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 path_energy0(y, x, U, mask=None):
"""Path energy without boundary potential handling."""
n_classes = K.shape(x)[2]
y_one_hot = K.one_hot(y, n_classes)
# Tag path energy
energy = K.sum(x * y_one_hot, 2)
energy = K.sum(energy, 1)
# Transition energy
y_t = y[:, :-1]
y_tp1 = y[:, 1:]
U_flat = K.reshape(U, [-1])
# Convert 2-dim indices (y_t, y_tp1) of U to 1-dim indices of U_flat:
flat_indices = y_t * n_classes + y_tp1
U_y_t_tp1 = K.gather(U_flat, flat_indices)
if mask is not None:
mask = K.cast(mask, K.floatx())
y_t_mask = mask[:, :-1]
y_tp1_mask = mask[:, 1:]
U_y_t_tp1 *= y_t_mask * y_tp1_mask
energy += K.sum(U_y_t_tp1, axis=1)
return energy
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 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 call(self, inputs):
inputs = [K.cast(x, dtype="int32") for x in inputs]
input_ids, input_mask, segment_ids = inputs
bert_inputs = dict(input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids)
if self.pooling == "first":
pooled = self.bert(inputs=bert_inputs,
signature="tokens",
as_dict=True)["pooled_output"]
elif self.pooling == "mean":
result = self.bert(inputs=bert_inputs,
signature="tokens",
as_dict=True)["sequence_output"]
mul_mask = lambda x, m: x * tf.expand_dims(m, axis=-1)
masked_reduce_mean = lambda x, m: tf.reduce_sum(
mul_mask(x, m), axis=1) / (tf.reduce_sum(
m, axis=1, keepdims=True) + 1e-10)
input_mask = tf.cast(input_mask, tf.float32)
pooled = masked_reduce_mean(result, input_mask)
else:
raise NameError(
f"Undefined pooling type (must be either first or mean, but is {self.pooling}"
)
return pooled
def call(self, inputs, output_shape=None):
updates, mask = inputs[0], inputs[1]
with tf.compat.v1.variable_scope(self.name):
mask = K.cast(mask, "int32")
input_shape = tf.shape(updates, out_type="int32")
# calculation new shape
if output_shape is None:
output_shape = (
input_shape[0],
input_shape[1] * self.size[0],
input_shape[2] * self.size[1],
input_shape[3],
)
self.output_shape1 = output_shape
# calculation indices for batch, height, width and feature maps
one_like_mask = K.ones_like(mask, dtype="int32")
batch_shape = K.concatenate([[input_shape[0]], [1], [1], [1]],
axis=0)
batch_range = K.reshape(tf.range(output_shape[0], dtype="int32"),
shape=batch_shape)
b = one_like_mask * batch_range
y = mask // (output_shape[2] * output_shape[3])
x = (mask // output_shape[3]) % output_shape[2]
feature_range = tf.range(output_shape[3], dtype="int32")
f = one_like_mask * feature_range
# transpose indices & reshape update values to one dimension
updates_size = tf.size(updates)
indices = K.transpose(
K.reshape(K.stack([b, y, x, f]), [4, updates_size]))
values = K.reshape(updates, [updates_size])
ret = tf.scatter_nd(indices, values, output_shape)
return ret
def loop_body(b, ignore_mask):
true_box = tf.boolean_mask(y_true[layer][b, ..., 0:4],
object_mask_bool[b, ..., 0])
iou = box_iou(pred_box[b], true_box)
best_iou = K.max(iou, axis=-1)
ignore_mask = ignore_mask.write(
b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b + 1, ignore_mask
def call(self, QKVs):
"""Core logic of multi-head self attention.
Args:
QKVs (list): inputs of multi-head self attention i.e. query, key and value.
Returns:
object: ouput tensors.
