Add a Data Manager Class

[talolard]

Hey is this something you'd like to add to the library ?

Rationale

Bookeeping which examples have been labeled is no fun and error prone and verbose. I propose a "Manager" class that will do it for us, so that labeling looks like this:

User Experience

for index in range(10):
 
 query_index, query_instance = learner.query(manager.unlabeld)
 ix = query_index[0]
 print(manager.remaining_sources[ix])
 label = (ix,int(input()))
 manager.add_labels(label)
 

In a notebook it works like this:

Idea

The idea is that the manager maintains masks for the labeled and unlabeled indices, and calculates offsets when adding new labels. End user shouldnt care about this stuff, and they can get the unlabeled examples from manager.unlabeled which is a view on the array.

This is some mock code, I haven't tested it properly yet

import typing
from typing import List,Tuple,Any,Union
import numpy as np
Label = Tuple[int, Any]
LabelList = List[Label]
Sources =List[Any]
class ALManager():
 def __init__(self, features:np.ndarray, labels:LabelList=None, sources:Sources=None):
 '''

 :param features: An array of the features that will be used for AL.
 :param labels: Any prexesiting labels. Each label is a tuple(idx,label)
 :param source: A list of the original data
 '''
 if labels is None:
 labels = []
 self.features = features
 if labels is None:
 self.labels = []
 self.labels = labels
 self.labeled_mask = np.zeros(self.features.shape[0],dtype=bool)
 self.unlabeled_mask = np.ones(self.features.shape[0],dtype=bool)
 self._update_masks(self.labels)
 self.sources = np.array(sources if sources else [])
 def _update_masks(self,labels:Union[LabelList,Label]):

 for label in labels:
 self.labeled_mask[label[0]] = True
 self.unlabeled_mask[label[0]] = False
 def _offset_new_labes(self,labels:LabelList):
 if len(self.labels)==0:
 # Nothing to correct in this case
 return labels
 correctLabels: LabelList = []
 labeledIndices =self.labeled_mask.nonzero()[0]

 for label in labels:
 #The argmax trick on a bool condition returns the index of the first labeled item which came before the new one
 newLabel :Label = (np.argmax(labeledIndices<=label[0])+1+label[0],label[1])
 correctLabels.append(newLabel)
 return correctLabels
 def add_labels(self, labels:LabelList):
 if isinstance(labels,tuple): # if this is a single example
 labels :LabelList = [labels]
 elif isinstance(labels,list):
 pass
 else:
 raise Exception("Malformed input. Please add either a tuple (ix,label) or a list [(ix,label),..]")
 labels = self._offset_new_labes(labels)
 self._update_masks(labels)
 self.labels +=(labels)

 @property
 def unlabeld(self):
 '''

 :return: A view (in the numpy sense) of the features restricted to those that aren't labeled
 '''
 return self.features[self.unlabeled_mask]

 @property
 def labeled(self):
 '''
 :return: A view (in the numpy sense) of the features restricted to those that aren't labeled
 '''
 return self.features[self.labeled_mask]

 @property
 def remaining_sources(self):
 '''

 :return: The original inputs, masked so that only unlabeled ones are returned
 '''
 return self.sources[self.unlabeled_mask]

[cosmic-cortex]

Hi!

Thanks for your suggestion and the already worked out code example! This manager is definitely a good idea and would significantly improve usability.

Just to give you a heads up, currently I am extremely busy since we are in the middle of a release sprint with my startup, but I'll try to make time for this during the weekend.