def f1_loss(y_true, y_pred):
tp = K.sum(K.cast(y_true*y_pred, 'float'), axis=0)
tn = K.sum(K.cast((1-y_true)*(1-y_pred), 'float'), axis=0)
fp = K.sum(K.cast((1-y_true)*y_pred, 'float'), axis=0)
fn = K.sum(K.cast(y_true*(1-y_pred), 'float'), axis=0)
p = tp / (tp + fp + K.epsilon())
r = tp / (tp + fn + K.epsilon())
f1 = 2*p*r / (p+r+K.epsilon())
f1 = tf.where(tf.is_nan(f1), tf.zeros_like(f1), f1)
return 1 - K.mean(f1)
def eval_loss(y_true, y_pred):
target = y_true
######output times weights here###############
mask = K.cast(K.expand_dims(weights, axis=-1), dtype='float32')
tensor_shape = y_pred.get_shape()
# x = tf.add(x, tf.constant(1, shape=x.shape))
yPred_stack = []
for i in range(tensor_shape[1]):
mask_i = K.cast(K.expand_dims(mask[i], axis=-1), dtype='float32')
yPred_i = K.cast(K.expand_dims(y_pred[:, i], axis=-1), dtype='float32')
yPred_stack.append(K.dot(yPred_i, mask_i))
output = tf.reshape(tf.stack(yPred_stack, axis=1, name='stack'),
[-1, tensor_shape[1]])
return y_pred[0, 7]
def K_sparse_categorical_crossentropy(target,
output,
from_logits=False,
axis=-1):
output_dimensions = list(range(len(output.get_shape())))
if axis != -1 and axis not in output_dimensions:
raise ValueError('{}{}{}'.format(
'Unexpected channels axis {}. '.format(axis),
'Expected to be -1 or one of the axes of `output`, ',
'which has {} dimensions.'.format(len(output.get_shape()))))
# If the channels are not in the last axis, move them to be there:
if axis != -1 and axis != output_dimensions[-1]:
permutation = output_dimensions[:axis] + output_dimensions[axis + 1:]
permutation += [axis]
output = tf.transpose(output, perm=permutation)
# Note: tf.nn.sparse_softmax_cross_entropy_with_logits
# expects logits, Keras expects probabilities.
if not from_logits:
_epsilon = _to_tensor(epsilon(), output.dtype.base_dtype)
output = tf.clip_by_value(output, _epsilon, 1 - _epsilon)
output = tf.log(output)
output_shape = output.get_shape()
targets = cast(flatten(target), 'int64')
logits = tf.reshape(output, [-1, tf.shape(output)[-1]])
res = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=targets,
logits=logits)
if len(output_shape) >= 3:
# if our output includes timestep dimension
# or spatial dimensions we need to reshape
return tf.reshape(res, tf.shape(output)[:-1])
else:
return res
def get_lr_decay(self):
lr = self.model.optimizer.lr
if self.model.optimizer.initial_decay > 0:
lr = lr * (1. / (1. + self.model.optimizer.decay * backend.cast(
self.model.optimizer.iterations,
backend.dtype(self.model.optimizer.decay))))
return backend.eval(lr), backend.eval(self.model.optimizer.decay)
def binary_PTA(y_true, y_pred, threshold=K.variable(value=0.5)):
y_pred = K.cast(y_pred >= threshold, 'float32')
# P = total number of positive labels
P = K.sum(y_true)
# TP = total number of correct alerts, alerts from the positive class labels
TP = K.sum(y_pred * y_true)
return TP / P
def binary_PFA(y_true, y_pred, threshold=K.variable(value=0.5)):
y_pred = K.cast(y_pred >= threshold, 'float32')
# N = total number of negative labels
N = K.sum(1 - y_true)
# FP = total number of false alerts, alerts from the negative class labels
TN = K.sum((1 - y_pred) * (1 - y_true))
return TN / N
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
