def call(self, inputs, source_mask=None, target_mask=None):
source, target = inputs
source_mask = rk.utils.convert_to_attention_mask(source, source_mask)
target_mask = rk.utils.convert_to_attention_mask(target, target_mask)
encoder_outputs = self.encoder(source, encoder_mask=source_mask)
do_pad = tf.equal(tf.mod(tf.shape(target)[1], 2), 1)
target = K.switch(
do_pad, tf.pad(target, [[0, 0], [0, 1], [0, 0]]), target)
encoder_outputs = K.switch(
do_pad, tf.pad(encoder_outputs, [[0, 0], [0, 1], [0, 0]]), encoder_outputs)
# also pad source mask, target_mask
output = target
log_det_W_list = []
log_s_list = []
for i in range(self.n_flows):
output, log_det_W = self.invertible_dense[i](output)
output, log_s = self.WN[i](
output, encoder_outputs=encoder_outputs,
encoder_mask=source_mask,
decoder_mask=target_mask)
log_s_list.append(log_s)
log_det_W_list.append(log_det_W)
return output, log_s_list, log_det_W_list
def BetheBlochGeant(
lnbg, kp0, kp1, kp2, kp3,
kp4): #kp0=2.33,kp1=0.20,kp2=3.00,kp3=173.0e-9,kp4=0.49848
bg = K.exp(lnbg)
mK = 0.307075e-3
me = 0.511e-3
rho = kp0
x0 = kp1 * 2.303
x1 = kp2 * 2.303
mI = kp3
mZA = kp4
bg2 = bg * bg
maxT = 2 * me * bg2
x = lnbg
lhwI = K.log(28.816e-9 * K.sqrt(K.cast(rho * mZA, dtype=float)) /
mI)
d2 = K.switch(
K.greater(x, x1), lhwI + x - 0.5,
K.switch(
K.greater(x, x0), lhwI + x - 0.5 + (0.5 - lhwI - x0) *
(((x1 - x) / (x1 - x0))**3), 0. * bg))
return mK * mZA * (1 +
bg2) / bg2 * (0.5 * K.log(2 * me * bg2 * maxT /
(mI * mI)) - bg2 /
(1 + bg2) - d2)
def get_updates(self, params, loss):
grads = self.get_gradients(loss, params)
c = 0
prev_grads = [
shared_zeros(p.shape, name=f"prev_grads") for p in params
]
c = 0
prev_steps = [
sharedX(np.full(p.shape, self.step_init), name=f"prev_stepd")
for p in params
]
self.updates = []
for p, grad, prev_grad, prev_step, c in zip(params, grads, prev_grads,
prev_steps):
grad_sgn = prev_grad * grad
new_step = K.switch(
K.ge(grad_sgn, 0.0),
K.minimum(prev_step * self.step_inc, self.step_max),
K.maximum(prev_step * self.step_dec, self.step_min))
self.updates.append((prev_step, new_step))
new_grad = K.switch(K.ge(grad_sgn, 0.0), grad, 0.0)
self.updates.append((prev_grad, new_grad))
new_p = p - K.sgn(new_grad) * new_step
self.updates.append((p, c(new_p)))
return self.updates
def get_updates(self, params, loss):
grads = self.get_gradients(loss, params)
shapes = [K.get_variable_shape(p) for p in params]
alphas = [
K.variable(K.ones(shape) * self.init_alpha) for shape in shapes
]
old_grads = [K.zeros(shape) for shape in shapes]
self.weights = alphas + old_grads
self.updates = []
for p, grad, old_grad, alpha in zip(params, grads, old_grads, alphas):
grad = K.sign(grad)
new_alpha = K.switch(
K.greater(grad * old_grad, 0),
K.minimum(alpha * self.scale_up, self.max_alpha),
K.switch(K.less(grad * old_grad, 0),
K.maximum(alpha * self.scale_down, self.min_alpha),
alpha))
grad = K.switch(K.less(grad * old_grad, 0), K.zeros_like(grad),
grad)
new_p = p - grad * new_alpha
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
self.updates.append(K.update(alpha, new_alpha))
self.updates.append(K.update(old_grad, grad))
return self.updates
