def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay *
K.cast(self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
# Learning rate multipliers
if self.multipliers:
multiplier = [
mult for mult in self.multipliers if mult in p.name
]
else:
multiplier = None
if multiplier:
new_lr_t = lr_t * self.multipliers[multiplier[0]]
if self.debug_verbose:
print('Setting {} to learning rate {}'.format(
multiplier[0], new_lr_t))
print(K.get_value(new_lr_t))
else:
new_lr_t = lr_t
if self.debug_verbose:
print('No change in learning rate {}'.format(p.name))
print(K.get_value(new_lr_t))
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = p - new_lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
p_t = p - new_lr_t * m_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
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay *
K.cast(self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
# Applies bounds on actual learning rate
step_size = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
final_lr = self.final_lr * lr / self.base_lr
lower_bound = final_lr * (1. - 1. / (self.gamma * t + 1.))
upper_bound = final_lr * (1. + 1. / (self.gamma * t))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsbound:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
# apply weight decay
if self.weight_decay != 0.:
g += self.weight_decay * K.stop_gradient(p)
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
if self.amsbound:
vhat_t = K.maximum(vhat, v_t)
denom = (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
denom = (K.sqrt(v_t) + self.epsilon)
# Compute the bounds
step_size_p = step_size * K.ones_like(denom)
step_size_p_bound = step_size_p / denom
bounded_lr_t = m_t * K.minimum(
K.maximum(step_size_p_bound, lower_bound), upper_bound)
p_t = p - bounded_lr_t
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def focal_loss(y_true, y_pred):
gamma = 2.0
alpha = 0.25
pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1)) - K.sum(
(1 - alpha) * K.pow(pt_0, gamma) * K.log(1. - pt_0))
def build_loss(self):
# Infinity norm
if np.isinf(self.p):
value = K.max(self.img)
else:
value = K.pow(K.sum(K.pow(K.abs(self.img), self.p)), 1. / self.p)
return normalize(self.img, value)
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
'''Bias corrections according to the Adam paper
'''
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
self.weights = [self.iterations] + ms + vs
for p, g, m, v in zip(params, grads, ms, vs):
####################################################
# Add a lr multiplier for vars outside excluded_vars
if p.name in self.excluded_vars:
multiplied_lr_t = lr_t
else:
multiplied_lr_t = lr_t * self.lr_mult
###################################################
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
'''Schedule multiplier eta_t = 1 for simple AdamW
According to the AdamW paper, eta_t can be fixed, decay, or
also be used for warm restarts (AdamWR to come).
'''
eta_t = 1.
p_t = p - eta_t * (multiplied_lr_t * m_t / (K.sqrt(v_t) + self.epsilon))
if self.weight_decay != 0:
'''Normalized weight decay according to the AdamW paper
'''
w_d = self.weight_decay * K.sqrt(self.batch_size / (self.samples_per_epoch * self.epochs))
p_t = p_t - eta_t * (w_d * p)
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_updates_Padam(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
base_lr = self._optimizer.learning_rate
if self.initial_decay > 0:
base_lr = base_lr * (1. / (1. + self.decay * K.cast(
self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
lr_t = base_lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
if self._get_multiplier(p) is None:
multiplier = 1.0
else:
multiplier = self._get_multiplier(p)
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
denom = (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
denom = (K.sqrt(v_t) + self.epsilon)
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
# Partial momentum adaption.
new_p = p - (lr_t * multiplier * (m_t /
(denom**(self.partial * 2))))
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def gelu(x):
"""
GELU activation, described in paper "Gaussian Error Linear Units (GELUs)"
https://arxiv.org/pdf/1606.08415.pdf
"""
c = math.sqrt(2 / math.pi)
return 0.5 * x * (1 + K.tanh(c * (x + 0.044715 * K.pow(x, 3))))
def binary_focal_loss_fixed(y_true, y_pred):
"""
:param y_true: A tensor of the same shape as `y_pred`
:param y_pred: A tensor resulting from a sigmoid
:return: Output tensor.
"""
pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
epsilon = K.epsilon()
# clip to prevent NaN's and Inf's
pt_1 = K.clip(pt_1, epsilon, 1. - epsilon)
pt_0 = K.clip(pt_0, epsilon, 1. - epsilon)
return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1)) \
-K.sum((1 - alpha) * K.pow(pt_0, gamma) * K.log(1. - pt_0))
def build_loss(self):
r"""Implements the N-dim version of function
$$TV^{\beta}(x) = \sum_{whc} \left ( \left ( x(h, w+1, c) - x(h, w, c) \right )^{2} +
\left ( x(h+1, w, c) - x(h, w, c) \right )^{2} \right )^{\frac{\beta}{2}}$$
to return total variation for all images in the batch.
