def triplet_loss(y_true, y_pred, alpha=0.6):
"""
Implementation of the triplet loss function
Arguments:
y_true -- true labels, required when you define a loss in Keras, you don't need it in this function.
y_pred -- python list containing three objects:
anchor -- the encodings for the anchor data
positive -- the encodings for the positive data (similar to anchor)
negative -- the encodings for the negative data (different from anchor)
Returns:
loss -- real number, value of the loss
"""
print('y_pred.shape = ', y_pred)
total_length = y_pred.shape.as_list()[-1]
# print('total_length=', total_length)
# total_length =12
anchor = y_pred[:, 0:int(total_length * 1 / 3)]
positive = y_pred[:, int(total_length * 1 / 3):int(total_length * 2 / 3)]
negative = y_pred[:, int(total_length * 2 / 3):int(total_length * 3 / 3)]
# distance between the anchor and the positive
pos_dist = K.abs(K.sum(K.square(anchor - positive), axis=1))
print(pos_dist)
# distance between the anchor and the negative
neg_dist = K.abs(K.sum(K.square(anchor - negative), axis=1))
print(neg_dist)
# compute loss
basic_loss = pos_dist - (neg_dist + alpha)
loss = K.maximum(basic_loss, 0.0)
return loss
def loss_functions(method="mse"):
""" loss_function returns the callable object to evaluate the loss.
# Arguments
method: String.
- "mse" for `Mean Squared Error` or
- "mae" for `Mean Absolute Error` or
- "se" for `Squared Error` or
- "ae" for `Absolute Error`.
# Returns
Callable function that gets (y_true, y_pred) as the input and
returns the loss value as the output.
# Raises
ValueError if anything other than "mse" or "mae" is passed.
"""
if method in ("mse", "mean_squared_error"):
return lambda y_true, y_pred: K.mean(K.square(y_true - y_pred),
axis=-1)
elif method in ("mae", "mean_absolute_error"):
return lambda y_true, y_pred: K.mean(K.abs(y_true - y_pred),
axis=-1)
elif method in ("se", "squared_error"):
return lambda y_true, y_pred: K.sum(K.square(y_true - y_pred),
axis=-1)
elif method in ("ae", "absolute_error"):
return lambda y_true, y_pred: K.sum(K.abs(y_true - y_pred),
axis=-1)
elif hasattr(k.losses, method):
return getattr(k.losses, method)
else:
raise ValueError(
'Supported losses: Keras loss function or (mse, mae, se, ae)')
def tol_equal(f, other, tol=1e-8):
"""Element-wise comparison applied to the `Functional` objects.
# Arguments
f: Functional object.
other: A python number or a tensor or a functional object.
tol: float - defaulted to 1e-8.
# Returns
A Functional.
"""
validate_functional(f)
assert isinstance(tol, float), "Expected a float for tolerance. "
inputs = f.inputs.copy()
if is_functional(other):
inputs += to_list(other.inputs)
lmbd = [
Lambda(lambda x: K.cast_to_floatx(
K.less_equal(K.abs(x[0] - x[1]), tol)),
name=graph_unique_name("tol_equal")) for X in f.outputs
]
else:
_warn_for_ndarray(other)
lmbd = [
Lambda(lambda x: K.cast_to_floatx(
K.less_equal(K.abs(x - other), tol)),
name=graph_unique_name("tol_equal")) for X in f.outputs
]
Functional = f.get_class()
res = Functional(inputs=unique_tensors(inputs),
outputs=_apply_operation(lmbd, f, other),
layers=lmbd)
return res
def _huber_loss(self, y_true, y_pred, clip_delta=1.0):
error = y_true - y_pred
cond = K.abs(error) <= clip_delta
squared_loss = 0.5 * K.square(error)
quadratic_loss = 0.5 * K.square(clip_delta) + clip_delta * (
K.abs(error) - clip_delta)
return K.mean(tf.where(cond, squared_loss, quadratic_loss))
def huber_loss(y_true, y_pred):
err = y_true - y_pred
cond = K.abs(err) < HUBER_LOSS_DELTA
L2 = 0.5 * K.square(err)
L1 = HUBER_LOSS_DELTA * (K.abs(err) - 0.5 * HUBER_LOSS_DELTA)
loss = tf.compat.v1.where(cond, L2, L1)
return K.mean(loss)
def DeltaLayer(encoded_l, encoded_r, negateDiffs=False):
"""
A Layer which computes all possible absolute differences of
all pixels. Input are two feature volumes, e.g. result of a conv layer
Hints:
- The Reshape reshapes a matrix row-wise, that means,
Reshape( (6,1) ) ([ 1 2 3
4 5 6]) is
1
2
3
4
5
6
- Algorithm:
- The left leg is reshaped to a w*h x 1 column vector (for each channel)
- The right leg is reshaped to a 1 x w*h row vector (for each channel)
- The left is tiled along colum axis, so from w*h x 1 to w*h x w*h (per channel)
- The right is tiled along row axis, so from 1 x w*h to w*h x w*h
- The absolute difference is calculated
Args:
encoded_l, encoded_r : left and right image tensor (batchsize,w,h,channels)
must have same size
negateDiffs: if True then not abs(diffs), but -abs(diffs) is returned.
