I’m training multiple convolutional neural networks to identify a specific disease from chest x-rays (similar to this paper). Two separate questions:
First, what is the best tool/library/package within the Pytorch ecosystem to systematically run experiments and log model results? Ideally, this tool would allow for training models with different sets of hyperparameters back-to-back (so I can leave it running and get all the results at once). I have tried Pytorch Lightning, but I am looking for a tool that allows more flexibility to modify architectures.
Second, I would like to implement multi-task learning - the second task is to predict the size of the heart. Any suggestions for modifying the architecture to accept this would be greatly appreciated.