This blogpost makes it sound as if using a deep learning library for a task like image classification is difficult, while it is just a matter of copy+paste, annotate, and run.

Dunno, maybe you aren't the target for the article. As someone who has never played with deep learning, I thought it was presented just fine.

it is a lot harder than what this library does.

The OP library specifically grabs the ImageNet model and uses transfer learning. That's just a general keras convolutional networks resource.

The entire library is only 100 lines of code, so you could copy and paste its own source.

I wonder if there is a BERT equivalent for NLP. Is BERT part of Spacy now?

We've been developing support for transformers in spaCy. See:

https://explosion.ai/blog/spacy-pytorch-transformers

https://github.com/explosion/spacy-pytorch-transformers

The Tf-Hub stuff is pretty close I guess https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1

huggingface offers a whole family of pretrained transformers.

https://github.com/huggingface/pytorch-transformers