def get_updates(self, loss, params):
sync_cond = K.equal((self.iterations + 1) // self.sync_period *
self.sync_period, (self.iterations + 1))
if TF_KERAS:
slow_params = [K.variable(K.get_value(p), name='sp_{}'.format(i))
for i, p in enumerate(params)]
self.updates = self.optimizer.get_updates(loss, params)
slow_updates = []
for p, sp in zip(params, slow_params):
sp_t = sp + self.slow_step * (p - sp)
slow_updates.append(K.update(sp, K.switch(
sync_cond,
sp_t,
sp,
)))
slow_updates.append(K.update_add(p, K.switch(
sync_cond,
sp_t - p,
K.zeros_like(p),
)))
else:
slow_params = {p.name: K.variable(K.get_value(
p), name='sp_{}'.format(i)) for i, p in enumerate(params)}
update_names = ['update', 'update_add', 'update_sub']
original_updates = [getattr(K, name) for name in update_names]
setattr(K, 'update', lambda x, new_x: ('update', x, new_x))
setattr(K, 'update_add', lambda x, new_x: ('update_add', x, new_x))
setattr(K, 'update_sub', lambda x, new_x: ('update_sub', x, new_x))
self.updates = self.optimizer.get_updates(loss, params)
for name, original_update in zip(update_names, original_updates):
setattr(K, name, original_update)
slow_updates = []
for i, update in enumerate(self.updates):
if isinstance(update, tuple):
name, x, new_x, adjusted = update + (update[-1],)
update_func = getattr(K, name)
if name == 'update_add':
adjusted = x + new_x
if name == 'update_sub':
adjusted = x - new_x
if x.name not in slow_params:
self.updates[i] = update_func(x, new_x)
else:
slow_param = slow_params[x.name]
slow_param_t = slow_param + \
self.slow_step * (adjusted - slow_param)
slow_updates.append(K.update(slow_param, K.switch(
sync_cond,
slow_param_t,
slow_param,
)))
self.updates[i] = K.update(x, K.switch(
sync_cond,
slow_param_t,
adjusted,
))
slow_params = list(slow_params.values())
self.updates += slow_updates
self.weights = self.optimizer.weights + slow_params
return self.updates
def merge_action_crossed_fn(x):
input = x[0]
is_crossed_orig = x[1]
is_crossed_orig = K.switch(is_crossed_orig > 0.1,
K.maximum(is_crossed_orig, 1),
is_crossed_orig * 0)
action_count = input.shape[1]
is_crossed = K.expand_dims(is_crossed_orig, axis=1)
is_crossed = K.expand_dims(is_crossed, axis=1)
is_crossed = K.temporal_padding(is_crossed, (0, action_count - 1))
is_crossed = K.squeeze(is_crossed, axis=2)
is_crossed_mask = K.expand_dims(is_crossed_orig, axis=1)
is_crossed_mask = K.repeat_elements(is_crossed_mask, action_count, axis=1)
res_crossed = (1 - is_crossed_mask) * input + is_crossed
carpisma_timer_orig = x[2]
carpisma_timer_orig = K.squeeze(carpisma_timer_orig, axis=2)
is_carpisma = K.sum(carpisma_timer_orig, axis=1)
is_carpisma = K.switch(is_carpisma > 0.1, K.maximum(is_carpisma, 1),
is_carpisma * 0)
not_carpisma = 1 - is_carpisma
print("carpisma timer", carpisma_timer_orig)
print("is carpisma", is_carpisma.shape)
print("not carpisma", not_carpisma.shape)
not_carpisma = K.expand_dims(not_carpisma, axis=1)
not_carpisma = K.repeat_elements(not_carpisma, action_count, axis=1)
res_crossed = res_crossed * not_carpisma
res = K.concatenate([res_crossed, carpisma_timer_orig], axis=1)
return res
def _init_cel(self, A_graph, b_graph, c_graph, y):