Default: False
Returns:
difference tensor, has size (batchsize, w*h, w*h, channels)
"""
w = encoded_l.shape[1]
h = encoded_l.shape[2]
chan = encoded_l.shape[3]
reshapel = Reshape((w * h, 1, chan)) # reshape layer
reshaped_l = reshapel(encoded_l)
reshaper = Reshape((1, w * h, chan))
reshaped_r = reshaper(encoded_r)
# 之所以是4个维度是因为第一个维度需要给batch,即Reshape输出的就是四维的
tiled_l = Lambda(lambda x: K.tile(x, [1, 1, w * h, 1]))(reshaped_l)
tiled_r = Lambda(lambda x: K.tile(x, [1, w * h, 1, 1]))(reshaped_r)
if negateDiffs:
diff = Lambda(lambda x: -K.abs(x[0] - x[1]))([tiled_l, tiled_r])
else:
diff = Lambda(lambda x: K.abs(x[0] - x[1]))([tiled_l, tiled_r])
# print("diff类型+++++++++++++", diff)
return diff
def call(self, inputs):
_, kernel_b = xnorize(self.kernel, self.H)
_, inputs_b = xnorize(inputs)
outputs = K.conv2d(inputs_b,
kernel_b,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
# calculate Wa and xa
# kernel_a
mask = K.reshape(
self.kernel,
(-1,
self.filters)) # self.nb_row * self.nb_col * channels, filters
kernel_a = K.stop_gradient(K.mean(K.abs(mask), axis=0)) # filters
# inputs_a
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
mask = K.mean(K.abs(inputs), axis=channel_axis, keepdims=True)
ones = K.ones(self.kernel_size + (1, 1))
inputs_a = K.conv2d(mask,
ones,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate
) # nb_sample, 1, new_nb_row, new_nb_col
if self.data_format == 'channels_first':
outputs = outputs * K.stop_gradient(inputs_a) * K.expand_dims(
K.expand_dims(K.expand_dims(kernel_a, 0), -1), -1)
else:
outputs = outputs * K.stop_gradient(inputs_a) * K.expand_dims(
K.expand_dims(K.expand_dims(kernel_a, 0), 0), 0)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, x):
if len(x) != 2:
raise Exception('input layers must be a list: mean and stddev')
if len(x[0].shape) != 2 or len(x[1].shape) != 2:
raise Exception('input shape is not a vector [batchSize, latentSize]')
mean = x[0]
stddev = x[1]
if self.reg == 'bvae':
# kl divergence:
latent_loss = -0.5 * K.mean(1 + stddev
- K.square(mean)
- K.exp(stddev), axis=-1)
# use beta to force less usage of vector space:
# also try to use <capacity> dimensions of the space:
latent_loss = self.beta * K.abs(latent_loss - self.capacity/self.shape.as_list()[1])
self.add_loss(latent_loss, x)
elif self.reg == 'vae':
# kl divergence:
latent_loss = -0.5 * K.mean(1 + stddev
- K.square(mean)
- K.exp(stddev), axis=-1)
self.add_loss(latent_loss, x)
epsilon = K.random_normal(shape=self.shape,
mean=0., stddev=1.)
if self.random:
# 'reparameterization trick':
return mean + K.exp(stddev) * epsilon
else: # do not perform random sampling, simply grab the impulse value
return mean + 0*stddev # Keras needs the *0 so the gradinent is not None
def call(self, x):
if len(x) != 2:
raise Exception('input layers must be a list: mean and stddev')
if len(x[0].shape) != 2 or len(x[1].shape) != 2:
raise Exception(
'input shape is not a vector [batchSize, latentSize]')
mean = x[0]
stddev = x[1]
if self.reg == 'bvae':
# kl divergence:
latent_loss = -0.5 * K.mean(
1 + stddev - K.square(mean) - K.exp(stddev), axis=-1)
# use beta to force less usage of vector space:
# also try to use <capacity> dimensions of the space:
latent_loss = self.beta * K.abs(latent_loss - self.capacity /
self.shape.as_list()[1])
self.add_loss(latent_loss, x)
elif self.reg == 'vae':
# kl divergence:
latent_loss = -0.5 * K.mean(
1 + stddev - K.square(mean) - K.exp(stddev), axis=-1)
self.add_loss(latent_loss, x)
if self.random:
# 'reparameterization trick':
epsilon = K.random_normal(shape=(self.batchSize, self.latentSize),
mean=0.,
stddev=1.)
