def get_variant_clipper(self, faulty_model, name=None):
model = self.get_model(name=name, training_variant='none')
def ranger_layer_factory(insert_layer_name):
return ClipperLayer(name=insert_layer_name, bounds=self.bounds)
model = insert_layer_nonseq(model, self.activation_name_pattern, ranger_layer_factory, 'dummy', model_name=name)
return model
def get_variant_profiler(self, faulty_model, name=None):
model = self.copy_model(faulty_model, name=(name or '') + '_base_copy')
def ranger_layer_factory(insert_layer_name):
return ProfileLayer(name=insert_layer_name)
model = insert_layer_nonseq(model, self.activation_name_pattern, ranger_layer_factory, 'dummy', model_name=name)
setattr(model, 'dropin', faulty_model.dropin)
return model
def get_variant_ranger(self, faulty_model, name=None):
model = self.copy_model(faulty_model, name=name + '_base_copy')
def ranger_layer_factory(insert_layer_name):
return RangerLayer(name=insert_layer_name, bounds=self.bounds)
model = insert_layer_nonseq(model,
self.activation_name_pattern,
ranger_layer_factory,
'dummy',
model_name=name)
return model
def __init__(self,
model,
representative_dataset=None,
a=None,
b=None,
r=0.5,
mode='worst',
regex='conv2d.*|dense.*',
perturb=lambda x, p: x + p,
count=1,
portion=None) -> None:
super().__init__()
self.model = model
self.representative_dataset = representative_dataset
self.r = r
self.mode = mode
self.regex = regex
self.perturb = perturb
self.count = count
self.portion = portion
if self.representative_dataset:
DropinProfiler.a, DropinProfiler.b = None, None
def profiler_layer_factory(insert_layer_name):
return DropinProfiler(name=insert_layer_name)
profiler = insert_layer_nonseq(model,
self.regex,
profiler_layer_factory,
'profiler',
only_last_node=True)
profiler.run_eagerly = True
train_data_size = len(representative_dataset)
for i, data in enumerate(self.representative_dataset):
x, y = data
profiler.predict(x)
logger.info('Done with {}/{} batches.'.format(
i, train_data_size))
self.a, self.b = DropinProfiler.a, DropinProfiler.b
else:
assert None not in (a, b)
self.a, self.b = a, b
self.perturbation_inputs = []