CarDS-Yale/PanEcho ? reverse-engineered prompt

Build me a PyTorch project for AI based echocardiogram interpretation that can take a short echo video clip and predict a bunch of reporting tasks from it.

I want the main model to work from a 16 frame clip by default, use a frame encoder plus a temporal Transformer, and return a dictionary of predictions for many tasks, with support for both classification and regression outputs. It should also let me change the clip length, keep only a subset of task heads, or strip the heads off entirely so I can use the backbone or just the image encoder for transfer learning.

Please include a simple torch hub style load entry point, a basic demo inference example, and the training and fine tuning scripts needed to reproduce the model and adapt it to new echo datasets. If you need current library details, look up the latest docs online first.