Source code for tmeasures.measure

from __future__ import annotations
from typing import List, Tuple
import numpy as np
import abc
import re
from .utils import get_all

ActivationsByLayer = List[np.ndarray]

# TODO change `layer` for `activation` in variable/methods to unify vocabulary

class MeasureResult:
    def __init__(self,layers:ActivationsByLayer,layer_names:List[str],measure:Measure,extra_values=dict()):
        assert (len(layers) == len(layer_names)), f"The number of layers {len(layers)} should match the length of the layer_names list {layer_names} "
        self.layers=layers
        self.layer_names=layer_names
        self.measure=measure
        self.extra_values=extra_values

    def __repr__(self):
        return f"MeasureResult({self.measure})"

    def all_1d(self):
        return np.any([ len(l.shape)==1 for l in self.layers])

    def per_layer_average(self) -> np.ndarray:
        result = []
        for layer in self.layers:
            layer=layer[:]
            layer_average=layer[np.isfinite(layer)].mean()
            result.append(layer_average)
        return np.array(result)

    def per_layer_mean_std(self) -> Tuple[np.ndarray,np.ndarray]:
        means = []
        stds = []
        for layer in self.layers:
            layer=layer[:]
            layer_mean=layer[np.isfinite(layer)].mean()
            layer_std=layer[np.isfinite(layer)].std()
            means.append(layer_mean)
            stds.append(layer_std)
        return np.array(means),np.array(stds)

    def remove_layers(self,remove_indices:List[int])->MeasureResult:
        n = len(self.layer_names)
        all_indices=set(list(range(n)))
        keep_indices = list(all_indices.difference(set(remove_indices)))
        layers = get_all(self.layers,keep_indices)
        layer_names = get_all(self.layer_names,keep_indices)
        return MeasureResult(layers,layer_names,self.measure,self.extra_values)

    def weighted_global_average(self):
        return self.per_layer_average().mean()

    def global_average(self)-> float:
        return self.per_layer_average().mean()

    def layers_dict(self):
        return dict(zip(self.layer_names,self.layers))



class StratifiedMeasureResult(MeasureResult):
    def __init__(self, layers:ActivationsByLayer, layer_names:List[str], measure:Measure,
                 results:List[MeasureResult], labels:List[str]):
        super().__init__(layers,layer_names,measure)
        self.results=results
        self.labels=labels

    def __repr__(self):
        return f"StratifiedMeasureResult {self.measure}"




class Measure(abc.ABC):
    def __repr__(self):
        return f"{self.__class__.__name__}"

    def id(self):
        return str(self)

    # @abc.abstractmethod
    # def name(self):
    #     pass
    #
    # @abc.abstractmethod
    # def abbreviation(self):
    #     pass

    # Returns the name of the measure, separated by spaces
    def name(self):
        # get first part before parenthesis (if any)
        class_name = self.__class__.__name__
        result = class_name.split("(")[0]
        # add a space before capitalized words
        result = re.sub( r"([A-Z])", r" \1", result)
        return result

    # generates an abbreviation of the measure, based on its name
    def abbreviation(self):
        initials=re.findall('[A-Z]', self.name())
        return "".join(initials)

    @abc.abstractmethod
    def eval(self)->MeasureResult:
        pass