def call(self, inputs, mask=None, **kwargs):
"""Core implemention of soft attention
Args:
inputs (object): input tensor.
Returns:
object: weighted sum of input tensors.
"""
attention = K.tanh(K.dot(inputs, self.W) + self.b)
attention = K.dot(attention, self.q)
attention = K.squeeze(attention, axis=2)
if mask is None:
attention = K.exp(attention)
else:
attention = K.exp(attention) * K.cast(mask, dtype="float32")
attention_weight = attention / (
K.sum(attention, axis=-1, keepdims=True) + K.epsilon())
attention_weight = K.expand_dims(attention_weight)
weighted_input = inputs * attention_weight
return K.sum(weighted_input, axis=1)
def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
"""Convert final layer features to bounding box parameters."""
num_anchors = anchors_per_level
# Reshape to batch, height, width, num_anchors, box_params.
anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])
grid_shape = K.shape(feats)[1:3] # height, width
grid_y = K.tile(
tf.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1],
name='yolo_head/tile/reshape/grid_y'),
[1, grid_shape[1], 1, 1])
grid_x = K.tile(
tf.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1],
name='yolo_head/tile/reshape/grid_x'),
[grid_shape[0], 1, 1, 1])
grid = tf.concat([grid_x, grid_y],
axis=-1,
name='yolo_head/concatenate/grid')
grid = K.cast(grid, K.dtype(feats))
feats = tf.reshape(feats, [
-1, grid_shape[0], grid_shape[1], num_anchors,
num_classes + 5 + NUM_ANGLES3
],
name='yolo_head/reshape/feats')
# Adjust predictions to each spatial grid point and anchor size.
box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(
grid_shape[..., ::-1], K.dtype(feats))
box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(
input_shape[..., ::-1], K.dtype(feats))
box_confidence = K.sigmoid(feats[..., 4:5])
box_class_probs = K.sigmoid(feats[..., 5:5 + num_classes])
polygons_confidence = K.sigmoid(feats[..., 5 + num_classes + 2:5 +
num_classes + NUM_ANGLES3:3])
polygons_x = K.exp(feats[...,
5 + num_classes:num_classes + 5 + NUM_ANGLES3:3])
dx = K.square(anchors_tensor[..., 0:1] / 2)
dy = K.square(anchors_tensor[..., 1:2] / 2)
d = K.cast(K.sqrt(dx + dy), K.dtype(polygons_x))
a = K.pow(input_shape[..., ::-1], 2)
a = K.cast(a, K.dtype(feats))
b = K.sum(a)
diagonal = K.cast(K.sqrt(b), K.dtype(feats))
polygons_x = polygons_x * d / diagonal
polygons_y = feats[...,
5 + num_classes + 1:num_classes + 5 + NUM_ANGLES3:3]
polygons_y = K.sigmoid(polygons_y)
if calc_loss == True:
return grid, feats, box_xy, box_wh, polygons_confidence
return box_xy, box_wh, box_confidence, box_class_probs, polygons_x, polygons_y, polygons_confidence
def yolo_head(feats, anchors, num_classes, input_shape):
"""Convert final layer features to bounding box parameters."""
num_anchors = len(anchors)
# Reshape to batch, height, width, num_anchors, box_params.
anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])
grid_shape = K.shape(feats)[1:3] # height, width
grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
[1, grid_shape[1], 1, 1])
grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
[grid_shape[0], 1, 1, 1])
grid = K.concatenate([grid_x, grid_y])
grid = K.cast(grid, K.dtype(feats))
feats = K.reshape(
feats,
[-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])
box_xy = K.sigmoid(feats[..., :2])
box_wh = K.exp(feats[..., 2:4])
box_confidence = K.sigmoid(feats[..., 4:5])
box_class_probs = K.sigmoid(feats[..., 5:])
# Adjust preditions to each spatial grid point and anchor size.
box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1],
K.dtype(feats))
return box_xy, box_wh, box_confidence, box_class_probs
def yolo_head(feats, anchors, num_classes):
# convert anchors to shape 1 , 1 , 1 , len of anchors , 2
num_anchors = len(anchors)
anchors_tensor = K.reshape(K.variable(anchors), [1, 1, 1, num_anchors, 2])
# conv_dims , width and hight of the grid
_, conv_height, conv_width, _ = K.int_shape(feats)
conv_dims = K.variable([conv_width, conv_height])
# reshape yolo network output to None , grid_width , grid_hight , num of amnchors , num of classes + 5
feats = K.reshape(
feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5])
# convert conv_dims after casting it to feats datatype to 1 , 1 , 1 , 2
conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats))
# create grid from (0 , 0 ) to (width , hight)
conv_index = np.array([_ for _ in np.ndindex(conv_width, conv_height)])
conv_index = conv_index[:, [1, 0]] # swap columns for YOLO ordering.
