Skip to content

Latest commit

 

History

History
5 lines (2 loc) · 942 Bytes

README.md

File metadata and controls

5 lines (2 loc) · 942 Bytes

Each patch-level downstream histology dataset used in evalauation is properly cited in the preprint can downloaded either from their authors github, website or other data hosting services easily, and most datasets folders can be formatted to leverage torch's ImageFolder Dataset class for ease of use, we provide a few custom dataset class for the datasets that arent easily parsed with ImageFolder .

For evaluation (zero-shot, retrieval, linear probing) we use Clip benchmark with modified classnames for each dataset based off of its labels. The templates for zero-shot classification can be found in the paper. For linear probing, the only addition to Clip benchmark was to run hyperparameter search and multiple seed runs for all models and baselines, as reported in the paper.