# Sanity Checks
y = tf.check_numerics(y, 'Problem with input y')
# Find intersection points between Ax-b and the line joining the c and y
Ac = tf.reduce_sum(A_graph * tf.expand_dims(c_graph, axis=-2), axis=-1)
bMinusAc = b_graph - Ac
yMinusc = y - c_graph
ADotyMinusc = tf.reduce_sum((A_graph * tf.expand_dims(yMinusc, -2)),
axis=-1)
intersection_alphas = bMinusAc / (ADotyMinusc + K.epsilon())
# Enforce intersection_alpha > 0 because the point must lie on the ray from c to y
less_equal_0 = K.less_equal(intersection_alphas,
K.zeros_like(intersection_alphas))
candidate_alpha = K.switch(
less_equal_0,
K.ones_like(intersection_alphas) *
tf.constant(np.inf, dtype='float32'), intersection_alphas)
# Find closest the intersection point closest to the interior point to get projection point
intersection_alpha = K.min(candidate_alpha, axis=-1, keepdims=True)
# If it is an interior point, y itself is the projection point
is_interior_point = K.greater_equal(intersection_alpha,
K.ones_like(intersection_alpha))
alpha = K.switch(is_interior_point, K.ones_like(intersection_alpha),
intersection_alpha)
# Return z = \alpha.y + (1 - \alpha).c
z = alpha * y + ((1 - alpha) * c_graph)
return z
def get_model(self):
input_shape = self.ops.INPUT_SIZE
input = Input(shape=input_shape, name='observation')
action_outputs = []
done_outputs = []
reward_outputs = []
goal_position = 0.5
for I in range(self.ops.ACTION_COUNT):
action_output = Lambda(mountaincar_next_state_fn,
arguments={'action': I})(input)
action_outputs.append(action_output)
#print(action_output.shape)
done_output = Lambda(lambda x: K.switch(
x[:, 0] > goal_position, 1 + x[:, 0:1] * 0, x[:, 0:1] * 0))(
action_output)
done_outputs.append(done_output)
# @ersin: -100: penalizer'li hali
#reward_output = Lambda(lambda x: x[:,0:1]*0 - 1)(done_output)
reward_output = Lambda(lambda x: K.switch(
x < 0.01, x[:, 0:1] * 0 - 0.01, x[:, 0:1] * 0 + 1))(
done_output)
reward_outputs.append(reward_output)
reward_output_concat = Concatenate()(reward_outputs)
model = Model(inputs=[input],
outputs=action_outputs + [reward_output_concat] +
done_outputs)
my_optimizer = Adam(lr=self.ops.LEARNING_RATE)
model.compile(optimizer=my_optimizer, loss='mse')
return model
def __init__(self, optimizer, steps_per_update=1, **kwargs):
assert float(
tf.__version__[:4]
) <= 1.13, "Please make sure that your tensorflow version is 1.13.x or lower."
super(AccumOptimizer, self).__init__(**kwargs)
self.optimizer = optimizer
with K.name_scope(self.__class__.__name__):
self.steps_per_update = steps_per_update
self.iterations = K.variable(0, dtype='int64', name='iterations')
self.cond = K.equal(self.iterations % self.steps_per_update, 0)
self.lr = self.optimizer.lr
self.optimizer.lr = K.switch(self.cond, self.optimizer.lr, 0.)
for attr in ['momentum', 'rho', 'beta_1', 'beta_2']:
if hasattr(self.optimizer, attr):
value = getattr(self.optimizer, attr)
setattr(self, attr, value)
setattr(self.optimizer, attr,
K.switch(self.cond, value, 1 - 1e-7))
for attr in self.optimizer.get_config():
if not hasattr(self, attr):
value = getattr(self.optimizer, attr)
setattr(self, attr, value)