conv_index = K.variable(
conv_index.reshape(1, conv_height, conv_width, 1, 2))
box_confidence = K.sigmoid(feats[..., 4:5])
box_xy = K.sigmoid(feats[..., :2])
box_wh = K.exp(feats[..., 2:4])
box_class_probs = K.softmax(feats[..., 5:])
box_xy = (box_xy + conv_index) / conv_dims
box_wh = box_wh * anchors_tensor / conv_dims
return box_confidence, box_xy, box_wh, box_class_probs
def vae_loss(self, x, z_decoded):
x = K.flatten(x)
z_decoded = K.flatten(z_decoded)
# Reconstruction loss (as we used sigmoid activation we can use binarycrossentropy)
recon_loss = keras.metrics.binary_crossentropy(x, z_decoded)
# KL divergence
kl_loss = -5e-4 * K.mean(1 + z_sigma - K.square(z_mu) - K.exp(z_sigma), axis=-1)
return K.mean(recon_loss + kl_loss)
def get_log_probability_density(pred, y):
mu_and_sigma = pred
mu = mu_and_sigma[:, :2]
sigma = mu_and_sigma[:, 2:]
variance = K.square(sigma)
pdf = 1. / K.sqrt(2. * np.pi * variance) * K.exp(-K.square(y - mu) /
(2. * variance))
log_pdf = K.log(pdf + K.epsilon())
return log_pdf
def full_affinity(input_x, scale):
"""Calculates the symmetrized full Gaussian affinity matrix, scaled by a provided scale.
Args:
input_x: input dataset of size n x d
scale: provided scale
Returns:
n x n affinity matrix
"""
sigma = K.variable(scale)
dist_x = squared_distance(input_x)
sigma_squared = K.expand_dims(K.pow(sigma, 2), -1)
weight_mat = K.exp(-dist_x / (2 * sigma_squared))
return weight_mat
def full_affinity(X, scale):
'''
Calculates the symmetrized full Gaussian affinity matrix, scaled
by a provided scale
X: input dataset of size n
scale: provided scale
returns: n x n affinity matrix
'''
sigma = K.variable(scale)
Dx = squared_distance(X)
sigma_squared = K.pow(sigma, 2)
sigma_squared = K.expand_dims(sigma_squared, -1)
Dx_scaled = Dx / (2 * sigma_squared)
W = K.exp(-Dx_scaled)
return W
def _yolo_head(self,
feats,
anchors,
num_classes,
input_shape,
calc_loss=False):
"""Convert final layer features to bounding box parameters"""
num_anchors = len(anchors)
anchors_tensor = K.reshape(K.constant(anchors),
[1, 1, 1, num_anchors, 2])
# height, width
grid_shape = K.shape(feats)[1:3]
grid_y = K.tile(
K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
[1, grid_shape[1], 1, 1],
)
grid_x = K.tile(
K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
[grid_shape[0], 1, 1, 1],
)
grid = K.concatenate([grid_x, grid_y])
grid = K.cast(grid, K.dtype(feats))
feats = K.reshape(
feats,
[-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])
# Adjust preditions to each spatial grid point and anchor size.
box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(
grid_shape[::-1], K.dtype(feats))
box_wh = (K.exp(feats[..., 2:4]) * anchors_tensor /
K.cast(input_shape[::-1], K.dtype(feats)))
box_confidence = K.sigmoid(feats[..., 4:5])
box_class_probs = K.sigmoid(feats[..., 5:])
if calc_loss == True:
return grid, feats, box_xy, box_wh
return box_xy, box_wh, box_confidence, box_class_probs
def loss(y_true, y_pred):
PPO_LOSS_CLIPPING = 0.2
PPO_ENTROPY_LOSS = 5 * 1e-3 # Does not converge without entropy penalty
log_pdf_new = get_log_probability_density(y_pred, y_true)
log_pdf_old = get_log_probability_density(old_prediction, y_true)
ratio = K.exp(log_pdf_new - log_pdf_old)
surrogate1 = ratio * advantage
clip_ratio = K.clip(ratio,
min_value=(1 - PPO_LOSS_CLIPPING),
max_value=(1 + PPO_LOSS_CLIPPING))
surrogate2 = clip_ratio * advantage
loss_actor = -K.mean(K.minimum(surrogate1, surrogate2))
sigma = y_pred[:, 2:]
variance = K.square(sigma)
loss_entropy = PPO_ENTROPY_LOSS * K.mean(
-(K.log(2 * np.pi * variance) + 1) / 2)
return loss_actor + loss_entropy
def __softmax_cosine_loss(y_true, y_pred, centers):
y_true_value = K.argmax(y_true)
base_loss = K.variable(0.0)
for label in range(CONFIG["num_classes"]):
center = centers[label]
indices = tf.where(tf.equal(y_true_value, label))
pred_per_class = tf.gather_nd(y_pred, indices=indices)
distances_per_class = __cosine(pred_per_class, center)
sum = K.sum(distances_per_class, axis=-1)
base_loss = tf.add(base_loss, sum)