"""
if len(QKVs) == 3:
Q_seq, K_seq, V_seq = QKVs
Q_len, V_len = None, None
elif len(QKVs) == 5:
Q_seq, K_seq, V_seq, Q_len, V_len = QKVs
Q_seq = K.dot(Q_seq, self.WQ)
Q_seq = K.reshape(Q_seq,
shape=(-1, K.shape(Q_seq)[1], self.multiheads,
self.head_dim))
Q_seq = K.permute_dimensions(Q_seq, pattern=(0, 2, 1, 3))
K_seq = K.dot(K_seq, self.WK)
K_seq = K.reshape(K_seq,
shape=(-1, K.shape(K_seq)[1], self.multiheads,
self.head_dim))
K_seq = K.permute_dimensions(K_seq, pattern=(0, 2, 1, 3))
V_seq = K.dot(V_seq, self.WV)
V_seq = K.reshape(V_seq,
shape=(-1, K.shape(V_seq)[1], self.multiheads,
self.head_dim))
V_seq = K.permute_dimensions(V_seq, pattern=(0, 2, 1, 3))
A = einsum("abij, abkj -> abik", Q_seq, K_seq) / K.sqrt(
K.cast(self.head_dim, dtype="float32"))
A = K.permute_dimensions(
A, pattern=(0, 3, 2, 1)
) # A.shape=[batch_size,K_sequence_length,Q_sequence_length,self.multiheads]
A = self.Mask(A, V_len, "add")
A = K.permute_dimensions(A, pattern=(0, 3, 2, 1))
if self.mask_right:
ones = K.ones_like(A[:1, :1])
lower_triangular = K.tf.matrix_band_part(ones,
num_lower=-1,
num_upper=0)
mask = (ones - lower_triangular) * 1e12
A = A - mask
A = K.softmax(A)
O_seq = einsum("abij, abjk -> abik", A, V_seq)
O_seq = K.permute_dimensions(O_seq, pattern=(0, 2, 1, 3))
O_seq = K.reshape(O_seq,
shape=(-1, K.shape(O_seq)[1], self.output_dim))
O_seq = self.Mask(O_seq, Q_len, "mul")
return O_seq
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 _backward(gamma, mask):
"""Backward recurrence of the linear chain crf."""
gamma = K.cast(gamma, 'int32')
def _backward_step(gamma_t, states):
y_tm1 = K.squeeze(states[0], 0)
y_t = batch_gather(gamma_t, y_tm1)
return y_t, [K.expand_dims(y_t, 0)]
initial_states = [K.expand_dims(K.zeros_like(gamma[:, 0, 0]), 0)]
_, y_rev, _ = K.rnn(_backward_step,
gamma,
initial_states,
go_backwards=True)
y = K.reverse(y_rev, 1)
if mask is not None:
mask = K.cast(mask, dtype='int32')
# mask output
y *= mask
# set masked values to -1
y += -(1 - mask)
return y
def viterbi_decode(x, U, b_start=None, b_end=None, mask=None):
"""Computes the best tag sequence y for a given input x, i.e. the one that
maximizes the value of path_energy."""
x = add_boundary_energy(x, b_start, b_end, mask)
alpha_0 = x[:, 0, :]
gamma_0 = K.zeros_like(alpha_0)
initial_states = [gamma_0, alpha_0]
_, gamma = _forward(
x,
lambda B: [K.cast(K.argmax(B, axis=1), K.floatx()),
K.max(B, axis=1)], initial_states, U, mask)
y = _backward(gamma, mask)
return y
def add_boundary_energy(x, b_start=None, b_end=None, mask=None):
"""Given the observations x, it adds the start boundary energy b_start (resp.
end boundary energy b_end on the start (resp. end) elements and multiplies
the mask."""