return mean + K.exp(stddev) * epsilon
else: # do not perform random sampling, simply grab the impulse value
return mean + 0 * stddev # Keras needs the *0 so the gradinent is not None
def define_deepDream_model_layerBased(model):
dream = model.input
print('Model loaded.')
# Get the symbolic outputs of each "key" layer (we gave them unique names).
layer_dict = dict([(layer.name, layer) for layer in model.layers])
# Define the loss.
loss = K.variable(0.)
for layer_name in settings['features']:
# Add the L2 norm of the features of a layer to the loss.
if layer_name not in layer_dict:
raise ValueError('Layer ' + layer_name + ' not found in model.')
coeff = settings['features'][layer_name]
x = layer_dict[layer_name].output
# We avoid border artifacts by only involving non-border pixels in the loss.
scaling = K.prod(K.cast(K.shape(x), 'float32'))
if K.image_data_format() == 'channels_first':
loss = loss + coeff * K.sum(K.square(x[:, :, 2:-2,
2:-2])) / scaling
else:
loss = loss + coeff * K.sum(K.square(x[:, 2:-2,
2:-2, :])) / scaling
# Compute the gradients of the dream wrt the loss.
grads = K.gradients(loss, dream)[0]
# Normalize gradients.
grads /= K.maximum(K.mean(K.abs(grads)), K.epsilon())
# Set up function to retrieve the value
# of the loss and gradients given an input image.
outputs = [loss, grads]
fetch_loss_and_grads = K.function([dream], outputs)
def __init__(self, model, layer_name):
self.model = model
self.layer_name = layer_name
dream = model.input
# Get the symbolic outputs of each "key" layer (we gave them unique names).
layers_all = [layer.name for layer in model.layers]
if layer_name not in layers_all:
raise ValueError('Layer ' + layer_name + ' not found in model.')
# Define the loss.
loss = K.variable(0.)
for layer_local in model.layers:
if layer_local.name == layer_name:
x = layer_local.output
# We avoid border artifacts by only involving non-border pixels in the loss.
if K.image_data_format() == 'channels_first':
scaling = K.prod(K.cast(K.shape(x), 'float32'))
loss = loss + K.sum(K.square(x[:, :, 2:-2,
2:-2])) / scaling
else:
scaling = K.prod(K.cast(K.shape(x), 'float32'))
loss = loss + K.sum(K.square(x[:, 2:-2,
2:-2, :])) / scaling
# Compute the gradients of the dream wrt the loss.
grads = K.gradients(loss, dream)[0]
# Normalize gradients.
grads /= K.maximum(K.mean(K.abs(grads)), K.epsilon())
