This repository implements PixelPick for low-annotation micrograph segmentation. The code is based on the original PixelPick paper (https://arxiv.org/abs/2104.06394).
pip install -r requirements.txtA sample query configuration can be found in configs/uhcs/default.yaml. The parameters for making queries are:
max_pixels: maximum number of pixels to query excluding the starting pixelsn_pixels_per_round: number of pixels to query in each round for each classn_init_pixels_per_class: number of pixels to query for each class at the beginningquery_strategy: the query strategy to use, can berandom,entropy,margin,least_confidence.top_n_percent: the top n percent of pixels to randomly sample from using the query strategy. For example, iftop_n_percentis 0.05 andquery_strategyisentropy, then the top 5% pixels with the highest entropy will be randomly sampled from.
The details of data preparation and training configuration can be found at https://github.com/leibo-cmu/MatSeg.
The following command
trains segmentation models with incremental queries using the entropy query strategy:
python train_pixelpick.py --config entropy.yaml