def clipped_relu(value):
"""Returns clipped relu, clip value set to 20."""
return K.relu(value, max_value=20)
def relu6(x):
return K.relu(x,max_value=6)
def train_step(self,
images,
style,
noise,
perform_gp=True,
perform_pl=False):
with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:
#Get style information
w_space = []
pl_lengths = self.pl_mean
for i in range(len(style)):
w_space.append(self.GAN.S(style[i]))
#Generate images
generated_images = self.GAN.G(w_space + [noise])
#Discriminate
real_output = self.GAN.D(images, training=True)
fake_output = self.GAN.D(generated_images, training=True)
#Hinge loss function
gen_loss = K.mean(fake_output)
divergence = K.mean(
K.relu(1 + real_output) + K.relu(1 - fake_output))
disc_loss = divergence
if perform_gp:
#R1 gradient penalty
disc_loss += gradient_penalty(images, real_output, 10)
if perform_pl:
#Slightly adjust W space
w_space_2 = []
for i in range(len(style)):
std = 0.1 / (K.std(w_space[i], axis=0, keepdims=True) +
1e-8)
w_space_2.append(w_space[i] +
K.random_normal(tf.shape(w_space[i])) /
(std + 1e-8))
#Generate from slightly adjusted W space
pl_images = self.GAN.G(w_space_2 + [noise])
#Get distance after adjustment (path length)
delta_g = K.mean(K.square(pl_images - generated_images),
axis=[1, 2, 3])
pl_lengths = delta_g
if self.pl_mean > 0:
gen_loss += K.mean(K.square(pl_lengths - self.pl_mean))
#Get gradients for respective areas
gradients_of_generator = gen_tape.gradient(
gen_loss, self.GAN.GM.trainable_variables)
gradients_of_discriminator = disc_tape.gradient(
disc_loss, self.GAN.D.trainable_variables)
#Apply gradients
self.GAN.GMO.apply_gradients(
zip(gradients_of_generator, self.GAN.GM.trainable_variables))
self.GAN.DMO.apply_gradients(
zip(gradients_of_discriminator, self.GAN.D.trainable_variables))
return disc_loss, gen_loss, divergence, pl_lengths
Y_test.sort(axis=1) #排一下序，因为只比较集合，不比较顺序
#搭建普通CNN分类模型
input_image = Input(shape=(None, None, 1))
cnn = Conv2D(64, (3, 3), activation='relu')(input_image)
cnn = Conv2D(64, (3, 3), activation='relu')(cnn)
cnn = AveragePooling2D((2, 2))(cnn)
cnn = Conv2D(128, (3, 3), activation='relu')(cnn)
cnn = Conv2D(128, (3, 3), activation='relu')(cnn)
cnn = GlobalAveragePooling2D()(cnn)
dense = Dense(128, activation='relu')(cnn)
output = Dense(10, activation='sigmoid')(dense)
model = Model(inputs=input_image, outputs=output)
model.compile(
loss=lambda y_true, y_pred: y_true * K.relu(0.9 - y_pred)**2 + 0.25 *
(1 - y_true) * K.relu(y_pred - 0.1)**2,
optimizer='adam',
metrics=['accuracy'])
model.summary()
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=20,
verbose=1,
validation_data=(x_test, y_test))
Y_pred = model.predict(X_test) #用模型进行预测
greater = np.sort(Y_pred, axis=1)[:, -2] > 0.5 #判断预测结果是否大于0.5
def call(self, x, mask=None):