# Set up function to retrieve the value
# of the loss and gradients given an input image.
outputs = [loss, grads]
self.fetch_loss_and_grads = K.function([dream], outputs)
def __init__(self):
tf.logging.set_verbosity(tf.logging.ERROR)
self.__DIMEN = 48
input_shape = ((self.__DIMEN**2) * 3, )
convolution_shape = (self.__DIMEN, self.__DIMEN, 3)
kernel_size_1 = (8, 8)
kernel_size_2 = (6, 6)
kernel_size_3 = (4, 4)
pool_size_1 = (6, 6)
pool_size_2 = (4, 4)
strides = 1
seq_conv_model = [
tf.keras.layers.Reshape(input_shape=input_shape,
target_shape=convolution_shape),
tf.keras.layers.Conv2D(32,
kernel_size=kernel_size_1,
strides=strides,
activation='relu'),
tf.keras.layers.MaxPooling2D(pool_size=pool_size_1,
strides=strides),
tf.keras.layers.Conv2D(64,
kernel_size=kernel_size_2,
strides=strides,
activation='relu'),
tf.keras.layers.MaxPooling2D(pool_size=pool_size_2,
strides=strides),
tf.keras.layers.Conv2D(128,
kernel_size=kernel_size_3,
strides=strides,
activation='relu'),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(3076,
activation=tf.keras.activations.sigmoid)
]
seq_model = tf.keras.Sequential(seq_conv_model)
input_x1 = tf.keras.layers.Input(shape=input_shape)
input_x2 = tf.keras.layers.Input(shape=input_shape)
output_x1 = seq_model(input_x1)
output_x2 = seq_model(input_x2)
distance_euclid = tf.keras.layers.Lambda(
lambda tensors: K.abs(tensors[0] - tensors[1]))(
[output_x1, output_x2])
outputs = tf.keras.layers.Dense(
1, activation=tf.keras.activations.sigmoid)(distance_euclid)
self.__model = tf.keras.models.Model([input_x1, input_x2], outputs)
self.__model.compile(loss=tf.keras.losses.binary_crossentropy,
optimizer=tf.keras.optimizers.Adam(lr=0.0001),
metrics=['accuracy'])
def create_model(self):
K.clear_session()
input0 = Input(shape=(self.c['sentencepad'], self.c['wordvectdim']))
input1 = Input(shape=(self.c['sentencepad'], self.c['wordvectdim']))
Convolt_Layer = []
MaxPool_Layer = []
Flatten_Layer = []
for kernel_size, filters in self.c['cnnfilters'].items():
Convolt_Layer.append(
Convolution1D(filters=filters,
kernel_size=kernel_size,
padding='valid',
activation=self.c['cnnactivate'],
kernel_initializer=self.c['cnninitial']))
MaxPool_Layer.append(
MaxPooling1D(pool_size=int(self.c['sentencepad'] -
kernel_size + 1)))
Flatten_Layer.append(Flatten())
Convolted_tensor0 = []
Convolted_tensor1 = []
for channel in range(len(self.c['cnnfilters'])):
Convolted_tensor0.append(Convolt_Layer[channel](input0))
Convolted_tensor1.append(Convolt_Layer[channel](input1))
MaxPooled_tensor0 = []
MaxPooled_tensor1 = []
for channel in range(len(self.c['cnnfilters'])):
MaxPooled_tensor0.append(MaxPool_Layer[channel](
Convolted_tensor0[channel]))
MaxPooled_tensor1.append(MaxPool_Layer[channel](
Convolted_tensor1[channel]))
Flattened_tensor0 = []
Flattened_tensor1 = []
for channel in range(len(self.c['cnnfilters'])):
Flattened_tensor0.append(Flatten_Layer[channel](
MaxPooled_tensor0[channel]))
Flattened_tensor1.append(Flatten_Layer[channel](
MaxPooled_tensor1[channel]))
if len(self.c['cnnfilters']) > 1:
Flattened_tensor0 = concatenate(Flattened_tensor0)
Flattened_tensor1 = concatenate(Flattened_tensor1)
else:
Flattened_tensor0 = Flattened_tensor0[0]
Flattened_tensor1 = Flattened_tensor1[0]
absDifference = Lambda(lambda X: K.abs(X[0] - X[1]))(
[Flattened_tensor0, Flattened_tensor1])
mulDifference = multiply([Flattened_tensor0, Flattened_tensor1])
allDifference = concatenate([absDifference, mulDifference])
for ilayer, densedimension in enumerate(self.c['densedimension']):
allDifference = Dense(
units=int(densedimension),
activation=self.c['denseactivate'],
kernel_initializer=self.c['denseinitial'])(allDifference)
output = Dense(
name='output',
units=self.c['num_classes'],
activation='softmax',
kernel_initializer=self.c['denseinitial'])(allDifference)
self.model = Model(inputs=[input0, input1], outputs=output)
self.model.compile(loss='mean_squared_error',
optimizer=self.c['optimizer'])
def additional_generator_losses(self):
loss_list = super(CGAN, self).additional_generator_losses()
l1_loss = self.l1_weight_penalty * K.mean(
K.abs(self.gt_image_placeholder - self.generator_output[0]))
loss_list.append(l1_loss)
self.generator_metric_names.append('l1')
return loss_list
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 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 smooth_l1_loss(y_true, y_pred):
"""Implements Smooth-L1 loss.
y_true and y_pred are typically: [N, 4], but could be any shape.
