ALMa: Active Learning (data) Manager
Active Learning is a popular technique to reduce annotation costs by using AI to decide what to label next. The subtle bookkeeping involved in keeping track of what has been labeled is tedious and error prone. We need to train our learner on Labeled data, but sample new examples to label from Unlabeled data. As we label data it moves around between the two subsets of our Dataset and managing the bookkeeping a chore that should be abstracted away.
Today we're happy to open-source ALMa the Active Learning Manager that abstracts away the bookkeeping. Whereas most implementations modify the data array in place, ALMa maintains views of Labeled and Unlabeled subsets of the original Dataset.
The Problem ALMa Solves
In a typical active learning setup we start with an Array Dataset of examples which naturally divides into two disjoint subsets Ulabeled data and Labeled data. When we work with our Unlabeled data as an array, it's indices don't correspond to the indices of our Dataset and they change every time we add new labeled data.
Active Learning without ALMa involves confusing Bookkeeping
In the example below, taken from the ModAL library, the original dataset is constantly modified with hard to read numpy code. While it is short, it is difficult to grock and harder still to ensure correctness.
for index in range(N_QUERIES): query_index, query_instance = learner.query(X_pool) #What is this reshape stuff doing ? X, y = X_pool[query_index].reshape(1, -1), y_pool[query_index].reshape(1, ) learner.teach(X=X, y=y) #Confusing np.delete X_pool, y_pool = np.delete(X_pool, query_index, axis=0), np.delete(y_pool, query_index)
Active Learning with ALMa is easy:
Here's the same code using ALMa, it has a few more lines, but the bookkeeping has effectively been abstracted away.
for index in range(N_QUERIES): index_to_label, query_instance = learner.query(manager.unlabeld) original_ix = manager.get_original_index_from_unlabeled_index(index_to_label) y = original_labels_train[original_ix] label = (index_to_label, y) manager.add_labels(labels) learner.teach(X=manager.labeled, y=manager.labels)
ALMa's solution: Simpler Bookkeeping with Views and Offsets
ALMa uses numpy's fancy indexing to maintain views of the Dataset which minimizes and simplifies the bookkeeping that needs to be done. ALMa relies on two numpy features to manage the bookkeeping, fancy indexing with mask index arrays to create views of the data, and the nonzero method to calculate index offsets for new labeled data (which comes in indexed relative to the Unlabeled subset)
Maintaining Views of the Labeld and Unlabeled data
ALMa uses numpy's mask index arrays to create "views" of Data that correspond to our Labeled and Unlabeled data.
When we initialize an ActiveLearningManager it creates a boolean array whose indices are True if the corresponding feature has been labeled, and False otherwise.
# Create a boolean array with the same length as features self.labeled_mask = np.zeros(self.features.shape, dtype=bool)
This makes getting the Unlabeled indices simple
def unlabeled_mask(self): return np.logical_not(self.labeled_mask)
We can then expose the views on the ActiveLearningManager as follows:
@property def labeled(self): return self.features[self.labeled_mask] @property def unlabeld(self): return self.features[self.unlabeled_mask]
Adding New Labels
With our views in place our active learning process boils down to:
- Sample some data from the Unlabeld subset
- Have the annotator label them and update ALMa
- Train the learner on the updated Labeled subset
But, when we sample from the Unlabeled, the examples are not indexed relative to the original dataset and so we need a way to recover to correct indices.
This would be easily solved if we had an array whose indices were the same as our Unlabeled data and values were the corresponding indices in the Dataset.
Mapping offsets with Numpy's nonzero method
This is actually easier done than said, numpy provides a nonzero() method that gives us exactly that.
import numpy as np a = np.zeros(10,dtype=np.bool) a.nonzero()
array(, dtype=int64) #Empty array
a =True a.nonzero()
array() # Maps the first true value to it's index in the original rray
a = True a.nonzero()
array([3, 7]) # Maps both True values to their correct place in the original array
Adding Labels With Numpy's nonzero method
ALMa holds a boolean array labeledmask_ that whose values are True when we already have a label for the example at the index. We calculate unlabeledmask_ by taking the logicalnot of _labeled_mask. So calling nonzero() on our unlabeledmask_ gives us a new array whose indices are the indices of our Unlabeled data and values are the indices of that example in our Dataset.
Armed with that, calculating the correct offsets is simple:
def _offset_new_labels(self, labels_for_unlabeled_dataset: LabelList): if len(self._labels) == 0: # Nothing to correct in this case return labels_for_unlabeled_dataset labels_for_dataset: LabelList =  unlabeled_indices_map = self.unlabeled_mask.nonzero() for label in labels_for_unlabeled_dataset: index_in_unlabeled, annotation = label index_in_dataset = unlabeled_indices_map[index_in_unlabeled] new_label: Label = (index_in_dataset, annotation) labels_for_dataset.append(new_label) return labels_for_dataset
And when we add one or more new labels ALMa does
def add_labels(self, labels: LabelList, offset_to_unlabeled=True): 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),..]" ) if offset_to_unlabeled: labels = self._offset_new_labes(labels) self._update_masks(labels) for label in labels: self._labels[label] = label
Managing state is generally difficult and error prone, and this is true for active learning as as well. By minimizing the state being muated and working with views of the data we can simplify the end users experience. We hope that using ALMa will help you focus on your research or production models by freeing you up from bookkeeping. Clone ALMa here