x_lo = K.relu(self.lo - x)
x_hi = K.relu(x - self.hi)
return x + self.gamma * x_lo - self.gamma * x_hi
def e_insensitive_loss(y_true, y_pred):
return K.relu(K.abs(y_true - y_pred) - .1)
def hard_swish(inputs):
return inputs * K.relu(inputs + 3.0, max_value=6.0) / 6.0
def call(self, x):
# print("!",x.shape)
E = K.reshape(x[:, :, 0], (-1, self.gs, 1))
p1 = K.reshape(x[:, :, 1], (-1, self.gs, 1))
p2 = K.reshape(x[:, :, 2], (-1, self.gs, 1))
p3 = K.reshape(x[:, :, 3], (-1, self.gs, 1))
pt = K.sqrt(p1**2 + p2**2)
p = K.sqrt(pt**2 + p3**2)
iszero = p**2 + E**2
iszero = 1 - K.relu(1 - self.numericC * iszero) + K.relu(
-self.numericC * iszero)
#return iszero
eta = iszero * 0.5 * K.log(0.0000000001 + (p + p3) /
(p - p3 + 0.0000000001))
#phi=iszero*t.math.acos(p3/(p+0.0000001))
phi = iszero * t.math.atan2(p2, p1)
#print("eta",eta.shape,"phi",phi.shape)
eta = K.reshape(eta, (-1, self.gs))
phi = K.reshape(phi, (-1, self.gs))
meta = K.mean(eta, axis=-1)
mp1 = K.mean(p1, axis=-2)
mp2 = K.mean(p2, axis=-2)
#print(p2.shape,p1.shape)
#print(mp1.shape,mp2.shape)
#exit()
mphi = t.math.atan2(mp2, mp1)
#print("meta",meta.shape,"mphi",mphi.shape)
#mphi=K.mean(phi,axis=-1)
#exit()
meta = K.reshape(K.repeat_elements(meta, self.gs, 0), (-1, self.gs))
mphi = K.repeat_elements(mphi, self.gs, 1)
#print("meta",meta.shape,"mphi",mphi.shape)
#deta=eta#-meta##not sure if abs here
#dphi=phi#-mphi##not sure here either
siszero = K.reshape(iszero, (-1, self.gs))
deta = K.reshape(siszero * (eta - meta), (-1, self.gs, 1))
dphi = K.reshape(siszero * (phi - mphi), (-1, self.gs, 1))
pi = t.constant(math.pi)
#dphi=K.min([t.math.floormod(dphi,2*pi),t.math.floormod(-dphi,2*pi)],axis=0)
opta = t.math.floormod(dphi, 2 * pi)
optb = t.math.floormod(-dphi, 2 * pi)
dphi = t.where(t.greater(opta, optb), optb, -opta)
#dphi=K.reshape(dphi,(-1,self.gs,1))#should actually be useless?
#dphi=K.min(K.concatenate((t.math.floormod(dphi,2*pi),t.math.floormod(-dphi,2*pi)),axis=-1),axis=-1)
#dphi=K.reshape(dphi,(-1,self.gs,1))
#deta=iszero*K.permute_dimensions(K.permute_dimensions(eta,(1,0,2))-meta,(1,0,2))#not sure if abs here required
#dphi=iszero*K.permute_dimensions(K.permute_dimensions(phi,(1,0,2))-mphi,(1,0,2))#also not sure here either
spt = K.sum(pt, axis=-2)
ppt = -iszero * K.permute_dimensions(
K.log(0.000000001 + K.permute_dimensions(pt, (1, 0, 2)) /
(spt + 0.0000001)), (1, 0, 2)
) #please note the added sign in comparison to the original paper
#print(iszero.shape,deta.shape,dphi.shape,ppt.shape)
#print(eta.shape,meta.shape)
#print(phi.shape,mphi.shape)
#exit()
#print(iszero.shape,deta.shape,dphi.shape,pt.shape)
#exit()
#meta=K.reshape(meta,(-1,self.gs,1))
#mphi=K.reshape(mphi,(-1,self.gs,1))
#phi=K.reshape(phi,(-1,self.gs,1))
