array_ops.size(inputs)
mean = array_ops.where(inputs_size > 0, mean, K.zeros_like(mean))
variance = array_ops.where(inputs_size > 0, variance,
K.zeros_like(variance))
return mean, variance
def _moments(self, inputs, reduction_axes, keep_dims):
mean, variance = nn.moments(inputs, reduction_axes, keep_dims=keep_dims)
# TODO(b/129279393): Support zero batch input in non DistributionStrategy
# code as well.
# TODO(b/130185866): Support zero batch input in graph mode.
if (ops.executing_eagerly_outside_functions() and
distribution_strategy_context.has_strategy()):
inputs_size = array_ops.size(inputs)
mean = array_ops.where(inputs_size > 0, mean, K.zeros_like(mean))
variance = array_ops.where(inputs_size > 0, variance,
K.zeros_like(variance))
return mean, variance
def _moments(self, inputs, reduction_axes, keep_dims):
mean, variance = nn.moments(inputs,
reduction_axes,
keep_dims=keep_dims)
# TODO(b/129279393): Support zero batch input in non DistributionStrategy
# code as well.
if distribution_strategy_context.has_strategy():
inputs_size = array_ops.size(inputs)
mean = tf_utils.smart_cond(inputs_size > 0, lambda: mean,
lambda: K.zeros_like(mean))
variance = tf_utils.smart_cond(inputs_size > 0, lambda: variance,
lambda: K.zeros_like(variance))
return mean, variance
def _moments(self, inputs, reduction_axes, keep_dims):
mean, variance = nn.moments(inputs,
reduction_axes,
keep_dims=keep_dims)
# TODO(b/129279393): Support zero batch input in non DistributionStrategy
# code as well.
if distribution_strategy_context.has_strategy(
) and not inputs.shape.is_fully_defined():
inputs_size = array_ops.size(inputs)
mean = array_ops.where(inputs_size > 0, mean, K.zeros_like(mean))
variance = array_ops.where(inputs_size > 0, variance,
K.zeros_like(variance))
return mean, variance
def get_initial_state(self, inputs):
"""
INPUTS: shape = (Batch, Time_steps, window_size, numberofSensors, 1 or filters )
This function will be called after build()
"""
initial_state = K.zeros_like(inputs)
# Because of all zeros, sum to delete the timestpes dimension ---->(Batch, Time_steps, window_size, numberofSensors, 1 or filters )
initial_state = K.sum(initial_state, axis=1)
# Through the convlution to inference the size of hidden state
for index, k_shape in enumerate(self.cell.kernel_shape):
shape = list(k_shape)
shape[-1] = self.cell.filters[index]
initial_state = self.cell.input_conv(initial_state,
array_ops.zeros(tuple(shape)),
padding=self.cell.padding)
#self.len_state = initial_state.shape[1]
if hasattr(self.cell.state_size, '__len__'):
return [initial_state for _ in self.cell.state_size]
else:
return [initial_state]
def subdiv_moments(self, inputs, reduction_axes, keep_dims):
# mean and variance only for the current batch
mean, net_sum, variance, squared_mean, input_batch_size = self._subdiv_calculate_mean_and_var(
inputs, reduction_axes, keep_dims)
if self._support_zero_size_input():
input_batch_size = 0 if input_batch_size is None else input_batch_size
mean = array_ops.where(input_batch_size > 0, mean,
K.zeros_like(mean))
net_sum = array_ops.where(input_batch_size > 0, net_sum,
K.zeros_like(net_sum))
variance = array_ops.where(input_batch_size > 0, variance,
K.zeros_like(variance))
squared_mean = array_ops.where(input_batch_size > 0, squared_mean,
K.zeros_like(squared_mean))
return mean, net_sum, variance, squared_mean, input_batch_size
def call(self, u_vecs, **kwargs):
if self.share_weights:
u_hat_vecs = K.conv1d(u_vecs, self.W)
else:
u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1])
