Repository containing FSDL project
The TorchXRayVision Project [1,2] provides several different models trained on a variety of X-ray datasets, such as NIH's ChestX-ray8. In this project, we aim to build an interface using the tools described in the lectures/labs to allow users / medical practitioners to explore, visualize, and interpret model predictions. For example, methods such as "gifsplanation" build animations to investigate counterfactual explanations for model outputs.
References: [1] https://huggingface.co/torchxrayvision [2] https://arxiv.org/abs/2111.00595 [3] https://mlmed.org/gifsplanation/
How does your project go beyond the usual "train a model to do well on a static benchmark" typical in ML class projects?
As described above, our focus is on using well-developed pre-trained models to investigate model predictions. Our proposed project highlights the deployment aspect of the class and we are also considering concepts related to data annotation / data flywheel feedback loop. In addition, we are interested in the possibility of using the baseline pre-trained models for transfer learning.
Installation: Run setup.py and then demo.py.
This requires a working version of ffmpeg that can be used via command-line.
Demo recording link: https://youtu.be/UyuGoUuO9C8?list=PL1T8fO7ArWle-HwX6SkoQ3j_ol19P7tGT