ret = K.concatenate((iszero, deta, dphi, ppt),
axis=-1) #adding iszero for numerical reasons
#print(ret.shape,x.shape)
#exit()
return ret
def call(self, x):
# print("!",x.shape)
E = K.reshape(x[:, :, 0], (-1, self.gs, 1))
p1 = K.reshape(x[:, :, 1], (-1, self.gs, 1))
p2 = K.reshape(x[:, :, 2], (-1, self.gs, 1))
p3 = K.reshape(x[:, :, 3], (-1, self.gs, 1))
pt = K.sqrt(p1**2 + p2**2)
p = K.sqrt(pt**2 + p3**2)
iszero = p**2 + E**2
iszero = 1 - K.relu(1 - self.numericC * iszero) + K.relu(
-self.numericC * iszero)
#return iszero
eta = iszero * 0.5 * K.log(0.0000000001 + (p + p3) /
(p - p3 + 0.0000000001))
#phi=iszero*t.math.acos(p3/(p+0.0000001))
phi = iszero * t.math.atan2(p2, p1)
#print("eta",eta.shape,"phi",phi.shape)
meta = K.mean(eta, axis=-2)
mphi = K.mean(phi, axis=-2)
#print("meta",meta.shape,"mphi",mphi.shape)
#deta=eta#-meta##not sure if abs here
#dphi=phi#-mphi##not sure here either
deta = iszero * K.permute_dimensions(
K.permute_dimensions(eta, (1, 0, 2)) - meta,
(1, 0, 2)) #not sure if abs here required
dphi = iszero * K.permute_dimensions(
K.permute_dimensions(phi, (1, 0, 2)) - mphi,
(1, 0, 2)) #also not sure here either
#print("deta",deta.shape,"dphi",dphi.shape)
lpt = iszero * K.log(pt + 0.0000001) ##
lE = iszero * K.log(E + 0.0000001) ##
#print("lpt",lpt.shape,"lE",lE.shape)
#rpt=K.reshape(pt,(-1,self.gs))
spt = K.sum(pt, axis=-2)
#print("spt",spt.shape)
#return K.permute_dimensions(K.log(1.0+K.abs(K.permute_dimensions(pt,(1,0,2))/(K.abs(spt)+1.0))),(1,0,2))
#ispt=1/(spt+0.000000001)
#print("ispt",ispt.shape)
ppt = -iszero * K.permute_dimensions(
K.log(0.000000001 + K.permute_dimensions(pt, (1, 0, 2)) /
(spt + 0.0000001)), (1, 0, 2)
) #please note the added sign in comparison to the original paper
#ppt=K.reshape(K.permute_dimensions(K.log(0.000000001+K.permute_dimensions(rpt,(1,0,2))/(spt+0.0000001)),(1,0)),(-1,self.gs,1))##
#ppt=K.reshape(K.permute_dimensions(K.log(0.000000001+K.permute_dimensions(rpt,(1,0))/(spt+0.0000001)),(1,0)),(-1,self.gs,1))##
#ppt=K.reshape(K.log(0.000000001+K.permute_dimensions(rpt,(1,0))/(spt+0.0000001)),(-1,self.gs,1))##
#ppt=iszero*K.log(pt/(spt+0.00000001))##
#print("ppt",ppt.shape)
sE = K.sum(E, axis=-2)
#print("sE",sE.shape)
pE = -iszero * K.permute_dimensions(
K.log(0.000000001 + K.permute_dimensions(E, (1, 0, 2)) /
(sE + 0.0000001)), (1, 0, 2)) #here was also a sign added
#pE=-iszero*K.log(sE/(E+0.00000001))##
#pE=-iszero
#print("pE",pE.shape)
dR = K.sqrt(deta**2 + dphi**2) ##
#print("dR",dR.shape)
ret = K.concatenate((iszero, deta, dphi, lpt, lE, ppt, pE, dR),
axis=-1) #adding iszero for numerical reasons
#print(ret.shape,x.shape)
#exit()
return ret
def call(self, x):
return self.w_2(self.dropout(K.relu(self.w_1(x))))
def _build_model(self, graph: Graph):
"""Return Siamese model."""