if mask is None:
if b_start is not None:
x = K.concatenate([x[:, :1, :] + b_start, x[:, 1:, :]], axis=1)
if b_end is not None:
x = K.concatenate([x[:, :-1, :], x[:, -1:, :] + b_end], axis=1)
else:
mask = K.cast(mask, K.floatx())
mask = K.expand_dims(mask, 2)
x *= mask
if b_start is not None:
mask_r = K.concatenate([K.zeros_like(mask[:, :1]), mask[:, :-1]],
axis=1)
start_mask = K.cast(K.greater(mask, mask_r), K.floatx())
x = x + start_mask * b_start
if b_end is not None:
mask_l = K.concatenate(
[mask[:, 1:], K.zeros_like(mask[:, -1:])], axis=1)
end_mask = K.cast(K.greater(mask, mask_l), K.floatx())
x = x + end_mask * b_end
return x
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 _correct_boxes(self, box_xy, box_wh, input_shape, image_shape):
"""Get corrected boxes"""
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape / image_shape))
offset = (input_shape - new_shape) / 2.0 / input_shape
scale = input_shape / new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
box_mins = box_yx - (box_hw / 2.0)
box_maxes = box_yx + (box_hw / 2.0)
boxes = K.concatenate([
box_mins[..., 0:1], # y_min
box_mins[..., 1:2], # x_min
box_maxes[..., 0:1], # y_max
box_maxes[..., 1:2], # x_max
])
# Scale boxes back to original image shape.
boxes *= K.concatenate([image_shape, image_shape])
return boxes
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 call(self, inputs, **kwargs):
padding = self.padding
pool_size = self.pool_size
strides = self.strides
if K.backend() == "tensorflow":
ksize = [1, pool_size[0], pool_size[1], 1]
padding = padding.upper()
strides = [1, strides[0], strides[1], 1]
output, argmax = tf.nn.max_pool_with_argmax(inputs,
ksize=ksize,
strides=strides,
padding=padding)
else:
errmsg = "{} backend is not supported for layer {}".format(
K.backend(),
type(self).__name__)
raise NotImplementedError(errmsg)
argmax = K.cast(argmax, K.floatx())
return [output, argmax]
def _forward(x, reduce_step, initial_states, U, mask=None):
"""Forward recurrence of the linear chain crf."""
def _forward_step(energy_matrix_t, states):
alpha_tm1 = states[-1]
new_states = reduce_step(K.expand_dims(alpha_tm1, 2) + energy_matrix_t)
return new_states[0], new_states
U_shared = K.expand_dims(K.expand_dims(U, 0), 0)
if mask is not None:
mask = K.cast(mask, K.floatx())
mask_U = K.expand_dims(K.expand_dims(mask[:, :-1] * mask[:, 1:], 2), 3)
U_shared = U_shared * mask_U
inputs = K.expand_dims(x[:, 1:, :], 2) + U_shared
inputs = K.concatenate([inputs, K.zeros_like(inputs[:, -1:, :, :])],
axis=1)
last, values, _ = K.rnn(_forward_step, inputs, initial_states)
return last, values
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5, print_loss=False):
'''Return yolo_loss tensor
Parameters
----------
yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_body
y_true: list of array, the output of preprocess_true_boxes
anchors: array, shape=(N, 2), wh
num_classes: integer
ignore_thresh: float, the iou threshold whether to ignore object confidence loss
Returns
-------
loss: tensor, shape=(1,)
'''
num_layers = len(anchors) // 3 # default setting
yolo_outputs = args[:num_layers]
y_true = args[num_layers:]
anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]
] if num_layers == 3 else [[3, 4, 5], [1, 2, 3]]
input_shape = K.cast(
K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
grid_shapes = [
K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0]))
for l in range(num_layers)
]
loss = 0
m = K.shape(yolo_outputs[0])[0] # batch size, tensor
mf = K.cast(m, K.dtype(yolo_outputs[0]))
for l in range(num_layers):
object_mask = y_true[l][..., 4:5]
true_class_probs = y_true[l][..., 5:]
grid, raw_pred, pred_xy, pred_wh = yolo_head(yolo_outputs[l],
anchors[anchor_mask[l]],
num_classes,
input_shape,
calc_loss=True)
pred_box = K.concatenate([pred_xy, pred_wh])
