My main takeaway from this thread (all perfectly good mistakes to make) is how many of these traps you will never fall into by using a suitable higher level API. Except 1 though. 1 is a process mistake and it's a fantastic lesson to learn.https://twitter.com/karpathy/status/1013244313327681536 …
80% of the user errors I come across are data preprocessing issues. Then, occasionally, optimization problems: bad learning rate, network doesn't train, etc.
