This package implements new dataset, loss and model operations in a
torch/luz framework to address the oracle data structure discussed
in my doctorate thesis (in
Portuguese).
The main objects of this package are:
captcha_oracle(): creates annotated datasets from an initial model and functions to access the oracle of a website.captcha_dataset_oracle(): This object is the same ascaptcha_dataset()from the{captcha}package, but deals with a different data structure that incorporates incomplete information provided by the websites.net_captcha_oracle(): This object is the same asnet_captcha()from the{captcha}package, but can consider an initial model as a parameter in its initialization method.oracle_loss(): This object implements the oracle loss proposed in the doctorate thesis, which incorporates the incomplete information provided by the oracle in the websites.captcha_accuracy_oracle(): calculates the accuracy of the model in the fitting process dealing with the new data structure provided bycaptcha_dataset_oracle().
The package also has an experimental feature to learn the parameters
online, by automatically accessing the oracle in the web. It is
implemented with the captcha_dataset_oracle_online() function. There
is a script in this
link
providing an example of this experimental feature.
This package is experimental and it was built as a tool to make simulations for my doctorate thesis. It may have breaking changes before it is considered as stable.
MIT