# Creating the inputs layers
inputs = [Input((1, ), dtype=tf.int32) for _ in range(4)]
edge_types = Input((1, ), dtype=tf.int32)
# Creating the embedding layer for the contexts
node_embedding_layer = Embedding(input_dim=graph.get_number_of_nodes(),
output_dim=self._embedding_size,
input_length=1,
name="node_embeddings")
# Get the node embedding
node_embeddings = [
UnitNorm(axis=-1)(node_embedding_layer(node_input))
for node_input in inputs
]
if self.requires_node_types():
max_node_types = graph.get_maximum_multilabel_count()
multilabel = graph.has_multilabel_node_types()
unknown_node_types = graph.has_unknown_node_types()
node_types_offset = int(multilabel or unknown_node_types)
node_type_inputs = [
Input((max_node_types, ), dtype=tf.int32) for _ in range(4)
]
node_type_embedding_layer = Embedding(
input_dim=graph.get_number_of_node_types() + node_types_offset,
output_dim=self._embedding_size,
input_length=max_node_types,
name="node_type_embeddings",
mask_zero=multilabel or unknown_node_types)
node_embeddings = [
Add()([
GlobalAveragePooling1D()(
node_type_embedding_layer(node_type_input)),
node_embedding
]) for node_type_input, node_embedding in zip(
node_type_inputs, node_embeddings)
]
else:
node_type_inputs = []
inputs.extend(node_type_inputs)
inputs.append(edge_types)
edge_types_number = graph.get_number_of_edge_types()
unknown_edge_types = graph.has_unknown_edge_types()
edge_types_offset = int(unknown_edge_types)
edge_type_embedding = GlobalAveragePooling1D()(Embedding(
input_dim=edge_types_number,
output_dim=self._embedding_size,
input_length=1 + edge_types_offset,
mask_zero=unknown_edge_types,
name="edge_type_embeddings",
)(edge_types))
(srcs_embedding, dsts_embedding, not_srcs_embedding,
not_dsts_embedding,
edge_type_embedding) = self._build_output(*node_embeddings,
edge_type_embedding,
edge_types)
loss = K.relu(self._relu_bias + tf.norm(
srcs_embedding + edge_type_embedding - dsts_embedding, axis=-1) -
tf.norm(not_srcs_embedding + edge_type_embedding -
not_dsts_embedding,
axis=-1))
# Creating the actual model
model = Model(inputs=inputs, outputs=loss, name=self.model_name())
model.add_loss(loss)
model.compile(optimizer=self._optimizer)
return model
def __call__(self,
loss,
seed_input,
penultimate_layer=-1,
seek_penultimate_conv_layer=True,
activation_modifier=lambda cam: K.relu(cam),
normalize_gradient=True,
expand_cam=True):
"""Generate a gradient based class activation map (CAM) by using positive gradient of
penultimate_layer with respect to loss.
For details on Grad-CAM, see the paper:
[Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via
Gradient-based Localization](https://arxiv.org/pdf/1610.02391v1.pdf).
# Arguments
loss: A loss function. If the model has multiple outputs, you can use a different
loss on each output by passing a list of losses.
seed_input: An N-dim Numpy array. If the model has multiple inputs,
you have to pass a list of N-dim Numpy arrays.
penultimate_layer: A number of integer or a tf.keras.layers.Layer object.
seek_penultimate_conv_layer: True to seek the penultimate layter that is a subtype of
`keras.layers.convolutional.Conv` class.
If False, the penultimate layer is that was elected by penultimate_layer index.
normalize_gradient: True to normalize gradients.
activation_modifier: A function to modify gradients.
expand_cam: True to expand cam to same as input image size.
![Note] Even if the model has multiple inputs, this function return only one cam
value (That's, when `expand_cam` is True, multiple cam images are generated from
a model that has multiple inputs).
# Returns
The heatmap image or a list of their images that indicate the `seed_input` regions
whose change would most contribute the loss value,
# Raises
ValueError: In case of invalid arguments for `loss`, or `penultimate_layer`.
"""
# Preparing
losses = self._get_losses_for_multiple_outputs(loss)
seed_inputs = self._get_seed_inputs_for_multiple_inputs(seed_input)
penultimate_output_tensor = self._find_penultimate_output(
penultimate_layer, seek_penultimate_conv_layer)
# Processing gradcam
model = tf.keras.Model(inputs=self.model.inputs,
outputs=self.model.outputs +
[penultimate_output_tensor])
with tf.GradientTape() as tape:
tape.watch(seed_inputs)
outputs = model(seed_inputs)
outputs, penultimate_output = outputs[:-1], outputs[-1]
loss_values = [loss(y) for y, loss in zip(outputs, losses)]
grads = tape.gradient(loss_values, penultimate_output)
if normalize_gradient:
grads = K.l2_normalize(grads)
weights = K.mean(grads,
axis=tuple(range(grads.ndim)[1:-1]),
keepdims=True)
cam = np.sum(penultimate_output * weights, axis=-1)
if activation_modifier is not None:
cam = activation_modifier(cam)
if not expand_cam:
return cam
# Visualizing
cam = self._zoom_for_visualizing(seed_inputs, cam)
if len(self.model.inputs) == 1 and not isinstance(seed_input, list):
cam = cam[0]
return cam
def relu6(x):
return K.relu(x)
def __call__(self,
loss,
seed_input,
penultimate_layer=-1,
seek_penultimate_conv_layer=True,
activation_modifier=lambda cam: K.relu(cam),
expand_cam=True,
training=False):
"""Generate gradient based class activation maps (CAM) by using positive gradient of
penultimate_layer with respect to loss.