batch_size = K.shape(u_vecs)[0]
input_num_capsule = K.shape(u_vecs)[1]
u_hat_vecs = K.reshape(u_hat_vecs,
(batch_size, input_num_capsule,
self.num_capsule, self.dim_capsule))
u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3))
# final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule]
b = K.zeros_like(
u_hat_vecs[:, :, :,
0]) # shape = [None, num_capsule, input_num_capsule]
for i in range(self.routings):
c = softmax(b, 1)
o = K.batch_dot(c, u_hat_vecs, [2, 2])
if K.backend() == 'theano':
o = K.sum(o, axis=1)
if i < self.routings - 1:
o = K.l2_normalize(o, -1)
b = K.batch_dot(o, u_hat_vecs, [2, 3])
if K.backend() == 'theano':
b = K.sum(b, axis=1)
return self.activation(o)
def create_inital_state(inputs, hidden_size):
# We are not using initial states, but need to pass something to K.rnn funciton
fake_state = K.zeros_like(inputs) # <= (batch_size, enc_seq_len, latent_dim
fake_state = K.sum(fake_state, axis=[1, 2]) # <= (batch_size)
fake_state = K.expand_dims(fake_state) # <= (batch_size, 1)
fake_state = K.tile(fake_state, [1, hidden_size]) # <= (batch_size, latent_dim
return fake_state
def reshape_connection(connection, shape):
if connection.shape[1] != shape[1]:
connection = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2))(connection)
if connection.shape[-1] != shape[-1]:
connection = Lambda(lambda x: K.concatenate([x, K.zeros_like(x)],
axis=-1))(connection)
return connection
def create_inital_state(inputs, hidden_size):
fake_state = K.zeros_like(
inputs) # (batch_size, enc_seq_len, latent_dim)
fake_state = K.sum(fake_state, axis=[1, 2]) # (batch_size)
fake_state = K.expand_dims(fake_state) # (batch_size, 1)
fake_state = K.tile(fake_state,
[1, hidden_size]) # (batch_size, latent_dim)
return fake_state
def inverse_root_via_eigenvalues(m):
ev, v = tf.linalg.eigh(m)
epsillon = 1e-8 # for numerical stability - clip
ev = tf.where(ev > epsillon, x=ev, y=K.ones_like(ev))
v = tf.where(ev > epsillon, x=v, y=K.zeros_like(v))
u = v
ev_inv_root = tf.math.reciprocal(tf.math.sqrt(ev))
res = tf.matmul(tf.matmul(u, tf.diag(ev_inv_root)), tf.transpose(v))
return res
def call(self, inputs, **kwargs):
input_shape = K.int_shape(inputs)
sequence_length, d_model = input_shape[-2:]
# output of the "sigmoid halting unit" (not the probability yet)
halting = K.sigmoid(
K.reshape(
K.bias_add(K.dot(K.reshape(inputs, [-1, d_model]),
self.act_weights['halting_kernel']),
self.act_weights['halting_biases'],
data_format='channels_last'),
[-1, sequence_length]))
if self.zeros_like_halting is None:
self.initialize_control_tensors(halting)
# useful flags
step_is_active = K.greater(self.halt_budget, 0)
no_further_steps = K.less_equal(self.halt_budget - halting, 0)
# halting probability is equal to
# a. halting output if this isn't the last step (we have some budget)
# b. to remainder if it is,
# c. and zero for the steps that shouldn't be executed at all
# (out of budget for them)
halting_prob = K.switch(
step_is_active, K.switch(no_further_steps, self.remainder,
halting), self.zeros_like_halting)
self.active_steps += K.switch(step_is_active, self.ones_like_halting,
self.zeros_like_halting)
# We don't know which step is the last, so we keep updating
# expression for the loss with each call of the layer
self.ponder_cost = (self.act_weights['time_penalty_t'] *
K.mean(self.remainder + self.active_steps))
# Updating "the remaining probability" and the halt budget
self.remainder = K.switch(no_further_steps, self.remainder,
self.remainder - halting)