# Darknet raw box to calculate loss.
raw_true_xy = y_true[l][..., :2] * grid_shapes[l][::-1] - grid
raw_true_wh = K.log(y_true[l][..., 2:4] / anchors[anchor_mask[l]] *
input_shape[::-1])
raw_true_wh = K.switch(object_mask, raw_true_wh,
K.zeros_like(raw_true_wh)) # avoid log(0)=-inf
box_loss_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4]
# Find ignore mask, iterate over each of batch.
ignore_mask = tf.TensorArray(K.dtype(y_true[0]),
size=1,
dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool')
def loop_body(b, ignore_mask):
true_box = tf.boolean_mask(y_true[l][b, ..., 0:4],
object_mask_bool[b, ..., 0])
iou = box_iou(pred_box[b], true_box)
best_iou = K.max(iou, axis=-1)
ignore_mask = ignore_mask.write(
b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b + 1, ignore_mask
_, ignore_mask = K.control_flow_ops.while_loop(lambda b, *args: b < m,
loop_body,
[0, ignore_mask])
ignore_mask = ignore_mask.stack()
ignore_mask = K.expand_dims(ignore_mask, -1)
# K.binary_crossentropy is helpful to avoid exp overflow.
xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(
raw_true_xy, raw_pred[..., 0:2], from_logits=True)
wh_loss = object_mask * box_loss_scale * 0.5 * K.square(
raw_true_wh - raw_pred[..., 2:4])
confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True)+ \
(1-object_mask) * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True) * ignore_mask
class_loss = object_mask * K.binary_crossentropy(
true_class_probs, raw_pred[..., 5:], from_logits=True)
xy_loss = K.sum(xy_loss) / mf
wh_loss = K.sum(wh_loss) / mf
confidence_loss = K.sum(confidence_loss) / mf
class_loss = K.sum(class_loss) / mf
loss += xy_loss + wh_loss + confidence_loss + class_loss
if print_loss:
loss = tf.Print(loss, [
loss, xy_loss, wh_loss, confidence_loss, class_loss,
K.sum(ignore_mask)
],
message='loss: ')
return loss
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 yolo_loss(args,
anchors,
num_classes,
rescore_confidence=False,
print_loss=False):
"""YOLO localization loss function.
Parameters
----------
yolo_output : tensor
Final convolutional layer features.
true_boxes : tensor
Ground truth boxes tensor with shape [batch, num_true_boxes, 5]
containing box x_center, y_center, width, height, and class.
detectors_mask : array
0/1 mask for detector positions where there is a matching ground truth.
matching_true_boxes : array
Corresponding ground truth boxes for positive detector positions.
Already adjusted for conv height and width.
anchors : tensor
Anchor boxes for model.
num_classes : int
Number of object classes.
rescore_confidence : bool, default=False
If true then set confidence target to IOU of best predicted box with
the closest matching ground truth box.
print_loss : bool, default=False
If True then use a tf.Print() to print the loss components.
Returns
-------
mean_loss : float
mean localization loss across minibatch
"""
(yolo_output, true_boxes, detectors_mask, matching_true_boxes) = args
num_anchors = len(anchors)
object_scale = 5
no_object_scale = 1
class_scale = 1
coordinates_scale = 1
pred_xy, pred_wh, pred_confidence, pred_class_prob = yolo_head(
yolo_output, anchors, num_classes)
# Unadjusted box predictions for loss.
# TODO: Remove extra computation shared with yolo_head.
yolo_output_shape = K.shape(yolo_output)
feats = K.reshape(yolo_output, [
-1, yolo_output_shape[1], yolo_output_shape[2], num_anchors,
num_classes + 5
])
pred_boxes = K.concatenate(
(K.sigmoid(feats[..., 0:2]), feats[..., 2:4]), axis=-1)
# TODO: Adjust predictions by image width/height for non-square images?
# IOUs may be off due to different aspect ratio.
# Expand pred x,y,w,h to allow comparison with ground truth.
# batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
pred_xy = K.expand_dims(pred_xy, 4)
pred_wh = K.expand_dims(pred_wh, 4)
pred_wh_half = pred_wh / 2.