def __metrics_base(self, y_true, y_pred):
""" Base for all the metrics defined below """
y_true, y_pred = K.flatten(tf.math.argmax(y_true, axis=-1)), K.flatten(
tf.math.argmax(y_pred, axis=-1))
con_mat = K.cast(tf.math.confusion_matrix(y_true, y_pred), K.floatx())
correct = tf.linalg.diag_part(con_mat)
total = K.sum(con_mat, axis=-1)
return correct, total, con_mat
def step(self, inputs, states):
states = list(states)
if self.teacher_force:
readout = states.pop()
ground_truth = states.pop()
assert K.ndim(ground_truth) == 3, K.ndim(ground_truth)
counter = states.pop()
if K.backend() == 'tensorflow':
with tf.control_dependencies(None):
zero = K.cast(K.zeros((1, ))[0], 'int32')
one = K.cast(K.zeros((1, ))[0], 'int32')
else:
zero = K.cast(K.zeros((1, ))[0], 'int32')
one = K.cast(K.zeros((1, ))[0], 'int32')
slices = [
slice(None), counter[0] - K.switch(counter[0], one, zero)
] + [slice(None)] * (K.ndim(ground_truth) - 2)
ground_truth_slice = ground_truth[slices]
readout = K.in_train_phase(
K.switch(counter[0], ground_truth_slice, readout), readout)
states.append(readout)
if self.decode:
model_input = states
else:
model_input = [inputs] + states
shapes = []
for x in model_input:
if hasattr(x, '_keras_shape'):
shapes.append(x._keras_shape)
del x._keras_shape # Else keras internals will get messed up.
model_output = _to_list(self.model.call(model_input))
for x, s in zip(model_input, shapes):
setattr(x, '_keras_shape', s)
if self.decode:
model_output.insert(1, model_input[0])
for tensor in model_output:
tensor._uses_learning_phase = self.uses_learning_phase
states = model_output[1:]
output = model_output[0]
if self.readout:
states += [output]
if self.teacher_force:
states.insert(-1, counter + 1)
states.insert(-1, ground_truth)
return output, states
def softmax_cross_entropy(target,
output,
from_logits=True,
axis=-1,
normalize=False):
""" Compute Softmax cross entropy loss for sparse target.
Args:
target (tensor): Target label. If 2D, shape is (w, h).
output (tensor): Logits or Probabilities. If 2D, shape is (w, h, ch).
from_logits (bool, optional): logits or softmax outputs? Defaults to True.
axis (int, optional): Specifying the channels axis. Defaults to -1.
normalize (bool, optional): Normalize loss across all instances. Defaults to False.
"""
_check_dtype(target, 'int32')
_check_dtype(output, 'float32')
output_dimensions = list(range(len(output.get_shape())))
if axis != -1 and axis not in output_dimensions:
raise ValueError('{}{}{}'.format(
'Unexpected channels axis {}. '.format(axis),
'Expected to be -1 or one of the axes of `output`, ',
'which has {} dimensions.'.format(len(output.get_shape()))))
# move the channels to be in the last axis:
if axis != -1 and axis != output_dimensions[-1]:
permutation = output_dimensions[:axis] + output_dimensions[axis + 1:]
permutation += [axis]
output = tf.transpose(output, perm=permutation)
# convert to the logits
if not from_logits:
_epsilon = _to_tensor(epsilon(), output.dtype.base_dtype)
output = tf.clip_by_value(output, _epsilon, 1 - _epsilon)
output = tf.log(output) # NOTE: log(exp(x)) = x
logits = output
# softmax_cross_entropy
output_shape = output.get_shape()
targets = cast(tf.reshape(target,
tf.shape(output)[:-1]), 'int32') # NOTE: cast...
res = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=targets,
logits=logits)
# reduce
if normalize:
return tf.reduce_mean(res)
else:
return tf.reduce_sum(tf.reduce_mean(res, axis=0)) # only batch-axis
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)
# 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])
# 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 call(self, inputs, mask=None, **kwargs):
input_len = K.shape(inputs)[1]
if self.attention_type == Attention.ATTENTION_TYPE_ADD:
e = self._call_additive_emission(inputs)
elif self.attention_type == Attention.ATTENTION_TYPE_MUL:
e = self._call_multiplicative_emission(inputs)
if self.attention_activation is not None:
e = self.attention_activation(e)
if self.attention_width is not None:
if self.history_only:
lower = K.arange(0, input_len) - (self.attention_width - 1)
else:
lower = K.arange(0, input_len) - self.attention_width // 2
lower = K.expand_dims(lower, axis=-1)
upper = lower + self.attention_width
indices = K.expand_dims(K.arange(0, input_len), axis=0)
e -= 10000.0 * (1.0 - K.cast(lower <= indices, K.floatx()) *
K.cast(indices < upper, K.floatx()))
if mask is not None:
mask = K.expand_dims(K.cast(mask, K.floatx()), axis=-1)
e -= 10000.0 * ((1.0 - mask) *
(1.0 - K.permute_dimensions(mask, (0, 2, 1))))
# a_{t} = \text{softmax}(e_t)
e = K.exp(e - K.max(e, axis=-1, keepdims=True))
a = e / K.sum(e, axis=-1, keepdims=True)
# l_t = \sum_{t'} a_{t, t'} x_{t'}
v = K.batch_dot(a, inputs)
if self.attention_regularizer_weight > 0.0:
self.add_loss(self._attention_regularizer(a))
if self.return_attention:
return [v, a]
return v
def yolo_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. / input_shape
scale = input_shape / new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
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_loss(self,
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 = self.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 call(self,
inputs,
initial_state=None,
initial_readout=None,
ground_truth=None,
mask=None,
training=None):
# input shape: `(samples, time (padded with zeros), input_dim)`
# note that the .build() method of subclasses MUST define
# self.input_spec and self.state_spec with complete input shapes.
if type(mask) is list:
mask = mask[0]
if self.model is None:
raise Exception('Empty RecurrentModel.')
num_req_states = self.num_states
if self.readout:
num_actual_states = num_req_states - 1
else:
num_actual_states = num_req_states
if type(inputs) is list:
inputs_list = inputs[:]
inputs = inputs_list.pop(0)
initial_states = inputs_list[:num_actual_states]
if len(initial_states) > 0:
if self._is_optional_input_placeholder(initial_states[0]):
initial_states = self.get_initial_state(inputs)
inputs_list = inputs_list[num_actual_states:]
if self.readout:
initial_readout = inputs_list.pop(0)
if self.teacher_force:
ground_truth = inputs_list.pop()
else:
if initial_state is not None:
if not isinstance(initial_state, (list, tuple)):
initial_states = [initial_state]
else:
initial_states = list(initial_state)
if self._is_optional_input_placeholder(initial_states[0]):
initial_states = self.get_initial_state(inputs)
elif self.stateful:
initial_states = self.states
else:
initial_states = self.get_initial_state(inputs)
if self.readout:
if initial_readout is None or self._is_optional_input_placeholder(
initial_readout):
output_shape = K.int_shape(_to_list((self.model.output))[0])
output_ndim = len(output_shape)
input_ndim = K.ndim(inputs)
initial_readout = K.zeros_like(inputs)
slices = [slice(None)] + [0] * (input_ndim - 1)
initial_readout = initial_readout[slices] # (batch_size,)
initial_readout = K.reshape(initial_readout,
(-1, ) + (1, ) * (output_ndim - 1))
initial_readout = K.tile(initial_readout,
(1, ) + tuple(output_shape[1:]))
initial_states.append(initial_readout)
if self.teacher_force:
if ground_truth is None or self._is_optional_input_placeholder(
ground_truth):
raise Exception(
'ground_truth must be provided for RecurrentModel with teacher_force=True.'