base_distances = K.zeros_like(y_true_value, dtype='float32')
for label in range(CONFIG["num_classes"]):
center = centers[label]
distances_with_all_classes = __cosine(y_pred, center)
exp_distances = K.exp(-distances_with_all_classes)
base_distances = tf.add(base_distances, exp_distances)
log_distances = K.log(base_distances)
sum_on_batch = K.sum(log_distances)
return base_loss + sum_on_batch
def call(self, x, mask=None):
features_dim = self.features_dim
step_dim = self.step_dim
e = K.reshape(
K.dot(K.reshape(x, (-1, features_dim)),
K.reshape(self.W, (features_dim, 1))),
(-1, step_dim)) # e = K.dot(x, self.W)
if self.bias:
e += self.b
e = K.tanh(e)
a = K.exp(e)
# apply mask after the exp. will be re-normalized next
if mask is not None:
# cast the mask to floatX to avoid float64 upcasting in theano
a *= K.cast(mask, K.floatx())
# in some cases especially in the early stages of training the sum may be almost zero
# and this results in NaN's. A workaround is to add a very small positive number ε to the sum.
a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx())
a = K.expand_dims(a)
c = K.sum(a * x, axis=1)
return c
def yolo_head(feats, anchors, num_classes):
"""Convert final layer features to bounding box parameters.
Parameters
----------
feats : tensor
Final convolutional layer features.
anchors : array-like
Anchor box widths and heights.
num_classes : int
Number of target classes.
Returns
-------
box_xy : tensor
x, y box predictions adjusted by spatial location in conv layer.
box_wh : tensor
w, h box predictions adjusted by anchors and conv spatial resolution.
box_conf : tensor
Probability estimate for whether each box contains any object.
box_class_pred : tensor
Probability distribution estimate for each box over class labels.
"""
num_anchors = len(anchors)
# Reshape to batch, height, width, num_anchors, box_params.
anchors_tensor = K.reshape(K.variable(anchors), [1, 1, 1, num_anchors, 2])
# Static implementation for fixed models.
# TODO: Remove or add option for static implementation.
# _, conv_height, conv_width, _ = K.int_shape(feats)
# conv_dims = K.variable([conv_width, conv_height])
# Dynamic implementation of conv dims for fully convolutional model.
conv_dims = K.shape(feats)[1:3] # assuming channels last
# In YOLO the height index is the inner most iteration.
conv_height_index = K.arange(0, stop=conv_dims[0])
conv_width_index = K.arange(0, stop=conv_dims[1])
conv_height_index = K.tile(conv_height_index, [conv_dims[1]])
# TODO: Repeat_elements and tf.split doesn't support dynamic splits.
# conv_width_index = K.repeat_elements(conv_width_index, conv_dims[1], axis=0)
conv_width_index = K.tile(
K.expand_dims(conv_width_index, 0), [conv_dims[0], 1])
conv_width_index = K.flatten(K.transpose(conv_width_index))
conv_index = K.transpose(K.stack([conv_height_index, conv_width_index]))
conv_index = K.reshape(conv_index, [1, conv_dims[0], conv_dims[1], 1, 2])
conv_index = K.cast(conv_index, K.dtype(feats))
feats = K.reshape(
feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5])
conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats))
# Static generation of conv_index:
# conv_index = np.array([_ for _ in np.ndindex(conv_width, conv_height)])
# conv_index = conv_index[:, [1, 0]] # swap columns for YOLO ordering.
# conv_index = K.variable(
# conv_index.reshape(1, conv_height, conv_width, 1, 2))
# feats = Reshape(
# (conv_dims[0], conv_dims[1], num_anchors, num_classes + 5))(feats)
box_xy = K.sigmoid(feats[..., :2])
box_wh = K.exp(feats[..., 2:4])
box_confidence = K.sigmoid(feats[..., 4:5])
box_class_probs = K.softmax(feats[..., 5:])
# Adjust preditions to each spatial grid point and anchor size.
# Note: YOLO iterates over height index before width index.
box_xy = (box_xy + conv_index) / conv_dims
box_wh = box_wh * anchors_tensor / conv_dims
return box_xy, box_wh, box_confidence, box_class_probs
def sample_z(args): z_mu, z_sigma = args eps = K.random_normal(shape=(K.shape(z_mu)[0], K.int_shape(z_mu)[1])) return z_mu + K.exp(z_sigma / 2) * eps