pred_mins = pred_xy - pred_wh_half
pred_maxes = pred_xy + pred_wh_half
true_boxes_shape = K.shape(true_boxes)
# batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
true_boxes = K.reshape(true_boxes, [
true_boxes_shape[0], 1, 1, 1, true_boxes_shape[1], true_boxes_shape[2]
])
true_xy = true_boxes[..., 0:2]
true_wh = true_boxes[..., 2:4]
# Find IOU of each predicted box with each ground truth box.
true_wh_half = true_wh / 2.
true_mins = true_xy - true_wh_half
true_maxes = true_xy + true_wh_half
intersect_mins = K.maximum(pred_mins, true_mins)
intersect_maxes = K.minimum(pred_maxes, true_maxes)
intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1]
pred_areas = pred_wh[..., 0] * pred_wh[..., 1]
true_areas = true_wh[..., 0] * true_wh[..., 1]
union_areas = pred_areas + true_areas - intersect_areas
iou_scores = intersect_areas / union_areas
# Best IOUs for each location.
best_ious = K.max(iou_scores, axis=4) # Best IOU scores.
best_ious = K.expand_dims(best_ious)
# A detector has found an object if IOU > thresh for some true box.
object_detections = K.cast(best_ious > 0.6, K.dtype(best_ious))
# TODO: Darknet region training includes extra coordinate loss for early
# training steps to encourage predictions to match anchor priors.
# Determine confidence weights from object and no_object weights.
# NOTE: YOLO does not use binary cross-entropy here.
no_object_weights = (no_object_scale * (1 - object_detections) *
(1 - detectors_mask))
no_objects_loss = no_object_weights * K.square(-pred_confidence)
if rescore_confidence:
objects_loss = (object_scale * detectors_mask *
K.square(best_ious - pred_confidence))
else:
objects_loss = (object_scale * detectors_mask *
K.square(1 - pred_confidence))
confidence_loss = objects_loss + no_objects_loss
# Classification loss for matching detections.
# NOTE: YOLO does not use categorical cross-entropy loss here.
matching_classes = K.cast(matching_true_boxes[..., 4], 'int32')
matching_classes = K.one_hot(matching_classes, num_classes)
classification_loss = (class_scale * detectors_mask *
K.square(matching_classes - pred_class_prob))
# Coordinate loss for matching detection boxes.
matching_boxes = matching_true_boxes[..., 0:4]
coordinates_loss = (coordinates_scale * detectors_mask *
K.square(matching_boxes - pred_boxes))
confidence_loss_sum = K.sum(confidence_loss)
classification_loss_sum = K.sum(classification_loss)
coordinates_loss_sum = K.sum(coordinates_loss)
total_loss = 0.5 * (
confidence_loss_sum + classification_loss_sum + coordinates_loss_sum)
if print_loss:
total_loss = tf.Print(
total_loss, [
total_loss, confidence_loss_sum, classification_loss_sum,
coordinates_loss_sum
],
message='yolo_loss, conf_loss, class_loss, box_coord_loss:')
return total_loss
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5):
'''Return yolo_loss tensor
Parameters
----------
yolo_outputs: list of tensor, the output of yolo_body
y_true: list of array, the output of preprocess_true_boxes
anchors: array, shape=(T, 2), wh
num_classes: integer
ignore_thresh: float, the iou threshold whether to ignore object confidence loss
Returns
-------
loss: tensor, shape=(1,)
'''
yolo_outputs = args[:3]
y_true = args[3:]
anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
input_shape = K.cast(
K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
grid_shapes = [
K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0]))
for l in range(3)
]
loss = 0
m = K.shape(yolo_outputs[0])[0]
for l in range(3):
object_mask = y_true[l][..., 4:5]
true_class_probs = y_true[l][..., 5:]
pred_xy, pred_wh, pred_confidence, pred_class_probs = yolo_head(
yolo_outputs[l], anchors[anchor_mask[l]], num_classes, input_shape)
pred_box = K.concatenate([pred_xy, pred_wh])
