def __init__(self, model, epsilon=1e-1, *args, **kwargs):
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
#TODO: fix. specify. extend.
("LRPSequentialPresetB is not advised "
"for networks with non-ReLU activations."),
check_type="warning",
)
class EpsilonProxyRule(rrule.EpsilonRule):
def __init__(self, *args, **kwargs):
super(EpsilonProxyRule, self).__init__(*args,
epsilon=epsilon,
bias=True,
**kwargs)
conditional_rules = [(kchecks.is_dense_layer, EpsilonProxyRule),
(kchecks.is_conv_layer, rrule.Alpha2Beta1Rule)]
bn_layer_rule = kwargs.pop("bn_layer_rule", rrule.AlphaBetaX2m100Rule)
super(LRPSequentialPresetB, self).__init__(model,
*args,
rule=conditional_rules,
bn_layer_rule=bn_layer_rule,
**kwargs)
def __init__(self, model, epsilon=1e-1, *args, **kwargs):
self._model_checks = [
# TODO: Check for non-linear output in general.
{
"check": lambda layer: kchecks.contains_activation(
layer, activation="softmax"),
"type": "exception",
"message": "Model should not contain a softmax.",
},
{
"check":
lambda layer: not kchecks.only_relu_activation(layer),
"type": "warning",
"message": (" is not advised for "
"networks with non-ReLU activations.") #TODO: fix. specify. extend.
}
]
class EpsilonProxyRule(rrule.EpsilonRule):
def __init__(self, *args, **kwargs):
super(EpsilonProxyRule, self).__init__(*args,
epsilon=epsilon,
bias=True,
**kwargs)
conditional_rules = [(kchecks.is_dense_layer, EpsilonProxyRule),
(kchecks.is_conv_layer, rrule.Alpha2Beta1Rule)
]
super(LRPSequentialPresetB, self).__init__(model,
*args,
rule = conditional_rules,
**kwargs )
def __init__(self, model, **kwargs):
super().__init__(model, **kwargs)
# Add and run model checks
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"Deconvnet is only specified for networks with ReLU activations.",
check_type="exception",
)
self._do_model_checks()
def __init__(self, model, percentile=98, **kwargs):
self._percentile = percentile
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"RectGrad is only specified for "
"networks with ReLU activations.",
check_type="exception",
)
super(RectGrad, self).__init__(model, **kwargs)
def __init__(self, model: Model, *args, **kwargs) -> None:
super().__init__(model, *args, **kwargs)
# Add and run model checks
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"This DeepTaylor implementation only supports ReLU activations.",
check_type="exception",
)
self._do_model_checks()
def __init__(self, model, *args, **kwargs):
self._add_model_check(
lambda layer: kchecks.is_batch_normalization_layer(layer),
"DeepTaylor (currently) does not support BachNormalization layer.",
check_type="exception",
)
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"This DeepTaylor implementation only supports ReLU activations.",
check_type="exception",
)
super(DeepTaylor, self).__init__(model, *args, **kwargs)
def __init__(self, model: Model, **kwargs):
# Inside function to not break import if Keras changes.
super().__init__(model, **kwargs)
# Add and run model checks
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"BaselineLRPZ only works with ReLU activations.",
check_type="exception",
)
self._add_model_check(
lambda layer: not isinstance(layer, BASELINE_LRPZ_LAYERS),
"BaselineLRPZ only works with a predefined set of layers.",
check_type="exception",
)
self._do_model_checks()
def __init__(self, model, patterns=None, pattern_type=None, **kwargs):
super().__init__(model, **kwargs)
# Add and run model checks
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
("PatternNet is not well defined for networks with non-ReLU activations."
),
check_type="warning",
)
self._add_model_check(
lambda layer: not kchecks.is_convnet_layer(layer),
("PatternNet is only well defined for convolutional neural networks."
),
check_type="warning",
)
self._add_model_check(
lambda layer: not isinstance(layer, SUPPORTED_LAYER_PATTERNNET),
("PatternNet is only well defined for conv2d/max-pooling/dense layers."
),
check_type="exception",
)
self._do_model_checks()
self._patterns = patterns
if self._patterns is not None:
# copy pattern references
self._patterns = list(patterns)
self._pattern_type = pattern_type
# Pattern projections can lead to +-inf value with long networks.
# We are only interested in the direction, therefore it is save to
# Prevent this by projecting the values in bottleneck layers to +-1.
if not kwargs.get("reverse_project_bottleneck_layers", True):
warnings.warn(
"The standard setting for 'reverse_project_bottleneck_layers'"
"is overwritten.")
