Lightly enables active learning with only a few lines of additional code. Learn here, how to get the most out of your data by maximizing the available information in your annotated dataset.
We designed Lightly Active Learning to give you the best performance while being very easy to use. Check out our tutorials to learn more:
Before you read on, make sure you have read the section on the The Lightly Platform. In particular, you should know how to create a dataset in the web-app. and how to upload images and embeddings to it. To do active learning, you will need such a dataset with embeddings (don’t worry, it’s free!).
Lightly makes use of the following concepts for active learning:
The ApiWorkflowClient is used to connect to our API. The API handles the selection of the images based on embeddings and active learning scores. To initialize the ApiWorkflowClient, you will need the datasetId and the token from the The Lightly Platform.
The ActiveLearningAgent builds the client interface of our active learning framework. It allows to indicate which images are preselected and which ones to sample from. Furthermore, one can query it to get a new batch of images. To initialize an ActiveLearningAgent you need an ApiWorkflowClient.
The SamplerConfig allows the configuration of a sampling request. In particular, you can set number of samples, the name of the resulting selection, and the SamplingMethod. Currently, you can set the SamplingMethod to one of the following:
Random: Selects samples uniformly at random.
Coreset: Greedily selects samples which are diverse.
Coral: Combines Coreset with scores to do active learning.
The Scorer takes as input the predictions of a pre-trained model on the set of unlabeled images. It offers a calculate_scores() method, which evaluates different scores based on how certain the model is about the images. When performing a sampling, the scores are passed to the API so the sampler can use them with Coral.
Continue reading to see how these components interact and how active learning is done with Lightly.
The goal of making an initial selection is to get a subdataset on which you can train an initial model. The output of the model can then be used to select new samples. That way, the model can be iteratively improved.
To make an initial selection, start off by adding your raw, unlabeled data and the according image embeddings to a dataset in the Lightly web-app. A simple way to do so is to use lightly-magic from the command-line. Don’t forget adapt the arguments input_dir, dataset_id and token.
# use trainer.max_epochs=0 to skip training lightly-magic input_dir='path/to/your/raw/dataset' dataset_id='xyz' token='123' trainer.max_epochs=0
Then, in your Python script, you will need to initialize the ApiWorkflowClient and the ActiveLearningAgent
import lightly from lightly.api import ApiWorkflowClient from lightly.active_learning.agents import ActiveLearningAgent api_client = ApiWorkflowClient(dataset_id='xyz', token='123') al_agent = ActiveLearningAgent(api_client)
It may not always be a good idea to sample from the full dataset. For example, it could be that a large portion of the images is blurry. In that case, it’s possible to create a tag in the web-app which only contains the sharp images and tell the ActiveLearningAgent to only sample from this tag. To do so, set the query_tag_name argument in the constructor of the agent.
Let’s configure the sampling request and request an initial selection next:
from lightly.active_learning.config import SamplerConfig from lightly.openapi_generated.swagger_client import SamplingMethod # we want an initial pool of 150 images config = SamplerConfig(n_samples=150, method=SamplingMethod.CORESET, name='initial-selection') al_agent.query(config) initial_selection = al_agent.labeled_set # initial_selection now contains 150 filenames assert len(initial_selection) == 150
The result of the query is a tag in the web-app under the name “initial-selection”. The tag contains the images which were selected by the sampling algorithm. Head there to scroll through the samples and download the selected images before annotating them. Alternatively, you can access the filenames of the selected images via the attribute labeled_set as shown above.
Active Learning Step¶
After you have annotated your initial selection of images, you can train a model on them. The trained model can then be used to figure out which images pose problems. This section will show you how these images can be added to the labeled dataset.
To continue with active learning with Lightly, you will need the ApiWorkflowClient and ActiveLearningAgent from before. If you perform the next selection step in a new file you have to initialize the client and agent again. If you have to re-initialize them, make sure to set the pre_selected_tag_name to your current selection (if this is the first iteration, this is the name you have passed to the sampler config when doing the initial selection). Note, that if you don’t have to re-initialize them, the tracking of the tags is taken care of for you.