# Darknet box loss.
xy_delta = (y_true[l][..., :2] - pred_xy) * grid_shapes[l][::-1]
wh_delta = K.log(y_true[l][..., 2:4]) - K.log(pred_wh)
# Avoid log(0)=-inf.
wh_delta = K.switch(object_mask, wh_delta, K.zeros_like(wh_delta))
box_delta = K.concatenate([xy_delta, wh_delta], axis=-1)
box_delta_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4]
# Find ignore mask, iterate over each of batch.
ignore_mask = tf.TensorArray(K.dtype(y_true[0]),
size=1,
dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool')
def loop_body(b, ignore_mask):
true_box = tf.boolean_mask(y_true[l][b, ..., 0:4],
object_mask_bool[b, ..., 0])
iou = box_iou(pred_box[b], true_box)
best_iou = K.max(iou, axis=-1)
ignore_mask = ignore_mask.write(
b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b + 1, ignore_mask
_, ignore_mask = K.control_flow_ops.while_loop(lambda b, *args: b < m,
loop_body,
[0, ignore_mask])
ignore_mask = ignore_mask.stack()
ignore_mask = K.expand_dims(ignore_mask, -1)
box_loss = object_mask * K.square(box_delta * box_delta_scale)
confidence_loss = object_mask * K.square(1-pred_confidence) + \
(1-object_mask) * K.square(0-pred_confidence) * ignore_mask
class_loss = object_mask * K.square(true_class_probs -
pred_class_probs)
loss += K.sum(box_loss) + K.sum(confidence_loss) + K.sum(class_loss)
return loss / K.cast(m, K.dtype(loss))
def call(self, inputs, **kwargs):
mask = K.not_equal(inputs, 0)
return K.cast(mask, K.floatx())
def chain_crf_loss(y, x, U, b_start=None, b_end=None, mask=None):
"""Variant of sparse_chain_crf_loss but with one-hot encoded tags y."""
y_sparse = K.argmax(y, -1)
y_sparse = K.cast(y_sparse, 'int32')
return sparse_chain_crf_loss(y_sparse, x, U, b_start, b_end, mask)
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5):
"""Return yolo_loss tensor
Parameters
----------
yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_body
y_true: list of array, the output of preprocess_true_boxes
anchors: array, shape=(N, 2), wh
num_classes: integer
ignore_thresh: float, the iou threshold whether to ignore object confidence loss
Returns
-------
loss: tensor, shape=(1,)
"""
num_layers = 1
yolo_outputs = args[:num_layers]
y_true = args[num_layers:]
g_y_true = y_true
input_shape = K.cast(
K.shape(yolo_outputs[0])[1:3] * grid_size_multiplier,
K.dtype(y_true[0]))
grid_shapes = [
K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0]))
for l in range(num_layers)
]
loss = 0
m = K.shape(yolo_outputs[0])[0] # batch size, tensor
mf = K.cast(m, K.dtype(yolo_outputs[0]))
for layer in range(num_layers):
object_mask = y_true[layer][..., 4:5]
vertices_mask = y_true[layer][..., 5 + num_classes + 2:5 +
num_classes + NUM_ANGLES3:3]
true_class_probs = y_true[layer][..., 5:5 + num_classes]
grid, raw_pred, pred_xy, pred_wh, pol_cnf = yolo_head(
yolo_outputs[layer],
anchors[anchor_mask[layer]],
num_classes,
input_shape,
calc_loss=True)
pred_box = K.concatenate([pred_xy, pred_wh])
raw_true_xy = y_true[layer][..., :2] * grid_shapes[layer][
..., ::-1] - grid
raw_true_polygon0 = y_true[layer][..., 5 + num_classes:5 +
num_classes + NUM_ANGLES3]
raw_true_wh = K.log(y_true[layer][..., 2:4] /
anchors[anchor_mask[layer]] *
input_shape[..., ::-1])
raw_true_wh = K.switch(object_mask, raw_true_wh,
K.zeros_like(raw_true_wh)) # avoid log(0)=-inf
raw_true_polygon_x = raw_true_polygon0[..., ::3]
raw_true_polygon_y = raw_true_polygon0[..., 1::3]
dx = K.square(anchors[anchor_mask[layer]][..., 0:1] / 2)
dy = K.square(anchors[anchor_mask[layer]][..., 1:2] / 2)
d = K.cast(K.sqrt(dx + dy), K.dtype(raw_true_polygon_x))
diagonal = K.sqrt(
K.pow(input_shape[..., ::-1][0], 2) +
K.pow(input_shape[..., ::-1][1], 2))
raw_true_polygon_x = K.log(raw_true_polygon_x / d * diagonal)
raw_true_polygon_x = K.switch(vertices_mask, raw_true_polygon_x,
K.zeros_like(raw_true_polygon_x))
box_loss_scale = 2 - y_true[layer][..., 2:3] * y_true[layer][..., 3:4]