self.halt_budget -= halting # OK to become negative
# If none of the inputs are active at this step, then instead
# of zeroing them out by multiplying to all-zeroes halting_prob,
# we can simply use a constant tensor of zeroes, which means that
# we won't even calculate the output of those steps, saving
# some real computational time.
if self.zeros_like_input is None:
self.zeros_like_input = K.zeros_like(inputs,
name='zeros_like_input')
# just because K.any(step_is_active) doesn't work in PlaidML
any_step_is_active = K.greater(K.sum(K.cast(step_is_active, 'int32')),
0)
step_weighted_output = K.switch(
any_step_is_active,
K.expand_dims(halting_prob, -1) * inputs, self.zeros_like_input)
if self.weighted_output is None:
self.weighted_output = step_weighted_output
else:
self.weighted_output += step_weighted_output
return [inputs, self.weighted_output]
def custom_loss(y_true, y_pred, loss_weights = loss_weights): # Verified
zero_index = K.zeros_like(y_true[:, 0])
ones_index = K.ones_like(y_true[:, 0])
# Classifier
labels = y_true[:, 0]
class_preds = y_pred[:, 0]
bi_crossentropy_loss = -labels * K.log(class_preds) - (1 - labels) * K.log(1 - class_preds)
classify_valid_index = tf.where(K.less(y_true[:, 0], 0), zero_index, ones_index)
classify_keep_num = K.cast(tf.cast(tf.reduce_sum(classify_valid_index), tf.float32) * SAMPLE_KEEP_RATIO, dtype = tf.int32)
# For classification problem, only pick 70% of the valid samples.
classify_loss_sum = bi_crossentropy_loss * tf.cast(classify_valid_index, bi_crossentropy_loss.dtype)
classify_loss_sum_filtered, _ = tf.nn.top_k(classify_loss_sum, k = classify_keep_num)
classify_loss = tf.where(K.equal(classify_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(classify_loss_sum_filtered))
# Bounding box regressor
rois = y_true[:, 1: 5]
roi_preds = y_pred[:, 1: 5]
roi_raw_mean_square_error = K.sum(K.square(rois - roi_preds), axis = 1) # mse
# roi_raw_smooth_l1_loss = K.mean(tf.where(K.abs(rois - roi_preds) < 1, 0.5 * K.square(rois - roi_preds), K.abs(rois - roi_preds) - 0.5)) # L1 Smooth Loss
roi_valid_index = tf.where(K.equal(K.abs(y_true[:, 0]), 1), ones_index, zero_index)
roi_keep_num = K.cast(tf.reduce_sum(roi_valid_index), dtype = tf.int32)
roi_valid_mean_square_error = roi_raw_mean_square_error * tf.cast(roi_valid_index, roi_raw_mean_square_error.dtype)
roi_filtered_mean_square_error, _ = tf.nn.top_k(roi_valid_mean_square_error, k = roi_keep_num)
roi_loss = tf.where(K.equal(roi_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(roi_filtered_mean_square_error))
# roi_valid_smooth_l1_loss = roi_raw_smooth_l1_loss * roi_valid_index
# roi_filtered_smooth_l1_loss, _ = tf.nn.top_k(roi_valid_smooth_l1_loss, k = roi_keep_num)
# roi_loss = K.mean(roi_filtered_smooth_l1_loss)
# Landmark regressor
pts = y_true[:, 5: 17]
pt_preds = y_pred[:, 5: 17]
pts_raw_mean_square_error = K.sum(K.square(pts - pt_preds), axis = 1) # mse
# pts_raw_smooth_l1_loss = K.mean(tf.where(K.abs(pts - pt_preds) < 1, 0.5 * K.square(pts - pt_preds), K.abs(pts - pt_preds) - 0.5)) # L1 Smooth Loss
pts_valid_index = tf.where(K.equal(y_true[:, 0], -2), ones_index, zero_index)