"""
image_dims = K.ndim(self.img) - 2
# Constructing slice [1:] + [:-1] * (image_dims - 1) and [:-1] * (image_dims)
start_slice = [slice(1, None, None)] + [
slice(None, -1, None) for _ in range(image_dims - 1)
]
end_slice = [slice(None, -1, None) for _ in range(image_dims)]
samples_channels_slice = [
slice(None, None, None),
slice(None, None, None)
]
# Compute pixel diffs by rolling slices to the right per image dim.
tv = None
for i in range(image_dims):
ss = tuple(samples_channels_slice + start_slice)
es = tuple(samples_channels_slice + end_slice)
diff_square = K.square(self.img[utils.slicer[ss]] -
self.img[utils.slicer[es]])
tv = diff_square if tv is None else tv + diff_square
# Roll over to next image dim
start_slice = np.roll(start_slice, 1).tolist()
end_slice = np.roll(end_slice, 1).tolist()
tv = K.sum(K.pow(tv, self.beta / 2.))
return normalize(self.img, tv)
def focal_loss(y_true, y_pred):
# Define espislon so that the backpropagation will not result int NaN
# for 0 divisor case
epsilon = K.epsilon()
# Add the epsilon to prediction value
# y_pred = y_pred + epsilon
# Clip the prediction value
y_pred = K.clip(y_pred, epsilon, 1.0 - epsilon)
alpha_factor = K.ones_like(y_true) * alpha
# Calculate p_t
p_t = tf.where(K.equal(y_true, 1), alpha_factor, 1 - alpha_factor)
# Calculate alpha_t
alpha_t = tf.where(K.equal(y_true, 1), alpha_factor, 1 - alpha_factor)
# Calculate cross entropy
cross_entropy = -K.log(p_t)
weight = alpha_t * K.pow((1 - p_t), gamma)
# Calculate focal loss
loss = weight * cross_entropy
# Sum the losses in mini_batch
loss = K.sum(loss, axis=1)
return loss
def smooth_l1(y_true, y_pred, sigma=3.0, axis=None):
"""Compute the smooth L1 loss of y_pred w.r.t. y_true.
Args:
y_true: Tensor from the generator of shape (B, N, 5).
The last value for each box is the state of the anchor
(ignore, negative, positive).
y_pred: Tensor from the network of shape (B, N, 4).
sigma: The point where the loss changes from L2 to L1.
Returns:
The smooth L1 loss of y_pred w.r.t. y_true.
"""
if axis is None:
axis = 1 if K.image_data_format(
) == 'channels_first' else K.ndim(y_pred) - 1
sigma_squared = sigma**2
# compute smooth L1 loss
# f(x) = 0.5 * (sigma * x)^2 if |x| < 1 / sigma / sigma
# |x| - 0.5 / sigma / sigma otherwise
regression_diff = K.abs(y_true - y_pred) # |y - f(x)|
regression_loss = tf.where(K.less(regression_diff, 1.0 / sigma_squared),
0.5 * sigma_squared * K.pow(regression_diff, 2),
regression_diff - 0.5 / sigma_squared)
return K.sum(regression_loss, axis=axis)
def continuity_loss(x, im_height, im_width):
assert K.ndim(x) == 4
a = K.square(x[:, :im_height - 1, :im_width - 1, :] -
x[:, 1:, :im_width - 1, :])
b = K.square(x[:, :im_height - 1, :im_width - 1, :] -
x[:, :im_height - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
def weighted_focal_loss(y_true, y_pred, n_classes=3, gamma=2., axis=None, from_logits=False):
"""Focal loss between an output tensor and a target tensor.
Automatically computes the class weights from the target image and uses
them to weight the cross entropy
Args:
y_true: A tensor of the same shape as y_pred.
y_pred: A tensor resulting from a softmax
(unless from_logits is True, in which
case y_pred is expected to be the logits).
from_logits: Boolean, whether y_pred is the
result of a softmax, or is a tensor of logits.
Returns:
tensor: Output tensor.