# Find ignore mask, iterate over each of batch.
ignore_mask = tf.TensorArray(K.dtype(y_true[0]),
size=1,
dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool')
def loop_body(b, ignore_mask):
true_box = tf.boolean_mask(y_true[layer][b, ..., 0:4],
object_mask_bool[b, ..., 0])
iou = box_iou(pred_box[b], true_box)
best_iou = K.max(iou, axis=-1)
ignore_mask = ignore_mask.write(
b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b + 1, ignore_mask
_, ignore_mask = tf.while_loop(lambda b, *args: b < m, loop_body,
[0, ignore_mask])
ignore_mask = ignore_mask.stack()
ignore_mask = K.expand_dims(ignore_mask, -1)
# K.binary_crossentropy is helpful to avoid exp overflow.
xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(
raw_true_xy, raw_pred[..., 0:2], from_logits=True)
wh_loss = object_mask * box_loss_scale * 0.5 * K.square(
raw_true_wh - raw_pred[..., 2:4])
confidence_loss = object_mask * K.binary_crossentropy(
object_mask, raw_pred[..., 4:5], from_logits=True
) + (1 - object_mask) * K.binary_crossentropy(
object_mask, raw_pred[..., 4:5], from_logits=True) * ignore_mask
class_loss = object_mask * K.binary_crossentropy(
true_class_probs,
raw_pred[..., 5:5 + num_classes],
from_logits=True)
polygon_loss_x = object_mask * vertices_mask * box_loss_scale * 0.5 * K.square(
raw_true_polygon_x -
raw_pred[..., 5 + num_classes:5 + num_classes + NUM_ANGLES3:3])
polygon_loss_y = object_mask * vertices_mask * box_loss_scale * K.binary_crossentropy(
raw_true_polygon_y,
raw_pred[..., 5 + num_classes + 1:5 + num_classes + NUM_ANGLES3:3],
from_logits=True)
vertices_confidence_loss = object_mask * K.binary_crossentropy(
vertices_mask,
raw_pred[..., 5 + num_classes + 2:5 + num_classes + NUM_ANGLES3:3],
from_logits=True)
xy_loss = K.sum(xy_loss) / mf
wh_loss = K.sum(wh_loss) / mf
class_loss = K.sum(class_loss) / mf
confidence_loss = K.sum(confidence_loss) / mf
vertices_confidence_loss = K.sum(vertices_confidence_loss) / mf
polygon_loss = K.sum(polygon_loss_x) / mf + K.sum(polygon_loss_y) / mf
diou_loss = K.sum(
object_mask * box_loss_scale *
(1 - box_diou(pred_box, y_true[layer][..., 0:4]))) / mf
loss += (xy_loss + wh_loss + confidence_loss + class_loss +
0.2 * polygon_loss + 0.2 * vertices_confidence_loss) / (
K.sum(object_mask) + 1) * mf
return loss
