def conv_bn_act(input_tensor, conv_layer, n_filters, filter_size=3, strides=1, activation="relu", bn=None, conv_args={}, name=""):
act = Activation(activation) if isinstance(activation, str) else activation
act.name = name + "_act"
x = conv_layer(n_filters, filter_size, name=name+"_conv", **conv_args, strides=strides)(input_tensor)
if bn is None:
x = act(x)
elif bn == "pre_act":
x = BatchNormalization(name=name+"_bn")(x)
x = act(x)
elif bn == "post_act":
x = act(x)
x = BatchNormalization(name=name+"_bn")(x)
return x
def dense_bn_act(input_tensor, n_dense, activation="relu", bn=None, dense_args={}, name=""):
"""
activation: either string or an instance of a keras activation that can be called
on a tensor
dense_args: dictionary containing keys and values for e.g. kernel_initializer, bias_initializer, ...
"""
act = Activation(activation) if isinstance(activation, str) else activation
act.name = name + "_act"
x = Dense(n_dense, name=name+"_dense", **dense_args)(input_tensor)
if bn is None:
x = act(x)
elif bn == "pre_act":
x = BatchNormalization(name=name+"_bn")(x)
x = act(x)
elif bn == "post_act":
x = act(x)
x = BatchNormalization(name=name+"_bn")(x)
return x
def __call__(self, act, name, *arg):
"""
Create an activation layer
act : type of act
name : name of layer
*arg : additional argument for activation
"""
if act == 'sigmoid':
la = Activation('sigmoid')
elif act == 'softplus':
la = Activation('softplus')
elif act == 'softmax':
la = Activation('softplus')
elif act == 'relu':
la = Activation('relu')
elif act == 'sigmoid':
la = Activation('sigmoid')
elif act == 'selu':
la = Activation('selu')
elif act == 'tanh':
la = Activation('tanh')
elif act == 'linear':
la = Activation('linear')
elif act == 'softmax':
la = Activation('softmax')
elif act == 'leakyrelu':
la = LeakyReLU(arg)
elif act == 'elu':
la = ELU(arg)
elif (act == 'swish'):
la = Activation('swish')
else:
print(act, "is not implemented")
assert (False)
la.name = name
return la
def modify_model_backprop(model, backprop_modifier):
"""Creates a copy of model by modifying all activations to use a custom op to modify the backprop behavior.
Args:
model: The `keras.models.Model` instance.
backprop_modifier: One of `{'guided', 'rectified'}`
Returns:
A copy of model with modified activations for backwards pass.
"""
# The general strategy is as follows:
# - Clone original model via save/load so that upstream callers don't see unexpected results with their models.
# - Modify all activations in the model as ReLU.
# - Save modified model so that it can be loaded with custom context modifying backprop behavior.
# - Call backend specific function that registers the custom op and loads the model under modified context manager.
# - Maintain cache to save this expensive process on subsequent calls.
#
# The reason for this round about way is because the graph needs to be rebuild when any of its layer builder
# functions are changed. This is very complicated to do in Keras and makes the implementation very tightly bound
# with keras internals. By saving and loading models, we dont have to worry about future compatibility.
#
# The only exception to this is the way advanced activations are handled which makes use of some keras internal
# knowledge and might break in the future.
# 0. Retrieve from cache if previously computed.
modified_model = _MODIFIED_MODEL_CACHE.get((model, backprop_modifier))
if modified_model is not None:
return modified_model
model_path = '/tmp/' + next(tempfile._get_candidate_names()) + '.h5'
try:
# 1. Clone original model via save and load.
model.save(model_path)
modified_model = load_model(model_path)
# 2. Replace all possible activations with ReLU.
for i, layer in utils.reverse_enumerate(modified_model.layers):
if hasattr(layer, 'activation'):
layer.activation = tf.nn.relu
if isinstance(layer, _ADVANCED_ACTIVATIONS):
# NOTE: This code is brittle as it makes use of Keras internal serialization knowledge and might
# break in the future.
modified_layer = Activation('relu')
modified_layer.inbound_nodes = layer.inbound_nodes
modified_layer.name = layer.name
modified_model.layers[i] = modified_layer
# 3. Save model with modifications.
modified_model.save(model_path)
# 4. Register modifier and load modified model under custom context.
modifier_fn = _BACKPROP_MODIFIERS.get(backprop_modifier)
if modifier_fn is None:
raise ValueError("'{}' modifier is not supported".format(backprop_modifier))
modifier_fn(backprop_modifier)
# 5. Create graph under custom context manager.
with tf.get_default_graph().gradient_override_map({'Relu': backprop_modifier}):
# This should rebuild graph with modifications.
modified_model = load_model(model_path)
# Cache to improve subsequent call performance.
_MODIFIED_MODEL_CACHE[(model, backprop_modifier)] = modified_model
return modified_model
finally:
os.remove(model_path)
def get_copy_of_layer(layer, verbose=False):
try:
from keras.layers import Activation
from keras import layers
except:
from tensorflow.keras.layers import Activation
from tensorflow.keras import layers
config = layer.get_config()
# Non-standard relu6 layer (from MobileNet)
if layer.__class__.__name__ == 'Activation':
if config['activation'] == 'relu6':
if get_keras_sub_version() == 1:
from keras.applications.mobilenet import relu6
else:
from keras_applications.mobilenet import relu6
layer_copy = Activation(relu6, name=layer.name)
return layer_copy
# DeepLabV3+ non-standard layer
if layer.__class__.__name__ == 'BilinearUpsampling':
from neural_nets.deeplab_v3_plus_model import BilinearUpsampling
layer_copy = BilinearUpsampling(upsampling=config['upsampling'],
output_size=config['output_size'],
name=layer.name)
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'UpsampleLike':
from keras_retinanet.layers import UpsampleLike
layer_copy = UpsampleLike(name=layer.name)
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'Anchors':
from keras_retinanet.layers import Anchors
layer_copy = Anchors(name=layer.name,
size=config['size'],
stride=config['stride'],
ratios=config['ratios'],
scales=config['scales'])
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'RegressBoxes':
from keras_retinanet.layers import RegressBoxes
layer_copy = RegressBoxes(name=layer.name,
mean=config['mean'],
std=config['std'])
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'PriorProbability':
from keras_retinanet.layers import PriorProbability
layer_copy = PriorProbability(name=layer.name,
mean=config['mean'],
std=config['std'])
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'ClipBoxes':
from keras_retinanet.layers import ClipBoxes
layer_copy = ClipBoxes(name=layer.name)
return layer_copy
# RetinaNet non-standard layer
if layer.__class__.__name__ == 'FilterDetections':
from keras_retinanet.layers import FilterDetections
layer_copy = FilterDetections(
name=layer.name,
max_detections=config['max_detections'],
nms_threshold=config['nms_threshold'],
score_threshold=config['score_threshold'],
nms=config['nms'],
class_specific_filter=config['class_specific_filter'],
trainable=config['trainable'],
parallel_iterations=config['parallel_iterations'])
return layer_copy
layer_copy = layers.deserialize({
'class_name': layer.__class__.__name__,
'config': config
})
try:
layer_copy.name = layer.name
except:
layer_copy._name = layer._name
return layer_copy