[nikonyrh]

I prefer to implement ideas "from scratch" (using Keras and Tensorflow) as much as possible, and quite rarely use pre-trained networks. But as pre-trained networks are getting more and more powerful, this isn't an ideal strategy :D Well I have deep-dived to Stable Diffusion.

I feel like if I only use pre-trained stuff and copy paste code, I couldn't solve novel problems, apply domain knowledge to the problem etc. But I mostly work on computer vision / image tasks, and at work we just use LightGBM for a regression problem on tabular data. I don't have much experience on NLP.