It's been a few years since I did Keras dev full time, but debugging was a nightmare. PyTorch- as long as you don't compile and you run on CPU, it's pretty much

It's been a few years since I did Keras dev full time, but debugging was a nightmare. PyTorch- as long as you don't compile and you run on CPU, it's pretty much like debugging standard code. Debugging CUDA is harder. Edit: Seems like `run_eagerly` solves a lot of the pain of debugging Keras code.