"""
diff = K.abs(y_true - y_pred)
less_than_one = K.cast(K.less(diff, 1.0), "float32")
loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5)
return loss
def call(self, x):
y_pred = x[0]
y_recont_gt = x[1]
y_prob_pred = tf.squeeze(x[2], axis=3)
y_prob_gt = x[3]
visible = tf.cast(y_prob_gt > 0.5, y_pred.dtype)
visible = tf.squeeze(visible, axis=3)
#generate transformed values using sym
if (len(self.sym) > 1):
#if(True):
for sym_id, transform in enumerate(self.sym): #3x3 matrix
tf_mat = tf.convert_to_tensor(transform, y_recont_gt.dtype)
y_gt_transformed = tf.transpose(
tf.matmul(tf_mat,
tf.transpose(tf.reshape(y_recont_gt, [-1, 3]))))
y_gt_transformed = tf.reshape(y_gt_transformed,
[-1, 128, 128, 3])
loss_xyz_temp = K.sum(K.abs(y_gt_transformed - y_pred),
axis=3) / 3
loss_sum = K.sum(loss_xyz_temp, axis=[1, 2])
if (sym_id > 0):
loss_sums = tf.concat(
[loss_sums,
tf.expand_dims(loss_sum, axis=0)], axis=0)
loss_xyzs = tf.concat(
[loss_xyzs,
tf.expand_dims(loss_xyz_temp, axis=0)],
axis=0)
else:
loss_sums = tf.expand_dims(loss_sum, axis=0)
loss_xyzs = tf.expand_dims(loss_xyz_temp, axis=0)
min_values = tf.reduce_min(loss_sums, axis=0, keepdims=True)
loss_switch = tf.cast(tf.equal(loss_sums, min_values),
y_pred.dtype)
loss_xyz = tf.expand_dims(tf.expand_dims(loss_switch, axis=2),
axis=3) * loss_xyzs
loss_xyz = K.sum(loss_xyz, axis=0)
else:
loss_xyz = K.sum(K.abs(y_recont_gt - y_pred), axis=3) / 3
prob_loss = K.square(y_prob_pred - K.minimum(loss_xyz, 1))
loss_invisible = (1 - visible) * loss_xyz
loss_visible = visible * loss_xyz
loss = loss_visible * 3 + loss_invisible + 0.5 * prob_loss
loss = K.mean(loss, axis=[1, 2])
return loss
def __call__(self, x):
regularization = 0
if self.l1:
regularization += K.sum(self.l1 * K.abs(x))
if self.l2:
regularization += K.sum(self.l2 * K.square(x))
if self.tv:
regularization += K.sum(self.tv * K.square(x[:, 1:] - x[:, :-1]))
return regularization
def define_rnn_network(hidden_state_dim, timepoints, timepoint_step):
rnn_input = layers.Input(shape=(40, ), name="g_input")
x = layers.Dense(64, activation="relu")(rnn_input)
x = layers.Dense(64, activation="relu")(x)
# The last output predicts the initial state for the RNN.
gru_state = layers.Dense(hidden_state_dim, activation="relu")(x)
# The RNN state is passed through a separate regressor to obtain the (mean,
# scale) for each of the 4 dimensions.
gru = layers.GRU(hidden_state_dim)
intermediate_regressor = layers.Dense(64, activation="relu")
mean_and_scale_regressor = layers.Dense(8,
activation="linear",
name="mean_and_scale_regressor")
# Ensure the regressed scale is positive; also reshape to Bx8 -> Bx1x8.
kSigmaMin = 1e-3 # avoid poor conditioning
absolute_scale_op = layers.Lambda(lambda x: K.concatenate(
(x[:, :4], K.abs(x[:, 4:]) + kSigmaMin), axis=1)[:, tf.newaxis, :])
output_regressor = lambda x: \
absolute_scale_op(mean_and_scale_regressor(intermediate_regressor(x)))
current_output = output_regressor(gru_state) # predict the first output
t = timepoint_step
next_timepoint_idx = 0
outputs = []
while t <= timepoints[-1] or np.isclose(t, timepoints[-1]):
if np.isclose(t, timepoints[next_timepoint_idx]):
outputs.append(current_output)
t = timepoints[next_timepoint_idx] # avoid any accumulative drift
next_timepoint_idx += 1
if next_timepoint_idx == len(timepoints):
break
#current_output_with_time = layers.Lambda(
# lambda x: K.concatenate((x, K.zeros_like(x)[:,:,:1] + t)))(
# current_output)
gru_state = gru(current_output, initial_state=gru_state)
current_output = output_regressor(gru_state)
t += timepoint_step
assert (len(outputs) == len(timepoints))
# Join the T [Bx1x8] outputs into a Bx8xT tensor, then split in half down
# the first axis and reform into a Bx4xTx2 tensor.
def rejoin_op(x):
x = K.stack(x, axis=-1)
return K.stack((x[:, 0, :4, :], x[:, 0, 4:, :]), axis=-1)
output = layers.Lambda(rejoin_op, name="transforms")(outputs)
M = models.Model(inputs=[rnn_input],
outputs=[output],
name='rnn_regressor')
return M
def _signed_sqrt(x):
'''Calculate element-wise signed square-root.
