threshold_decoder.py

# Copyright 2019 Mycroft AI Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Code for converting network output to confidence level
"""
importnumpyasnp
fromtypingimportTuple

fromprecise.functionsimportasigmoid, sigmoid, pdf


classThresholdDecoder:
"""
 Decode raw network output into a relatively linear threshold using
 This works by estimating the logit normal distribution of network
 activations using a series of averages and standard deviations to
 calculate a cumulative probability distribution

 Background:
 We could simply take the output of the neural network as the confidence of a given
 prediction, but this typically jumps quickly between 0.01 and 0.99 even in cases where
 the network is less confident about a prediction. This is a symptom of the sigmoid squashing
 high values to values close to 1. This ThresholdDecoder measures the average output of
 the network over a dataset and uses that to create a smooth distribution so that an output
 of 80% means that the network output is greater than roughly 80% of the dataset
 """
def__init__(self, mu_stds: Tuple[Tuple[float, float]], center=0.5, resolution=200, min_z=-4, max_z=4):
self.min_out=int(min(mu+min_z*stdformu, stdinmu_stds))
self.max_out=int(max(mu+max_z*stdformu, stdinmu_stds))
self.out_range=self.max_out-self.min_out
self.cd=np.cumsum(self._calc_pd(mu_stds, resolution))
self.center=center

defdecode(self, raw_output: float) ->float:
ifraw_output==1.0orraw_output==0.0:
returnraw_output
ifself.out_range==0:
cp=int(raw_output>self.min_out)
else:
ratio= (asigmoid(raw_output) -self.min_out) /self.out_range
ratio=min(max(ratio, 0.0), 1.0)
cp=self.cd[int(ratio* (len(self.cd) -1) +0.5)]
ifcp<self.center:
return0.5*cp/self.center
else:
return0.5+0.5* (cp-self.center) / (1-self.center)

defencode(self, threshold: float) ->float:
threshold=0.5*threshold/self.center
ifthreshold<0.5:
cp=threshold*self.center*2
else:
cp= (threshold-0.5) *2* (1-self.center) +self.center
ratio=np.searchsorted(self.cd, cp) /len(self.cd)
returnsigmoid(self.min_out+self.out_range*ratio)

def_calc_pd(self, mu_stds, resolution):
points=np.linspace(self.min_out, self.max_out, resolution*self.out_range)
returnnp.sum([pdf(points, mu, std) formu, stdinmu_stds], axis=0) / (resolution*len(mu_stds))