For details on GradCAM++, see the paper:
[GradCAM++: Improved Visual Explanations for Deep Convolutional Networks]
(https://arxiv.org/pdf/1710.11063.pdf).
# Arguments
loss: A loss function. If the model has multiple outputs, you can use a different
loss on each output by passing a list of losses.
seed_input: An N-dim Numpy array. If the model has multiple inputs,
you have to pass a list of N-dim Numpy arrays.
penultimate_layer: A number of integer or a tf.keras.layers.Layer object.
seek_penultimate_conv_layer: True to seek the penultimate layter that is a subtype of
`keras.layers.convolutional.Conv` class.
If False, the penultimate layer is that was elected by penultimate_layer index.
activation_modifier: A function to modify gradients.
expand_cam: True to expand cam to same as input image size.
![Note] Even if the model has multiple inputs, this function return only one cam
value (That's, when `expand_cam` is True, multiple cam images are generated from
a model that has multiple inputs).
training: A bool whether the model's trainig-mode turn on or off.
# Returns
The heatmap image or a list of their images that indicate the `seed_input` regions
whose change would most contribute the loss value,
# Raises
ValueError: In case of invalid arguments for `loss`, or `penultimate_layer`.
"""
# Preparing
losses = self._get_losses_for_multiple_outputs(loss)
seed_inputs = self._get_seed_inputs_for_multiple_inputs(seed_input)
penultimate_output_tensor = self._find_penultimate_output(
penultimate_layer, seek_penultimate_conv_layer)
# Processing gradcam
model = tf.keras.Model(inputs=self.model.inputs,
outputs=self.model.outputs +
[penultimate_output_tensor])
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(seed_inputs)
outputs = model(seed_inputs, training=training)
outputs, penultimate_output = outputs[:-1], outputs[-1]
loss_values = [loss(y) for y, loss in zip(outputs, losses)]
grads = tape.gradient(
loss_values,
penultimate_output,
unconnected_gradients=tf.UnconnectedGradients.ZERO)
score = sum([K.exp(tf.reshape(v, (-1, ))) for v in loss_values])
score = tf.reshape(score,
(-1, ) + tuple(np.ones(grads.ndim - 1, np.int)))
first_derivative = score * grads
second_derivative = first_derivative * grads
third_derivative = second_derivative * grads
global_sum = K.sum(penultimate_output,
axis=tuple(
np.arange(len(penultimate_output.shape))[1:-1]),
keepdims=True)
alpha_denom = second_derivative * 2.0 + third_derivative * global_sum
alpha_denom = alpha_denom + tf.cast(
(second_derivative == 0.0), second_derivative.dtype)
alphas = second_derivative / alpha_denom
alpha_normalization_constant = K.sum(
alphas,
axis=tuple(np.arange(len(alphas.shape))[1:-1]),
keepdims=True)
alpha_normalization_constant = alpha_normalization_constant + tf.cast(
(alpha_normalization_constant == 0.0),
alpha_normalization_constant.dtype)
alphas = alphas / alpha_normalization_constant
if activation_modifier is None:
weights = first_derivative
else:
weights = activation_modifier(first_derivative)
deep_linearization_weights = weights * alphas
deep_linearization_weights = K.sum(
deep_linearization_weights,
axis=tuple(np.arange(len(deep_linearization_weights.shape))[1:-1]),
keepdims=True)
cam = K.sum(deep_linearization_weights * penultimate_output, axis=-1)
if activation_modifier is not None:
cam = activation_modifier(cam)
if not expand_cam:
return cam
factors = (zoom_factor(cam.shape, X.shape) for X in seed_inputs)
cam = [zoom(cam, factor) for factor in factors]
if len(self.model.inputs) == 1 and not isinstance(seed_input, list):
cam = cam[0]
return cam
def hard_sigmoid(x):
return backend.relu(x + 3.0, max_value=6.0) / 6.0
def call(self, inputs):