# Cover the original get_gradients method with accumulative gradients.
def get_gradients(loss, params):
return [ag / self.steps_per_update for ag in self.accum_grads]
self.optimizer.get_gradients = get_gradients
def get_updates(self, loss, params):
# Create accumulated gradients
grads = self.get_gradients(loss, params)
self.updates = []
with tf.control_dependencies([self.iterations.assign_add(1)]):
update_cond = K.equal(self.iterations % self.accumulation_steps, 0)
sub_step = (self.iterations - 1) % self.accumulation_steps + 1
fake_iterations = (self.iterations - 1) // self.accumulation_steps + 1
acc_grads = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
for grad, acc_grad in zip(grads, acc_grads):
ave_grad = grad / K.cast(self.accumulation_steps, K.floatx())
self.updates.append(K.update(
acc_grad,
K.switch(
K.equal(sub_step, 1),
ave_grad,
acc_grad + (ave_grad - acc_grad) / K.cast(sub_step, K.floatx())
),
))
self.optimizer.get_gradients = lambda _loss, _params: \
[K.switch(update_cond, grad, K.zeros_like(grad))
for grad in acc_grads]
# Use fake iterations
original_iterations = self.optimizer.iterations
original_assign_add = getattr(state_ops, 'assign_add')
setattr(
state_ops,
'assign_add',
lambda ref, val: original_assign_add(ref, val) if ref is not fake_iterations
else original_assign_add(original_iterations, val)
)
self.optimizer.iterations = fake_iterations
# Use fake learning rate
self.optimizer.learning_rate = K.switch(update_cond, self.lr, 0.0)
# Freeze momentum
momentum = {}
for name in self.momentum_names:
if hasattr(self.optimizer, name):
momentum[name] = getattr(self.optimizer, name)
setattr(self.optimizer, name, K.switch(update_cond, momentum[name], (1.0 - K.epsilon())))
for update in self.optimizer.get_updates(loss, params):
if update is not None:
self.updates.append(update)
# Restore variables
for name, value in momentum.items():
setattr(self.optimizer, name, value)
self.optimizer.learning_rate = self._lr
self.optimizer.iterations = original_iterations
setattr(state_ops, 'assign_add', original_assign_add)
return self.updates
def _get_update_list(self, kernel):
update_list = super(E2EFSSoft, self)._get_update_list(kernel)
update_list += [
(self.moving_factor, K.switch(K.less(self.moving_T, self.warmup_T),
self.start_alpha,
K.minimum(self.alpha_M, self.start_alpha + (1. - self.start_alpha) * (self.moving_T - self.warmup_T) / self.T))),
(self.moving_T, self.moving_T + 1),
(self.moving_decay, K.switch(K.less(self.moving_factor, self.alpha_M), self.moving_decay, K.maximum(.75, self.moving_decay + self.epsilon)))
]
return update_list
def loss_units(x):
t = x / K.max(K.abs(x))
x = K.switch(K.less(t, K.epsilon()), K.zeros_like(x), x)
m = K.sum(K.cast(K.greater(x, 0.), K.floatx()))
sum_x = K.sum(x)
moving_units = K.switch(K.less_equal(m, self.units), m,
(1. - self.moving_decay) * self.moving_units)
epsilon_minus = 0.
epsilon_plus = K.switch(K.less_equal(m, self.units), self.moving_units, 0.)
return K.relu(moving_units - sum_x - epsilon_minus) + K.relu(sum_x - moving_units - epsilon_plus)
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.halting_kernel),
self.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.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 wrapper(y_true, y_pred):
if (usesoftmax):
y_pred = tf.keras.activations.softmax(y_pred)
y_pred = backend.clip(y_pred, _EPSILON, 1.0 - _EPSILON)