"""
if from_logits:
raise Exception('weighted_focal_loss cannot take logits')
if axis is None:
axis = 1 if K.image_data_format() == 'channels_first' else K.ndim(y_pred) - 1
reduce_axis = [x for x in list(range(K.ndim(y_pred))) if x != axis]
# scale preds so that the class probas of each sample sum to 1
y_pred = y_pred / K.sum(y_pred, axis=axis, keepdims=True)
# manual computation of crossentropy
_epsilon = tf.convert_to_tensor(K.epsilon(), y_pred.dtype.base_dtype)
y_pred = tf.clip_by_value(y_pred, _epsilon, 1. - _epsilon)
y_true_cast = K.cast(y_true, K.floatx())
total_sum = K.sum(y_true_cast)
class_sum = K.sum(y_true_cast, axis=reduce_axis, keepdims=True)
class_weights = 1.0 / K.cast_to_floatx(n_classes) * tf.divide(total_sum, class_sum + 1.)
temp_loss = (K.pow(1. - y_pred, gamma) * K.log(y_pred) * class_weights)
focal_loss = - K.sum(y_true * temp_loss, axis=axis)
return focal_loss
def total_variation_loss(x):
assert K.ndim(x) == 4
a = K.square(x[:, :, 1:, :img_width - 1] -
x[:, :, :img_height - 1, :img_width - 1])
b = K.square(x[:, :, :img_height - 1, 1:] -
x[:, :, :img_width - 1, :img_height - 1])
#a = K.square(x[:, :, :img_width-1, :img_height-1] - x[:, :, 1:, :img_height-1])
#b = K.square(x[:, :, :img_width-1, :img_height-1] - x[:, :, :img_width-1, 1:])
return K.sum(K.pow(a + b, 1.25))
def total_variation_loss(x):
assert K.ndim(x) == 4
if K.image_data_format() == "channels_first":
a = K.square(x[:, :, :img_width - 1, :img_height - 1] - x[:, :, 1:, :img_height - 1])
b = K.square(x[:, :, :img_width - 1, :img_height - 1] - x[:, :, :img_width - 1, 1:])
else:
a = K.square(x[:, :img_width - 1, :img_height - 1, :] - x[:, 1:, :img_height - 1, :])
b = K.square(x[:, :img_width - 1, :img_height - 1, :] - x[:, :img_width - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay *
K.cast(self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
beta_1_t = K.pow(self.beta_1, t)
beta_2_t = K.pow(self.beta_2, t)
rho = 2 / (1 - self.beta_2) - 1
rho_t = rho - 2 * t * beta_2_t / (1 - beta_2_t)
r_t = K.sqrt(
K.relu(rho_t - 4) * K.relu(rho_t - 2) * rho / ((rho - 4) *
(rho - 2) * rho_t))
flag = K.cast(rho_t > 4, K.floatx())
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
self.weights = [self.iterations] + ms + vs
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)
mhat_t = m_t / (1 - beta_1_t)
vhat_t = K.sqrt(v_t / (1 - beta_2_t))
p_t = p - lr * mhat_t * (flag * r_t / (vhat_t + self.epsilon) +
(1 - flag))
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def custom_loss(self, y_true, y_pred):
"""
GloVe's loss function, view section 3.1 on the original paper for details.
:param y_true: The actual values, y_true = X_ij.
:param y_pred: The predicted occurrences from the model ( w_i^T*w_j ).
:return: The loss associated with this batch.
"""
x_max = self.x_max
alpha = self.alpha
fxij = k.pow(k.clip(y_true / x_max, 0.0, 1.0), alpha)
return k.sum(fxij * k.square(y_pred - k.log(y_true)), axis=-1)
def total_variation_loss(x):
assert 4 == K.ndim(x)
if K.image_dim_ordering() == 'th':
a = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, 1:, :img_ncols - 1])
b = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, :img_nrows - 1, 1:])
else:
a = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, 1:, :img_ncols - 1, :])
b = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, :img_nrows - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
def total_variation_loss(x, img_nrows, img_ncols):
assert 4 == K.ndim(x)
if K.image_data_format() == 'channels_first':
a = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, 1:, :img_ncols - 1])
b = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, :img_nrows - 1, 1:])
else:
a = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, 1:, :img_ncols - 1, :])
b = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, :img_nrows - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
def focal_crossentropy(y_true, y_pred):
bce = K.binary_crossentropy(y_true, y_pred)
y_pred = K.clip(y_pred, K.epsilon(), 1.- K.epsilon())
p_t = (y_true*y_pred) + ((1-y_true)*(1-y_pred))
alpha_factor = 1
modulating_factor = 1
alpha_factor = y_true*alpha + ((1-alpha)*(1-y_true))
modulating_factor = K.pow((1-p_t), gamma)
# compute the final loss and return
return K.mean(alpha_factor*modulating_factor*bce, axis=-1)
def call(self, inputs, **kwargs):
length = K.shape(inputs[0])[1] + K.shape(inputs[1])[1]
inputs = K.tile(
K.expand_dims(K.arange(length - 1, -1, -1, dtype=K.floatx()), axis=0),
[K.shape(inputs[0])[0], 1],
)
if self.clamp_len is not None:
inputs = K.clip(inputs, min_value=0, max_value=self.clamp_len)
inputs = K.expand_dims(inputs, axis=-1)
output_dim = K.cast(self.output_dim, K.floatx())
ranges = K.expand_dims(K.arange(0.0, self.output_dim, 2.0), axis=0) / output_dim
inverse = 1.0 / K.pow(10000.0, ranges)
positions = inputs * inverse
return K.concatenate([K.sin(positions), K.cos(positions)], axis=-1)
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
# decoupled weight decay (4/6)
wd = self.wd
lr = self.lr
if self.initial_decay > 0:
lr *= (1. / (1. + self.decay * K.cast(self.iterations,
K.dtype(self.decay))))
# decoupled weight decay (5/6)
eta_t = lr / self.init_lr
t = K.cast(self.iterations, K.floatx()) + 1
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
self.weights = [self.iterations] + ms + vs
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)
# decoupled weight decay (6/6)
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon) - eta_t * wd * p
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def __call__(self, x):
assert K.ndim(x.output) == 4
x_out = x.output
shape = K.shape(x_out)
img_width, img_height,channel = (shape[1],shape[2], shape[3])
size = img_width * img_height * channel
if image_dim_ordering() == 'th':
a = K.square(x_out[:, :, :img_width - 1, :img_height - 1] - x_out[:, :, 1:, :img_height - 1])
b = K.square(x_out[:, :, :img_width - 1, :img_height - 1] - x_out[:, :, :img_width - 1, 1:])
else:
a = K.square(x_out[:, :img_width - 1, :img_height - 1, :] - x_out[:, 1:, :img_height - 1, :])
b = K.square(x_out[:, :img_width - 1, :img_height - 1, :] - x_out[:, :img_width - 1, 1:, :])
loss = self.weight * K.sum(K.pow(a + b, 1.25))
return loss
def total_variation_loss(x):
img_nrows = self.img_shape[1]
img_ncols = self.img_shape[2]
assert K.ndim(x) == 4
if K.image_data_format() == 'channels_first':
a = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, 1:, :img_ncols - 1])
b = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] -
x[:, :, :img_nrows - 1, 1:])
else:
a = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, 1:, :img_ncols - 1, :])
b = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] -
x[:, :img_nrows - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
def call(self, x, **kwargs):
if (self.size is None) or (self.mode == 'sum'):
self.size = int(x.shape[-1])
batch_size, seq_len = K.shape(x)[0], K.shape(x)[1]
position_j = 1. / K.pow(
10000., 2 * K.arange(self.size / 2, dtype='float32') / self.size)
position_j = K.expand_dims(position_j, 0)
# K.arange不支持变长，只好用这种方法生成
position_i = K.cumsum(K.ones_like(x[:, :, 0]), 1) - 1
position_i = K.expand_dims(position_i, 2)
position_ij = K.dot(position_i, position_j)
position_ij = K.concatenate(
[K.cos(position_ij), K.sin(position_ij)], 2)
if self.mode == 'sum':
return position_ij + x
elif self.mode == 'concat':
return K.concatenate([position_ij, x], 2)
def ssd_loss(y_true, y_pred): #ssd_loss
num_classes = 11 # tf.shape(y_true)[2] - 4 # openCV에서 동적 shape를 지원안함.