# re-initializing the ApiWorkflowClient and ActiveLearningAgent api_client = ApiWorkflowClient(dataset_id='xyz', token='123') al_agent = ActiveLearningAgent(api_client, preselected_tag_name='initial-selection')
The next part is what differentiates active learning from simple subsampling; the trained model is used to get predictions on the data and the sampler then decides based on these predictions. To get a list of all filenames for which predictions are required, you can use the query_set:
# get all filenames in the query set query_set = al_agent.query_set
Use this list to get predictions on the unlabeled images.
Important: The predictions need to be in the same order as the filenames in the list returned by the ActiveLearningAgent.
For classification, the predictions need to be in a numpy array and normalized, such that the rows sum to one. Then, create a scorer object like so:
from lightly.active_learning.scorers import ScorerClassification scorer = ScorerClassification(predictions)
Now you have everything to get the next batch of images. One important thing to mention here is that the argument n_samples always refers to the total size of the labeled set.
# we want a total of 200 images after the first iteration (50 new samples) # this time, we use the CORAL sampler and provide a scorer to the query config = SamplerConfig(n_samples=200, method=SamplingMethod.CORAL, name='al-iteration-1') al_agent.query(sampler_config, scorer) labeled_set_iteration_1 = al_agent.labeled_set added_set_iteration_1 = al_agent.added_set assert len(labeled_set_iteration_1) == 200 assert len(added_set_iteration_1) == 50
As before, there will be a new tag named al-iteration-1 visible in the web-app. Additionally, you can access the filenames of all the images in the labeled set and the filenames which were added by this query via the attributes labeled_set and added_set respectively. You can repeat the active learning step until the model achieves the required accuracy.
As the web-app allows viewing the active learning scores in the embedding view, there are usecases where only active learning scores should be uploaded to the web-app, but without performing a sampling. This is also easily possible:
Lightly has so called scorers for the common computer vision tasks such as image classification, detection and others. Depending on the task you are working on you can use a different scorer.
Use this scorer when working on a classification problem (binary or multiclass).
Currently we offer three uncertainty scorers,which are taken from http://burrsettles.com/pub/settles.activelearning.pdf, Section 3.1, page 12f and also explained in https://towardsdatascience.com/uncertainty-sampling-cheatsheet-ec57bc067c0b They all have in common, that the score is highest if all classes have the same confidence and are 0 if the model assigns 100% probability to a single class. The differ in the number of class confidences they take into account.
This score is 1 - the highest confidence prediction. It is high when the confidence about the most probable class is low.
This score is 1 - the margin between the highest confidence and second highest confidence prediction. It is high when the model cannot decide between the two most probable classes.
This scorer computes the entropy of the prediction. The confidences for all classes are considered to compute the entropy of a sample.
For more information about how to use the classification scorer have a look here:
Use this scorer when working on an object detection problem using bounding boxes. The object detection scorers require the input to be in the ObjectDetectionOutput format.
We expect the model predictions to contain
bounding boxes of shape (x0, y0, x1, y1)
objectness_probability for each bounding box
classification_probabilities for each bounding box
You can find more about the format here:
We also provide a helper method to work with the model output format consisting
of only a probability per bounding box and the associated label.
Currently, the following scorers are available:
object_frequency This score measures the number of objects in the image. Use this scorer if you want scenes with lots of objects in them. This is suited for computer vision tasks such as perception in autonomous driving.
objectness_least_confidence This score is 1 - the mean of the highest confidence prediction. Use this scorer to select images where the model is insecure about both whether it found an object at all and the class of the object.
classification_scores These scores are computed for each object detection per image out of the class probability prediction for this detection. Then, they are reduced to one score per image by taking the maximum. In particular we support: - uncertainty_least_confidence - uncertainty_margin - uncertainty_entropy The scores are computed using the scorer for classification.
For more information about how to use the object detection scorer have a look here:
Use this scorer when you’re training a model for semantic segmentation. The semantic segmentation scorer expects a list or generator of pixelwise label predictions.
We expect the model predictions to be of shape W x H x C, where
W is the width of the image
H is the height of the image
C is the number of segmentation classes (e.g. 2 for background vs foreground)
Currently, the following scorers are available:
classification_scores These scores treat segmentation as a pixelwise classification task. The classification uncertainty scores are computed per pixel and then reduced to a single score per image by taking the mean. In particular, we support: - uncertainty_least_confidence - uncertainty_margin - uncertainty_entropy The scores are computed using the scorer for classification.
For more information about how to use the semantic segmentation scorer have a look here:
Check out our tutorial about how to use Lightly Active Learning: