https://github.com/google/magika

via code:

https://zhuanlan.zhihu.com/p/390826470

from transformers import AutoModelForMaskedLM, AutoTokenizer

checkpoint = “camembert-base”

model = AutoModelForMaskedLM.from_pretrained(checkpoint)

tokenizer = AutoTokenizer.from_pretrained(checkpoint)

model.push_to_hub(“dummy-model”)

tokenizer.push_to_hub(“dummy-model”)

config.push_to_hub(““)

currently we can use hyperopt as the online optimizer. of course for offline optimization there’s better option or prediction for it.

For a sequence of models, use classic logic solver to find best logic combination.

select model 0.

select model 1 with 0, iterate through 16 different situations(0, not 0, 1, not 1, 0 and 1, 0 or 1, 0 and not 1, 0 or not 1, not 0 and 1,not 0 or 1, not 0 and not 1, not 0 or not 1), choose the best one. mark it as model A.

select model 2, use the same optimizer to generate model B.

finally iterate through all models. generte model X as a combination of best logic models.