else:
kwargs["reverse_project_bottleneck_layers"] = True
def __init__(self, model, **kwargs):
# Inside function to not break import if Keras changes.
BASELINELRPZ_LAYERS = (
keras.engine.topology.InputLayer,
keras.layers.convolutional.Conv1D,
keras.layers.convolutional.Conv2D,
keras.layers.convolutional.Conv2DTranspose,
keras.layers.convolutional.Conv3D,
keras.layers.convolutional.Conv3DTranspose,
keras.layers.convolutional.Cropping1D,
keras.layers.convolutional.Cropping2D,
keras.layers.convolutional.Cropping3D,
keras.layers.convolutional.SeparableConv1D,
keras.layers.convolutional.SeparableConv2D,
keras.layers.convolutional.UpSampling1D,
keras.layers.convolutional.UpSampling2D,
keras.layers.convolutional.UpSampling3D,
keras.layers.convolutional.ZeroPadding1D,
keras.layers.convolutional.ZeroPadding2D,
keras.layers.convolutional.ZeroPadding3D,
keras.layers.core.Activation,
keras.layers.core.ActivityRegularization,
keras.layers.core.Dense,
keras.layers.core.Dropout,
keras.layers.core.Flatten,
keras.layers.core.Lambda,
keras.layers.core.Masking,
keras.layers.core.Permute,
keras.layers.core.RepeatVector,
keras.layers.core.Reshape,
keras.layers.core.SpatialDropout1D,
keras.layers.core.SpatialDropout2D,
keras.layers.core.SpatialDropout3D,
keras.layers.local.LocallyConnected1D,
keras.layers.local.LocallyConnected2D,
keras.layers.Add,
keras.layers.Concatenate,
keras.layers.Dot,
keras.layers.Maximum,
keras.layers.Minimum,
keras.layers.Subtract,
keras.layers.noise.AlphaDropout,
keras.layers.noise.GaussianDropout,
keras.layers.noise.GaussianNoise,
keras.layers.normalization.BatchNormalization,
keras.layers.pooling.GlobalMaxPooling1D,
keras.layers.pooling.GlobalMaxPooling2D,
keras.layers.pooling.GlobalMaxPooling3D,
keras.layers.pooling.MaxPooling1D,
keras.layers.pooling.MaxPooling2D,
keras.layers.pooling.MaxPooling3D,
)
self._add_model_softmax_check()
self._add_model_check(
lambda layer: not kchecks.only_relu_activation(layer),
"BaselineLRPZ only works with ReLU activations.",
check_type="exception",
)
self._add_model_check(
lambda layer: not isinstance(layer, BASELINELRPZ_LAYERS),
"BaselineLRPZ only works with a predefined set of layers.",
check_type="exception",
)
super(BaselineLRPZ, self).__init__(model, **kwargs)
def __init__(self, model, allow_lambda_layers=False, **kwargs):
# Inside function to not break import if Keras changes.
BASELINELRPZ_LAYERS = (
keras.engine.topology.InputLayer,
keras.layers.convolutional.Conv1D,
keras.layers.convolutional.Conv2D,
keras.layers.convolutional.Conv2DTranspose,
keras.layers.convolutional.Conv3D,
keras.layers.convolutional.Conv3DTranspose,
keras.layers.convolutional.Cropping1D,
keras.layers.convolutional.Cropping2D,
keras.layers.convolutional.Cropping3D,
keras.layers.convolutional.SeparableConv1D,
keras.layers.convolutional.SeparableConv2D,
keras.layers.convolutional.UpSampling1D,
keras.layers.convolutional.UpSampling2D,
keras.layers.convolutional.UpSampling3D,
keras.layers.convolutional.ZeroPadding1D,
keras.layers.convolutional.ZeroPadding2D,
keras.layers.convolutional.ZeroPadding3D,
keras.layers.core.Activation,
keras.layers.core.ActivityRegularization,
keras.layers.core.Dense,
keras.layers.core.Dropout,
keras.layers.core.Flatten,
keras.layers.core.Lambda,
keras.layers.core.Masking,
keras.layers.core.Permute,
keras.layers.core.RepeatVector,
keras.layers.core.Reshape,
keras.layers.core.SpatialDropout1D,
keras.layers.core.SpatialDropout2D,
keras.layers.core.SpatialDropout3D,
keras.layers.local.LocallyConnected1D,
keras.layers.local.LocallyConnected2D,
keras.layers.Add,
keras.layers.Concatenate,
keras.layers.Dot,
keras.layers.Maximum,
keras.layers.Minimum,
keras.layers.Subtract,
keras.layers.noise.AlphaDropout,
keras.layers.noise.GaussianDropout,
keras.layers.noise.GaussianNoise,
keras.layers.normalization.BatchNormalization,
keras.layers.pooling.GlobalMaxPooling1D,
keras.layers.pooling.GlobalMaxPooling2D,
keras.layers.pooling.GlobalMaxPooling3D,
keras.layers.pooling.MaxPooling1D,
keras.layers.pooling.MaxPooling2D,
keras.layers.pooling.MaxPooling3D,
)
self._model_checks = [
# todo: Check for non-linear output in general.
{
"check": lambda layer: kchecks.contains_activation(
layer, activation="softmax"),
"type": "exception",
"message": "Model should not contain a softmax.",
},
{
"check":
lambda layer: not kchecks.only_relu_activation(layer),
"type": "exception",
"message": ("BaselineLRPZ is not working for "
"networks with non-ReLU activations."),
},
{
"check":
lambda layer: not isinstance(layer, BASELINELRPZ_LAYERS),
"type": "exception",
"message": ("BaselineLRPZ is only defined for "
"certain layers."),
},
{
"check":
lambda layer: (not allow_lambda_layers and
isinstance(layer, keras.layers.core.Lambda)),
"type": "exception",
"message": ("Lamda layers are not allowed. "
"To allow use allow_lambda_layers kw."),
},
]
self._allow_lambda_layers = allow_lambda_layers
super(BaselineLRPZ, self).__init__(model, **kwargs)
def test_only_relu_activation(activation, layer, expected):
assert ichecks.only_relu_activation(layer) == expected