Args:
x: input tensor.
Returns:
Element-wise signed square-root tensor.
'''
return keras_backend.sign(x) * keras_backend.sqrt(keras_backend.abs(x) + 1e-9)
def weighted_bce_loss(y_true, y_pred, weight):
# avoiding overflow
epsilon = 1e-7
y_pred = K.clip(y_pred, epsilon, 1. - epsilon)
logit_y_pred = K.log(y_pred / (1. - y_pred))
# https://www.tensorflow.org/api_docs/python/tf/nn/weighted_cross_entropy_with_logits
loss = (1. - y_true) * logit_y_pred + (1. + (weight - 1.) * y_true) * \
(K.log(1. + K.exp(-K.abs(logit_y_pred))) + K.maximum(-logit_y_pred, 0.))
return K.sum(loss) / K.sum(weight)
def _compare(row1, row2):
### 比较行与行
# shape (5),(5)
# 比较方式: 类型+个数
_sum = tf.constant(0., dtype=tf.float32)
_cnt = tf.constant(0., dtype=tf.float32)
# 类型--两两相乘/总次数
r = 0
while (r < board_size):
c = 0
vr = tf.slice(row1, [r], [1])
vr = K.squeeze(vr, -1)
while (c < board_size):
vc = tf.slice(row2, [c], [1])
vc = K.squeeze(vc, -1)
calc = K.abs(vr + vc) / (K.abs(vr - vc) + K.epsilon())
calc = K.clip(calc, 1, 9) #裁剪
calc = calc - 0.5 + K.abs(vr + vc) * 0.001
_sum = tf.cond(
tf.logical_or(tf.not_equal(vc, 0), tf.not_equal(vr, 0)),
lambda: _sum + calc, lambda: _sum)
_cnt = tf.cond(
tf.logical_or(tf.not_equal(vc, 0), tf.not_equal(vr, 0)),
lambda: _cnt + 1, lambda: _cnt)
c = c + 1
r = r + 1
cnt3 = _compare_cnt(row1, row2)
cnt3 = tf.cond(tf.equal(cnt3, 0), lambda: cnt3 + 1, lambda: cnt3)
output = tf.cond(tf.equal(_cnt, 0), lambda: _sum,
lambda: tf.squeeze(_sum * cnt3 / _cnt))
### 查看行与行比较结果
# if EmbeddingsLayer.debug:
# logging.getLogger().info("\n--row1 %s" % (row1))
# logging.getLogger().info("--row2 %s" % (row2))
# logging.getLogger().info("--sum %s" % (_sum))
# logging.getLogger().info("--cnt %s" % (_cnt))
# logging.getLogger().info("--compare %s" % (output))
###
### 直接输出概率:
### 手动测试0.7比较好
output = output * 0.7
return output
def __init__(self):
tf.logging.set_verbosity(tf.logging.ERROR)
self.__DIMEN = 128
input_shape = ((self.__DIMEN**2) * 3, )
convolution_shape = (self.__DIMEN, self.__DIMEN, 3)
kernel_size_1 = (4, 4)
kernel_size_2 = (3, 3)
pool_size_1 = (3, 3)
pool_size_2 = (2, 2)
strides = 1
seq_conv_model = [
Reshape(input_shape=input_shape, target_shape=convolution_shape),
Conv2D(32,
kernel_size=kernel_size_1,
strides=strides,
activation=activations.leaky_relu),
Conv2D(32,
kernel_size=kernel_size_1,
strides=strides,
activation=activations.leaky_relu),
MaxPooling2D(pool_size=pool_size_1, strides=strides),
Conv2D(64,
kernel_size=kernel_size_2,
strides=strides,
activation=activations.leaky_relu),
Conv2D(64,
kernel_size=kernel_size_2,
strides=strides,
activation=activations.leaky_relu),
MaxPooling2D(pool_size=pool_size_2, strides=strides),
Flatten(),
Dense(64, activation=activations.sigmoid)
]
seq_model = tf.keras.Sequential(seq_conv_model)
input_x1 = Input(shape=input_shape)
input_x2 = Input(shape=input_shape)
output_x1 = seq_model(input_x1)
output_x2 = seq_model(input_x2)
distance_euclid = Lambda(
lambda tensors: K.abs(tensors[0] - tensors[1]))(
[output_x1, output_x2])
outputs = Dense(1, activation=activations.sigmoid)(distance_euclid)
self.__model = models.Model([input_x1, input_x2], outputs)
self.__model.compile(loss=losses.binary_crossentropy,
optimizer=optimizers.Adam(lr=0.0001))
def cycle_loss(y_true, y_pred):
if k.image_data_format() is 'channels_first':
x_w = 2
x_h = 3
else:
x_w = 1
x_h = 2
loss = k.abs(y_true - y_pred)
loss = k.sum(loss, axis=x_h)
loss = k.sum(loss, axis=x_w)
return loss
def not_equal(f, other, tol=None):
"""Element-wise comparison applied to the `Functional` objects.