# The first part of the input is the pointcloud.
point_cloud = inputs[0]
assert_shape_is(point_cloud, (1024, 3))
# The second part of the input are the KNNs.
knn = inputs[1]
assert_shape_is(knn, (1024, 20, 3))
# Reshape the pointcloud if necessary.
if len(point_cloud.shape) == 4:
pass
elif len(point_cloud.shape) == 3:
point_cloud = K.expand_dims(point_cloud, axis=2)
else:
raise Exception("Invalid shape!")
assert_shape_is(point_cloud, (1024, 1, 3))
# Tile the pointcloud to make it compatible with KNN.
point_cloud_tiled = K.tile(point_cloud, [1, 1, self.k, 1])
assert_shape_is(point_cloud_tiled, (1024, 20, 3))
# Compute difference between tiled pointcloud and knn.
point_cloud_knn_difference = point_cloud_tiled - knn
assert_shape_is(point_cloud_knn_difference, (1024, 20, 3))
# MLP 1 for self attention including batch normalization.
self_attention = self.self_attention_mlp1(point_cloud)
if self.batch_normalization == True:
self_attention = self.self_attention_bn1(self_attention)
assert_shape_is(self_attention, (1024, 1, 16))
# MLP 2 for self attention including batch normalization.
self_attention = self.self_attention_mlp2(self_attention)
if self.batch_normalization == True:
self_attention = self.self_attention_bn2(self_attention)
assert_shape_is(self_attention, (1024, 1, 1))
# MLP 1 for neighbor attention including batch normalization.
neighbor_attention = self.neighbor_attention_mlp1(point_cloud_knn_difference)
if self.batch_normalization == True:
neighbor_attention = self.neighbor_attention_bn1(neighbor_attention)
assert_shape_is(neighbor_attention, (1024, 20, 16))
# Graph features are the ouput of the first MLP.
graph_features = neighbor_attention
# MLP 2 for neighbor attention including batch normalization.
neighbor_attention = self.neighbor_attention_mlp2(neighbor_attention)
if self.batch_normalization == True:
neighbor_attention = self.neighbor_attention_bn2(neighbor_attention)
assert_shape_is(neighbor_attention, (1024, 20, 1))
# Merge self attention and neighbor attention to get attention coefficients.
logits = self_attention + neighbor_attention
assert_shape_is(logits, (1024, 20, 1))
logits = K.permute_dimensions(logits, (0, 1, 3, 2))
assert_shape_is(logits, (1024, 1, 20))
# Apply leaky relu and softmax to logits to get attention coefficents.
logits = K.relu(logits, alpha=0.2)
attention_coefficients = K.softmax(logits)
assert_shape_is(attention_coefficients, (1024, 1, 20))
# Compute attention features from attention coefficients and graph features.
attention_features = tf.matmul(attention_coefficients, graph_features)
attention_features = tf.add(attention_features, self.output_bias)
attention_features = K.relu(attention_features)
assert_shape_is(attention_features, (1024, 1, 16))
# Reshape graph features.
#graph_features = K.expand_dims(graph_features, axis=2)
assert_shape_is(graph_features, (1024, 20, 16))
# Done.
return attention_features, graph_features, attention_coefficients
def __call__(self,
loss,
seed_input,
penultimate_layer=-1,
activation_modifier=lambda cam: K.relu(cam),
normalize_gradient=True):
"""Generate a gradient based class activation map (CAM) by using positive gradient of
penultimate_layer with respect to loss.
For details on Grad-CAM, see the paper:
[Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via
Gradient-based Localization](https://arxiv.org/pdf/1610.02391v1.pdf).
# Arguments
loss: A loss function. If the model has multipul outputs, you can use a different
loss on each output by passing a list of losses.
seed_input: An N-dim Numpy array. If the model has multipul inputs,
you have to pass a list of N-dim Numpy arrays.
penultimate_layer: A number of integer or a tf.keras.layers.Layer object.
normalize_gradient: True to normalize gradients.
activation_modifier: A function to modify gradients.