pts_keep_num = K.cast(tf.reduce_sum(pts_valid_index), dtype = tf.int32)
pts_valid_mean_square_error = pts_raw_mean_square_error * tf.cast(pts_valid_index, tf.float32)
pts_filtered_mean_square_error, _ = tf.nn.top_k(pts_valid_mean_square_error, k = pts_keep_num)
pts_loss = tf.where(K.equal(pts_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(pts_filtered_mean_square_error))
# pts_valid_smooth_l1_loss = pts_raw_smooth_l1_loss * pts_valid_index
# pts_filtered_smooth_l1_loss, _ = tf.nn.top_k(pts_valid_smooth_l1_loss, k = pts_keep_num)
# pts_loss = K.mean(pts_filtered_smooth_l1_loss)
loss = classify_loss * loss_weights[0] + roi_loss * loss_weights[1] + pts_loss * loss_weights[2]
return loss
def _assign_moving_average(self, variable, value, momentum, inputs_size):
with K.name_scope('AssignMovingAvg') as scope:
with ops.colocate_with(variable):
decay = ops.convert_to_tensor_v2(1.0 - momentum, name='decay')
if decay.dtype != variable.dtype.base_dtype:
decay = math_ops.cast(decay, variable.dtype.base_dtype)
update_delta = (
variable - math_ops.cast(value, variable.dtype)) * decay
if inputs_size is not None:
update_delta = array_ops.where(inputs_size > 0, update_delta,
K.zeros_like(update_delta))
return state_ops.assign_sub(variable, update_delta, name=scope)
def _assign_moving_average(self, variable, value, momentum, inputs_size):
with K.name_scope('AssignMovingAvg') as scope:
with ops.colocate_with(variable):
# decay = ops.convert_to_tensor(1.0 - momentum, name='decay')
# if decay.dtype != variable.dtype.base_dtype:
# decay = math_ops.cast(decay, variable.dtype.base_dtype)
if 'moving_mean' in variable.name:
if self.bn_state == 1:
self.agg_mean += value
elif self.bn_state == 2:
if self.bn_update_cntr > 0:
self.agg_mean = self.agg_mean / self.bn_update_cntr
update_delta = math_ops.cast(
self.agg_mean, variable.dtype)
if inputs_size is not None:
update_delta = array_ops.where(
inputs_size > 0, update_delta,
K.zeros_like(update_delta))
return state_ops.assign(variable,
update_delta,
name=scope)
elif 'moving_variance' in variable.name:
if self.bn_state == 1:
self.agg_var += value
self.bn_update_cntr += 1
elif self.bn_state == 2:
if self.bn_update_cntr > 0:
self.agg_var = self.agg_var / self.bn_update_cntr
update_delta = math_ops.cast(
self.agg_var, variable.dtype)
if inputs_size is not None:
update_delta = array_ops.where(
inputs_size > 0, update_delta,
K.zeros_like(update_delta))
return state_ops.assign(variable,
update_delta,
name=scope)
return state_ops.assign(variable, variable, name=scope)
def get_initial_states(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)
depthwise_shape = list(self.depthwise_kernel_shape)
pointwise_shape = list(self.pointwise_kernel_shape)
initial_state = self.input_conv(
initial_state, K.zeros(tuple(depthwise_shape)),
K.zeros(tuple(pointwise_shape)), padding=self.padding)
initial_states = [initial_state for _ in range(2)]
return initial_states
def initialize_control_tensors(self, halting):
"""
Initializes constants and some step-tracking variables
during the first call of the layer (since for the Universal Transformer
all the following calls are supposed to be with inputs of identical
shapes).