""" here all calculations will be based on the class greater than 0, except accuracy"""
avgIOU = 0.0
for i in range(batch_size):
numUnion = 1.0
recall = 0.0
numClass = 0.0
IOU = 0.0
mask = backend.argmax(y_true[i], -1)
pred = backend.argmax(y_pred[i], -1)
for c in np.arange(1, num_classes, 1):
msk_equal = backend.cast(backend.equal(mask, c),
dtype='float32')
masks_sum = backend.sum(msk_equal)
predictions_sum = backend.sum(
backend.cast(backend.equal(pred, c), 'float32'))
numTrue = backend.sum(
backend.cast(backend.equal(pred, c), 'float32') *
backend.cast(backend.equal(mask, c), 'float32'))
unionSize = masks_sum + predictions_sum - numTrue
maskhaslabel = backend.greater(masks_sum, 0)
predhaslabel = backend.greater(predictions_sum, 0)
predormaskexistlabel = backend.any(backend.stack(
[maskhaslabel, predhaslabel], axis=0),
axis=0)
IOU = backend.switch(predormaskexistlabel,
lambda: IOU + numTrue / unionSize,
lambda: IOU)
numUnion = backend.switch(predormaskexistlabel,
lambda: numUnion + 1,
lambda: numUnion)
recall = backend.switch(maskhaslabel,
lambda: recall + numTrue / masks_sum,
lambda: recall)
numClass = backend.switch(maskhaslabel, lambda: numClass + 1,
lambda: numClass)
IOU = IOU / numUnion
avgIOU = avgIOU + IOU
avgIOU = avgIOU / batch_size
iou_loss = 1.0 - avgIOU
# print(np.shape(y_true), np.shape(y_pred))
main_loss = backend.mean(weighted_loss_fn(y_true, y_pred))
# dice_loss = soft_dice_loss(y_true, y_pred)
return main_loss + 0.1 * iou_loss
def weighted_sum(first,
second,
sigma,
first_threshold=-np.inf,
second_threshold=np.inf):
logit_probs = first * sigma + second * (1.0 - sigma)
infty_tensor = K.ones_like(logit_probs) * INFTY
logit_probs = K.switch(K.greater(first, first_threshold), logit_probs,
infty_tensor)
logit_probs = K.switch(K.greater(second, second_threshold), logit_probs,
infty_tensor)
return logit_probs
def _get_update_list(self, kernel):
super(E2EFSSoft, self)._get_update_list(kernel)
self.moving_factor.assign(
K.switch(
K.less(self.moving_T, self.warmup_T), self.start_alpha,
K.minimum(
self.alpha_M, self.start_alpha + (1. - self.start_alpha) *
(self.moving_T - self.warmup_T) / self.T)))
self.moving_T.assign_add(1.)
self.moving_decay.assign(
K.switch(K.less(self.moving_factor, self.alpha_M),
self.moving_decay,
K.maximum(.75, self.moving_decay + self.epsilon)))
def _get_update_list(self, kernel):
update_list = super(E2EFSRanking, self)._get_update_list(kernel)
update_list += [
(self.moving_factor, K.switch(K.less_equal(self.moving_T, self.warmup_T),
self.start_alpha,
K.minimum(self.alpha_M, self.start_alpha + (1. - self.start_alpha) * (self.moving_T - self.warmup_T) / self.T))),
(self.moving_T, self.moving_T + 1),
(self.moving_units, K.switch(K.less_equal(self.moving_T, self.warmup_T),
K.cast_to_floatx((1. - self.start_alpha) * np.prod(K.int_shape(kernel))),
K.maximum(self.alpha_M, np.prod(K.int_shape(kernel)) * K.pow(K.cast_to_floatx(1. / np.prod(K.int_shape(kernel))), self.speedup * (self.moving_T - self.warmup_T) / self.T)))),
# K.maximum(1., (self.T - self.start_alpha - self.speedup * (self.moving_T - self.warmup_T)) * np.prod(K.int_shape(kernel)) / self.T))),
]
return update_list
def new_update(x, new_x):
if x is var and self._do_layer_adaptation(x):
dx = new_x - x
lr_t = self._decayed_lr(x.dtype.base_dtype)
lr_t = K.clip(lr_t, K.epsilon(), 1e10)
x_norm = tf.norm(x)
g_norm = tf.norm(dx / lr_t)
ratio = K.switch(
x_norm > 0.,
K.switch(g_norm > K.epsilon(), x_norm / g_norm, 1.),
1.)