)
if K.backend() == 'tensorflow':
with tf.control_dependencies(None):
counter = K.zeros((1, ))
else:
counter = K.zeros((1, ))
counter = K.cast(counter, 'int32')
initial_states.insert(-1, counter)
initial_states[-2]
initial_states.insert(-1, ground_truth)
num_req_states += 2
if len(initial_states) != num_req_states:
raise ValueError('Layer requires ' + str(num_req_states) +
' states but was passed ' +
str(len(initial_states)) + ' initial states.')
input_shape = K.int_shape(inputs)
if self.unroll and input_shape[1] is None:
raise ValueError('Cannot unroll a RNN if the '
'time dimension is undefined. \n'
'- If using a Sequential model, '
'specify the time dimension by passing '
'an `input_shape` or `batch_input_shape` '
'argument to your first layer. If your '
'first layer is an Embedding, you can '
'also use the `input_length` argument.\n'
'- If using the functional API, specify '
'the time dimension by passing a `shape` '
'or `batch_shape` argument to your Input layer.')
preprocessed_input = self.preprocess_input(inputs, training=None)
constants = self.get_constants(inputs, training=None)
if self.decode:
initial_states.insert(0, inputs)
preprocessed_input = K.zeros((1, self.output_length, 1))
input_length = self.output_length
else:
input_length = input_shape[1]
if self.uses_learning_phase:
with learning_phase_scope(0):
last_output_test, outputs_test, states_test, updates = rnn(
self.step,
preprocessed_input,
initial_states,
go_backwards=self.go_backwards,
mask=mask,
constants=constants,
unroll=self.unroll,
input_length=input_length)
with learning_phase_scope(1):
last_output_train, outputs_train, states_train, updates = rnn(
self.step,
preprocessed_input,
initial_states,
go_backwards=self.go_backwards,
mask=mask,
constants=constants,
unroll=self.unroll,
input_length=input_length)
last_output = K.in_train_phase(last_output_train,
last_output_test,
training=training)
outputs = K.in_train_phase(outputs_train,
outputs_test,
training=training)
states = []
for state_train, state_test in zip(states_train, states_test):
states.append(
K.in_train_phase(state_train,
state_test,
training=training))
else:
last_output, outputs, states, updates = rnn(
self.step,
preprocessed_input,
initial_states,
go_backwards=self.go_backwards,
mask=mask,
constants=constants,
unroll=self.unroll,
input_length=input_length)
states = list(states)
if self.decode:
states.pop(0)
if self.readout:
states.pop()
if self.teacher_force:
states.pop()
states.pop()
if len(updates) > 0:
self.add_update(updates)
if self.stateful:
updates = []
for i in range(len(states)):
updates.append((self.states[i], states[i]))
self.add_update(updates, inputs)
# Properly set learning phase
if 0 < self.dropout + self.recurrent_dropout:
last_output._uses_learning_phase = True
outputs._uses_learning_phase = True
if self.return_sequences:
y = outputs
else:
y = last_output
if self.return_states:
return [y] + states
else:
return y
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.initial_decay > 0:
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))
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]
self.weights = [self.iterations] + ms + vs
for p, g, m, v in zip(params, grads, ms, vs):
# if a weight tensor (len > 1) use weight normalized parameterization
# this is the only part changed w.r.t. keras.optimizers.Adam
ps = K.get_variable_shape(p)
if len(ps) > 1:
# get weight normalization parameters
V, V_norm, V_scaler, g_param, grad_g, grad_V = get_weightnorm_params_and_grads(
p, g)
# Adam containers for the 'g' parameter
V_scaler_shape = K.get_variable_shape(V_scaler)
m_g = K.zeros(V_scaler_shape)
v_g = K.zeros(V_scaler_shape)
# update g parameters
m_g_t = (self.beta_1 * m_g) + (1. - self.beta_1) * grad_g
v_g_t = (self.beta_2 *
v_g) + (1. - self.beta_2) * K.square(grad_g)
new_g_param = g_param - lr_t * m_g_t / (K.sqrt(v_g_t) +
self.epsilon)
self.updates.append(K.update(m_g, m_g_t))
self.updates.append(K.update(v_g, v_g_t))
# update V parameters
m_t = (self.beta_1 * m) + (1. - self.beta_1) * grad_V
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(grad_V)
new_V_param = V - 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))
# if there are constraints we apply them to V, not W
# if p in constraints:
# c = constraints[p]
# new_V_param = c(new_V_param)
# wn param updates --> W updates
add_weightnorm_param_updates(self.updates, new_V_param,
new_g_param, p, V_scaler)
else: # do optimization normally
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 / (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 p in constraints:
# c = constraints[p]
# new_p = c(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