# Arguments
f: Functional object.
other: A python number or a tensor or a functional object.
tol: (float) If you need a tolerance measure.
# Returns
A Functional.
"""
validate_functional(f)
assert isinstance(
tol, (type(None), float)), 'Expected a floating value for `tol`.'
inputs = f.inputs.copy()
if is_functional(other):
inputs += to_list(other.inputs)
if tol is None:
lambda_opr = lambda x: K.cast_to_floatx(K.not_equal(x[0], x[1]))
else:
lambda_opr = lambda x: K.cast_to_floatx(
K.greater(K.abs(x[0] - x[1]), tol))
else:
_warn_for_ndarray(other)
if tol is None:
lambda_opr = lambda x: K.cast_to_floatx(K.not_equal(x, other))
else:
lambda_opr = lambda x: K.cast_to_floatx(
K.greater(K.abs(x - other), tol))
lmbd = [
Lambda(lambda_opr, name=graph_unique_name("not_equal"))
for X in f.outputs
]
Functional = f.get_class()
res = Functional(inputs=unique_tensors(inputs),
outputs=_apply_operation(lmbd, f, other),
layers=lmbd)
return res
def custom_loss(y_true, y_pred, loss_weights = loss_weights): # Verified
zero_index = K.zeros_like(y_true[:, 0])
ones_index = K.ones_like(y_true[:, 0])
# Classifier
labels = y_true[:, 0]
class_preds = y_pred[:, 0]
bi_crossentropy_loss = -labels * K.log(class_preds) - (1 - labels) * K.log(1 - class_preds)
classify_valid_index = tf.where(K.less(y_true[:, 0], 0), zero_index, ones_index)
classify_keep_num = K.cast(tf.cast(tf.reduce_sum(classify_valid_index), tf.float32) * SAMPLE_KEEP_RATIO, dtype = tf.int32)
# For classification problem, only pick 70% of the valid samples.
classify_loss_sum = bi_crossentropy_loss * tf.cast(classify_valid_index, bi_crossentropy_loss.dtype)
classify_loss_sum_filtered, _ = tf.nn.top_k(classify_loss_sum, k = classify_keep_num)
classify_loss = tf.where(K.equal(classify_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(classify_loss_sum_filtered))
# Bounding box regressor
rois = y_true[:, 1: 5]
roi_preds = y_pred[:, 1: 5]
roi_raw_mean_square_error = K.sum(K.square(rois - roi_preds), axis = 1) # mse
# roi_raw_smooth_l1_loss = K.mean(tf.where(K.abs(rois - roi_preds) < 1, 0.5 * K.square(rois - roi_preds), K.abs(rois - roi_preds) - 0.5)) # L1 Smooth Loss
roi_valid_index = tf.where(K.equal(K.abs(y_true[:, 0]), 1), ones_index, zero_index)
roi_keep_num = K.cast(tf.reduce_sum(roi_valid_index), dtype = tf.int32)
roi_valid_mean_square_error = roi_raw_mean_square_error * tf.cast(roi_valid_index, roi_raw_mean_square_error.dtype)
roi_filtered_mean_square_error, _ = tf.nn.top_k(roi_valid_mean_square_error, k = roi_keep_num)
roi_loss = tf.where(K.equal(roi_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(roi_filtered_mean_square_error))
# roi_valid_smooth_l1_loss = roi_raw_smooth_l1_loss * roi_valid_index
# roi_filtered_smooth_l1_loss, _ = tf.nn.top_k(roi_valid_smooth_l1_loss, k = roi_keep_num)
# roi_loss = K.mean(roi_filtered_smooth_l1_loss)
# Landmark regressor
pts = y_true[:, 5: 17]
pt_preds = y_pred[:, 5: 17]
pts_raw_mean_square_error = K.sum(K.square(pts - pt_preds), axis = 1) # mse
# pts_raw_smooth_l1_loss = K.mean(tf.where(K.abs(pts - pt_preds) < 1, 0.5 * K.square(pts - pt_preds), K.abs(pts - pt_preds) - 0.5)) # L1 Smooth Loss
pts_valid_index = tf.where(K.equal(y_true[:, 0], -2), ones_index, zero_index)
pts_keep_num = K.cast(tf.reduce_sum(pts_valid_index), dtype = tf.int32)