new_x = x + dx * ratio
return old_update(x, new_x)
def get_model(self):
input_shape = self.ops.INPUT_SIZE
input = Input(shape=input_shape, name='observation')
x = input
#x = Dense(24,activation="relu", kernel_initializer='he_uniform')(x)
action_outputs = []
done_outputs = []
reward_outputs = []
losses = []
loss_weights = []
for I in range(self.ops.ACTION_COUNT):
action_output = Dense(
input_shape[0],
kernel_initializer=Constant(1),
bias_initializer=Constant(0.05 if I == 0 else -0.05))(x)
action_output = Lambda(lambda x: K.clip(x, -1, 1))(action_output)
action_outputs.append(action_output)
print(action_output.shape)
done_output = Lambda(lambda x: K.switch(
K.abs(x - 0.5) < 0.05, 1 + x * 0, x * 0))(action_output)
#done_output = Dense(1, kernel_initializer='he_uniform', activation='sigmoid')(x)
done_outputs.append(done_output)
losses.append('mse')
loss_weights.append(1)
#reward_output = Dense(20, activation='sigmoid')(input)
#reward_output = RBFLayer(100, initializer=RandomUniform(-1.0, 1.0), betas=0.02)(input)
#reward_output = Dense(20, activation='relu')(input)
#reward_output = Dense(20, activation='hard_sigmoid')(input)
#reward_output = Dense(1)(reward_output)
reward_output = Lambda(lambda x: K.switch(x < 0.01, x * 0 - 0.01, x
* 0 + 1))(done_output)
reward_outputs.append(reward_output)
#reward_hidden = Dense(256,activation="relu", kernel_initializer='he_uniform')(x)
#reward_output = Dense(1, kernel_initializer=keras.initializers.random_uniform(-0.02, 0.02))(reward_hidden)
#reward_outputs.append(reward_output)
losses.append('mse')
loss_weights.append(5)
reward_output_concat = Concatenate()(reward_outputs)
for I in range(self.ops.ACTION_COUNT):
losses.append('binary_crossentropy')
loss_weights.append(0.3)
model = Model(inputs=[input],
outputs=action_outputs + [reward_output_concat] +
done_outputs)
my_optimizer = Adam(lr=self.ops.LEARNING_RATE)
#my_optimizer = SGD(lr=self.ops.LEARNING_RATE)
#my_optimizer = RMSprop(lr=self.ops.LEARNING_RATE) #epsilon=None, , rho=0.90, decay=0.0
model.compile(optimizer=my_optimizer,
loss=losses,
loss_weights=loss_weights)
return model
def nmse(y_true, y_pred, a, c):
dim = len(c) - 1
loss_classes = [[] for i in range(dim)]
cond = [[] for i in range(dim)]
loss = [[] for i in range(dim)]
for i in range(0, dim):
loss_classes[i] = a[i] * K.square((y_pred - y_true) / (y_true))
cond[i] = K.less(y_true, c[i + 1]) & K.greater(y_true, c[i])
loss[0] = K.switch(cond[0], loss_classes[0], loss_classes[1])
for i in range(1, dim):
loss[i] = K.switch(cond[i], loss_classes[i], loss[i - 1])
return K.mean(loss[dim - 1], axis=-1)
def check_carpisma(x):
cars = tf.strided_slice(x, [0, 0, 0, 3], [1, 210, 160, 40], [1, 1, 1, 4])
cars = K.sum(cars, axis=3, keepdims=True)
tavuks = tf.strided_slice(x, [0, 0, 0, 43], [1, 210, 160, 52],
[1, 1, 1, 4])
tavuks = K.sum(tavuks, axis=3, keepdims=True)
#cars = K.sum(x[:,:,:,3:4:40],axis=3,keepdims=True)
#tavuks = K.sum(x[:,:,:,43:4:52],axis=3,keepdims=True)
#cars_capped = K.switch(K.greater(cars, 0.1), K.ones_like(cars), K.zeros_like(cars))
#tavuks_capped = K.switch(K.greater(tavuks, 0.1), K.ones_like(tavuks), K.zeros_like(tavuks))
cars_capped = K.switch(cars > 0.1, K.maximum(cars, 1), cars * 0)
tavuks_capped = K.switch(tavuks > 0.1, K.maximum(tavuks, 1), tavuks * 0)
carpisma = K.sum(cars_capped * tavuks_capped, axis=[1, 2, 3])
carpisma = K.minimum(carpisma, 1)
return carpisma
def get_updates(self, params, loss):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.inital_decay > 0:
lr *= (1. / (1. + self.decay * self.iterations))
t = self.iterations + 1
lr_t = lr * K.sqrt(1. - K.pow(self.beta_2, t)) / (
1. - K.pow(self.beta_1, t))
shapes = [K.get_variable_shape(p) for p in params]
ms = [K.zeros(shape) for shape in shapes]
vs = [K.zeros(shape) for shape in shapes]
f = K.variable(0)
d = K.variable(1)
self.weights = [self.iterations] + ms + vs + [f, d]
cond = K.greater(t, K.variable(1))
small_delta_t = K.switch(K.greater(loss, f), self.small_k + 1,
1. / (self.big_K + 1))
big_delta_t = K.switch(K.greater(loss, f), self.big_K + 1,
1. / (self.small_k + 1))
c_t = K.minimum(K.maximum(small_delta_t, loss / (f + self.epsilon)),
big_delta_t)
f_t = c_t * f
r_t = K.abs(f_t - f) / (K.minimum(f_t, f))
d_t = self.beta_3 * d + (1 - self.beta_3) * r_t
f_t = K.switch(cond, f_t, loss)
d_t = K.switch(cond, d_t, K.variable(1.))