# Returns
The heatmap image or a list of their images that indicate the `seed_input` regions
whose change would most contribute the loss value,
# Raises
ValueError: In case of invalid arguments for `loss`, or `penultimate_layer`.
"""
losses = self._prepare_losses(loss)
seed_inputs = [
x if tf.is_tensor(x) else tf.constant(x)
for x in listify(seed_input)
]
seed_inputs = [
tf.expand_dims(seed_input, axis=0)
if X.shape == input_tensor.shape[1:] else X
for X, input_tensor in zip(seed_inputs, self.model.inputs)
]
if len(seed_inputs) != len(self.model.inputs):
raise ValueError('')
penultimate_output_tensor = self._find_penultimate_output(
self.model, penultimate_layer)
model = tf.keras.Model(inputs=self.model.inputs,
outputs=self.model.outputs +
[penultimate_output_tensor])
with tf.GradientTape() as tape:
tape.watch(seed_inputs)
outputs = model(seed_inputs)
outputs = listify(outputs)
loss_values = [loss(y) for y, loss in zip(outputs[:-1], losses)]
penultimate_outputs = outputs[-1]
grads = tape.gradient(loss_values, penultimate_outputs)
if normalize_gradient:
grads = K.l2_normalize(grads)
weights = K.mean(grads, axis=tuple(np.arange(len(grads.shape))[1:-1]))
cam = np.asarray([
np.sum(o * w, axis=-1)
for o, w in zip(penultimate_outputs, weights)
])
if activation_modifier is not None:
cam = activation_modifier(cam)
input_dims_list = (X.shape[1:-1] for X in seed_inputs)
output_dims = penultimate_outputs.shape[1:-1]
zoom_factors = ([
i / (j * 1.0) for i, j in iter(zip(input_dims, output_dims))
] for input_dims in input_dims_list)
cams = [
np.asarray([zoom(v, factor) for v in cam])
for factor in zoom_factors
]
if len(self.model.inputs) == 1 and not isinstance(seed_input, list):
cams = cams[0]
return cams
def margin_loss(y_true, y_pred):
lamb, margin = 0.5, 0.1
return K.sum((y_true * K.square(K.relu(1 - margin - y_pred)) + lamb *
(1 - y_true) * K.square(K.relu(y_pred - margin))),
axis=-1)
def relu6(inputs):
return K.relu(inputs, max_value=6.0)
def call(self, x, **kwargs):
return (K.concatenate((K.relu(x, alpha=0), K.relu(-x, alpha=0)),
axis=self.channel_axis) * self._get_coef())
def HS(self, x):
""" Construct Hard Swish activation function
x : input to activation function
"""
return (x * K.relu(x + 3, max_value=6.0)) / 6.0
def call(self, inputs, **kwargs):
inputs -= K.mean(inputs, axis=1, keepdims=True)
inputs = K.l2_normalize(inputs, axis=1)
pos = K.relu(inputs)
neg = K.relu(-inputs)
return K.concatenate([pos, neg], axis=1)
def call(self, inputs):
if K.dtype(inputs) != 'float32':
inputs = K.cast(inputs, 'float32')
inner_out = K.relu(K.dot(inputs, self.weights_inner) + self.bais_inner)
outputs = K.dot(inner_out, self.weights_out) + self.bais_out
return outputs
def call(self, inputs, **kwargs):
pos = K.relu(inputs)
neg = K.relu(-inputs)
return K.concatenate([pos, neg])
def discriminator_loss_fn(self, real, fake):
loss = K.mean(K.relu(1. - real), axis=0)
loss += K.mean(K.relu(1. + fake), axis=0)
return loss
def sparse_softmax(x):
x = K.relu(x)
e = K.exp(x - K.max(x, axis=(1, 2, 3), keepdims=True))
s = K.sum(e, axis=(1, 2, 3), keepdims=True)
return e / s
def relu6(x):
return backend.relu(x, max_value=6)
def HS(x):
""" Hard Swish activation """
return (x * K.relu(x + 3, max_value=6.0)) / 6.0
def hinge_d(y_true, y_pred):
return K.mean(K.relu(1.0 + (y_true * y_pred)))
def hard_swish(x):
return x * K.relu(x + 3.0, max_value=6.0) / 6.0