pts_valid_mean_square_error = pts_raw_mean_square_error * tf.cast(pts_valid_index, tf.float32)
pts_filtered_mean_square_error, _ = tf.nn.top_k(pts_valid_mean_square_error, k = pts_keep_num)
pts_loss = tf.where(K.equal(pts_keep_num, 0), tf.constant(0, dtype = tf.float32), K.mean(pts_filtered_mean_square_error))
# pts_valid_smooth_l1_loss = pts_raw_smooth_l1_loss * pts_valid_index
# pts_filtered_smooth_l1_loss, _ = tf.nn.top_k(pts_valid_smooth_l1_loss, k = pts_keep_num)
# pts_loss = K.mean(pts_filtered_smooth_l1_loss)
loss = classify_loss * loss_weights[0] + roi_loss * loss_weights[1] + pts_loss * loss_weights[2]
return loss
def dice_coef_loss(y_true, y_pred, smooth=1e-07, label_smoothing=0):
y_pred = ops.convert_to_tensor(y_pred)
y_true = math_ops.cast(y_true, y_pred.dtype)
label_smoothing = ops.convert_to_tensor(label_smoothing, dtype=K.floatx())
def _smooth_labels():
return y_true * (1.0 - label_smoothing) + 0.5 * label_smoothing
y_true = smart_cond.smart_cond(label_smoothing, _smooth_labels,
lambda: y_true)
return (2. * K.sum(K.abs(y_true * y_pred), axis=-1) + smooth) / (
K.sum(K.square(y_true), -1) + K.sum(K.square(y_pred), -1) + smooth)
def call(self, inputs, **kwargs):
W = K.tanh(self.W_hat) * K.sigmoid(self.M_hat)
a = K.dot(inputs, W)
if self.nac_only:
outputs = a
else:
m = K.exp(K.dot(K.log(K.abs(inputs) + self.epsilon), W))
g = K.sigmoid(K.dot(inputs, self.G))
outputs = g * a + (1. - g) * m
return outputs
def my_sigmoid_loss(y_true, y_pred):
""" Loss function in form of a mean sigmoid. Used for overlap.
This is an alternative for mean squard error where
- the loss for small differences is smaller than squared diff
- the loss for large errors is kind of equal
In Matlab: 1./(1+exp(-((diff+0.25)*24-12))), diff is absolute difference
"""
diff = K.abs(y_pred - y_true)
sigmoidx = (diff + 0.25) * 24 - 12
loss = K.mean(1 / (1 + K.exp(-sigmoidx)))
return loss
def define_poly_network(poly_order, timepoints, past_frames):
poly_input = layers.Input(shape=(past_frames * 4, ), name="g_input")
x = layers.Dense(64, activation="relu")(poly_input)
x = layers.Dense(64, activation="relu")(x)
x = layers.Dense(64, activation="relu")(x)
#x = layers.Dense(64, activation="relu")(x)
# coeffs: for each output dimension, (a, b, c, ..., sigma), where sigma is
# the confidence value
coeffs = layers.Dense(4 * (poly_order + 2),
activation="linear",
name="coeffs")(x)
coeffs = layers.Reshape((4, poly_order + 2))(coeffs)
# timepoints: PxT for P polynomial coefficients, i.e.
# [ t_0 t_1 ... ]
# [ t_0^2 t_1^2 ... ]
# [ ... ... ... ]
timepoints = K.constant([[pow(t, i) for t in timepoints]
for i in range(1, poly_order + 1)])
# generate distribution mean and standard deviation
# the mean is computed as c_1 * t^P + c_2 * t^{P-1} + ... + c_P * t
# the std. dev. is computed as |d_0 + d_1 * t| + eps
kSigmaMin = 1e-3 # avoid poor conditioning
mu = layers.Lambda(lambda x: K.dot(x[..., :-2], timepoints))(coeffs)
sigma = layers.Lambda(lambda x: K.abs(x[..., -1, tf.newaxis] * timepoints[
0, :]) + K.abs(x[..., -2, tf.newaxis]) + kSigmaMin)(coeffs)
output = layers.Lambda(lambda x: K.stack(x, axis=-1),
name="transforms")([mu, sigma])
M = models.Model(inputs=[poly_input],
outputs=[output, coeffs],
name='poly_regressor')
return M