self.updates.append(K.update(f, f_t))
self.updates.append(K.update(d, d_t))
for p, g, m, v in zip(params, grads, ms, vs):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
p_t = p - lr_t * m_t / (d_t * K.sqrt(v_t) + self.epsilon)
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
self.updates.append(K.update(p, new_p))
return self.updates
def piecewise_linear(t, schedule, from_zero=True):
"""分段线性函数
其中schedule是形如{1000: 1, 2000: 0.1}的字典，
表示 t ∈ [0, 1000]时，输出从0均匀增加至1，而
t ∈ [1000, 2000]时，输出从1均匀降低到0.1，最后
t > 2000时，保持0.1不变。
"""
schedule = sorted(schedule.items())
if from_zero and schedule[0][0] != 0:
schedule = [(0, 0.0)] + schedule
t = K.cast(t, K.floatx())
x = (t * 0 + 1) * schedule[0][1]
for i in range(len(schedule)):
t_begin = schedule[i][0]
x_begin = x
if i != len(schedule) - 1:
dx = schedule[i + 1][1] - schedule[i][1]
dt = schedule[i + 1][0] - schedule[i][0]
slope = 1.0 * dx / dt
x = schedule[i][1] + slope * (t - t_begin)
else:
x = (t * 0 + 1) * schedule[i][1]
x = K.switch(t >= t_begin, x, x_begin)
return x
def new_get_updates(self, loss, params):
self.accumulated_grads = [
K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params
]
iters = K.update_add(self.accumulated_iters, 1)
new_grads = orig_get_gradients(loss, params)
if do_mean:
new_grads = [g / K.cast(iter_size, K.dtype(g)) for g in new_grads]
self.updated_grads = [
K.update_add(p, g)
for p, g in zip(self.accumulated_grads, new_grads)
]
def update_function():
with tf.control_dependencies(orig_get_updates(loss, params)):
reset_grads = [
K.update(p, K.zeros(K.int_shape(p), dtype=K.dtype(p)))
for p in self.accumulated_grads
]
return tf.group(*(reset_grads + [iters]))
def just_store_function():
return tf.group(*[iters])
update_switch = K.equal(iters % iter_size, 0)
with tf.control_dependencies(self.updated_grads):
self.updates = [
K.switch(update_switch, update_function, just_store_function)
]
return self.updates
def rpn_bbox_loss_graph(config, target_bbox, rpn_match, rpn_bbox):
"""Return the RPN bounding box loss graph.
config: the model config object.
target_bbox: [batch, max positive anchors, (dy, dx, log(dh), log(dw))].
Uses 0 padding to fill in unsed bbox deltas.
rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
-1=negative, 0=neutral anchor.
rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))]
"""
# Positive anchors contribute to the loss, but negative and
# neutral anchors (match value of 0 or -1) don't.
rpn_match = K.squeeze(rpn_match, -1)
indices = tf.where(K.equal(rpn_match, 1))
# Pick bbox deltas that contribute to the loss
rpn_bbox = tf.gather_nd(rpn_bbox, indices)
# Trim target bounding box deltas to the same length as rpn_bbox.
batch_counts = K.sum(K.cast(K.equal(rpn_match, 1), tf.int32), axis=1)
target_bbox = batch_pack_graph(target_bbox, batch_counts,
config.IMAGES_PER_GPU)
loss = smooth_l1_loss(target_bbox, rpn_bbox)
loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
return loss
def frcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
"""Loss for Faster R-CNN bounding box refinement.
target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
target_class_ids: [batch, num_rois]. Integer class IDs.
pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
"""
# Reshape to merge batch and roi dimensions for simplicity.
target_class_ids = K.reshape(target_class_ids, (-1,))
target_bbox = K.reshape(target_bbox, (-1, 4))
pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))
# Only positive ROIs contribute to the loss. And only
# the right class_id of each ROI. Get their indices.
positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
positive_roi_class_ids = tf.cast(
tf.gather(target_class_ids, positive_roi_ix), tf.int64)
indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)
# Gather the deltas (predicted and true) that contribute to loss
target_bbox = tf.gather(target_bbox, positive_roi_ix)
pred_bbox = tf.gather_nd(pred_bbox, indices)
# Smooth-L1 Loss
loss = K.switch(tf.size(target_bbox) > 0,
smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
tf.constant(0.0))
loss = K.mean(loss)
return loss
def mrcnn_mask_loss_graph(target_masks, target_class_ids, pred_masks):
"""Mask binary cross-entropy loss for the masks head.
target_masks: [batch, num_rois, height, width].
A float32 tensor of values 0 or 1. Uses zero padding to fill array.
target_class_ids: [batch, num_rois]. Integer class IDs. Zero padded.
pred_masks: [batch, proposals, height, width, num_classes] float32 tensor
with values from 0 to 1.
"""
# Reshape for simplicity. Merge first two dimensions into one.
target_class_ids = K.reshape(target_class_ids, (-1, ))
mask_shape = tf.shape(target_masks)
target_masks = K.reshape(target_masks, (-1, mask_shape[2], mask_shape[3]))
pred_shape = tf.shape(pred_masks)
pred_masks = K.reshape(pred_masks,
(-1, pred_shape[2], pred_shape[3], pred_shape[4]))
pred_masks = tf.transpose(pred_masks, [0, 3, 1, 2])
positive_ix = tf.where(target_class_ids > 0)[:, 0]
positive_class_ids = tf.cast(tf.gather(target_class_ids, positive_ix),
tf.int64)
indices = tf.stack([positive_ix, positive_class_ids], axis=1)
y_true = tf.gather(target_masks, positive_ix)
y_pred = tf.gather_nd(pred_masks, indices)
loss = K.switch(
tf.size(y_true) > 0, K.binary_crossentropy(target=y_true,
output=y_pred),
tf.constant(0.0))
loss = K.mean(loss)
return loss
def huber_loss(y, y_pred, delta: float = 1.0):
"""
Return the Huber loss between tensors.
Reference:
https://en.wikipedia.org/wiki/Huber_loss
https://web.stanford.edu/class/cs20si/2017/lectures/slides_03.pdf
https://keras.io/backend/
Args:
y: ground truth y labels
y_pred: predicted y labels
delta: the separating constant between MSE and MAE
Returns:
a scalar loss between the ground truth and predicted labels
"""
# calculate the residuals
residual = K.abs(y_pred - y)
# determine the result of the logical comparison to delta
condition = K.less_equal(residual, delta)
# calculate the two possible returns (MSE and MAE)
then_this = 0.5 * K.square(residual)
else_this = delta * residual - 0.5 * K.square(delta)
# use the condition to determine the resulting tensor
return K.switch(condition, then_this, else_this)
def dropped_mask():
drop_mask = K.switch(
K.random_uniform(K.shape(inputs)) < self.drop_rate,
K.ones_like(inputs, K.floatx()),
K.zeros_like(inputs, K.floatx()),
)
return target_mask * drop_mask
def new_get_updates(self, loss, params):
self.accumulated_grads = [
K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params
]
update_iter = [K.update_add(self.accumulated_iterations, 1)]
new_grads = self.orig_get_gradients(loss, params)
if self.do_mean:
new_grads = [
g / K.cast(self.iter_size, K.dtype(g)) for g in new_grads
]
update_grads = [
K.update_add(p, g)
for p, g in zip(self.accumulated_grads, new_grads)
]
def update_func():
with tf.control_dependencies(self.orig_get_updates(loss, params)):
reset_grads = [
K.update(p, K.zeros(K.int_shape(p), dtype=K.dtype(p)))
for p in self.accumulated_grads
]
return tf.group(*(reset_grads + update_iter))
def just_iter_func():
return tf.group(*update_iter)
# do the original get_updates() computations only once every
# 'iter_size' iterations
update_switch = K.equal(self.accumulated_iterations % self.iter_size,
0)
with tf.control_dependencies(update_grads):
self.updates = [
K.switch(update_switch, update_func, just_iter_func)
]
return self.updates