activeInstanceSegment

A repository to benchmark active learning strategies on microscopy data.

Active Learning

The following active learning strategies are implemented:

• random: randomly sample data points from unlabeled pool (used as benchmark)
• mc_dropout: sample data points based on uncertainty quantified using Monte Carlo dropout.
• tta: sample data points based on uncertainty quantified using test time augmentation.
• hybrid: sample data points based on uncertainty quantified using mc dropout and clustering over the datas latent space representation.

The al strategy can be specified in the cfg.yaml in ./pipeline_configs. Benchmarking can be done with

python -m src.al_main -c <path_to_config>

Installation

installation on a ubuntu 22.04 can be done with the following script

$ ./shell_scripts/install.sh

Data

all used datasets need to follow the COCO format

How to Add a New Dataset

to add a new dataset, the following steps need to be performed

1. create a directory in data/ named after the new dataset
2. the new directory contains folders for every data split e.g. train, test
3. each data split folder contains the data stored in COCO format in the cell_acdc_coco_ds.json and the images in the images folder
4. the dataset name together with the split types must be added in src/globals.py to the DATASETS_DSPLITS dictionary
5. the dataset can be specified in the configuration yamls as datasetname_splittype new datasets can be added using the Data2cocoConverter class in utils.datapreprocessing.data2coco

Model Architecture

The active learning is built ontop of the detectron2 implementation of Mask R-CNN. Training a model without active learning can be done by running

python -m src.pipeline_runner -c <path_to_config>

Virtual Enviroment

this project uses a venv which is initailized by running

$ python -m venv ac_acdc_env

the venv is activated using the following command

$ source ac_acdc_env/bin/activate 

requirements can be installed by running

$ pip install -r requirements.txt 

Configuration

hyperparameters used for model training, testing and during active learning are specified in configuration.yaml files in the pipeline_config directory. The configuration file containes detectron2 configurations and custom active learning configurations.

Active Learning

The following active learning configs can be specified

• DATASETS.TRAIN_UNLABELED: name of training dataset used for active learning
• INCREMENT_SIZE: number of data points which are annotated each active learning iteration
• INIT_SIZE: size of initial training dataset
• MAX_LOOPS: maximal number of active learning loops
• NUM_MC_SAMPLES: number of monte carlo samples used for uncertianty estivation
• OBJECT_TO_IMG_AGG: aggregation of object uncertainties to an uncertainty value for the entire image (mean, max,min,sum)
• OUTPUT_DIR: path to output directory
• QUERY_STRATEGY: used active learning strategy (random, mc_drop, tta, hybrid)
• TTA_MAX_NOISE: max intensity of gaussian noise applied during test-time augmentation (only applied in QUERY_STRATEGY:tta)
• SAMPLE_EVERY: number of images which form the subset used for AL sampling
• MAX_TRAINING_EPOCHS: number of training epochs during active learning
• RETRAIN: flag if model should be retrained from scratch every active learning iteration (only true implemented)