"""
self.zeros_like_halting = K.zeros_like(halting,
name='zeros_like_halting')
self.ones_like_halting = K.ones_like(halting, name='ones_like_halting')
self.remainder = self.ones_like_halting
self.active_steps = self.zeros_like_halting
self.halt_budget = self.ones_like_halting - self.halt_epsilon
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:
depthwise_shape = list(self.depthwise_kernel_shape)
pointwise_shape = list(self.pointwise_kernel_shape)
ones = K.zeros_like(inputs)
ones = K.sum(ones, axis=1)
ones = self.input_conv(ones, K.zeros(depthwise_shape),
K.zeros(pointwise_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 call(self, u_vecs, scores=None):
# if self.share_weights:
# u_hat_vecs = K.conv1d(u_vecs, self.W)
# else:
# u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1])
u_hat_vecs = u_vecs
batch_size = K.shape(u_vecs)[0]
input_num_capsule = K.shape(u_vecs)[1]
if scores is not None:
scores = K.permute_dimensions(scores, (0, 2, 1))
u_hat_vecs = u_hat_vecs * scores
u_hat_vecs = K.reshape(u_hat_vecs,
(batch_size, input_num_capsule,
self.num_capsule, self.dim_capsule))
u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3))
b = K.zeros_like(
u_hat_vecs[:, :, :,
0]) # shape = [None, num_capsule, input_num_capsule]
# biases = self.add_weight(name='capsule_kernel',
# shape=(batch_size1, self.num_capsule, self.dim_capsule),
# # shape=self.kernel_size,
# dtype=tf.float32,
# initializer='glorot_uniform',
# trainable=True)
# biases = tf.get_variable(name='bias',
# shape=(self.num_capsule, self.dim_capsule), initializer='glorot_uniform',)
for i in range(self.routings):
# b = K.permute_dimensions(b, (0, 2, 1)) # shape = [None, input_num_capsule, num_capsule]
# c = K.softmax(b)
leak = tf.zeros_like(b, optimize=True)
leak = tf.reduce_sum(leak, axis=1, keep_dims=True)
leaky_logits = tf.concat([leak, b], axis=1)
leaky_routing = tf.nn.softmax(leaky_logits, dim=1)
c = tf.split(leaky_routing, [1, self.num_capsule], axis=1)[1]
# c = K.permute_dimensions(c, (0, 2, 1))
# b = K.permute_dimensions(b, (0, 2, 1))
o = K.batch_dot(c, u_hat_vecs, [2, 2]) # + self.biases
outputs = self.activation(o)
if i < self.routings - 1:
b = K.batch_dot(outputs, u_hat_vecs, [2, 3])
# self.c = scores
return outputs
def get_initial_state(self, inputs):
# (samples, timesteps, rows, cols, z, filters)
initial_state = K.zeros_like(inputs)
d = [1,2,1,1,1] if self.cell.data_format == 'channels_first' else [1,1,1,1,2]
initial_state = K.tile(initial_state, d)
shape = list(self.cell.kernel_shape)
shape[-1] = self.cell.filters
initial_state = self.cell.input_conv(initial_state,
array_ops.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 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 get_initial_state(self, inputs):
initial_state = K.zeros_like(inputs)
initial_state = K.sum(initial_state, axis=(1, 2))
initial_state = K.expand_dims(initial_state)
initial_state = K.tile(initial_state, [1, self.units]) # (samples, output_dim)
n = K.identity(initial_state)
d = K.identity(initial_state)
h = K.identity(initial_state)
dtype = initial_state.dtype.name
min_value = np.array([1E38]).astype(dtype).item()
a_max = K.identity(initial_state) - min_value
h = h + self.cell.recurrent_activation(K.expand_dims(self.cell.initial_attention, axis=0))
return [n, d, h, a_max]
def get_initial_state(self, inputs):
# (samples, timesteps, rows, cols, z, filters)
initial_state = K.zeros_like(inputs)
# (samples, rows, cols, z, 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,
array_ops.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 check_the_config_valid(para, window_size, feature):
initial_state = np.zeros((1, window_size, feature, 1))
initial_state = tf.cast(initial_state, 'float32')
initial_state = K.zeros_like(initial_state)
channel = 1
try:
for i in range(para["preprocessing_layers"]):
shape = (para["pre_kernel_width"], 1, channel,
para["pre_number_filters"])
channel = para["pre_number_filters"]
initial_state = K.conv2d(
initial_state, array_ops.zeros(tuple(shape)),
(para["pre_strides"],
1)) #,dilation_rate=(para["pre_dilation_rate"],1))
for i in range(1, 4):
assert len(para["eclstm_{}_recurrent_activation".format(i)]) == len(para["eclstm_{}_conv_activation".format(i)]) == \
len(para["eclstm_{}_number_filters".format(i)]) == len(para["eclstm_{}_kernel_width".format(i)])== \
len(para["eclstm_{}_fusion".format(i)]), "Archtecture Parameters of {} layer should be in same length".format(i)
for j in range(
len(para["eclstm_{}_recurrent_activation".format(i)])):
if para["eclstm_{}_recurrent_activation".format(i)][0] is None:
break
if para["eclstm_{}_fusion".format(i)][j] == "early":
shape = (para["eclstm_{}_kernel_width".format(i)][j],
feature, channel,
para["eclstm_{}_number_filters".format(i)][j])
feature = 1
channel = para["eclstm_{}_number_filters".format(i)][j]
else:
shape = (para["eclstm_{}_kernel_width".format(i)][j], 1,
channel,
para["eclstm_{}_number_filters".format(i)][j])
channel = para["eclstm_{}_number_filters".format(i)][j]
initial_state = K.conv2d(
initial_state, array_ops.zeros(tuple(shape)),
(para["eclstm_{}_strides".format(i)], 1))
print("valid Configuration!")
return True
except:
print(
"Invalid Configuration! Try smaller strides or kernel size or greater window size!"
)
return False
def call(self, y):
# Sanity Check
if isinstance(y, list):
raise ValueError('TSG layer has only 1 input')
# y = tf_print(y, [y], message='{}: The unconstrained action is:'.format(y.name.split('/')[0]), summarize=-1)
y = check_numerics(y, 'Problem with input y')
# Calculate A.c
Ac = tensordot(self.A_graph, self.c_graph, 1)
# Calculate b - Ac
bMinusAc = self.b_graph - Ac
# Calculate y - c
yMinusc = y - self.c_graph
# Calculate A.(y - c)
ADotyMinusc = K.sum((self.A_graph * expand_dims(yMinusc, -2)), axis=2)
# Do elem-wise division
intersection_points = bMinusAc / (ADotyMinusc + K.epsilon()
) # Do we need the K.epsilon()?
# Enforce 0 <= intersection_points <= 1 because the point must lie between c and y
greater_1 = K.greater(intersection_points,
K.ones_like(intersection_points))
candidate_alpha = K.switch(greater_1,
K.ones_like(intersection_points) + 1,
intersection_points)
less_0 = K.less(candidate_alpha, K.zeros_like(intersection_points))
candidate_alpha = K.switch(less_0,
K.ones_like(intersection_points) + 1,
candidate_alpha)
# Find farthest intersection point from y to get projection point
alpha = K.min(candidate_alpha, axis=-1, keepdims=True)
# If it is an interior point, y itself is the projection point
interior_point = K.greater(alpha, K.ones_like(alpha))
alpha = K.switch(interior_point, K.ones_like(alpha), alpha)
# alpha = tf_print(alpha, [alpha], message="{}: The value of alpha is: ".format(alpha.name.split('/')[0]))
# Return \alpha.y + (1 - \alpha).c
z = alpha * y + ((1 - alpha) * self.c_graph)
# z = tf_print(z, [z], message='{}: The constrained action is:'.format(z.name.split('/')[0]), summarize=-1)
return z
def _assign_moving_average(self, variable, value, momentum, inputs_size):
with ops.name_scope(None, 'AssignMovingAvg',
[variable, value, momentum]) as scope:
with ops.colocate_with(variable):
decay = ops.convert_to_tensor(1.0 - momentum, name='decay')
if decay.dtype != variable.dtype.base_dtype:
decay = math_ops.cast(decay, variable.dtype.base_dtype)
update_delta = (variable -
math_ops.cast(value, variable.dtype)) * decay
# TODO(b/129279393): Support zero batch input in non
# DistributionStrategy code as well.
if distribution_strategy_context.has_strategy():
update_delta = tf_utils.smart_cond(
inputs_size > 0, lambda: update_delta,
lambda: K.zeros_like(update_delta))
return state_ops.assign_sub(variable, update_delta, name=scope)
def myCrossEntropy(y_true, y_pred, e=0.3):
loss = K.sparse_categorical_crossentropy(y_true, y_pred)
loss0 = K.sparse_categorical_crossentropy(K.zeros_like(y_true), y_pred)
loss1 = K.sparse_categorical_crossentropy(K.ones_like(y_true), y_pred)
loss2 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 2, y_pred)
loss3 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 3, y_pred)
loss4 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 4, y_pred)
loss5 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 5, y_pred)
loss6 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 6, y_pred)
loss7 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 7, y_pred)
loss8 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 8, y_pred)
loss9 = K.sparse_categorical_crossentropy(K.ones_like(y_true) * 9, y_pred)
return ((100.0 - 5.765 - 1.359 - 1.000 - 1.348 - 1.554 - 1.995 - 3.042 -
6.347 - 10.431 - 17.632) * loss + 5.765 * loss0 + 1.359 * loss1 +
1.000 * loss2 + 1.348 * loss3 + 1.553 * loss4 + 1.995 * loss5 +
3.042 * loss6 + 6.347 * loss7 + 10.421 * loss8 + 17.632 * loss9)
def _assign_subdiv_moving_average(self, variable, value, momentum,
subdivsions, count):
with K.name_scope('AssignSubDivMovingAvg') as scope:
with ops.colocate_with(variable):
decay = ops.convert_to_tensor_v2_with_dispatch(1.0 - momentum,
name='decay')
if decay.dtype != variable.dtype.base_dtype:
decay = math_ops.cast(decay, variable.dtype.base_dtype)
# get the aggregated update
update_delta = (variable -
math_ops.cast(value, variable.dtype)) * decay
# update at the end of last step
update_delta = array_ops.where((count + 1) % subdivisions == 0,
update_delta,
K.zeros_like(update_delta))
return state_ops.assign_sub(variable, update_delta, name=scope)
def _assign_subdiv_rotating_sum(self, variable, value, subdivisions, count,
inputs_size):
with K.name_scope('AssignSubDivRotatedSum') as scope:
with ops.colocate_with(variable):
# reduce it for the current
update_delta = value #/subdivisions
# if the input size is 0
if inputs_size is not None:
update_delta = array_ops.where(inputs_size > 0,
update_delta,
K.zeros_like(update_delta))
# if we are starting a new batch set the variable to 0 by removing it
# from update delta then add the delta to the variable to get
# rid of the value variable
update_delta = array_ops.where(count % subdivisions == 0,
update_delta - variable,
update_delta)
return state_ops.assign_add(variable, update_delta, name=scope)
def multi_inputs_multi_outputs_model():
a = keras.layers.Input(shape=(16,), name='input_a')
b = keras.layers.Input(shape=(16,), name='input_b')
m = keras.layers.Input(shape=(8,), dtype='bool', name='input_m')
dense = keras.layers.Dense(8, name='dense_1')
a_2 = dense(a)
# Apply a mask
s_2 = keras.layers.Lambda(lambda k:
K.switch(k[0], k[1], K.zeros_like(k[1])))([m, a_2])
b_2 = dense(b)
merged = keras.layers.concatenate([s_2, b_2], name='merge')
c = keras.layers.Dense(3, activation='softmax', name='dense_2')(merged)
d = keras.layers.Dense(2, activation='softmax', name='dense_3')(merged)
model = keras.models.Model(inputs=[a, b, m], outputs=[c, d])
model.compile(
loss='categorical_crossentropy',
optimizer='rmsprop',
metrics={
'dense_2': 'categorical_accuracy',
'dense_3': 'categorical_accuracy'
})
return model
