I’ve watched the full Youtube course and tasks series; I’ve read the entire Transformers How-To and Conceptual Guide, and a bunch of the API docs, and still every little basic thing seems to result in errors whose solutions I can’t find in documentation.
For example (working with GPT2), one doc said to create your
from_pretrained() and then simply call it with
model(**input_ids), but this produced a complex
ModelOutput object that the tokenizer couldn’t decode. Randomly I found another doc that said to call
model.generate(**input_ids) instead, which worked (lucky me!).
As another example, I was trying to set up a
GenerationConfig, but when I pass that in (either
model.generate(generation_config=gen_config, inputs=input_ids) or
model.generate(generation_config=gen_config, **input_ids)) it fails deep inside the tokenizer, or the model’s forward pass.
And there have been so many more…
The use of complex structures all of which are indexable, and Python’s total lack of clear typing, makes this even harder to “guess” my way through…
I know this isn’t the recommended format for asking for support. I could file an issue for every one of these things with complete code, but there would be dozens and I think that’s a waste of everyone’s time. I feel like I’m just missing some basic documentation somewhere.