For my research I usually build my experiments with PyTorch. Anyone here have pros and cons of writing with FastAI? It seems you can get experiments running very quickly, but is there also indepth control of your code if say you want to write your own loss functions or change parts of your model, etc.? Wondering if it’s worth spending time learning although I’m already very conformatble writing directly with PyTorch. Thanks!
This is really a matter of how comfortable you are with building something given just the documentation.
I found the documentation of fastai to be a bit lacking in certain respects which made it difficult for me to build something of my own without having a reference tutorial. But that might be a limitation of my own current capabilities hence my comment above on this being dependent on your current skill set.
Fastai did, for one project, allow me to quickly create a prototype and train on it. However, it was still a relatively straightforward classification task. But loading in the data and then using transfer learning were very easy, straightforward and not at all time consuming.
The library is pytorch based so you shouldn’t have any issue writing your own functions or changing parts of your model as per me. I think you should be fine with fastai. I would recommend either checking out their free book - https://github.com/fastai/fastbook or maybe this - https://www.youtube.com/playlist?list=PLFDkaGxp5BXDvj3oHoKDgEcH73Aze-eET (corresponding website - https://walkwithfastai.com/)