y_true = tf.reshape(y_true, [-1, num_classes + 4])
y_pred = tf.reshape(y_pred, [-1, num_classes - 1 + 4])
eps = K.epsilon()
# Split Classification and Localization output
y_true_clf, y_true_loc = tf.split(y_true, [num_classes, 4], axis=-1)
y_pred_clf, y_pred_loc = tf.split(y_pred, [num_classes - 1, 4], axis=-1)
# split foreground & background
mask = y_true_clf[:, -1]
ignore_mask = tf.where(tf.equal(mask, -1.),
tf.zeros_like(mask),
tf.ones_like(mask))
neg_mask = tf.where(tf.equal(mask, 1.),
tf.ones_like(mask),
tf.zeros_like(mask))
pos_mask = tf.where(tf.equal(mask, 0.),
tf.ones_like(mask),
tf.zeros_like(mask))
y_true_clf = tf.where(tf.not_equal(y_true_clf, 0),
tf.ones_like(y_true_clf),
tf.zeros_like(y_true_clf))
num_pos = tf.reduce_sum(pos_mask)
num_neg = tf.reduce_sum(neg_mask)
# Focal Loss
y_pred_clf = K.clip(y_pred_clf, eps, 1. - eps)
pt = tf.where(tf.equal(y_true_clf[:, :num_classes - 1], 1.),
y_pred_clf, 1. - y_pred_clf)
loss = -K.pow(1. - pt, gamma) * tf.log(pt)
clf_loss = tf.reduce_sum(alpha * loss, axis=-1)
clf_loss = tf.reduce_sum(ignore_mask * clf_loss) / (num_pos + num_neg + eps)
# smooth l1 loss
l1_loss = tf.abs(y_true_loc - y_pred_loc)
l2_loss = 0.5 * (y_true_loc - y_pred_loc) ** 2
loc_loss = tf.where(tf.less(l1_loss, 1.0),
l2_loss,
l1_loss - 0.5)
loc_loss = tf.reduce_mean(loc_loss, axis=-1)
loc_loss = tf.reduce_sum(loc_loss * pos_mask) / (num_pos + eps)
# total loss
return alpha * clf_loss + loc_loss
def focal_loss(y_true, y_pred):
# Define epsilon so that the backpropagation will no result in NaN
# for o divisor case
epsilon = K.epsilon()
# Add the epsilon to prediction value
# y_pred = y_pred + epsilon
# Clip the prediction value
y_pred = K.clip(y_pred, epsilon, 1.0 - epsilon)
# Calculate cross entropy
cross_entropy = -y_true * K.log(y_pred)
# Calculate weight that consists of modulating factor and weighting factor
weight = alpha * y_true * K.pow((1 - y_pred), gamma)
# Calculate focal loss
loss = weight * cross_entropy
# Sum the losses in mini_batch
loss = K.sum(loss, axis=1)
return loss
def call(self, x, mask=None):
if (self.size == None) or (self.mode == 'sum'):
self.size = int(x.shape[-1])
position_j = 1. / \
K.pow(10000., 2 * K.arange(self.size / 2, dtype='float32') / self.size)
position_j = K.expand_dims(position_j, 0)
position_i = tf.cumsum(K.ones_like(x[:, :, 0]), 1) - 1
position_i = K.expand_dims(position_i, 2)
position_ij = K.dot(position_i, position_j)
outputs = K.concatenate([K.cos(position_ij), K.sin(position_ij)], 2)
if self.mode == 'sum':
if self.scale:
outputs = outputs * self.size**0.5
return x + outputs
elif self.mode == 'concat':
return K.concatenate([outputs, x], 2)
def categorical_focal_loss_fixed(y_true, y_pred):
"""
:param y_true: A tensor of the same shape as `y_pred`
:param y_pred: A tensor resulting from a softmax
:return: Output tensor.
"""
# Scale predictions so that the class probas of each sample sum to 1
y_pred /= K.sum(y_pred, axis=-1, keepdims=True)
# Clip the prediction value to prevent NaN's and Inf's
epsilon = K.epsilon()
y_pred = K.clip(y_pred, epsilon, 1. - epsilon)
# Calculate Cross Entropy
cross_entropy = -y_true * K.log(y_pred)
# Calculate Focal Loss
loss = alpha * K.pow(1 - y_pred, gamma) * cross_entropy
# Sum the losses in mini_batch
return K.sum(loss, axis=1)
def focal_loss(y_true, y_pred, gamma=2., alpha=.25):
pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
return K.mean(-alpha * K.pow(1 - pt_1, gamma) * K.log(pt_1) -
(1 - alpha) * K.pow(pt_0, gamma) * K.log(1 - pt_0),
axis=-1)